Sensors 2022, 22,2012 3085 and shaft cracks in the vibration signals, including but not limited to Continuous Wavelet Transform (CWT) and Hilbert Huang Transform (HHT). The latter is also used to effectively monitor defects in bearing [35], while the former is used for early stage fault detection in the outer race [6]. Envelop detection with autocorrelation is used to detect faulty patterns at the inception stage in low SNR (Signal to Noise Ratio) AE signals [6], with wavelet transform for de-noising of the AE signals. Spectral components of the Intrinsic Mode Functions (IMF) are used to analyse both the AE and vibration signals for the detection of, Dearing defects, broken bar and unbalanced load distribution in the indoor motors [39]. However, it needs the a priori knowledge of the number of modes in which the signal is required to be decomposed. Features from raw vibration signals were extracted by [8], using Empirical Mode Decomposition algorithm and CNN. These features were fed to SVM and CNN training algorithms for the classification of faults into the outer race, inner race and ball faults. In another study, Khan et al. used EMD with KNN to classify the different ‘machine states into normal, cracking, offset pulley and wear states [40]. In [41], the EMD is combined with deep neural networks to classify the faults into roller, inner and outer races with an accuracy of 98.5%, The authors have detected and classified faults using ensemble empirical mode decomposition (EEMD) and SVM. The EMD is used to decompose the vibration signal, isolating and denoising high frequency IMFs using Pearson correlation coefficients and the wavelet semi-soft threshold, respectively. The Eigen vector of the signal is used as a feature vector for the SVM to classify faults into inner race, outer race and rolling elements, with an accuracy of 100% [12]; however, the 100% accuracy seems to be on the higher end. Delprete et al. [43] used orthogonal empirical mode decomposition analysis (a time-frequency) to detect faults in the inner and outer raceway of the bearings using vibration signals. Informative signatures can be extracted form vibration signals in the form of wavelets transform. Morlet wavelet transform is used to extract features from vibration signals that were then used with artificial neural networks and SVM to classify the signals into ball fault, inner and outer race fault, with 95.271% accuracy for SVM and 87.25% for ANN [44]. The authors of [45] combined features like vibration severity, dyadic wavelet energy time- spectrum and coefficients power spectrum of maximum wavelet energy level and fed it to SVM for classification into a normal state, eccentric axle fault (RAF), bearing pedestal fault (BPF), and sealing ring wear fault. The SVM parameters were optimized using the modified shuffled frog-leaping algorithm, claiming a maximum accuracy of 96% [45] . A Deep Fault Diagnosis (DFD) method is proposed for rotating machinery with scarce data labels. In this procedure, discriminative STFT data are obtained from the spectrogram of, the vibration signals. Several SVM models are trained with different features selected from the pool with scarce labels and the most discriminative features and the best SVM models were selected, hence forming an augmented training set. This augmented training set was then forwarded to a 2-D deep CNN. The proposed algorithm classified different faults with an accuracy of 98.4% [46]. The research community has also used image analysis for diagnosing faults; here, image sparse representation is proposed to extract meaningful features from redundant information present in images using orthogonal matching pursuit and K-singular value decomposition algorithms along with 2-D PCA. The features are used with a minimum distance to classify into the inner race, outer race and ball faults with an average accuracy of 99.92% [47]. In [18], Moussa et al. proposed an algorithm for bearing fault diagnosis based on the probability of image recognition techniques under constant and variable speed conditions using average PCA. The paper used vibration spectrum imaging of the vibration signal obtained from faulty and normal bearings with CNN for classification. In [49], the time domain vibration signal was first segmented using a time-moving segmentation window and was then transformed into a spectral image for training and testing with CNN. ‘The proposed scheme provided good accuracy at different levels of noises and speeds. Similarly, the time domain, frequency domain and time-frequency domain parameters are used to detect faults in axial piston pumps with deep belief networks [50].Sensors 2022, 22,2012 40815 The literature review as discussed above does not cover the complete review of the topic; however, it provides quite a clear picture of the work being contributed by the research community. The statistical approach, frequency domain and time-frequency domain provide different levels of accuracy with conventional machine learning techniques and deep neural networks. In this research work, the statistical and frequency domain techniques are analysed to detect and classify the faults in rotating machinery using conventional machine learning algorithms. The vibration signal from the faulty and healthy rotating machinery was recorded. From these vibration signals, the statistical features, such as Skewness, Kurtosis, Average and root mean square (RMS) values of time domain vibration signals, are considered. The same features were then extracted from the second