Friday 23 May 2025
A team of researchers has developed a new method for generating synthetic data that mimics real-world sensor readings, specifically designed for inertial measurement units (IMUs) used in smartphones and other devices. This innovation has significant implications for fields like human activity recognition, pedestrian dead reckoning, and device position detection.
The challenge lies in creating realistic synthetic data that captures the complex patterns and variations found in real-world sensor readings. Traditional methods often rely on simple generative models or manual feature engineering, which can result in inaccurate or incomplete data. The new approach uses a diffusion-based model, specifically designed for time-series data, to generate high-quality synthetic IMU signals.
The researchers transformed the raw IMU data into image representations using a technique called delay embedding. This allowed them to leverage the power of vision-based diffusion models, which excel at generating realistic images. By conditioning the generation process on device placement labels, the model produced diverse and realistic synthetic data that accurately captures the distinctive characteristics of different smartphone carrying positions.
To evaluate the quality of their generated data, the researchers employed a dual-classifier framework, using both image- and signal-based CNNs to classify real and synthetic test sets. The results showed impressive performance on both real and synthetic data, with minimal accuracy gaps between the two.
This breakthrough has significant implications for various applications that rely on high-quality sensor data. For instance, it can enable more efficient and effective training of machine learning models in fields like activity recognition, pedestrian dead reckoning, and device position detection. The generated data can also be used to augment existing datasets, reducing the need for extensive data collection and improving overall model robustness.
The team’s innovative approach has opened up new possibilities for generating high-quality synthetic sensor data, demonstrating the potential of diffusion-based models in time-series applications. As the field continues to evolve, this breakthrough is likely to have far-reaching impacts on various domains where accurate sensor data is crucial.
Cite this article: “Synthetic Sensor Data Generation for IMUs Using Diffusion-Based Models”, The Science Archive, 2025.
Synthetic Data, Inertial Measurement Units, Imus, Human Activity Recognition, Pedestrian Dead Reckoning, Device Position Detection, Diffusion-Based Model, Time-Series Data, Image Representation, Delay Embedding
