Synthetic Leak Data Generation Using Variational Autoencoders to Address Data Imbalance in Acoustic Emission-Based Pipe Leak Detection

A synthetic data generation method is proposed to mitigate data imbalance in pipeline leak detection using acoustic emission (AE) sensors. Collecting sufficient AE signals in the leak state is challenging due to the rarity of leaks and safety concerns. The rarity of leaks leads to highly imbalanced datasets. The performance of leak detection methods may be degraded because the models tend to be biased towards the normal state. The proposed method utilizes a variational autoencoder (VAE) to probabilistically model the difference between the normal-state and leak-state spectrograms. After training the VAE with the spectrogram differences, the decoder of the VAE generates spectrogram differences from random latent vectors. Synthetic leak-state spectrograms are created by adding the generated spectrogram differences to normal-state spectrograms. The effectiveness of the proposed method is evaluated by comparing the leak detection performance of models trained with and without the proposed method. A leak detection model trained with synthetic leak data generated by the proposed method shows improved detection performance compared to models trained using existing oversampling methods.