Cloud-edge collaborative model automatic training and online learning integration method and system

Through the integrated method of cloud-edge collaboration model automation training and online learning, combined with the improved Bayesian optimization algorithm and homoeconomic optimization method, the deep learning model is solved in the ocean data without adapting to data changes and cold start problems, achieving high efficiency, Stable marine data prediction and detection.

CN119250162APending Publication Date: 2025-01-03SHANDONG COMP SCI CENTNAT SUPERCOMP CENT IN JINAN +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411278919.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-12
Publication Date
2025-01-03

AI Technical Summary

Technical Problem

Existing deep learning models have problems in marine data that do not adapt to real-time changes in data distribution, cold start and catastrophic forgetting problems, and the design process requires a lot of resources and time.

Method used

The integrated method of cloud-edge collaboration model automation training and online learning is adopted. The hyperparameters are optimized by offline pre-training and edge-end online learning in the cloud, combined with the improved Bayesian optimization algorithm, and the generalized additive model and homoethic optimization are used. Methods: Nelder-Mead algorithm is used to track local minimum values ​​to realize automated training and online update of the model.

Benefits of technology

It improves the training efficiency and accuracy of the marine model, solves the problems of cold start and catastrophic forgetting, reduces resource consumption and time, and realizes the real-time adaptation and stability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119250162A_ABST
    Figure CN119250162A_ABST
Patent Text Reader

Abstract

The invention provides a cloud-edge collaborative model automatic training and online learning integration method and system, and the method comprises the steps: carrying out the offline training of an ocean model through a cloud end, and publishing a pre-trained ocean model to an edge end for online learning reasoning, so as to solve a cold start problem of an online learning algorithm in an initial model prediction stage; a generalized addition model is used as an agent model of a Bayesian optimization algorithm, a homotopy optimization method is adopted to track a local minimum value along a deformation path, hyper-parameter optimization is carried out on an ocean model through the improved Bayesian optimization algorithm, the hyper-parameter optimization efficiency and effect are improved, the ocean model training efficiency is improved, and the time needed by the whole process is shortened.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of ocean research technology, and in particular to a cloud-edge collaborative model automation training and online learning integration method and system. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, deep learning has played an increasingly important role in ocean intelligence and has been widely used in multiple scenarios. Artificial intelligence can help solve many difficult problems in various fields of the ocean. For example, the use of artificial intelligence algorithms to predict and detect anomalies in ocean observation data can not only better control ocean changes, but also abnormal data has an important impact on the accuracy of scientific research and decision-making. Therefore, the prediction and anomaly detection of ocean observation data are crucial. By applying advanced data analysis and machine learning techniques, real-time monitoring of ocean observation data streams can be achieved, and abnormal data points can be automatically detected and identified. Once an anomaly is detected, the system will issue an alarm so that corrective measures can be taken in time to ensure the quality and reliability of ocean monitoring data.

[0004] In the field of marine intelligence, most deep learning models are currently built based on the offline learning paradigm, which is mainly applicable to scenarios with relatively stable data distribution. However, in the face of real-time changes in marine data, the performance of these models is difficult to guarantee. For example, the original intention of the AWAC algorithm is to achieve deep reinforcement learning through offline pre-training and online adjustment, but due to the existence of distribution bias, the agent strategy trained offline cannot be directly used for online reasoning. Figure 1 As shown in Figure 1, the offline model learns the data distribution before the mutation, while the data after the mutation is regarded as abnormal data. Therefore, the offline model is difficult to adapt to the anomaly detection and prediction tasks of data changes.

[0005] Online learning algorithms encounter a cold start problem during the initial model prediction phase. Due to the lack of historical information in the model, the prediction and detection performance in the initial phase fluctuates greatly, limiting its ability to make accurate predictions and decisions.

[0006] Moreover, online learning systems need to run for a long time, continuously collect data, evolve models, and update parameters. Over time, the model may forget previous knowledge, leading to the problem of catastrophic forgetting. This means that the model may forget previously acquired knowledge while learning new knowledge. In addition, deep learning models are usually manually designed by experts through a trial-and-error process, which requires a lot of resources and time, even for experts.

[0007] In summary, the current deep learning models applied to ocean data have at least the following problems:

[0008] 1. The deep learning model based on offline learning is not suitable for ocean observation data with real-time changes in data distribution. The offline model cannot adapt to the changing data anomaly detection and prediction tasks;

[0009] 2. Online learning algorithms have cold start and catastrophic forgetting problems;

[0010] 3. Deep learning models are usually manually designed by experts through a trial-and-error process, which requires a lot of resources and time. Summary of the invention

[0011] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a cloud-edge collaborative model automatic training and online learning integration method and system, which combines offline training and online learning, not only adapts to the changing real-time ocean data, but also solves the cold start problem of online learning; through the improved Bayesian optimization algorithm, hyperparameter optimization is performed to improve the efficiency and effect of hyperparameter optimization, thereby improving the efficiency of ocean model training.

[0012] To achieve the above objectives, the first aspect of the present invention provides a cloud-edge collaborative model automatic training and online learning integration method, including:

[0013] Use historical ocean observation data to pre-train the ocean model offline in the cloud, and publish the pre-trained ocean model to the edge.

[0014] Based on real-time ocean observation data at the edge, the pre-trained ocean model is learned online to obtain the optimal ocean model.

[0015] Among them, in the offline training of the ocean model, an improved Bayesian optimization algorithm is used to optimize the hyperparameters of the ocean model. Specifically, the generalized additive model is used as a proxy model of the Bayesian optimization algorithm; the continuous deformation path of the generalized additive model is constructed using the homotopy optimization method, and the local minimum of the homotopy function is tracked along the continuous deformation path using the Nelder-Mead algorithm.

[0016] A second aspect of the present invention provides a cloud-edge collaborative model automated training and online learning integrated system, including:

[0017] The offline training unit is used to pre-train the ocean model offline in the cloud using historical ocean observation data, and publish the pre-trained ocean model to the edge. In the offline training of the ocean model, the hyperparameters of the ocean model are optimized by using an improved Bayesian optimization algorithm, specifically: the generalized additive model is used as a proxy model of the Bayesian optimization algorithm; the continuous deformation path of the generalized additive model is constructed by using the homotopy optimization method, and the local minimum of the homotopy function is tracked along the continuous deformation path by using the Nelder-Mead algorithm;

[0018] The online learning unit is used to perform online learning of the pre-trained ocean model based on real-time ocean observation data at the edge to obtain the optimal ocean model.

[0019] The third aspect of the present invention provides a computer device, comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, a cloud-edge collaborative model automatic training and online learning integration method is performed.

[0020] The fourth aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the computer program executes a cloud-edge collaborative model automation training and online learning integration method.

[0021] One or more of the above technical solutions have the following beneficial effects:

[0022] In the present invention, the ocean model is trained offline through the cloud, and then the pre-trained ocean model is published to the edge for online learning and reasoning to solve the cold start problem encountered by the online learning algorithm in the initial model prediction stage; the generalized additive model is used as a proxy model of the Bayesian optimization algorithm, and the homology optimization method is used to track the local minimum along the deformation path. The ocean model is hyperparameter optimized through the improved Bayesian optimization algorithm, which improves the efficiency and effect of hyperparameter optimization, improves the efficiency of ocean model training, and shortens the time required for the entire process.

[0023] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0025] Figure 1 Schematic diagram of anomaly detection for offline models;

[0026] Figure 2 This is a block diagram of the overall processing of the ocean model in the first embodiment of the present invention;

[0027] Figure 3 This is a cold start response flow chart based on AutoML in Example 1 of the present invention;

[0028] Figure 4This is a schematic diagram of mixed time series cross validation in Embodiment 1 of the present invention;

[0029] Figure 5 This is a schematic diagram of the ocean model selection process in the first embodiment of the present invention;

[0030] Figure 6 This is a schematic diagram of homotopy optimization in Embodiment 1 of the present invention;

[0031] Figure 7 This is a block diagram of online training of the edge model in Example 1 of the present invention. DETAILED DESCRIPTION

[0033] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0034] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0035] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0036] Embodiment 1

[0037] This embodiment discloses a cloud-edge collaborative model automated training and online learning integration method, including:

[0038] Use historical ocean observation data to pre-train the ocean model offline in the cloud, and publish the pre-trained ocean model to the edge.

