A Convolutional Autoencoder for Fast Compressive Sensing Reconstruction of Vibration Signals

Abstract

n many health monitoring applications, large volumes of high-frequency measurement data must be acquired and processed to extract reliable health indicators for fault detection and identification. Compressive sensing (CS) provides an effective framework to reduce data dimensionality at the acquisition stage by exploiting signal sparsity, enabling sub-Nyquist sampling and lowering storage and transmission requirements. However, practical CS deployment is often limited by the reconstruction step, which typically relies on iterative optimization algorithms that are computationally expensive and difficult to implement in real-time monitoring systems. This work proposes a learned reconstruction strategy that replaces conventional CS solvers with a convolutional autoencoder based approach. The sensing process follows the standard CS formulation, where the original signal is projected onto a lower-dimensional measurement space using a fixed random sensing matrix. During training, the autoencoder is constrained so that its encoder reproduces the measurement operation, while the decoder learns a data-driven inverse mapping to reconstruct the original signal from compressed measurements. At inference time, compressed measurements are directly fed into the decoder, eliminating iterative reconstruction. Experimental results obtained on simulated gearbox signals and real vibration measurements demonstrate that the proposed method significantly reduces reconstruction time compared with classical CS algorithms while preserving diagnostically relevant information for fault detection.