Data augmentation of optical time series signals for small samples

Area of Science:

• Signal Processing
• Machine Learning
• Optical Sensing

Background:

• Deep learning applications in analyzing one-dimensional optical time series signals are hindered by the scarcity of labeled data.
• Acoustic emission (AE) data from optical sensors is often limited, posing a challenge for model training.

Purpose of the Study:

• To propose a novel deep convolutional generative adversarial network (DCGAN) model for augmenting optical time series signals.
• To address the data scarcity bottleneck in deep learning for optical time series analysis.

Main Methods:

• Development of a DCGAN model tailored for optical time series signal augmentation.
• Training the DCGAN on a small dataset of acoustic emission (AE) signals obtained from an optical sensor.
• Generating new AE signal fragments using the trained model.

Main Results:

• The proposed DCGAN model successfully generated stable and diverse AE signal fragments by epoch 500.
• No model collapse was observed during the generation process.
• Statistical analysis confirmed no significant difference (at the 0.05 level) between the features of the generated and original data.

Conclusions:

• The developed DCGAN model effectively generates synthetic optical time series signals, overcoming data limitations.
• This approach provides a viable solution for enhancing datasets in deep learning applications involving optical time series.
• The generated data preserves key features of the original signals, validating the method's effectiveness.