As science and technology advance, rotary machines in modern industries become more advanced and complex, making maintenance more dificult. Multiple failures can occur in machine components, requiring monitoring and periodic maintenance. Multiple types of failure may occur, causing severe damage. As a result, modern technologies for early fault detection have developed. Vibration analysis of rotating machinery is one of the most often used condition monitoring techniques. Artificial intelligence (AI) approaches are widely used for selecting the features afected by faults. This paper discusses three diferent types of artificial intelligence methods, which are Artificial Neural Networks, K-nearest neighbor, and Support Vector Machine, and presents a comprehensive review of recent studies on fault diagnosis for various rotary machine elements, comprising the type of failure, feature extraction method, and classification technique performance.
fault. Elaborating that the authenticity and uniqueness
of the extracted features are of major concern [
The use of rotating machinery in modern industries is This paper discusses the applications of three types of
growing more sophisticated and complicated as science artificial intelligence methods in machinery faults
diagand technology development, and maintenance becomes nosis, which are the Artificial Neural Networks (ANN),
more dificult as the sophistication of these machines in- K-nearest neighbor (KNN), and Support Vector Machine
creases. Due to the extensive working hours with severe (SVM), and presents a comprehensive review of previous
loads of this type of equipment, its components are sub- studies on fault diagnosis techniques for several
comjected to multiple faults, requiring monitoring and peri- ponents of rotary machines, including the type of
failodic maintenance. Multiple types of failure may occur in ure, feature extraction method, and the performance of
various parts, such as bearings, gears, pulleys, shafts, etc. classification technique. The remaining of this paper is
These faults afect the machine’s performance and might organized as follows. Section 2 introduces a theoretical
result in major issues and financial losses. As a result, basis of AI techniques. Section 3 reviews the applications
modern technologies for early detection of faults, includ- of AI techniques in fault diagnosis of rotating machinery.
ing condition monitoring and fault diagnostics are being Finally, the conclusions are drawn in Section 5.
investigated and developed to ensure that the machines
operate efectively. The main goals of fault diagnostic
investigation are to determine the machine’s normal op- 2. Theoretical basis of Artificial
erating condition, identify the type of fault, and predict Intelligence techniques
the fault’s progression. One of the most common types
of condition monitoring is the vibration analysis of rotat- 2.1. Artificial neural networks (ANN)
ing machines. However, vibration signals are typically
non-stationary, non-linear, and intermixed with noise. The ability of biological neurons in the human brain has
consequently, recent research has forecasted the domi- inspired researchers to invent a computational structure,
nance of AI over technological innovations [
K-NN method is one of the most basic and
straightforward classification approaches. It is an instance-based
learning technique based on grouping components with
similar features; it determines the class category of a
test case based on its k closest neighbors. Evelyn Fix
and Joseph Hodges developed it in the early 1950s, and
Thomas Cover extended it subsequently[
Support vector machines are a model of artificial intelligence technology commonly used for data classification and regression. SVMs, which Vapnik introduced in the middle of the 90s, are supervised learning techniques based on statistical learning theory. Because of their greater capacity to construct an accurate representation of the connection between the input and output from a
Mohammed et al. [
