Edge Deep Learning Model Adaptation for Local Weather Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large-scale deep learning models face limitations when applied directly to edge computing devices due to constrained computing resources, and methods to make them lightweight face challenges in providing accurate weather predictions due to data dependency and environmental differences.
Innovation Solution
An edge computing device optimizes a deep learning model by measuring similarity between installation environment data and training data, generating training data based on reliability information, and updating the model using a loss function that incorporates meteorological agency information when available.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing large-scale deep learning models are applied directly to edge computing devices, then prediction accuracy is maintained, but computing resource constraints are violated
Solution Approach 1:
The patent segments the deep learning model into a lightweight base model deployed on the edge device and a separate fine-tuning mechanism that adapts to local environmental data. This allows the core prediction functionality to remain on the constrained edge device while complex computations are performed remotely or in advance, resolving the contradiction between maintaining accuracy and respecting resource constraints.
Solution Approach 2:
The patent changes the parameters of the deep learning model by performing fine-tuning with local environmental data specific to each edge device's installation environment. This adapts the model to local conditions without requiring the full large-scale model to be deployed, thus maintaining accuracy while reducing computational burden on the edge device.
2Device complexity
If deep learning models are made lightweight for edge devices, then computing resource constraints are satisfied, but prediction accuracy deteriorates due to lack of environmental context
Solution Approach 1:
The patent performs preliminary actions by collecting and processing environmental data specific to each edge device's installation location before the model is deployed. This pre-processing of local environmental information allows the lightweight model to achieve higher accuracy without requiring complex computations during actual prediction, as the environmental context is already prepared and integrated into the model's training data.
Solution Approach 2:
The patent implements feedback mechanisms where the lightweight model's predictions are continuously refined using local environmental data and performance information. This feedback loop allows the model to adapt to local conditions over time, improving accuracy without increasing the model's base complexity or requiring additional computational resources for the core prediction function.
3Adaptability or versatility
If deep learning models are trained with general training data, then model versatility is improved, but adaptability to specific installation environments deteriorates
Solution Approach 1:
The patent applies local quality by training the lightweight model with general data for versatility while simultaneously fine-tuning it with locally-specific environmental data from each edge device's installation environment. This dual approach ensures the model maintains both general adaptability and local reliability, as it incorporates both universal patterns and environment-specific characteristics into its predictions.
Data Source
AI summary
Proposed is a method of optimizing a deep learning model, which is performed by an edge computing device. The method may include measuring a similarity between data collected from an installation environment of the edge computing device and training data used for training a deep learning model installed in the edge computing device. The method may also include determining whether to perform optimization of the deep learning model based on a measurement result of the similarity. The method may further include generating training data for performing the optimization of the deep learning model based on reliability information, and training and generating a deep learning model based on the training data. The method may further include updating the existing deep learning model applied to the edge computing device with the generated deep learning model.