In modern industries, fault detection of rotating machinery is a fundamental issue since it helps to reduce the unnecessary expenditures in repairs while improving machine performance. On this matter, the main challenge is to determine the current state of the machine from a set of measurements, termed Condition-Based Maintenance (CBM). The machine state is assessed usually using vibration analysis because it gives a high precision and has a low economic cost. Nonetheless, the following two problems arise: The first one appears when the machine conditions are time-varying, either by changes in speed or load, inasmuch as the vibration signals are realizations of non-stationary processes. The second problem is associated to the amount of available data characterizing different machine states, since, in most of the cases, recordings of damaged machine are not available. The latter fact hinders the application of conventional classification techniques due to strong imbalance of the faulty/normal classes (machine states).
With regard to the former problem
, some authors have used time-frequency representations (TFR) in describing the machine dynamic behavior under non-stationary operating conditions. In particular, Sedjic et
al in  summarize different methods for estimating energy concentration of several TFR extracted from a set of test rig faults. But they only identify visually the qualitative difference between several faults instead of carrying out a quantitative automated classification procedure. However, inclusion of the classification stage implies high computational cost since TFR map comprises a lot of non-relevant information that should be avoided . To reduce the computational cost, authors  and  use basic statistical features (mean, standard deviation
, kurtosis, and root mean square) estimated from a time series and its frequency representation. Nevertheless, those features do not describe properly the dynamic behavior generated by non-stationary operating conditions of the machine. Therefore, there is a need to carry out a methodology to characterize machine dynamic behavior
, but preserving a low computational cost.
With regard to the latter problem, one-class classification (OCC) techniques have been used to determine whether the machine state ceases to be normal or first damage symptoms appear. Thus, Tax and Duin in  compare several commonly used one-class classifiers such as the normal distribution classifier, the k
-nearest neighbor classifier, and support-vector data description (SVDD). Considered classifiers are trained and tested employing vibration signals at different constant speeds by using a set of statistical-based features
, however, achieved classification performance is low. With the aim of improving the classification performance, some authors have proposed different methodologies based on estimation of statistical features from piecewise segmented non-stationary vibration signals. Among other approaches are the following: weighted SVDD , moving-average model , wavelet packet transform , and subspace reduction by principal component analysis (PCA) . These methodologies reach high classification performance, but in practical cases like as high speed fluctuations
, inherent signal segmentation implies loss of relevant information .
In this paper, a novel methodology for mechanical systems description having non-stationary behavior is introduced. In particular, the proposed approach uses the spectral sub-band centroids and the linear frequency cepstral coefficients; all of them extracted from a TFR for machine dynamic characterization. Due to the large number of features obtained from the TFR
, a feature selection process is carried out to determine contribution of most relevant dynamic characteristics. Finally, resulting dynamic features are validated by the used OCC, whose performance is compared when extracting the whole TFR map. Proposed methodology is tested with a dataset collected in a test rig for normal, unbalanced, and misaligned assemblies. Recordings are acquired under variable speed conditions including machine start-up and coast-down. The proposed methodology is also validated over recordings of a real ship driveline.
The agenda of this paper is as follows: in Section 2, brief overview is given to describe TFR-based dynamic features and one-class approaches employed in this work for analysis of non-stationary signals. Sections 3 and 4 bring numerical experiments of achieved performance that is compared against other state-of-the-arttechniques. Discussion of results is given inSection 5, while Section 6providesthe related conclusions and future research.
Time-frequency representation determines the energy concentration along the frequency axis at a given time instant. Particularly, the Short Time Fourier Transform (STFT) introduces time localization by using a sliding window function
going along with the signal