Unsupervised Deep Wavelet-domain Recovery for On-sensor Compressed Data Acquisition

The next generation of oil and gas exploration requires seismic acquisition systems capable of operating at a large scale while maintaining high levels of flexibility, automation, and scalability. Within the emerging paradigm of Industry 5.0, which promotes collaborative and AI-driven industrial processes, these requirements become even more critical. One of the main challenges arises from the need to transmit a large number of measurements collected by each sensor/geophone through bandwidth-limited communication links without overwhelming either the sensors or the central processing infrastructure. Addressing this challenge calls for lightweight data reduction techniques that can operate directly at the sensing stage. In this work, we introduce an efficient in-field compression framework for seismic data that combines sparsity-aware sensing with an unsupervised deep learning–based reconstruction strategy in the wavelet-domain. Two structured reconstruction algorithms are used and evaluated. By leveraging known noise statistics, the proposed framework jointly learns both the signal prior and the reconstructed signal without relying on labeled training data. The learning procedure is implemented through an unrolled optimization framework in which alternating minimization is used to iteratively update both the neural network parameters and the signal estimates. A notable advantage of the proposed approach is its general applicability, as it does not rely on specific assumptions about the underlying statistics of the seismic data. This allows the framework to adapt to a wide variety of exploration environments. Experimental results obtained from real seismic datasets show that the method achieves an effective compromise between compression and high-quality signal recovery. The results further demonstrate improvements in both compression efficiency and reconstruction performance compared with existing techniques.