Sparse decomposition has been generally emphasized in signal processing theory. In this paper, a nonelectrical signal feature dataset of key components of high-voltage DC converter valve is established by using principal component analysis to streamline the data volume. The compression-aware feature extraction algorithm based on polynomial matrix sparse coding is used to extract and collect the nonelectrical signal parametric data. Through the performance over the experimental signal analysis, it can be known that the eigenvalues of a total of 10 parameters, including the infrared temperature measurement results, the appearance, the presence of corrosion or dirt, and the presence of abnormal vibration and sound, are all greater than 1. Therefore, these 10 parameters are identified as the key parameters. When the number of measurement points is between 64 and 200, the algorithm in this paper can satisfy the need of feature extraction when the signal length is insufficient, compared with the traditional approach. In the empirical analysis of the vibration signal as an example, the method of this paper can effectively extract the frequency and time domain of the vibration signal.