Analog In-memory Computing (AIMC) based on Embedded Phase-change Memory (ePCM) has the potential to outbreak the performances of sensor-near computing frameworks thanks to its unique capability to parallelize multiply-and-accumulate operations. The latter are, among the many others, at the basis of the Compressed Sensing (CS) theory. This work aims at exploiting the characteristics of an AIMC prototype based on an ePCM array, both designed in a 90-nm CMOS technology by STMicroelectronics, for the purpose of CS-driven acceleration data compression in the context of vibration inspection. This application domain, indeed, necessitates from novel architectures and computing paradigms in order to be compatible with real-time and energy-efficient implementations.To this end, we have evaluated the impact of input datum quantization and conductance drift on the quality of the reconstructed spectra, prioritizing its suitability for damage detection, along with its electrical and computational performances. Experimental analyses conducted on a representative laboratory testbed revealed that the exploitation of an ePCM-based AIMC prototype can be 4 \(\times \) less power demanding and 6 \(\times \) faster than standard 32-bit MCU architectures. Importantly, spectral peak identification necessary for structural assessment is demonstrated even in presence of deeply quantized (8 bit) data and against the influence of cell drift.