Autoencoder Restoration of Missing Time-Series Values with Limited Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for missing value restoration, such as historical average and linear interpolation, are inadequate in environments with highly variable energy consumption patterns or when large amounts of data are missing, while deep learning-based methods like GANs are time-consuming and unstable, especially with limited data.
Innovation Solution
A method using a trained autoencoder that generates and processes additional missing values based on patterns in time series data, minimizing mean squared error to restore missing values accurately, integrating denoising and masked autoencoder techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If GAN-based method is used for missing value restoration, then restoration accuracy is improved, but training time increases and learning stability deteriorates
Solution Approach 1:
The patent replaces the complex GAN-based deep learning system with a simplified autoencoder architecture that uses mean squared error optimization. This substitution maintains restoration accuracy while dramatically reducing training time and improving learning stability, as the autoencoder does not require the adversarial training process that characterizes GANs.
Solution Approach 2:
The patent changes the optimization parameter from the complex adversarial loss function used in GANs to the simpler mean squared error function. This parameter change simplifies the training process, reduces computational complexity, and eliminates the instability associated with GAN training while maintaining effective missing value restoration.
2Productivity
If DAE is used for missing value restoration, then learning speed is improved, but applicability deteriorates when complete DLP data is insufficient
Solution Approach 1:
Instead of requiring complete data as input and outputting restored data (conventional DAE approach), the patent inverts the problem by inputting data with missing values and directly optimizing to restore them using mean squared error. This inversion allows the model to train effectively even when complete DLP data is insufficient, as it learns from the actual incomplete data structures rather than requiring synthetic complete data.
Solution Approach 2:
The patent creates a universal solution that works across different data completeness scenarios. The autoencoder with mean squared error optimization can handle environments with varying amounts of complete data availability, making it universally applicable unlike DAE which specifically requires sufficient complete DLP data for training.
3Stability of the object's composition
If MAE is used for missing value restoration, then training stability is improved, but restoration performance deteriorates when similar patterns are difficult to find
Solution Approach 1:
The patent replaces the pattern-matching mechanism of MAE with a direct optimization approach using mean squared error. Instead of relying on finding similar patterns in the data, the model directly minimizes the reconstruction error, providing stable training while maintaining high restoration performance even when similar patterns are scarce.
Data Source
AI summary
An operating method of an electronic device may include preprocessing time series data including a first missing value, obtaining output data by inputting the preprocessed time series data into an autoencoder, and training the autoencoder based on the output data and the time series data. The preprocessed time series data may include the first missing value and a second missing value generated as a preprocessing result.