Seismic data compression using auto-associative neural network and restricted Boltzmann machine

In a geophysical exploration survey, thousands of geophones are deployed where each geophone must transmit hundreds of recording over a narrow band channel to a fusion center. A lightweight compression technique for geophones is required to reduce the data traffic to the center. We present an efficient implementation of neural network for real-time seismic data compression for geophones. We use an auto-associative neural network architecture with a single linear hidden layer. The neural network is trained in two stages. First, we apply unsupervised learning with restricted Boltzmann machine (RBM) to obtain good initial weights. Secondly, the neural network with the initial weights is further fine-tuned in a supervised fashion with scaled conjugate gradient (SCG). Experimental results with real data have shown that the trained neural network achieves a peak signal-to-noise ratio (PSNR) of more than 30dB with a compression ratio of 10:1. The RBM is also proven to speed up the training up to nine times than that of without RBM. The proposed method is also compared with the linear predictive coding (LPC) and shows significant superiority in terms of compression error and reconstruction quality.