Variational Auto-Encoder for Synthetic Time-Series Label Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in accurately analyzing time-series data without access to sufficient annotated or labeled ground-truth data, which is often unavailable or limited, especially in dynamic learning environments where data accrues over time.
Innovation Solution
A method and system for label generation in time-series classification, utilizing a variational auto-encoder (VAE) to iteratively generate synthetic time-series data and associated labels, with a label quality checker model to assess and improve label quality through self-correction mechanisms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If domain experts manually annotate time-series data for training and validation, then label quality is improved, but time consumption and cost increase significantly
Solution Approach 1:
The patent uses Variational Auto-Encoders to generate synthetic copies of time-series data that replicate the statistical properties and patterns of real data. These synthetic copies serve as training samples with automatically generated labels, eliminating the need for manual annotation while preserving data characteristics needed for model training
Solution Approach 2:
The system implements self-supervised learning where the model generates its own training data and labels through the VAE framework. The iterative process automatically creates labeled samples without external human intervention, with the model improving its own training dataset through synthetic data generation and label propagation
2Adaptability or versatility
If more time-series data is accrued in dynamic learning environments, then model adaptability is improved, but the need for continuous manual labeling increases
Solution Approach 1:
The patent implements an iterative process where synthetic data generation and label creation continue automatically as new real data arrives. The system continuously updates the training dataset by generating synthetic samples from the evolving data distribution, maintaining model adaptability without interrupting the data accrual process for manual labeling
Solution Approach 2:
The system adapts dynamically to changing data distributions by retraining the VAE on accumulated data and generating new synthetic samples that reflect current patterns. The iterative framework allows the model to evolve with accruing data, adjusting the synthetic data generation process to match changing temporal patterns and characteristics
3Quantity of substance
If synthetic data is generated using VAE in multiple iterations, then data volume is improved, but computational complexity increases
Solution Approach 1:
The patent generates synthetic data in controlled iterations, creating only the necessary amount of synthetic samples needed to reach target data volumes. Each iteration generates a subset of synthetic data that is progressively added to the training set, avoiding unnecessary computational overhead while achieving sufficient data quantity for model training
Solution Approach 2:
The synthetic data generation process is divided into discrete iterative steps, where each iteration handles a manageable portion of data generation. This segmentation allows the system to process and integrate synthetic samples incrementally, reducing memory requirements and enabling parallel processing of different iteration batches
4Productivity
If labels are generated automatically with iterative appending, then labeling speed is improved, but noisy labels increase
Solution Approach 1:
The patent incorporates a label quality checker that provides feedback on the accuracy of generated labels. The system monitors label quality metrics and uses this feedback to adjust the synthetic data generation process, filtering out low-quality synthetic samples and their associated noisy labels while retaining high-quality additions to the training set
Data Source
AI summary
This disclosure relates generally to method and system for time series classification. Conventional methods for time-series classification requires substantial amount of annotated data for classification and label generation. The disclosed method and system are capable of generating accurate labels for time-series data by utilizing a small amount of representative data for each class. In an embodiment, the disclosed method generates a time-series data synthetically and associated labels by using a portion of the representative time-series data in each iteration, and self-correcting the generated labels based on a determination of quality of the generated labels using label quality checker models.