[0039] Based on real-time ocean observation data at the edge, the pre-trained ocean model is learned online to obtain the optimal ocean model.

[0040] Among them, in the offline training of the ocean model, an improved Bayesian optimization algorithm is used to optimize the hyperparameters of the ocean model. Specifically, the generalized additive model is used as a proxy model of the Bayesian optimization algorithm; the continuous deformation path of the generalized additive model is constructed using the homotopy optimization method, and the local minimum of the homotopy function is tracked along the continuous deformation path using the Nelder-Mead algorithm.

[0041] Combine the following Figure 2 The cloud-edge collaborative model automatic training and online learning integration method proposed in this embodiment is described in detail.

[0042] This embodiment establishes an efficient and real-time ocean data intelligent analysis and processing architecture based on AutoML. By combining cloud-side automatic machine learning with edge-side online learning, an efficient and accurate end-cloud-edge architecture for ocean time series prediction and detection is achieved.

[0043] Step 1: Use historical ocean data to train the ocean model offline in the cloud, and publish the pre-trained ocean model to the edge for online learning. The offline and online methods are combined to deal with the "cold start" problem of edge prediction.

[0044] Online learning algorithms often encounter cold start problems in the initial prediction stage of the model. Due to the lack of sufficient historical information, the initial performance of the algorithm will be greatly reduced. The lack of sufficient training data prevents the model from fully learning the patterns and rules in the data, thus limiting its ability to achieve accurate predictions and decision-making. To solve this problem, Figure 3 In this embodiment, a cloud server is used to pre-train the ocean model, and the ocean model is initialized through large-scale ocean data. Then, the trained ocean model is published to the edge to better deal with such cold start problems. With the high scalability of cloud computing resources, this strategy can perform deep learning and large-scale training of the model in the cloud, make full use of historical data, and improve the initial performance of the model.

[0045] This embodiment proposes to improve the intelligent analysis of edge data based on online learning methods. By introducing online learning to the edge for ocean data prediction, real-time online updates and real-time reasoning and prediction of the model can be achieved, thereby coping with the "concept drift" problem caused by offline learning. According to the changes in online data, the model is adjusted in real time so that it can accurately reflect online changes and improve the accuracy of online predictions.

[0046] This embodiment adopts the strategy of using a large amount of data to train the initial model in the cloud and deploying it to the edge for online prediction. By performing offline training in the cloud and publishing the model to the edge for online learning, the combination of offline and online effectively addresses the "cold start" problem of edge prediction.

[0047] In order to reduce costs, improve efficiency, and achieve full automation of the machine learning process, this embodiment optimizes the automatic machine learning strategy. Since the existing model selection method cannot adapt to the online learning mode, this embodiment proposes a hybrid time series cross-validation technology for model selection. By combining the rolling window and the extended window method, a hybrid cross-validation strategy is formed, focusing on the performance of the model over time, and supporting hyperparameter optimization over time to better select the best performing model. In terms of hyperparameter optimization, the traditional grid search has high computational cost and the Bayesian optimization starts slowly. This embodiment proposes a hyperparameter optimization method based on a GAM (generalized additive model) surrogate model. This strategy enhances the established optimization method (Bayesian optimization), by using the GAM surrogate model for learning, adjusting to the desired target, and then feeding back the hyperparameters to the Bayesian optimization for performance evaluation, and the two interact with each other to adjust to the optimal state. Subsequently, the optimal path is found using the homology method, in which the Nelder-Mead method is used to fit the path to ensure that each step is optimal. By this method, the optimization process can be quickly guided.

[0048] In the cloud offline training process, this embodiment also proposes two automatic machine learning (AutoML) optimization strategies to better realize process automation and improve cloud operation efficiency and the performance of the entire architecture.

[0049] In the cloud process, in order to improve the training efficiency and automate the entire machine learning process, such as Figure 4 As shown in Figure 1, a hybrid time series cross-validation strategy is proposed to achieve ocean model selection. This is a cross-validation method suitable for time series data. Unlike traditional cross-validation methods, time series cross-validation needs to consider the time dependency of the data, so it cannot randomly shuffle the data for segmentation like traditional methods.

