Abstract: Compressed sensing can achieve effective dimension-reduction of data (i.e., compressed measurements) by constructing observation matrix that satisfies the restricted isometry property (RIP). However, along with this comes the challenge of reconstructing the original signal from the compressed signal with high quality and efficient manner. To circumvent the tedious reconstruction process, this paper proposed a fault diagnosis method based on compressed domain feature identification. Within the basic framework of compressed sensing, the well-known Gaussian random measurement matrix was replaced by a row ladder observation matrix to achieve compressed measurement of the original signal. Aiming at the interference of random noise on the compressed observation signal, a weak fault feature extraction method based on maximum correlation kurtosis deconvolution (MCKD) and 1.5-dimensional spectrum was established. In which the periodic shock component in the compressed signal was enhanced by MCKD to eliminate the interference of transmission path and background noise, and the 1.5-dimensional envelope spectrum was subsequently used to extract the fault feature frequency. The simulation and experimental results show that the proposed method not only circumvents the complex reconstruction process needed by classical compression sensing, but also achieves accurate fault diagnosis results under conditions of strong noise interference.