Data acquisition instruments were included in the DSP program for acquiring data from the model and are supported by Matlab software. The peak position component method (PPCM) was used to identify the highest peaks and their positions from the PSD analysis and form a matrix for the input data of ANN. The Figure 4 shows the vibration spectrum of the first accelerometer. A multilayer chine (EDM) was used to create the faults in the bearing’s elements. A 14-bit USB-6009 National Instrument data acquisition is used to collect signals from three calibrated Figure 2: K-NN single-axis ADXL001 accelerometers. Labview, Matlab, and C language software were used for data acquisition to extract and analyze vibration data. After the DWT, perceptron ANN for each accelerometer was employed to the statistical features were extracted and used for declassify four health cases: normal shaft and a shaft with signing ANN. Only the standard deviation (STD) was 40%, 50%, and 60% pre-crack. The Levenerg-Marquart chosen as an input to ANN as it was the most sensitive learning algorithm was used as a training method. ANN feature for bearing faults. A multilayer perceptron neural for the data of the first and the third accelerometer distin- network was utilized for fault detection. Table 1 shows guished four conditions with 100% accuracy, and for the the parameters used for designing the neural network. second accelerometer data, it achieved 99.7% accuracy in ANN was trained to diferentiate between various types fault classification. of robot-bearing faults and achieved a remarkable level
Ali et al. [
Jaber and Bicker [
Sharma et al. extracted [32]frequency-domain features from the vibration signals of a three-phase induction motor. The vibration signal was collected from a single-axis accelerometer. The extracted features are used to classify three bearing conditions: healthy, inner race defect, and outer race defect. The electric discharge machine (EDM) was used to create 2 mm holes in the inner and outer races. For each condition, five features are extracted from the frequency domain: mean frequency, median frequency, lower band power, upper band power, and band power ratio. The acquired features were normalized to increase classification accuracy and distinguish any bias. As a result, ten sets of five normalized characteristics were collected for each bearing condition. Six sets were used for training ANN, while the remaining four were used for testing. A single-layer ANN was utilized for fault classification. To train ANN, two diferent techniques, scaled conjugate gradient (SCG) and Levenberg–Marquardt (LM), are used, and their performance is compared. For greater network sizes, the SCG algorithm outperforms the LM algorithm, and the proposed ANN could identify each fault condition with great accuracy.
Rao and Reddy [33]used wavelet transforms to identify a method for detecting irregularities such as open cracks or grooves on a rotating stepped shaft with several discs. The vibration signals were extracted from the displacement sensor and transformed into discrete and continuous wavelet transforms (DWT and CWT) at a specific rotor speed. The detailed process and techniques of analysis are shown in the Figure 9. The diference in wavelet coeficients of rotors with and without grooves is analyzed to identify the damage or groove locations.
A reduction in the shaft’s diameter was used to model the cracks and grooves classified as radial cracks. A numerical analysis of various cases was used to find each case’s first five natural frequencies. In this research, a three-layer feed-forward ANN was utilized. The DWT coeficient of various crack locations and depths was used to train ANN. The Levenberg-Marquardt algorithm is employed to train ANN. Consequently, the overall prediction accuracy of the developed ANN is 99.53%. It was able to detect cracks as small as 1% of their diameter.
Espinoza and Sinha [34]developed a smart vibrationbased machine learning model that uses ANN. The model is then tested blindly to detect the rig’s healthy and faulty conditions when running at diferent speeds. The test rig has four accelerometers placed on four bearings. Four scalar features were extracted from vibration data samples and arranged in three scenarios. In the first scenario, each sample is obtained randomly from an accelerometer from only one bearing. The second scenario considers the measurement from only one bearing with a fixed location.
The third scenario considers a simultaneous vibration signal collection from all four bearings. The scalar features are: root mean square (RMS),kurtosis, variance, and skewness. Spectral analysis is also used to examine the dynamic behavior of the rig under various conditions.
ANN of multilayer perceptron (MLA) based on the back propagation was utilized to classify acquired data. Depending on the results obtained in the particular scenario, either Bayesian regularization functions or scaled conjugate gradient are used to train the network. The results found that the first scenario observed has an almost 25% chance that the health conditions will be misdiagnosed as faulty. The second scenario performs relatively better than the first because consistent data from a specific bearing point exposes some features of machine behavior.