[0050] This example combines the rolling window and expanding window methods to form a hybrid cross-validation strategy that focuses on the performance of the ocean model over time. This evaluation method supports hyperparameter optimization over time, thereby improving the prediction accuracy and stability of the ocean model. At the same time, it achieves automation and efficiency of the entire machine learning process.

[0051] like Figure 3 As shown, at each time step, the training data and validation data are used to optimize the hyperparameters of the ocean model; then, the ocean model with the best hyperparameters is trained on the combined training and validation data and evaluated on the test data. Finally, the performance graph of the ocean model is obtained by combining these time-varying metrics.

[0052] The existing cross-validation method divides all data sets, and as the amount of training data increases, the model performance may decline. Therefore, this embodiment adopts a hybrid method of the expansion strategy and the fixed window strategy. First, a fixed window size of n is proposed, and the collected ocean data is continuously input into the model according to the time node (minute or hour level). Initially, it starts with the first batch of data and expands until the window size is n. The window size will be fixed, and then move backward according to this window size of n. Not only can the initial data be effectively utilized to improve the initial prediction accuracy of the model, but also the stability and generalization ability of the model can be maintained after the window is fixed, thereby achieving more efficient and reliable time series prediction.

[0053] like Figure 4 As shown, for each time step, this embodiment divides the data into a training subset, a validation subset, and a test portion. The test portion uses the next batch of data, the validation subset uses the current batch of data, and the training subset uses all the data that can be contained in the window before the validation subset. Next, the model hyperparameters are searched for the training subset and the validation subset, and the best hyperparameters found are used to train and evaluate the model. The next batch of data is then evaluated as a test set, and the evaluation score is used as the final model performance measure. The best parameters are searched through Bayesian, and it is iterated multiple times for each subset; in the hyperparameter optimization process, such as Figure 5 As shown in Figure 1, the MSE indicator is used for model selection; after all iterations are completed, the average of the evaluation indicators is calculated to evaluate the overall performance of the model. Assume that K iterations are performed and the evaluation indicator for each iteration is MSE k , the average performance indicators are as follows:

[0054]

[0055]

[0056] In the large-scale training of cloud models, hyperparameter optimization is an indispensable part. Existing hyperparameter optimization techniques rely on naive search methods or assume that the loss function is smooth and continuous, but this is not always the case. The computational cost of traditional grid search is very high, and the startup speed of Bayesian optimization may be slow. Since the search space of hyperparameters is usually high-dimensional and non-convex, it is often challenging to effectively find the global minimum. In order to improve efficiency and better find the best model, this embodiment proposes a hyperparameter optimization method based on the GAM surrogate model. This strategy enhances the established optimization method, which is enhanced here on the Bayesian optimization.

[0057] First, GAM is a flexible regression model used to capture the nonlinear relationship between independent variables and dependent variables. Each independent variable has a corresponding smooth function, which makes the model more interpretable, and the influence of the independent variable can be understood by visualizing each smooth function. GAM builds a model by adding multiple nonlinear functions, so it has high flexibility when dealing with complex data relationships. Therefore, this embodiment selects the GAM model as the proxy model.

[0058] The basic form of a GAM is as follows:

[0059] g(E(Y))=β0+f1(x1)+f2(x2)+...+f p (x p ) (3)

[0060] Where Y is the dependent variable, E(Y) is the expected value, and g is the link function used to transform the expected value to the linear predictor. β0 is the intercept term, which is used when all x p When both are zero, it represents the expected value of the response variable Y. p (x p ) is an unknown smooth function used to capture the nonlinear relationship between the independent variable and the dependent variable. Each term f in the GAM model p (x p ) can be regarded as the corresponding hyperparameter x p The goal of Bayesian optimization is to adjust x p The value of is used to minimize the objective function E(Y) to find the optimal model configuration or hyperparameter combination.

[0061] Next, we build a hyperparameter optimization method based on the GAM model, construct continuous deformation between proxy models through the input of new ocean data, and use homotopy optimization to track local minima along the deformation. Figure 6 Homotopy optimization is a progressive optimization method that gradually approaches the optimal solution by defining a continuous transformation path from the initial problem to the target problem. The basic idea is to start from an easy-to-solve initial problem and gradually transform it to a complex target problem. At each step, the transformation process is guided by solving the local minimum of the current problem.

