Source-linked AI summary
Time Series Data Augmentation for Deep Learning: A Survey
Qingsong Wen, Liang Sun, Fan Yang, Xiaomin Song, Jingkun Gao, Xue Wang, Huan Xu
TL;DR
Deep learning time-series applications often lack sufficient labeled data, motivating task- and structure-aware augmentation. This paper builds a taxonomy and systematic survey, evaluates methods across classification, anomaly detection, and forecasting, and identifies future research directions. Overall, the review reports promising augmentation benefits but also task- and setting-dependent limitations.
Problem
Many time-series applications lack enough labeled data, while existing augmentation methods do not fully use temporal properties or generalize across tasks.
Method
The paper proposes a taxonomy, systematically surveys basic and advanced augmentation methods, and empirically compares them across classification, anomaly detection, and forecasting.
Results
Augmentation improves anomaly detection and produces promising average forecasting results, although negative outcomes occur for specific data-model pairs.
Takeaways & Limitations
The survey organizes methods and identifies five future directions, including time-frequency augmentation, augmentation selection and combination, Gaussian processes, and deep generative models.
Takeaways & Limitations
Flipping retains labels only when upward and downward directions are symmetric.
Abstract
from arXiv · showhide
Deep learning performs remarkably well on many time series analysis tasks recently. The superior performance of deep neural networks relies heavily on a large number of training data to avoid overfitting. However, the labeled data of many real-world time series applications may be limited such as classification in medical time series and anomaly detection in AIOps. As an effective way to enhance the size and quality of the training data, data augmentation is crucial to the successful application of deep learning models on time series data. In this paper, we systematically review different data augmentation methods for time series. We propose a taxonomy for the reviewed methods, and then provide a structured review for these methods by highlighting their strengths and limitations. We also empirically compare different data augmentation methods for different tasks including time series classification, anomaly detection, and forecasting. Finally, we discuss and highlight five future directions to provide useful research guidance.
1 Introduction
Time-series deep learning needs substantial training data, yet many applications lack enough labels and require augmentation tailored to temporal structure and task differences. The paper addresses these gaps with a taxonomy, systematic review across common tasks, empirical evaluations, and future directions.
- Deep learning for time-series tasks relies on abundant training data to avoid overfitting, but many real-world applications lack enough labeled examples.
- Time-series augmentation must account for temporal dependency, transformed frequency and time-frequency domains, and complex multivariate dynamics.Methods transferred directly from image or speech processing may not generate valid synthetic time series.
- Existing reviews had not comprehensively covered time-series augmentation across classification, forecasting, and anomaly detection, nor highlighted future research opportunities.
- The paper proposes a taxonomy and systematically reviews augmentation methods for forecasting, anomaly detection, and classification.
- The taxonomy covers time, frequency, and time-frequency transformations alongside decomposition-based, model-based, embedding-space, generative, and automated methods.The paper also evaluates methods on three typical tasks and identifies five future directions.
2 Basic Data Augmentation Methods
Basic time-series augmentation methods operate in the time, frequency, and time-frequency domains, using direct perturbations, spectral modifications, and feature-space transformations. The reviewed methods include task-specific techniques and report improvements, while also exposing assumptions and inconsistent effects across settings.
- Time Domain: Time-domain methods directly manipulate input series through noise injection and other perturbations, while label expansion targets anomaly detection.Label expansion changes points near labeled anomalies into anomalous examples when temporal and value proximity makes them likely anomalous.
- Time Domain: Window cropping randomly extracts continuous subsamples, preserving the original labels for classification and slicing anomaly labels with the value series.Classification testing uses majority voting across slices.
- Time Domain: Window warping randomly compresses or extends a selected time range, changes total sequence length, and is therefore paired with window cropping for deep learning models.
- Time Domain: Flipping negates each time-series value while retaining labels under the assumption that upward and downward directions are symmetric.
- Frequency Domain: Frequency-domain augmentation perturbs amplitude and phase spectra after Fourier transformation, and APP combined with time-domain methods improves time-series anomaly detection.Amplitude segments receive Gaussian-noise replacements, while phase segments receive zero-mean Gaussian noise.
