Diagnosis and prognosis of rolling element bearings at low speeds and varying load conditions using higher order statistics and artificial intelligence

Abstract

Condition monitoring (CM) is commonly used in determining the operational states and health of rotating machines. The rise in the complexity of modern machines have led to advances in CM technologies to increase and improve product reliability and reduced downtime. Vibration and acoustic emission signals generated by these modern and complex machines are often immersed in background noise making it difficult to detect faults. Extracting signal features that are sensitive to faults still attracts considerable attention to detect and identify faults in rotating machines. This is especially so for low speed machinery under varying load and speed conditions. Such conditions are found in many industrial applications and include draglines in the mining industry and large rolling mills in many materials processing environments. Condition monitoring techniques for stationary systems are inadequate at accurately detecting and diagnosing faults under such conditions. Using acoustic emission transducers are better options to the use of accelerometers for data generation for the analysis under these conditions because of its higher sensitivity to detect low energy level response because of the low speed condition. While skewness and kurtosis have been used extensively in the condition monitoring of bearings and gears, higher order statistical (HOS) techniques have not found wide application in machine condition monitoring. This is because if a process is Gaussian then HOS provide no additional information that can be obtained from the second or higher order statistics. There is however reason to believe that these HOS techniques could play an important role in condition monitoring, provided appropriate care is taken. In problems that are non-Gaussian, non-minimum in phase, nonlinear in behavior and robust to additive noise, HOS techniques like the 6th order statistical moment (hyperflatness) could play an important role in its condition monitoring (CM). By applying this method, processing of acoustic emission signal at low speed and varying load condition could be very useful by providing details about the signal which the conventional second order statistics cannot. HOS techniques are extensions of the better-known concepts of correlation (in time or space) power spectra. Higher order spectra are higher order Fourier spectral representations of third and higher order correlations or moments. With this approach simple representation and interpretation of the online extracted information could be made possible. For example, different colours of light emitting diode (LED) could be used to indicate the types of faults, such that a non-expert or a simple classification algorithm may interpret the result. Empirical models include the fields of regression and classification which are also collectively referred to as supervised learning and is dependent on its construction or optimization on large sets of representative data. In classifying faults in rolling element bearings (REBs) at low speeds and under varying load conditions, support vector machines for regression and genetic algorithm (SVMGA) which is a supervised machine learning algorithm can be used for its classification or regression problems. Using a technique called the kernel, transforms the data and finds an optimal boundary between the possible outputs. It does some complex data transformations, and then separate the data based on the labels or outputs defined. The Hidden Markov Model has also found application in the diagnosis of fault in rolling element bearings, as it has been proven to diagnose incipient faults better but often requires a large data set. For this reason, the Bayesian Robust New Hidden Markov Model (BRNHMM) will also be used here to diagnose faults of two different categories: debris induced fault on roller bearing and a fault induced on the outer race of a roller bearing. The model setup in this work was formulated and validated with the use of data generated from an experimental test rig originally designed by Aye for his PhD research, and was further improved on as part of this work. Simulated data was also used to validate the result obtained from the test rig. Fault diagnosis and prognosis is achieved with the use of eXtended Takagi-Sugeno (xTS) fuzzy and recursive least square algorithm (exTSFRLSA) and support vector data descriptive (SVDD) method in this work. The (exTSFRLSA) has many applications of which one is the prediction of the sequence of state change, based on the sequence of observations. SVDD belong to the statistical learning theory class which is used here in this work to show the remaining useful life (RUL) of the bearing under study. Although many models have recently been applied with good generalization of results in predicting the RUL of bearings where they integrate the statistics (like the Kaplan-Meier estimator, Mahanalobis distance, principal component analysis (PCA) etc.) and artificial intelligent (AI) methods (e.g. artificial neural network (ANN), feed forward neural network (FFNN), recurrent neural network (RNN) etc.), it has been proven that using only the statistical method is not sufficient to give good predictions, but a hybrid of both methods often yield good results. exTSFRLSA is used on most occasions for tuning, adjusting parameters and for adaptation in the propagation model by comparing predicted and measured defect sizes as in, hence the instantaneous rate of defect propagation of the bearing can be captured despite defect growth behavior variation.