Transforming Time Series into Images for Convolutional Networks

26 concepts18 papers1 proposal

Time-series data is inherently sequential, traditionally analyzed using 1D convolutional networks, recurrent architectures, or statistical signal processing. However, 2D Convolutional Neural Networks (CNNs) possess an incredibly powerful inductive bias for spatial pattern recognition, perfected over a decade of computer vision breakthroughs.

To exploit this, a growing body of research bypasses 1D architectures entirely by transforming 1D and multi-channel time-series signals into 2D images. This shift allows models to extract non-linear characteristics, structural complexities, and cross-channel relationships that traditional sequence models often overlook.

Why 2D Convolutions Excel on 1D Data

In a standard 1D signal, a convolutional kernel slides exclusively along the temporal axis. If the data contains multiple channels, a 1D kernel typically aggregates information across all channels simultaneously, treating them as a flat vector at each time step.

When a sequence is transformed into a 2D image, the vertical and horizontal axes can be mapped to represent distinct, interacting dimensions—such as time versus frequency, or sensor channel versus sensor channel. A 2D convolutional kernel can then capture localized patterns across both dimensions simultaneously.

As demonstrated by Maggipinto et al. in DeepVM: A Deep Learning-based approach with automatic feature extraction for 2D input data Virtual Metrology, treating multi-dimensional industrial sensor data as an image allows a 2D network to capture complex interdependencies between time and wavelength. Instead of treating these dimensions independently, the 2D kernel acts as a localized pattern detector across both domains, capturing structural features that independent statistical methods miss.

Strategies for Time-to-Image Transformation

Researchers have developed several distinct paradigms to convert temporal waveforms into visual structures, depending on the nature of the underlying data.

Direct Grid and Waveform Plotting

The most direct approach reshapes raw data points into a grid or plots the waveform directly into a pixel matrix. For instance, Lima et al. in Heartbeat Anomaly Detection using Adversarial Oversampling and Ullah et al. in An Automatic Premature Ventricular Contraction Recognition System Based on Imbalanced Dataset and Pre-Trained Residual Network Using Transfer Learning on ECG Signal segment raw 1D ECG signals around individual heartbeat peaks and map them into 2D spaces to isolate visual anomalies.

Similarly, Sui et al. in Epileptic Focus Localization Based on iEEG Plot Images by Using Convolutional Neural Network convert raw intracranial EEG signals into plot images to leverage the visual pattern recognition strengths of CNNs. The authors of Using Deep and Convolutional Neural Networks for Accurate Emotion Classification on DEAP Data also reshape EEG time-series data directly into grid-like structures to leverage standard pooling and convolutional filters.

Time-Frequency and Statistical Mapping

Instead of raw values, signals can be transformed into mathematical matrices before imaging. Vilamala et al. in Neural Networks for Interpretable Analysis of EEG Sleep Stage Scoring convert time-frequency spectral data into RGB images using colormaps, while Desai et al. in Non-linear Embedding Methods for Identifying Similar Brain Activity in 1 Million iEEG Records Captured From 256 RNS System Patients use spectrogram images to match complex intracranial EEG patterns across massive patient cohorts.

To capture non-linear dynamics, Zhao et al. in Enhancing Detection Accuracy for Clinical Heart Failure Utilizing Pulse Transit Time Variability and Machine Learning transform physiological signals into 2D distance distribution matrices, exposing structural patterns of dysfunction that traditional time-domain statistics miss.

Preserving Spatial and Physical Topology

When dealing with multi-sensor arrays, the image configuration can mirror the physical layout of the sensors. In neurological applications, Mokatren et al. in EEG Classification based on Image Configuration in Social Anxiety Disorder preserve the spatial configuration of electrodes when building the image, ensuring that adjacent pixels correspond to adjacent brain regions.

Cheng et al. expand on this in Temporal EEG Imaging for Drowsy Driving Prediction by stacking these spatial maps across multiple time points into a single "temporal EEG image." Similarly, Kwon et al. construct an asymmetry matrix image in Deep-Asymmetry: Asymmetry Matrix Image for Deep Learning Method in Pre-Screening Depression to explicitly capture spatial imbalances in brain activity across hemispheres.

Broad Generalization: From Medicine to Machinery

The intuition behind this technique extends far beyond clinical physiology. Any high-dimensional or noisy time-series data can benefit from the visual pattern-recognition capabilities of CNNs.

For example, Tripathi et al. in ImAiR: Airwriting Recognition Framework Using Image Representation of IMU Signals convert IMU motion sensor data into images to capture complex spatial handwriting gestures. In industrial diagnostics, Mayaki et al. in Machinery Anomaly Detection using artificial neural networks and signature feature extraction use signature feature extraction to turn sensor data into 2D images for faster fault detection. This is mirrored in energy systems by Liu et al. in Battery Fault: A Comprehensive Dataset and Benchmark for Battery Fault Diagnosis, where raw voltage time-series are converted into grayscale images to expose subtle visual signatures of battery degradation.

Bridging the Gap with Native Visual Formats

An interesting corollary to this research direction occurs when clinical data is natively stored as an image (such as printed paper or PDF reports). Rather than treating this as a limitation, researchers leverage it.

Xu et al. demonstrate in Identifying Hypertrophic or Dilated Cardiomyopathy: Development and Validation of a Fine-Tuned ResNet50 Model Based on Electrocardiogram Image that fine-tuning a standard ResNet50 directly on ECG images bypasses complex signal preprocessing entirely, utilizing the visual format directly for diagnosis. Conversely, when 1D models are preferred for legacy workflows, Sau et al. show in A comparison of artificial intelligence–enhanced electrocardiography approaches for the prediction of time to mortality using electrocardiogram images that digitization pipelines can extract raw waveforms from these images, proving that the visual and digital representations contain deeply compatible diagnostic signatures.

By framing time-series analysis as a computer vision problem, these diverse approaches unlock pre-trained architectures (like ResNet) and advanced spatial pooling techniques. Whether through direct plotting, spectrogram generation, or topologically aware matrix mapping, transforming sequences into images allows deep learning models to see patterns that are otherwise hidden in the timeline.

Key Foundational Work

Using Deep and Convolutional Neural Networks for Accurate Emotion Classification on DEAP Data (2017) — contributed 3 concept elements to this direction.
A Convolutional Neural Network Based Auto Features Extraction Method for Tea Classification with Electronic Tongue (2019) — contributed 2 concept elements to this direction.
Heartbeat Anomaly Detection using Adversarial Oversampling (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_time_series_to_image_transformation_eeg__9c746ed6,
  title = {time-series to image transformation (EEG to image conversion for CNNs)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/time-series-to-image-transformation-eeg-to-image-conversion-for-cnns/rdr_9c746ed6909846d7a6a9f191179e26d4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_time_series_to_image_transformation_eeg__9c746ed6_v1,
  title = {time-series to image transformation (EEG to image conversion for CNNs)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/time-series-to-image-transformation-eeg-to-image-conversion-for-cnns/rdr_9c746ed6909846d7a6a9f191179e26d4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake