Artificial neural networks are currently the most commonly found data-driven techniques in fault diagnosis literature (Heng et al. 2009) and they have been widely implemented in the fault diagnosis of rotating machinery. Examples of ANN applications in fault diagnosis are listed in Table 3.1, where applications to common machine groups are reported.
Table 3.1 - Neural Networks Applications in Fault Diagnosis of Rotating Machinery Neural
Networks
Motors Pumps Bearings Turbines Gears & Gearboxes Shafts BPFFN Mahamad and Hiyama (2011) , Li et al. (2004) Ilott and Griffiths (1997) Li and Wu (1989), Liu and Mengel (1992), Samanta and Al- Balushi (2003), Gebraeel et al. (2004), Sreejith, et al. (2008) Wang et al. (2010), Prieto et al. (2013) Yang et al. (2008), Wu and Chan (2009) McCormick and Nandi (1997), Kuoet al. (2002) RNN Yam et al. (2001), Malhi et al. (2011) Mohamm- adi et al. (2011) RBF Selaimia (2006), Onel and Benbouzid (2008) Lu et al. (2011) Rong et al. (2009) Li et al. (2009) MLP Paya et al. (1997) Meesad and Yen (2000), Senguler et al. (2010) Meesad and Yen (2000) Meesad and Yen (2000) Kohonen Kowalski and Kowalska (2003), Bay and Bayir (2005)
SOM Yang et al.
(2004), Premrudee preechach arn et al. (2002) Zhang and Ganesan (1997), Zhong et al. (2005), Hu et al. (2003), Wu et al. (2002), Donat et al. (2008) Liao et al. (2005) LVQ Zareiet al. (2008) Abu- Mahfouz (2005) Meesad and Yen (2000)
Several selected applications listed in Table 3.1 are explained below.
ANN systems have been used to support classification of fault analysis. An earlier application of ANNs in the field of bearing fault analysis can be found in Li and Wu (1989). In the investigation, a perceptron-type network was used to analyse the
Chapter 3 – Review of Artificial Intelligence (AI) Systems in Fault Diagnosis 57 experimental data from ball bearings. The results showed that the network recognised faults made on the outer race of the bearings with a percentage of error smaller than the one of the conventional methods. It was reported that the proposed technique achieved a 14 per cent better rate than the conventional methods.
In an investigation by Liu and Mengel (1992), it was shown that the perceptron network was capable of distinguishing between six different cases of ball bearing faults. The fault detection used the variations of the peak amplitude in the frequency domain, the peak RMS and the power spectrum parameters as the training data for the perceptron network.
Baillie and Mathew (1994) diagnosed rolling element bearing faults using artificial neural networks and a bearing fault diagnostics system was developed. The incoming vibration signal was presented to each neural network model in the system and the network model that best approximated the signal was chosen to indicate a type of fault. The system was trained to diagnose fault conditions such as imbalance, outer race faults, inner race faults and normal conditions. The neural network-based diagnostics system was tested and it was shown that the system achieved accuracy in 95 per cent in all the test data set.
Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF) neural networks utilised to classify the condition of rotating machines were used by McCormick and Nandi (1997). In the classification tasks, the findings were that similar success rates were achieved by both MLP and RBF networks. Further detail shows that the RBF networks needed significantly shorter time for training compared to the time needed by the MLP network. However, the MLP network achieved faster operation time and used fewer neurons.
Application of MLP and Learning Vector Quantization (LVQ) as classifiers used to diagnose faults in gears, bearings and shaft was carried out by Meesad and Yen (1997). Both networks performed successfully. Off-line training and iterative data feed were needed to achieve a successful fault classification process.
Zhang and Ganesan (1997) applied a multi-variable trend estimation of fault development to predict RUL of a bearing system by using a self-organising neural network. Condition monitoring was performed via online vibration measurements and fault quantification was formulated into a multivariate trend analysis. Self- organising neural networks were used to perform the multivariate trending of the fault development. It was found that the accuracy of the proposed prediction algorithm was the same as one of the SOM algorithms.
Yam et al. (2001) investigated the trend in predicting machine condition by using a recurrent neural-network system.
Kowalski and Kowalska (2003) demonstrated a Neural Networks application for induction motor fault diagnosis. In this research work two kinds of NN were proposed as multilayer perceptron networks and self-organising Kohonen networks. The results of the experimental tests showed that neural networks could be effectively used for the recognition of stator, rotor, rolling element bearing and supply asymmetric faults by appropriate measurements and interpretation of FFT analysis of current vibration spectra.
Gebraeel et al. (2004) carried out investigations on thrust bearing prognosis in an attempt to determine the prediction of the actual bearing failure time. The investigation aimed to develop neural-network-based models for predicting bearing failures. An experimental setup was developed to perform accelerated bearing tests where vibration information was collected from a number of bearings that were run until failure. This information was then used to train neural network models for the prediction of bearing operating times. Vibration data from a set of validation bearings was then applied to these network models. The resulting predictions were then used to estimate the bearing failure time.
Sreejith et al. (2008) proposed an application of neural networks for automated diagnosis of localised faults in rolling element bearings. Kurtosis and log-likelihood classification (Goumas et al. 2001 and Abbasion et al. 2007) extracted from time domain vibration signals were used as an input feature for the neural network. The
Chapter 3 – Review of Artificial Intelligence (AI) Systems in Fault Diagnosis 59 results showed that the trained neural network was able to classify different states of bearing faults with an accuracy rate of 100%.
Wang et al. (2010) used the autoregressive (AR) method combined with the back- propagation neural network (BPNN) in rotating machinery fault diagnosis. A new fault diagnosis method was studied by using the differences in AR coefficients with BPNN. The obtained diagnosis results were compared with three methods, BPNN with AR coefficients, BPNN with AR coefficient differences and BPNN with AR coefficient distances. It was found that the diagnosis results obtained by using BPNN with AR coefficient differences were superior to the other two methods.
Prieto et al. (2013) combined statistical-time features and neural networks in the detection of bearing faults in electric motors. Statistical-time features such as root mean square (RMS), standard deviation, variance and crest factorwere calculated from acquired vibration signals. The discriminant analysis (DA) value was used for the purpose of feature selection. The final classification tasks were carried out using a hierarchical neural network structure and the effectiveness of the method was verified by experimental results obtained from different operating conditions. The proposed method achieved a 95 percent classification rate of the overall test set.