PVS-GEN: Systematic Approach for Universal Synthetic Data Generation Involving Parameterization, Verification, and Segmentation

Kyung-Min Kim, Jong Wook Kwak2024 · DOI

TL;DR

This paper introduces PVS-GEN, a novel, automated system for generating universal synthetic time-series sensor data. By parameterizing, verifying, and segmenting data, PVS-GEN achieves high reproducibility across diverse sensor types, addressing limitations in current synthetic data generation methods.

What Problem Does This Paper Address?

Generating high-quality synthetic data for time-series sensors is crucial for various applications, from improving AI model robustness to addressing privacy concerns. However, existing methods struggle with:

  • Lack of Universality: Many approaches are tailored to specific data types or sensor characteristics, making them inflexible.
  • Complexity of Real-World Data: Time-series data, especially from sensors with high frequency, significant variation, and irregular patterns (Type-II data), is challenging to model accurately.
  • Fragmented Research: The field is fragmented, with a lack of standardized metrics for evaluating synthetic data quality, hindering consistent comparisons.
  • Manual Effort: Current methods often require significant human intervention for parameter tuning and model selection.

This paper aims to bridge this gap by proposing a systematic, automated, and universally applicable approach to synthetic time-series sensor data generation.

Methods

The core of PVS-GEN is a three-step process:

  1. Parameterization: This initial step configures a "SyNode" data frame, which acts as a foundational structure for the synthetic data. It extracts ARIMA parameters from empirical data to generate initial synthetic data, minimizing human intervention.
  2. Verification: PVS-GEN introduces a novel metric called the Possibility of Reproducibility (PoR). This metric quantifies the similarity between empirical and synthetic data by comparing their frequency domain representations (obtained via Discrete Fourier Transform - DFT). A PoR score close to 1 indicates high similarity.
  3. Segmentation: If the generated data fails to meet the PoR criterion, the system employs a change-point detection algorithm to segment the time-series data. This allows for more localized and accurate modeling of segments with distinct characteristics. This process is recursive, iteratively refining parameters and segmentation until satisfactory reproducibility is achieved.

The system also incorporates strategies to manage segmentation overhead, such as setting a maximum depth and allowing flexible adjustment of the PoR threshold.

Datasets

The study evaluated PVS-GEN using several real-world time-series datasets from the UCI Machine Learning Repository:

  • Gas Sensor Array Dataset: Data from a PTFE test chamber exposed to CO and humid synthetic air.
  • Low-Energy House Dataset: Energy usage data from appliances, including temperature and humidity measurements.
  • EEG Alcoholism Dataset: EEG signals from subjects under various conditions.
  • Heterogeneity Activity Recognition Dataset: Readings from accelerometer and gyroscope sensors in smartphones and smartwatches.

These datasets represent a mix of Type-I (low volatility) and Type-II (high volatility) time-series data, allowing for a comprehensive evaluation of PVS-GEN's universality.

Key Insights and Ideas

  • The PoR Metric: The introduction of the PoR metric is a significant contribution, providing a standardized, quantitative way to assess synthetic data quality that goes beyond simple descriptive statistics. It effectively captures temporal characteristics.
  • Segmentation for Complexity: The iterative segmentation approach is key to handling complex, high-volatility time-series data (Type-II). By breaking down data into more homogeneous segments, PVS-GEN can model intricate patterns that single models often miss.
  • Automation is Key: The system's ability to automate parameter extraction and segmentation significantly reduces the manual effort typically required in synthetic data generation.
  • Universal Applicability: The PVS-GEN framework is designed to be independent of specific sensor types, demonstrating its potential for broad application across various IoT and sensor data scenarios.

Key Results

  • PVS-GEN achieved a high average PoR value of 0.83, indicating strong reproducibility.
  • Compared to other methods (SES, ARIMA, SVR, LSTM), PVS-GEN showed an increase in PoR values of up to 37.1% and 19.1% on average.
  • The method demonstrated superior performance on more complex Type-II data types where existing methodologies struggled.
  • PVS-GEN maintained a reasonable computation time, outperforming machine learning methods in terms of efficiency while achieving higher reproducibility.

Limitations and Future Work

  • Computational Complexity: While PVS-GEN is more efficient than some ML methods, the iterative segmentation process can still introduce computational overhead, especially for very long time-series. The paper notes that more segmentations can improve reproducibility but increase execution time.
  • Optimal Segmentation Length: The study identified an approximate optimal segmentation length of 250 points for balancing execution time and reproducibility. However, finding this optimal length might require further tuning.

Future work will focus on further optimizing the modeling approach, exploring detailed components of data generation, parameter extraction, and change-point detection to minimize time overhead and enhance reproducibility.

Summary updated

BibTeX
@misc{kim2024pvs,
  author = {Kyung-Min Kim and Jong Wook Kwak},
  title = {{PVS-GEN: Systematic Approach for Universal Synthetic Data Generation Involving Parameterization, Verification, and Segmentation}},
  year = {2024},
  doi = {10.3390/s24010266},
}

Loading related content...

Figure