- Time-Frequency Domain: Time-frequency augmentation applies transformations to representations such as STFT features, including local averaging, feature-vector shuffling, warping, and masking.STFT-based augmentation was used for human activity classification with a deep LSTM, while SpecAugment targets speech recognition.
3 Advanced Data Augmentation Methods
The survey organizes advanced time series augmentation into decomposition-based, statistical generative, and learning-based methods. These methods manipulate components, model temporal dynamics, or generate samples in learned spaces and through automated policy search.
- 3.1 Decomposition-based Methods: Decomposition methods separate time series into trend, seasonal, and remainder signals before generating or augmenting samples.The decomposition is represented as x_t = τ_t + s_t + r_t.
- 3.1 Decomposition-based Methods: Decomposition-based augmentation has improved forecasting and anomaly detection, including augmentation applied to robustly decomposed residuals.The survey reports significant anomaly-detection gains versus the same method without augmentation.
- 3.2 Statistical Generative Models: Statistical generative models simulate time series by modeling their dynamics and conditional distributions over successive values.Reviewed examples include mixture of Gaussian trees, LGT, and mixture autoregressive models.
- 3.3 Learning-based Methods: Learning-based methods include embedding-space transformations, deep generative models, and automated augmentation policies.Embedding-space methods interpolate or extrapolate between nearby encoded samples, while automated methods search for transformation compositions.
- 3.3 Learning-based Methods: Recurrent GANs and TimeGAN generate realistic time series, with TimeGAN combining an embedding mapping, adversarial training, and supervised stepwise losses.RCGAN can also be trained with differential privacy for sensitive domains.
4 Preliminary Evaluation
The preliminary evaluations examine augmentation in classification, anomaly detection, and forecasting. Across these tasks, augmentation generally improves performance, although forecasting results can vary by data and model pair.
- 4.1 Time Series Classification: Classification evaluation uses 5,000 labeled one-week time series and tests robustness under injected spike, step, and slope outliers.A fully convolutional network is trained with an 80% training split.
- 4.2 Time Series Anomaly Detection: Anomaly-detection comparisons evaluate raw data, decomposed residuals, and augmented decomposed residuals using precision, recall, and F1 score.The summarized results use a U-Net-based network on the Yahoo! dataset.
- 4.3 Time Series Forecasting: Augmentation produces promising average MASE improvements across DeepAR and Transformer models on electricity, traffic, and M4 datasets.Negative results remain for specific data/model pairs, motivating automated policies that stabilize forecasting effects.
5 Discussion for Future Opportunities
The survey identifies future opportunities in time-frequency augmentation, imbalance-aware methods, augmentation selection, alternative generative models, and computational efficiency. These directions address underexplored transforms, class imbalance, policy design, and model diversity.
- Time-Frequency Augmentation: Wavelet transforms such as CWT, DWT, and MODWT are proposed as underexplored tools for time-frequency augmentation.The survey notes their ability to characterize time-varying properties and handle non-stationary series and non-Gaussian noise.
- Imbalanced Classification: Jointly applying data augmentation and class weighting is proposed for imbalanced time series classification.The survey cautions that oversampling can change the raw-data distribution and cause overfitting.
- Augmentation Selection and Efficiency: Future work should develop time-series-specific strategies for selecting and combining augmentation methods, potentially using reinforcement or meta learning.The survey also identifies algorithm efficiency as an important practical consideration.
- Generative Models: Alternative deep generative models, including DARNs, normalizing flows, and VAEs, remain less-investigated opportunities beyond GAN-based augmentation.Deep autoregressive networks are highlighted as naturally aligned with sequential generation and causal temporal structure.
6 Conclusion
The paper provides a comprehensive survey of time series data augmentation across tasks, organizes methods into a taxonomy, empirically compares them, and identifies future research directions.
- 6 Conclusion: The survey covers augmentation methods for forecasting, anomaly detection, and classification through a taxonomy of basic and advanced approaches.It summarizes representative methods, compares them empirically in typical tasks, and highlights future research directions.