Wavelet-based multi-frequency image transformation for time series classification
Öz
Context—Time series classification (TSC) plays a critical role in many application domains such as healthcare, fault diagnosis, and signal analysis, where signals are often non-stationary and noisy. Although convolutional neural networks (CNNs) have achieved strong performance, their effectiveness is limited when operating directly on raw signals, particularly in capturing multi-frequency characteristics. Image-based representations such as Gramian Angular Field (GAF), Markov Transition Field (MTF), and Recurrence Plot (RP) have enabled CNNs to exploit temporal structures more effectively. However, most existing approaches generate these images directly from raw signals, leaving the potential of frequency-aware representations underexplored.
Objective—The objective of this study is to develop a hybrid time series classification framework tailored for sensor-based signals with rich frequency content. In particular, the study investigates whether generating GAF, MTF, and RP images from different Discrete Wavelet Transform (DWT) components can enhance the representation quality of frequency-dependent time series. By assigning distinct image encodings to approximation and detail coefficients, the proposed approach aims to improve dataset representations by capturing complementary information across multiple frequency bands commonly observed in sensor data. The effectiveness of the framework is evaluated on benchmark datasets from the UCR archive.
Method—In the proposed method, each time series is decomposed using a two-level Discrete Wavelet Transform with the db4 wavelet to obtain approximation and detail coefficients. These coefficients are then independently transformed into image representations: MTF from approximation coefficients, GAF from medium-frequency components, and RP from high-frequency components. The resulting images are combined into multi-channel inputs and classified using lightweight CNN architecture. Additional experiments using direct image transformation and multi-branch and multi-channel CNN designs are conducted to assess the contribution of wavelet-based representations.
Results—Experimental results on multiple UCR benchmark datasets demonstrate that the proposed DWT-based image representations improve classification accuracy, particularly for long and noisy time series. The wavelet-based approach shows clear advantages in datasets characterized by complex temporal dynamics and low signal-to-noise ratios. For short and relatively clean signals, CNNs trained on raw image representations achieve comparable or slightly better performance. These findings indicate that frequency-aware image encoding enhances robustness without increasing model complexity.
Conclusion—This study shows that integrating Discrete Wavelet Transform with image-based time series representations provides a more balanced and informative feature space for CNN-based classification. By mapping different DWT components to GAF, MTF, and RP, the proposed framework captures temporal patterns across multiple frequency scales. The results highlight the importance of frequency-specific encoding, especially for challenging real-world signals. Future work may explore deeper architectures and adaptive wavelet-image assignments to further improve classification performance.
Anahtar Kelimeler
Kaynakça
- H. Sharabiani, S. Darabi, S. Harford, E. Douzali, F. Karim, H. Johnson, S. Chen, “Asymptotic dynamic time warping calculation with utilizing value repetition”, Knowledge and Information Systems, 57(2), 359–388, 2018. https://doi.org/10.1007/s10115-018-1163-4.
- W. Chen, K. Shi, “Multi-scale attention convolutional neural network for time series classification”, Neural Networks, 136, 126–140, 2021. https://doi.org/10.1016/j.neunet.2021.01.001.
- H. Kang, T. H. Lee, J. Lee, “A graph convolutional network for time series classification using recurrence plots”, Applied Intelligence, 55(15), 972, 2025. https://doi.org/10.1007/s10489-025-06841-3.
- H. I. Fawaz, B. Lucas, G. Forestier, C. Pelletier, D. F. Schmidt, J. Weber, G. Webb, L. Idoumghar, P. A. Muller, F. Petitjean, “InceptionTime: Finding AlexNet for time series classification”, Data Mining and Knowledge Discovery, 34(6), 1936–1962, 2020. https://doi.org/10.1007/s10618-020-00710-y.
- C. Alagöz, “Crossfire: Cross-domain feature integration for robust time series classification”, PeerJ Computer Science, 11, e3328, 2025. https://doi.org/10.7717/peerj-cs.3328.
- X. T. Li, T. Y. Li, Y. Wang, “GW-DC: A deep clustering model leveraging two-dimensional image transformation and enhancement”, Algorithms, 14(12), 349, 2021. https://doi.org/10.3390/a14120349.
- X. T. Li, K. Zhou, F. Xue, Z. B. Chen, Z. Q. Ge, X. Chen, K. Song, “A wavelet transform-assisted convolutional neural network multi-model framework for monitoring large-scale fluorochemical engineering processes”, Processes, 8(11), 1480, 2020. https://doi.org/10.3390/pr8111480.
- X. Y. Lu, Y. Li, X. Chen, Y. Q. Li, Y. B. Liu, “Discrete wavelet transform assisted convolutional neural network equalizer for PAM VLC system”, Optics Express, 32(6), 10429–10443, 2024. https://doi.org/10.1364/OE.516195.
Ayrıntılar
Birincil Dil
İngilizce
Konular
Örüntü Tanıma, Sınıflandırma algoritmaları
Bölüm
Araştırma Makalesi
Erken Görünüm Tarihi
13 Nisan 2026
Yayımlanma Tarihi
-
Gönderilme Tarihi
28 Ocak 2026
Kabul Tarihi
17 Mart 2026
Yayımlandığı Sayı
Yıl 2026 Sayı: Advanced Online Publication