derivative of the time domain vibration signals, making a vector of eight features. For the second and third steps, these features were calculated by first taking the Fourier Transform and Power Spectral Density of the vibration signals. Alll three sets of feature vectors were concatenated creating time domain, frequency domain and spectral power domain feature vectors. SVM, KNN and KLDA classifiers are used for the classification of signals into outer race fault, inner race fault, ball fault and healthy signal. Simulation results showed that the KLDA resulted in an accuracy of 99.13% for our proposed method using PSD, followed by the statistical features with KLDA giving an accuracy of 98.275%, showing the strength of, our proposed algorithms. ‘These results were then compared with those of EMD, which has a very good accuracy as given in the literature review discussed above. In this case, the accuracy of the EMD feature vector was 97.01% with SVM. It is important to note that the size of the feature vector for EMD is 14 x 160,000 in this particular case and that of the Statistical, FT and PSD is 8 x 228. This also shows the effectiveness of the proposed approach in terms of size of, the feature vector and hence less computation. ‘The primary contributions of this research are enumerated below: 1. We exploit the behaviour of feature extraction based on the performance of selected classifiers; 2. We propose to utilise the statistical features including kurtosis, skewness, average and root mean square of the selected window; 3. Toextract the more meaningful information, we extracted the same features from the second derivative; 4. Weutilise both frequency and time domain signal information for feature extraction. ‘employing the moving window concept; 5. We propose a system which generates an accuracy of more than 95%—utilising less than 5% of information and achieving the reduction percentage of 95%. In the rest of the paper, signal processing techniques for extracting the proposed features and calculating FFT and PSD and so forth are discussed in Section 2. The discussion of these features for fault detection is given Section 3. The vibration signals are analysed using signal graphs of different faults. A discussion on the classifier and classification of faults into outer race, inner race, and ball faults is presented in Section 4, with the concluding remarks given in Section 5. 2, Feature Extraction In this section, the digital signal processing (DSP) techniques that are used for feature extraction are discussed. These techniques can be categorised under three main headings, the statistical features of time domain signal, the statistical features of signal in Fourier do- ‘main and the statistical features of signal's Power Spectral Density. The raw vibration data ‘was used to extract statistical features like maximum value, minimum value, standard devi- ation, mean, median, variance, skewness, kurtosis, range, Fisher Information Ratio [51,52], Petrosian Fractal Dimension [51], and entropy. The results of skewness, kurtosis and stan- dard deviation and second derivative and so forth, to name a few. The features that were selected are mean value, standard deviation, skewness, kurtosis and second derivative. ‘These features were calculated for time domain vibration signal, for the signal in FourierSensors 2022, 22,2012 50815 domain and for the same signal after calculating its PSD. These features were then fed to the SVM, KNN and KLDA algorithms for classification into outer race fault, inner race fault, and ball fault as given in the block diagram of Figure 1. As is shown, the raw data are extracted, pre-processed and its statistical features like skewness, kurtosis, average and RMS values are calculated. It is important to note that extracting these statistical features in the Fourier domain and from the PSD of the vibration signal is not used previously, to the best of our knowledge. This proposed method has also reduced the number of data points and hence computational requirements as discussed in Section 4. The vibration signal of, healthy and faulty bearings were recorded, using vibration sensors, shaft rotating at a rate ‘of 800 revolutions per minute and a sampling rate of 40,000 samples per second. Figure 2 shows the specifications and block diagram of the test rig with ball bearing model, fault types and specifications of the data acquisition board. (at = ae Beats zit steoromd sonra) (a) Tet Rig (b) Test Rig Seitngs Instrumentation Settings DAA board Ni PCi-44728, 24-bit Samoling ate 40000 carnnlac/Ser Computer \7 Computer , 868 RAM Op. system windows? (6) Instrumentation Settings Figure 2. Test rig block diagram and specifications ‘The equations for skewness, kurtosis, standard deviation and second derivative are given below. The skewness is a measure of symmetry or the lack of symmetry, and is zero for the data with normal distribution and should be zero for any symmetric data. The kurtosis is a measure of whether the data are heavy-tailed or light-tailed relative to a normal distribution. Both provide important statistical characteristics of a population. The mean and the standard deviation provide the central value; however, some times it is required to find how far the data are spread, and that is measured by variance and standard deviation. ‘The mathematical representation of all these statistical measures is given here for ready reference. For a 1-Dimensional data ¥4, Y2,... Yj, the mathematical representation of these statistical measures is given as: