AI flywheel operation method

Abstract

The invention belongs to the technical field of prediction model training, and relates to an AI flywheel operation method, a model generates a preliminary prediction result through a prediction step, compares a prediction value with an actual value through incremental learning, calculates an error and feeds back the error to the model, and through a feature updating step, the model can automatically evaluate and update the correlation of input features, so that the prediction accuracy is improved. According to the method, input data is ensured to be consistent with a current environment, a new data sample is generated through a generative adversarial network (GAN), a training data set is expanded, the generalization ability of a model is improved, time sequence difference optimization further enables the model to adapt to prediction deviation caused by time change through weight adjustment for different time periods, the processing effect of time sequence data is optimized, and the prediction accuracy is improved. The long-term adaptability and prediction accuracy of the model are improved on the whole, the problem that the prediction capability of the AI model in a dynamic environment is reduced is effectively solved, it is guaranteed that the model is continuously optimized and adapted under continuously changing data and environments, and efficient prediction performance is kept.

Description

Technical Field

[0001] The present application belongs to the field of prediction model training technology, and more specifically, to an AI flywheel operation method. Background Art

[0002] In the application of artificial intelligence (AI) and machine learning, the accuracy and generalization ability of predictive models are the core criteria for evaluating model performance. However, traditional AI models usually rely on static data sets for training. Once deployed, their predictive ability may decline over time, especially when facing a changing environment and dynamic data. For example, the distribution of input data may change over time. This phenomenon is called data drift or concept drift. Data drift can cause the prediction results of the original model to deviate from the actual situation, because the model is trained under a fixed data distribution in the past. When the new data distribution is inconsistent with the data distribution during training, the prediction accuracy of the model will decrease; and in many application scenarios, the initial training data set may not fully cover all possible situations or anomalies. This means that after the model is deployed, it may encounter situations that have not been seen in the training data set. These new or rare situations require the model to have the ability to learn and adapt quickly to avoid deviations in model predictions. Summary of the invention

[0003] The present invention provides an AI flywheel operation method, which is intended to solve the current technical problem that the prediction ability of the model decreases over time when faced with a constantly changing environment and dynamic data.

[0004] An AI flywheel operation method comprises the following steps:

[0005] Model prediction: Make predictions based on the prediction model and obtain prediction results;

[0006] Incremental learning: Compare the predicted value with the actual value, calculate the prediction error, and return the error to the prediction model based on the calculated error to decide whether to perform incremental learning. If incremental learning is required, the incremental learning method is used to feed the error back to the model and adjust the weight of the prediction model.

[0007] Feature update: Automatically evaluate the impact of input features on the model based on error changes, and automatically update the input features of the model;

[0008] Data enhancement: Generate new data samples based on the generative adversarial network GAN to train the prediction model;

[0009] Time series difference optimization: According to different time periods, the output of the model is optimized by adjusting the weights of different time periods;

[0010] Adaptive adjustment: Dynamically adjust the model structure and training strategy based on the feedback results of incremental learning, feature updates, data enhancement, and timing difference optimization.

[0011] The AI ​​flywheel operation method of the present invention provides a continuous optimization solution for the problem of the decline in the prediction ability of AI models in the face of changing environments and dynamic data through a series of dynamic technical steps. First, the model generates preliminary prediction results through the prediction step, and compares the predicted value with the actual value through incremental learning, calculates the error and feeds it back to the model, so as to determine whether the parameters of the model need to be adjusted according to the error. This incremental learning mechanism enables the model to continuously optimize itself according to real-time feedback, avoiding the decline in prediction accuracy caused by fixed training of traditional models. Secondly, through the feature update step, the model can automatically evaluate and update the relevance of input features to ensure that the input data is consistent with the current environment, thereby improving the adaptability of the model. Data enhancement generates new data samples through generative adversarial networks (GANs), expands the training data set, and improves the generalization ability of the model, especially when the sample is insufficient or the data distribution changes. Time series difference optimization further enables the model to adapt to the prediction deviation caused by time changes and optimizes the processing effect of time series data by adjusting the weights for different time periods. Finally, the adaptive adjustment mechanism combines the feedback of incremental learning, feature update, data enhancement and time series optimization to dynamically adjust the structure and training strategy of the model, and improve the long-term adaptability and prediction accuracy of the model as a whole. Through the collaborative work of these steps, the present invention can effectively solve the problem of reduced predictive ability of AI models in dynamic environments, ensure that the model is continuously optimized and adapted to changing data and environments, and maintain efficient predictive performance.