[0062] Bayesian optimization simulates the distribution of the objective function in the hyperparameter space by constructing a proxy function. Homotopy functions are mainly used to gradually guide the Bayesian optimization process, so that in the initial stage, the optimized objective function is relatively simple and easy to handle, and then gradually becomes more complex until it is finally optimized to the objective function g(x).

[0063] First, to construct the proxy model GAM, consider F(x) as the objective function of the unconstrained optimization problem, assuming that the local minimum x of F(x) has been calculated. Starting from the GAMf(x) at time point T1, when the ocean data is updated, a GAM g(x) suitable for time point T2 is created. Therefore, a continuous deformation path from f(x) to g(x) can be defined, and linear deformation can be used:

[0064] h(x,λ)=(1-λ)f(x)+λg(x) (4)

[0065] Among them, λ gradually changes from 0 to 1, h(x,0)=f(x), h(x,1)=g(x), h(x,λ) represents the homotopy function, which constructs a continuous change from function f(x) to function g(x).

[0066] The homology parameter λ gradually changes from 0 to 1, controlling how the function H(x,λ) transforms from f(x) to g(x). x is the optimization variable or parameter, usually representing the optimal solution to be found. In Bayesian optimization or other hyperparameter optimization processes, x represents possible hyperparameter combinations.

[0067] Using the homotopy optimization method, a continuous path between the local minimum x1 of the current proxy model f(x) and the local minimum x2 of the proxy model g(x) is found through the deformation path h(λ,λ). At each step, λ is updated and the local minimum of h(x,λ) is solved, gradually deforming to the target model g(x).

[0068] The homology function h(x,λ) gradually changes from the initial model to an updated model that includes the new ocean data. At each step, the Nelder-Mead optimization is used to minimize h(x,λ) to find the local minimum, and the optimal point x in the previous step is obtained. (k-1) Start by getting the optimal point x for the current step (k) , this process is repeated in homotopy optimization until λ reaches 1, thus finding the minimum value of the final objective function.

[0069] Nelder-Mead optimization is a non-gradient optimization algorithm that gradually adjusts the vertices of the simplex through reflection, expansion, and contraction operations to eventually find the local minimum of the objective function. In Bayesian hyperparameter optimization, Nelder-Mead optimization can be used as part of homotopy optimization to minimize the objective function on the homotopy path, thereby improving the efficiency and effectiveness of hyperparameter optimization.

[0070] At each fixed value of λ, the Nelder-Mead algorithm is used to optimize the current homotopy function h(x,λ).

[0071] Starting from λ=0 (the simpler auxiliary function f(x) is optimized at this time), the Nelder-Mead method is used to find the optimal solution of the current homology function.

[0072] Once a local optimum is found for the current value of λ, slightly increase the value of λ and re-apply the Nelder-Mead algorithm to optimize the new homotopy function h(x,λ).

[0073] Each time the value of λ is reduced, the optimization objective is gradually transitioned from a simple function f(x) to a complex function g(x).

[0074] By applying Nelder-Mead optimization on each transition step, the optimal solution of the gradually changing homotopy function is found.

[0075] The Nelder-Mead algorithm is a local search method that does not require the calculation of gradient information. It gradually approaches the local minimum of the objective function, i.e., the homology function h(x,λ), by moving vertices of simple shapes (such as reflection, expansion, contraction, and shrinkage).

[0076] Among them, the Nelder-Mead algorithm includes four main operations:

[0077] x r =x c +α(x c -x h ) (5)

[0078] The purpose of the reflection operation is to calculate the reflection point xr based on the worst point xh and the centroid xc. This exploration can help the algorithm get out of the local optimum and expand the search range. r : Reflection point, obtained by reflecting the worst point, x c is to remove the worst point x h The center of the remaining vertices, x h It is the worst point, and α is usually taken as 1.

[0079] x e =x c +γ(x r -x c ) (6)

[0080] When the reflection point has shown the potential to be better than the current solution, the expansion operation will be triggered. Its purpose is to explore a wider range of search directions to ensure that no potential optimal points are missed. e : Extension point. When the reflection point is better than the optimal point x1, the point is obtained by further extending outward. γ is usually taken as 2.