The third scenario indicates the best performance, with 100% accuracy. The model was trained at a rotational speed of 1800 rpm and was tested blindly using test rig data at 2400 rpm without training. The results showed 100% accuracy in machine condition diagnosis, as shown and used for training the KNN to detect the remaining in Figure 10. faults. The results showed an accuracy of 100% for the ifrst step, and the overall accuracy exceeded 90% for the
Pandya et al.[36] presented a fault classification He et al. [35] proposed a two-step plastic-bearing fault di- method based on acoustic emission for bearing health agnostics method. This approach investigated cage fault, monitoring. Five conditions were examined: healthy rolling element fault, and surface contact faults in the bearing, outer race, inner race, ball, and combined defect. inner and outer race. Two accelerometers attached to the The data was collected using an acoustic emission sensurface of the bearing house were used to collect vibra- sor installed on the housing of the test bearing and an tion data with NI PCI-4472B data acquisition. Frequency OROS 3 SERIES acoustic analyzer, and it was processed domain features were extracted from vibration signals by using NVGATE software. The Hilbert–Huang Transform the envelope analysis technique and fed to the statistical (HHT) was employed in this method. Band-pass filtering classification method. The first step only used the statis- with empirical mode decomposition was performed on tical classification method to classify the outer race fault, the collected signals, and IMFs were optimized. IMFs as shown in Figure 11. As a second step, entropy and were used to extract nine time-frequency domain feadistance (time-domain features) were extracted by EMD tures: peak, mean, RMS, kurtosis, crest factor, impulse factor, shape factor, ring down count, and energy. Multiple supervised machine learning techniques were used features (RMS, standard deviation, skewness, crest factor, for the fault classification using WEKA software, and the kurtosis, peak level, K factor, mean value, variance, and weighted KNN was the most eficient technique. Then, a median). PCA was used to reduce features. The waterfall modified KNN method based on an asymmetric proximity fusion model was used as the fusion approach to merge function (APF) was developed to improve classification sensor data efectively. The results of the experiments accuracy even further. The results demonstrate that the indicate that the proposed method improves fault identiproposed APF-KNN algorithm with optimized features ifcation and diagnostic accuracy, as shown in the Table 2. surpasses the KNN method by 96.6667%.
Wang et al.[37] developed a bearing fault diagnosis Tain et al.[39] proposed a method for detecting and method based on combining the Kernel Principal Com- monitoring bearing faults in an electric motor. Four faults ponent Analysis (KPCA) with the KNN algorithm. KPCA were investigated: the outer race, inner race, rolling elis a method for applying the kernel method to generalize ement, and cage fault. Vibration signals were extracted linear Principal Component Analysis (PCA) to non-linear from two accelerometers installed on the bearing houscases. An electro-discharge machine was used to artifi- ing and decomposed in the frequency domain to obtain a cially seed a single point fault in the bearing inner race, set of sub-signals. Diferent fault features were extracted outer race, and rolling element. The accelerometers were based on cross-correlation and spectral kurtosis (SK). The used to sample vibration signals positioned at the drive principal component analysis (PCA) technique was utiend. The extracted features are the clearance factor, the lized to reduce the redundancy of fault features. KNN is greatest peak, impulse factor, kurtosis, mean absolute used to combine the features with a health index, which is diference, peak factor, peak-to-peak value, rad ampli- further analyzed for defect detection. Following various tude, RMS value, skewness, variance, waveform index, experiments, a gearbox was mounted to create signals and kurtosis value. KPCA was applied to the extracted that mask bearing signals and cause false-negative designal and used to define KNN. As a result, this method tection. As a result, this method was efective in fault could classify bearing faults with an accuracy of 96.67%, classification and isolating bearing signals from gear sigas shown in Figure 12. nals to detect previously unknown faults.
Safizadeh and Latifi[ 38] developed a unique bearing Gohari and Eydi [40] used KNN to identify shaft unfault diagnostic approach based on KNN and a combina- balance in multi-disc rotors and compared the results tion of an accelerometer and a load cell. This method was with the Decision Tree (DT) Algorithm. Various masses used to classify three bearing conditions: a healthy, outer were mounted in three radiuses to create an unbalanced race defect, and ball defect. Spark erosion techniques condition. Two ADXL335 accelerometers attached to were employed to create the defects in bearing elements. the shaft were employed to collect vibration data from A Piezoelectric IMI 608A111 accelerometer was used to the test setup. A data acquisition device (ADVANTECH collect vibration data, and a SEWHACNM SM601 load 4711A) was used for data collection. A program was decell was used with a DACELL DNAM100 amplifier to mea- veloped in Labview Software for signal processing and sure the load. The output of each sensor was connected recording, and filtering of unwanted noises. Eight stato the NI-USB-9233 DAQ. Two frequency domain features tistical features were extracted from time and frequency were extracted (amplitude in ball pass frequency of the domain signals: the peak value, skewness, average, RMS outer race and ball spin frequency) with ten time-domain error, absolute average, the peak of average ratio, crest factor, standard deviation, kurtosis factor, and peak in the frequency domain. The extracted features were fed to the KNN and the DT algorithms. The study revealed that the KNN surpasses the Decision Tree in estimating unbalance parameters and can classify the faults with an average accuracy of 86.6%. The Figure 13 shows the KNN performance of unbalanced locating.