[0012] Preferably, the incremental learning comprises the following steps:

[0013] Error calculation: Calculate the error e for each prediction result t :

[0014] e t =y ac tual ,t -y pre d ,t ;

[0015] Where: e t represents the error of the t-th prediction; y actual,t represents the actual value of the t-th prediction; y pred,t represents the model output of the t-th prediction;

[0016] Incremental learning judgment: If the error exceeds the preset threshold, incremental learning is triggered, where the error threshold is dynamically adjusted based on the error fluctuation history of the prediction model;

[0017] Error-based dynamic weight adjustment: Dynamically adjust the weight of the prediction model based on error feedback:

[0018]

[0019] Where: represents the weight after the tth adjustment; represents the weight before the tth time; η represents the learning rate; e t Indicates the current error; x t represents the input feature vector;

[0020] Incremental update: Combine the gradient weighted energy update method and update the model weights through back propagation:

[0021]

[0022] In the formula: Δw t Represents the weight update amount for the tth time; Represents the loss function L on weight w t gradient.

[0023] Preferably, the feature update comprises the following steps:

[0024] Error feedback and feature influence evaluation: For each time step t, calculate each input feature The prediction error e t Contribution

[0025]

[0026] Where: Indicates the influence of the i-th feature on the error in the t-th prediction; Represents the error to feature The partial derivative of y actual,t represents the actual value of the t-th prediction; y pred,t represents the model output of the t-th prediction;

[0027] Adaptive update of feature importance: Dynamically adjust the weight of features based on error feedback:

[0028]

[0029] Where: Represents the features before the tth update The weight of Represents the features after the tth update The weight of ; η represents the learning rate;

[0030] Feature elimination: For features If its influence is less than the threshold, the feature is removed from the model. After the feature is removed, the model maintains the prediction accuracy by retraining and updating the new feature set.

[0031] Preferably, the feature update further comprises the following steps:

[0032] The importance of the features is dynamically updated through the gain and loss adjustment strategy:

[0033]

[0034]

[0035] Where: represents the t-th feature gain, which indicates the contribution of this feature to the error; represents the loss degree of feature i; n represents the total number of features; if the gain of a feature is lower than the gain threshold and the loss is greater than the loss threshold during each incremental learning process, the weight is reduced until it is completely eliminated.

[0036] Preferably, the data enhancement comprises the following steps:

[0037] Error-driven sample generation strategy: Set an error threshold ∈, for each error e t , if |e t |>∈, data enhancement is triggered; otherwise, no data enhancement is required;

[0038] Error-based feature space expansion: Generate samples based on the generator network G of the generative adversarial network:

[0039]

[0040] Where: represents the sample generated by the tth augmentation; z t represents a sample from the random noise space; e t represents the error of the t-th prediction; G(·) represents the generator part of the generative adversarial network;

[0041] Enhanced sample quality assessment and screening: For each sample Calculate the quality assessment score through the discriminator

[0042]

[0043] Where: represents the quality score of the enhanced sample; D(·) represents the discriminator network;

[0044] like If it is greater than the preset threshold, the sample is considered valid and added to the training set, otherwise it is discarded.

[0045] Preferably, the data enhancement further includes weight adjustment of enhanced samples;

[0046] The weight of the enhanced sample Based on the quality assessment score generated To make dynamic adjustments:

[0047]

[0048] Where: Represents enhanced samples The weights during training; λ represents the global adjustment factor.

[0049] The beneficial effects of the present invention include:

[0050] The AI ​​flywheel operation method of the present invention provides a continuous optimization solution for the problem of the decline in the prediction ability of AI models in the face of changing environments and dynamic data through a series of dynamic technical steps. First, the model generates preliminary prediction results through the prediction step, and compares the predicted value with the actual value through incremental learning, calculates the error and feeds it back to the model, so as to determine whether the parameters of the model need to be adjusted according to the error. This incremental learning mechanism enables the model to continuously optimize itself according to real-time feedback, avoiding the decline in prediction accuracy caused by fixed training of traditional models. Secondly, through the feature update step, the model can automatically evaluate and update the relevance of input features to ensure that the input data is consistent with the current environment, thereby improving the adaptability of the model. Data enhancement generates new data samples through generative adversarial networks (GANs), expands the training data set, and improves the generalization ability of the model, especially when the sample is insufficient or the data distribution changes. Time series difference optimization further enables the model to adapt to the prediction deviation caused by time changes and optimizes the processing effect of time series data by adjusting the weights for different time periods. Finally, the adaptive adjustment mechanism combines the feedback of incremental learning, feature update, data enhancement and time series optimization to dynamically adjust the structure and training strategy of the model, and improve the long-term adaptability and prediction accuracy of the model as a whole. Through the collaborative work of these steps, the present invention can effectively solve the problem of reduced predictive ability of AI models in dynamic environments, ensure that the model is continuously optimized and adapted to changing data and environments, and maintain efficient predictive performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0052] Figure 1 An overall step block diagram provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0054] See also Figure 1 As shown, the best embodiment of the present invention is further described;

[0055] An AI flywheel operation method comprises the following steps:

[0056] Model prediction: Make predictions based on the prediction model (such as LSTM model prediction load) to obtain prediction results;

[0057] Incremental learning: Compare the predicted value with the actual value, calculate the prediction error, and return the error to the prediction model based on the calculated error to decide whether to perform incremental learning. If incremental learning is required, the incremental learning method is used to feed the error back to the model and adjust the weight of the prediction model.

[0058] Feature update: Automatically evaluate the impact of input features on the model based on error changes, and automatically update the input features of the model;

[0059] Data enhancement: Generate new data samples based on the generative adversarial network GAN to train the prediction model;

[0060] Time series difference optimization: According to different time periods, the output of the model is optimized by adjusting the weights of different time periods;

[0061] Adaptive adjustment: Dynamically adjust the model structure and training strategy based on the feedback results of incremental learning, feature updates, data enhancement, and timing difference optimization.

[0062] The AI ​​flywheel operation method of the present invention provides a continuous optimization solution for the problem of the decline in the prediction ability of AI models in the face of changing environments and dynamic data through a series of dynamic technical steps. First, the model generates preliminary prediction results through the prediction step, and compares the predicted value with the actual value through incremental learning, calculates the error and feeds it back to the model, so as to determine whether the parameters of the model need to be adjusted according to the error. This incremental learning mechanism enables the model to continuously optimize itself according to real-time feedback, avoiding the decline in prediction accuracy caused by fixed training of traditional models. Secondly, through the feature update step, the model can automatically evaluate and update the relevance of input features to ensure that the input data is consistent with the current environment, thereby improving the adaptability of the model. Data enhancement generates new data samples through generative adversarial networks (GANs), expands the training data set, and improves the generalization ability of the model, especially when the sample is insufficient or the data distribution changes. Time series difference optimization further enables the model to adapt to the prediction deviation caused by time changes and optimizes the processing effect of time series data by adjusting the weights for different time periods. Finally, the adaptive adjustment mechanism combines the feedback of incremental learning, feature update, data enhancement and time series optimization to dynamically adjust the structure and training strategy of the model, and improve the long-term adaptability and prediction accuracy of the model as a whole. Through the collaborative work of these steps, the present invention can effectively solve the problem of reduced predictive ability of AI models in dynamic environments, ensure that the model is continuously optimized and adapted to changing data and environments, and maintain efficient predictive performance.

[0063] As a possible implementation of this embodiment, the incremental learning includes the following steps:

[0064] Error calculation: Calculate the error e for each prediction result t :

[0065] e t =y actual,t -y pred,t ;

[0066] Where: e t represents the error of the t-th prediction; y actual,t represents the actual value of the t-th prediction; y pred,t represents the model output of the t-th prediction;

[0067] Incremental learning judgment: If the error exceeds the preset threshold, incremental learning is triggered, where the error threshold is dynamically adjusted based on the error fluctuation history of the prediction model;

[0068] Error-based dynamic weight adjustment: Dynamically adjust the weight of the prediction model based on error feedback:

[0069]

[0070] Where: represents the weight after the tth adjustment; represents the weight before the tth time; η represents the learning rate; e t Indicates the current error; x t represents the input feature vector;

[0071] Incremental update: Combine the gradient weighted energy update method and update the model weights through back propagation:

[0072]

[0073] In the formula: Δw t Represents the weight update amount for the tth time; Represents the loss function L on weight w t gradient.

[0074] In this embodiment, by introducing the error weighting factor in incremental learning, the model can more strongly adjust the weights of features and samples with large errors, which helps to improve the model's adaptability to areas with large errors. Dynamically adjusting feature weights based on error feedback avoids fixed feature selection strategies and improves the model's ability to adapt to the importance of different features in different data contexts.

[0075] As a possible implementation of this embodiment, the feature update includes the following steps:

[0076] Error feedback and feature influence evaluation: For each time step t, calculate each input feature The prediction error e t Contribution

[0077]

[0078] Where: Indicates the influence of the i-th feature on the error in the t-th prediction; Represents the error to feature The partial derivative of y actual,t represents the actual value of the t-th prediction; y pred,t represents the model output of the t-th prediction;

[0079] Adaptive update of feature importance: Dynamically adjust the weight of features based on error feedback:

[0080]

[0081] Where: Represents the features before the tth update The weight of Represents the features after the tth update The weight of ; η represents the learning rate;

[0082] The dynamic weighted update mechanism based on error feedback and feature influence can improve the model's attention to important features and reduce the impact of unimportant features.

[0083] Feature elimination: For features If its influence is less than the threshold, the feature is removed from the model. After the feature is removed, the model maintains the prediction accuracy by retraining and updating the new feature set.

[0084] In this embodiment, by eliminating features that contribute less to the model, the computational complexity is reduced while maintaining the efficiency and accuracy of the model.

[0085] As a possible implementation of this embodiment, the feature update further includes the following steps:

[0086] The importance of the features is dynamically updated through the gain and loss adjustment strategy:

[0087]

[0088]

[0089] Where: represents the t-th feature gain, which indicates the contribution of this feature to the error; represents the loss degree of feature i; n represents the total number of features; if the gain of a feature is lower than the gain threshold and the loss is greater than the loss threshold during each incremental learning process, the weight is reduced until it is completely eliminated.

[0090] In this embodiment, the feature gain and loss are combined to dynamically adjust the feature weights, so that the model can more flexibly cope with the changing data distribution and feature importance.

[0091] As a possible implementation of this embodiment, the data enhancement includes the following steps:

[0092] Error-driven sample generation strategy: Set an error threshold ∈, for each error e t , if |e t |>∈, data enhancement is triggered; otherwise, no data enhancement is required;

[0093] Error-based feature space expansion: Generate samples based on the generator network G of the generative adversarial network:

[0094]

[0095] Where: represents the sample generated by the tth augmentation; z t represents a sample from the random noise space; e t represents the error of the t-th prediction; G(·) represents the generator part of the generative adversarial network;

[0096] In this embodiment, e t As a control factor, we can generate samples with diversity in the area with large errors by changing the input noise space. This process helps the model to conduct intensive training in the area with large errors, thereby improving the generalization ability in these areas.

[0097] Enhanced sample quality assessment and screening: For each sample Calculate the quality assessment score through the discriminator

[0098]

[0099] Where: represents the quality score of the enhanced sample; D(·) represents the discriminator network;

[0100] like If it is greater than the preset threshold, the sample is considered valid and added to the training set, otherwise it is discarded.

[0101] As a possible implementation of this embodiment, the data enhancement further includes weight adjustment of the enhanced samples;

[0102] The weight of the enhanced sample Based on the quality assessment score generated To make dynamic adjustments:

[0103]

[0104] Where: Represents enhanced samples The weights during training; λ represents the global adjustment factor.

[0105] In this embodiment, an adaptive data enhancement scheme driven by error feedback is used. This scheme dynamically generates new training samples based on the prediction error and error fluctuation of the model, and performs sample expansion in areas where the model performs poorly, further improving the accuracy and generalization ability of the model in different scenarios.

[0106] As a possible implementation of this embodiment, the timing difference optimization includes the following steps:

[0107] Time segmentation and feature extraction: First, the entire time series data is divided by time to obtain multiple time segments. Each time segment can represent a different time interval, usually divided according to the periodicity or volatility of the data. For example, it can be divided according to granularity such as hours, days, weeks, months, etc., or customized according to actual business scenarios.

[0108] For example, if we are dealing with the problem of electricity demand forecasting, we can divide the data according to different time periods of the day, such as "day" and "night", and calculate the features of each time period separately.

[0109] Weight adjustment mechanism for time series differences: In each time period, the model may have different responses to input features in different time periods; in order to optimize the model's prediction output, a time period weight w can be assigned to each time period i , and dynamically adjust the weights according to the error feedback of the time period; in this way, the model will adjust the weights according to different time periods during prediction, enhancing the model's adaptability to time series data;

[0110] Time series weight adjustment formula: The weight w of each time period i (t), dynamically adjusted according to the error change:

[0111] w i (t) = w i (t-1)+η·Δe i (t) ΔX i (t);

[0112] Where: w i (t) represents the weight of time segment i at time t; w i (t-1) represents the time period weight of the previous moment; η represents the learning rate, which controls the update speed; Δe i (t) = e i (t)-e i (t-1), represents the error change at time t in time period i; ΔX i (t) represents the increment of input feature change in time period i;

[0113] Weighted optimization of timing errors: In order to further optimize the timing differences, in addition to adjusting the weights, weighting factors can also be added to the model output. i (t), weight the output of the model and strengthen the model's focus on important time periods:

[0114]

[0115] Where: y adjusted (t) represents the final adjusted model output; yi (t) represents the predicted output of the model at time i; w i (t) represents the weight of time period i;

[0116] In this formula, the output y of the model is adjusted (t) is obtained by weighting the prediction results of each time period, thereby improving the prediction accuracy of the model for a specific time period.

[0117] Adjustment of loss function for time series differences: During the training process, we can add a time series weight factor to the loss function so that the optimization process not only focuses on the overall error, but also considers the different error weights of each time period, further optimizing the model's time series prediction ability.

[0118]

[0119] Where: K sdjusted represents the adjusted loss function; L i (t) represents the basic loss at time t in time period i; w i (t) weighting factor for time period i;

[0120] By introducing the time period weight w into the loss function i (t), the model will dynamically adjust parameter updates according to the importance of each time period during training, so as to better adapt to the time differences in time series data.

[0121] Adaptive adjustment of time series data: Time series difference optimization is not just about adjusting static weight values, but also requires dynamic adjustment of time period division strategies, error calculation methods, and feature extraction methods based on data flow and model feedback. By combining with feedback mechanisms such as incremental learning and feature updates, the model can better adapt to changes in long time series and improve the prediction capabilities of non-stationary time series data.

[0122] For example, the length of the time period can be dynamically adjusted based on the error change trend or the volatility of the time series data, or different feature extraction strategies (such as sliding windows, Fourier transforms, etc.) can be used to cope with different types of time series data.

[0123] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An AI flywheel operation method, characterized in that: The following steps are involved: Model prediction: Make predictions based on the prediction model and obtain prediction results; Incremental learning: Compare the predicted value with the actual value, calculate the prediction error, and return the error to the prediction model based on the calculated error to decide whether to perform incremental learning; If incremental learning is required, the incremental learning method is used to feed the error back to the model and adjust the weight of the prediction model; Feature update: Automatically evaluate the impact of input features on the model based on error changes, and automatically update the input features of the model; Data enhancement: Generate new data samples based on the generative adversarial network GAN to train the prediction model; Time series difference optimization: According to different time periods, the output of the model is optimized by adjusting the weights of different time periods; Adaptive adjustment: Dynamically adjust the model structure and training strategy based on the feedback results of incremental learning, feature updates, data enhancement, and timing difference optimization.

2. The AI ​​flywheel operation method according to claim 1, characterized in that: Incremental learning The following steps are involved: Error calculation: Calculate the error e for each prediction result t : e t =the ac approx l ,t -y pre d ,t ; Where: e t represents the error of the t-th prediction; y actual,t represents the actual value of the t-th prediction; y pred,t represents the model output of the t-th prediction; Incremental learning judgment: If the error exceeds the preset threshold, incremental learning is triggered, where the error threshold is dynamically adjusted based on the error fluctuation history of the prediction model; Error-based dynamic weight adjustment: Dynamically adjust the weight of the prediction model based on error feedback: Where: represents the weight after the tth adjustment; represents the weight before the tth time; η represents the learning rate; e t Indicates the current error; x t represents the input feature vector; Incremental update: Combine the gradient weighted energy update method and update the model weights through back propagation: In the formula: Δw t Represents the t-th weight update amount; Represents the loss function L on weight w t gradient.

3. The AI ​​flywheel operation method according to claim 1, characterized in that: The feature update comprises the following steps: Error feedback and feature influence evaluation: For each time step t, calculate each input feature The prediction error e t Contribution Where: Indicates the influence of the i-th feature on the error in the t-th prediction; Represents the error to feature The partial derivative of y actual,t represents the actual value of the t-th prediction; y pred,t represents the model output of the t-th prediction; Adaptive update of feature importance: Dynamically adjust the weight of features based on error feedback: Where: Represents the features before the tth update The weight of Represents the features after the tth update The weight of ; η represents the learning rate; Feature elimination: For features If its influence is less than the threshold, the feature is removed from the model. After the feature is removed, the model maintains the prediction accuracy by retraining and updating the new feature set.

4. The AI ​​flywheel operation method according to claim 3, characterized in that: The feature update further comprises the following steps: The importance of the features is dynamically updated through the gain and loss adjustment strategy: Where: represents the t-th feature gain, which indicates the contribution of this feature to the error; represents the loss degree of feature i; n represents the total number of features; if the gain of a feature is lower than the gain threshold and the loss is greater than the loss threshold during each incremental learning process, the weight is reduced until it is completely eliminated.

5. The AI ​​flywheel operation method according to claim 1, characterized in that: The data enhancement includes the following steps: Error-driven sample generation strategy: Set an error threshold ∈, for each error e t , if |e t |>∈, data enhancement is triggered; otherwise, no data enhancement is required; Error-based feature space expansion: Generate samples based on the generator network G of the generative adversarial network: Where: represents the sample generated by the tth augmentation; z t represents a sample from the random noise space; e t represents the error of the t-th prediction; G(·) represents the generator part of the generative adversarial network; Enhanced sample quality assessment and screening: For each sample Calculate the quality assessment score through the discriminator Where: represents the quality score of the enhanced sample; D(·) represents the discriminator network; like If it is greater than the preset threshold, the sample is considered valid and added to the training set, otherwise it is discarded.

6. The AI ​​flywheel operation method according to claim 5, characterized in that: The data enhancement also includes weight adjustment of enhanced samples; The weight of the enhanced sample Based on the quality assessment score generated To make dynamic adjustments: Where: Represents enhanced samples The weights during training; λ represents the global adjustment factor.

Images