[0081] x c =xc +ρ(x h -x c ) (7)

[0082] When reflection or expansion fails to improve the objective function value, the exploration range is reduced through contraction operation to avoid invalid search process and make it easier for the algorithm to focus on the possible optimal solution. c : represent the contraction point respectively, and ρ is usually taken as 0.5.

[0083] x i =x l +σ(x i -x l ) (8)

[0084] If the previous operation does not improve the situation, the overall contraction is performed to reduce the size of the search area, focus the algorithm on the current optimal area, and improve the stability and convergence of the algorithm. i : The coordinates of the current search point. All points form a simple shape (e.g., triangle, tetrahedron), x l is the optimal point, and σ is usually taken as 0.5.

[0085] x1, x2, …, x h : The vertices in the simple shape sorted by the objective function value. x l is the optimal point (the function value is the smallest), x h These points are calculated by the Nelder-Mead algorithm and returned to the homotopy function, which is then used to calculate the objective function value f(x) at point x.

[0086] In the Bayesian optimization process, we hope to find the hyperparameter combination that makes the objective function the best. However, actually evaluating the objective function can be very expensive, so GAM is used as a proxy model to approximate the behavior of the objective function in order to quickly evaluate the performance of different hyperparameter combinations. GAM model-based hyperparameter optimization is used as an external strategy in the Bayesian optimization method to achieve efficient space exploration and converge to the best solution faster.

[0087] Take the hyperparameters as input, train a GAM model, and evaluate the performance of the model (such as MSE) on the training set as the output of the objective function. Use the Bayesian optimizer to select the next most promising hyperparameter combination for evaluation based on the objective function evaluation of the previous step. Whenever there is a new hyperparameter combination and its evaluation result, update the GAM surrogate model to more accurately estimate the objective function in the next round of Bayesian optimization.

[0088] Step 2: A dynamic ensemble learning method based on stacking is proposed to deploy the optimal model to the edge. The edge model is uploaded to the cloud and integrated with the best model trained offline with full data on the cloud to address the "catastrophic forgetting" problem in online learning.

[0089] There is a phenomenon of "concept drift" in ocean data, that is, the distribution of data changes unpredictably over time, resulting in the model's inability to adapt to such changes in real time, which in turn affects the accuracy of the results. If not handled, concept drift may have a significant impact on the analysis and prediction of ocean data. Traditional deep learning methods inject new knowledge into the model by regularly updating the model, but the distribution of data may change in the period between updates, causing the model's performance on new data to decline, thus posing a great threat to the ocean observation system. Therefore, online learning, which performs fast and instant learning through data streams, came into being.

[0090] This embodiment improves the intelligent analysis capability of ocean data based on online learning methods, introduces online learning to the edge to predict ocean data, and addresses the "concept drift" problem caused by offline learning.

[0091] Figure 7 In this embodiment, edge online training, updating, and prediction are used to cope with real-time changes. The ocean model can be continuously trained by processing real-time data to adapt to the changing data distribution.

[0092] In addition, this embodiment also achieves the continuous stability of the edge model by combining it with offline learning in the cloud, with MSE and R 2 As a standard for measuring model performance, when both indicators fail to reach the standard values, recent data is immediately obtained to re-perform model training and model selection in the cloud, and a model with the best performance is selected, so that the best network model with dynamic optimization can be obtained at the edge.

[0093] In addition, this embodiment prevents outliers from having a negative impact on the online learning model selection decision. Outliers often interfere with MSE and R 2 Beyond normal standards, use z-score to identify outliers.

[0094]

[0095] Among them, x is the error value between the predicted value and the actual value, μ is the mean of the error, and σ is the standard deviation of the error. Usually, data points whose z-score exceeds a certain threshold (such as 3 or -3) are considered outliers. In order to calculate the z-score online, it is necessary to incrementally update the mean and standard deviation of the error. For each newly arrived data point, calculate its z-score and determine whether it is an outlier. If an outlier is detected, you can choose to ignore the outlier until you actually detect a performance drop before reselecting the model.

[0096] Long-term online learning may lead to catastrophic forgetting. The continuous loss of useful knowledge may threaten the long-term performance and reliability of the ocean prediction system, resulting in unstable and inaccurate prediction results. In order to deal with this problem, this embodiment proposes a method that combines offline training with online learning. The model on the edge is uploaded to the cloud and integrated with the best model trained offline with full data on the cloud. This embodiment adopts a dynamic integrated learning method based on stacking to deploy the optimal model to the edge to deal with the problem of "catastrophic forgetting" in online learning.