Lu et al. [41] proposed a fault diagnosis method for
rotating machinery called Enhanced K-Nearest Neighbor
(EKNN), based on KNN and spare coding. This method
investigated four bearing conditions: normal conditions,
inner race faults, outer race faults, and roller faults. An 3.3. SVM Applications
accelerometer attached to the bearing was utilized to
collect vibration signals. Fast Fourier Transform (FFT) was Jiang et al. [
Li et al. [42] presented a bearing fault detection method domain features were extracted for each condition, inbased on the Orthogonal Wavelet Transform K-Nearest cluding mean, peak, amplitude square, RMS, root ampliNeighbor Algorithm (OWT-KNN). The OWT can decom- tude, standard deviation, skewness, kurtosis, waveform pose the signals into the equivalent local detail signals at factor, pulse factor, and margin factor. The multi-sensor each scale in terms of time and frequency. The Figure 15 information fusion model was utilized to establish a mulshows the flowchart of (OWT-KNN). EDM was used to implant faults in the inner race and the ball. The vibration signals were collected from three accelerometers placed in diferent positions and subjected to a multilayer orthogonal wavelet transform. Peak-to-peak values were found on each scale to create the feature vectors. Lately, the created feature vectors are utilized to generate the classification model and train the KNN classifier. The classification results show that this approach can achieve a 100% fault classification. tion of bearing faults. This method studied the following cases: healthy, inner and outer race degradation, and inner and outer race failure. A run-to-failure test provided the data for this investigation. Vibration data was collected by four accelerometers attached to four bearings.
The RMS was calculated from the collected time-domain
signals; the Hilbert transformation was used to detect the
envelope of a time-domain signal. Then, the envelope
signal is converted into the frequency domain using the fast
Fourier transformation. Fault-specific frequencies are
found in the frequency spectrum, and energy associated
with each frequency band is collected. Four frequency
doFigure 15: OWTKNN Flowchart[42] main features were extracted in addition to RMS, and all
these features were used to train the SVM classifier. The
obtained classification accuracy for healthy, inner and
outer race degradation and inner and outer race failure
tidimensional vector by extracting the same character is 99.3%, 86.2%, 97.7%, 87.8%, and 84.2%, respectively.
from diferent sensors. The SVM was applied to each Huo et al. [
Tabrizi et al. [
Senanayaka et al. [
and abnormal vibration data. The SVM was applied to the collected data without preprocessing, and its classiifcation accuracy was 49.71%. Then, the FFT technique was used to extract the power spectrum from the time domain data, and the PCA algorithm was used to reduce the dimensions (see Figure 18). The SVM algorithm was applied to the processed data and could predict the normal and abnormal conditions with an average accuracy of 98.8%.
Rotating machinery fault diagnosis plays an important role in saving maintenance costs, downtime, and safety risks. In this study, a variety of AI approaches for rotating machinery diagnostics are discussed. The theoretical approach and fault diagnostic application of ANN, K-NN, and SVM have been reviewed. This study advises performing feature selection on the feature vector before employing AI classification techniques to identify the most sensitive feature for each fault case. Further, this study suggest using other AI classification techniques such as, Random Forest, Decision Tree, Naive Bayes, and Deep learning. As AI techniques grow more sophisticated, it is expected that AI approaches will remain interesting and efective for detecting faults in rotating machinery.