[0097] During the online learning process, the best performing model is dynamically evaluated and selected, and stored as a base learner for ensemble learning. The meta-learner obtains the real-time model of online learning through this ensemble learning to ensure the consistency and stability of online learning. A dynamic meta-learner ensemble learning method is formed to solve the problem of catastrophic forgetting. In order to ensure that the model remains stable and adaptable in long-term operation, a regular learning strategy can be adopted, such as systematic model update and optimization once a year.

[0098] Based on stacking ensemble learning, the previous knowledge is retained through the base learners of the first layer (such as LSTM, TCN, RNN, etc.), and the input data is predicted through the base learners. These prediction results are used as features to train the meta-learner (such as CNN, etc.) so that it can integrate these features for prediction. After these learning results are input into the second-layer meta-learner (such as CNN, etc.), the meta-learner is published to the edge for online learning. This method can effectively solve the problem of catastrophic forgetting. Specifically, the base learner is responsible for capturing the features and knowledge in historical tasks, and the meta-learner is responsible for integrating these features and using them for learning new tasks. In the online learning stage, the meta-learner can be used alone for the prediction and update of new data.

[0099] This embodiment uses a combination of offline and online methods to address the "cold start" problem of edge prediction. At the same time, in order to automate the process, this embodiment proposes mixed time cross-validation to effectively evaluate the performance of the time series data model, maintain the time order, make full use of the data, evaluate the stability of the model, identify trends and seasonal changes in the data, and improve the robustness of the model. Through methods such as rolling cross-validation and extended window cross-validation, the consistency and reliability of the model in different time periods can be ensured, providing strong support for time series data modeling. A hyperparameter optimization method based on GAM is also proposed. This method combines the flexibility of GAM and the progressiveness of homology optimization. Through gradual deformation and optimization, it can effectively deal with objective functions with high computational costs. In the process of continuous data updating, it can smoothly transition from the current model to the updated model containing new data, thereby improving the stability and adaptability of the model.

[0100] This embodiment introduces online training of edge models. By combining online training, updating, and prediction on the edge with offline learning on the cloud, the model can be continuously optimized and stabilized, real-time prediction performance can be improved, concept drift problems can be solved, and the robustness and adaptability of the system can be enhanced. This approach makes full use of edge and cloud resources to ensure the efficiency and stability of the model in long-term operation.

[0101] This embodiment uses stacking-based ensemble learning to dynamically reselect and store the best model through online learning, combined with a periodic learning strategy, to effectively improve the adaptability and stability of the model, solve the problem of catastrophic forgetting, enhance the robustness of the system, and improve the overall prediction performance. This method not only ensures that the model responds to new data in a timely manner, but also retains the memory of historical data, thereby maintaining efficient and stable prediction capabilities in long-term operation.

[0102] Embodiment 2

[0103] The purpose of this embodiment is to provide a cloud-edge collaborative model automated training and online learning integrated system, including:

[0104] The offline training unit is used to pre-train the ocean model offline in the cloud using historical ocean observation data, and publish the pre-trained ocean model to the edge. In the offline training of the ocean model, the improved Bayesian optimization algorithm is used to optimize the hyperparameters of the ocean model, specifically: the generalized additive model is used as a proxy model of the Bayesian optimization algorithm; when the hyperparameters are updated, the deformation path of the generalized additive model is constructed, and the local minimum value is tracked along the deformation path using the homotopy optimization method;

[0105] The online learning unit is used to perform online learning of the pre-trained ocean model based on real-time ocean observation data at the edge to obtain the optimal ocean model.

[0106] Embodiment 3

[0107] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0108] Embodiment 4

[0109] The purpose of this embodiment is to provide a computer-readable storage medium.

[0110] A computer-readable storage medium stores a computer program, which executes the steps of the above method when executed by a processor.

[0111] =The steps involved in the apparatus of the above embodiments 2, 3 and 4 correspond to the method embodiment 1. For the specific implementation, please refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method of the present invention.

[0112] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0113] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. The cloud-edge collaborative model automated training and online learning integration method is characterized by: include: Use historical ocean observation data to pre-train the ocean model offline in the cloud, and publish the pre-trained ocean model to the edge. Based on real-time ocean observation data at the edge, the pre-trained ocean model is learned online to obtain the optimal ocean model. Among them, in the offline training of the ocean model, an improved Bayesian optimization algorithm is used to optimize the hyperparameters of the ocean model. Specifically, the generalized additive model is used as a proxy model of the Bayesian optimization algorithm; the continuous deformation path of the generalized additive model is constructed using the homotopy optimization method, and the local minimum of the homotopy function is tracked along the continuous deformation path using the Nelder-Mead algorithm.

2. The cloud-edge collaborative model automated training and online learning integration method according to claim 1, characterized in that: For the ocean model trained online at the edge, the mean square error index and the coefficient of determination index are used to measure the performance of the ocean model. If the ocean model trained online does not meet the mean square error index and the coefficient of determination index, the cloud obtains real-time ocean observation data to train the ocean model and obtain the optimal ocean model.

3. The cloud-edge collaborative model automated training and online learning integration method according to claim 1, characterized in that: The ocean model that meets the evaluation criteria during the online learning process is uploaded to the cloud as a base learner. The base learner and the best ocean model trained using the full amount of ocean observation data in the cloud are integrated, and the optimal ocean model is deployed to the edge using a dynamic integrated learning method called stacking.

4. The cloud-edge collaborative model automated training and online learning integration method according to claim 1, characterized in that: In offline training, a hybrid time series cross-validation strategy is used to select the ocean model, specifically: Determine the size of the fixed window; Expand the training data input at each time step based on the size of the fixed window; Divide the expanded training data into training subset and validation subset at each step; The training subset and validation subset are used to search for the ocean model hyperparameters, and the best hyperparameters found are used to train and evaluate the ocean model.

5. The cloud-edge collaborative model automated training and online learning integration method according to claim 1, characterized in that: The continuous deformation path of the generalized additive model is constructed using the homotopy optimization method, and the local minimum of the homotopy function is tracked along the continuous deformation path using the Nelder-Mead algorithm, specifically: Using homotopy optimization methods, we create continuous deformation paths before and after the ocean observation data is updated; The Nelder-Mead algorithm is used to optimize the current homotopy function under each fixed homotopy parameter to find the local optimal solution.

6. The cloud-edge collaborative model automated training and online learning integrated method as described in claim 1 or 5, characterized in that: Under each fixed homotopy parameter, the local minimum of the objective function is gradually approached through the reflection, expansion, contraction and shrinkage operations of the Nelder-Mead algorithm.

7. Cloud-edge collaborative model automated training and online learning integrated system, characterized by including: The offline training unit is used to pre-train the ocean model offline in the cloud using historical ocean observation data, and publish the pre-trained ocean model to the edge. In the offline training of the ocean model, the improved Bayesian optimization algorithm is used to optimize the hyperparameters of the ocean model, specifically: the generalized additive model is used as a proxy model of the Bayesian optimization algorithm; when the hyperparameters are updated, the deformation path of the generalized additive model is constructed, and the local minimum value is tracked along the deformation path using the homotopy optimization method; The online learning unit is used to perform online learning of the pre-trained ocean model based on real-time ocean observation data at the edge to obtain the optimal ocean model.

8. The cloud-edge collaborative model automated training and online learning integrated system as claimed in claim 7, characterized in that: In the offline training unit, a continuous deformation path of the generalized additive model is constructed using a homotopy optimization method, and a local minimum value of the homotopy function is tracked along the continuous deformation path using a Nelder-Mead algorithm, specifically: Using homotopy optimization methods, we create continuous deformation paths before and after the ocean observation data is updated; The Nelder-Mead algorithm is used to optimize the current homotopy function under each fixed homotopy parameter to find the local optimal solution.

9. A computer device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate through the bus. When the machine-readable instructions are executed by the processor, the cloud-edge collaborative model automatic training and online learning integration method as described in any one of claims 1 to 6 is performed.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the cloud-edge collaborative model automation training and online learning integration method as described in any one of claims 1 to 6.

Citation Information

Cited By

  • Underwater robot energy consumption optimization path planning method

    CN119714304A