CHAPTER 2 Literature Review
2.2 Study of structural integrity monitoring techniques
2.2.3 Artificial intelligence techniques used for crack identification
2.2.3.2 Artificial neural network techniques
In curent section different types of Artificial Neural Network (ANN) based techniques used for crack detection are described. ANN is used as a capable technique for damage detection.
A Back Propagation Neural Network (BPNN) technique for condition monitoring of cracked beam has been designed by Rakideh et al. [68]. They have extracted the first three natural frequencies of the beam using analytical method and fed these natural frequencies to BPNN model to predict the crack location and crack depth. They have concluded that the neural network is a powerful method to determine the location and depth of the crack. Also, the capability of prediction accuracy has increased with increasing the numbers of the natural frequencies. But an experimental verification is required in above paper. An artificial neural network based approach for identification of damage in an industrial welding robot has been proposed by Eski et al. [69]. To achieve the objective, an experimental setup is fabricated to accumulate the related data and the accelerations of welding robot, which has six degrees of freedom, are examined. They have concluded that RBNN is a capability to analyze the accelerations of manipulator joints during a prescribed trajectory. Schlechtingen and Santos [70] have proposed a comparison of results obtained from three different models, are the regression based model and two artificial neural network based models, which are a full signal reconstruction and an autoregressive normal behavior model used for fault identification of a wind turbine bearing. After the comparison of results, they have found all three proposed models were capable to identify initial faults. They have determined the neural network based approaches give best fault visibility with less computational time with comparison to regression based approach.
Multi Layers Perceptron (MLP) and Self Organizing Map (SOM) neural network based classifier for prognosis of fault of three phase induction motor and evaluated the performance of classifiers have been developed by Ghate and Dudul [71]. The different number of learning rules and transfer functions has investigated for different number of hidden layers. The simple statistical parameters used as input feature space and principal
component analysis are used for reduction of input dimensionality. They have also tested their approach with noise and found the performance of the proposed method satisfactory.
As the further analysis RBFNN may give better results than MLP and SOM. Fan et al.
[72] have presented a fault detection and diagnosis strategy for local system of air handling units. The strategy consists of two stages which are the fault detection stage and the fault diagnosis stage, respectively. In the first stage, the neural network fault detection model is used by them for generating estimates of sensor values and they have compared to actual values to produce residuals. The design neural network fault detection model has trained using an abundance of characteristic information from the historical data in the system. They have concluded that the trained neural network model can detect the abnormal condition in the system.
Saravanan et al. [73] have evaluated the effectiveness of wavelet-based features for fault diagnosis of a gear box using artificial neural network (ANN) and Proximal Support Vector Machines (PSVM). They have found PSVM has superiority over ANN in classification of features. Paviglianiti et al. [74] have proposed a detection and isolation sensor faults in a robotic manipulator. The proposed methods can trackle the influence of outer disturbances and uncertainties of the model. The dynamics of the proposed model have enhanced by using a radial basis function type of neural network. A new damage detection technique by using the Auto Regressive (AR) with a back propagation neural network has been developed by Wang et al. [75]. They have found the difference in the values of AR coefficients, which indicates AR coefficients of ideal signal for normal machine are deducted from faulty machines. The results obtained by them are compared with the three methods, which include the difference of AR coefficients with BPNN, the AR coefficients with BPNN and the distance of AR coefficients method for 23 samples.
The authors have found that the difference of AR coefficients with BPNN were superior to AR coefficients with BPNN and distance of AR coefficient methods.
A fault identification and health monitoring techniques for a gear-set using continuous wavelet transform and neural network technique has been presented by Wu and Chan [76]. In the proposed fault diagnosis technique, sound emission of the gear-set is used for evaluation. A continuous wavelet transform technique combined with a feature selection of energy spectrum is used for examining fault signals in a gear-set of machines. They have concluded that the sound emission from the system can be used for promising fault diagnosis and condition monitoring of the rotating machines. Mehrjoo et al. [77] have proposed a damage detection methodology to assess the damage intensities of joints in
truss bridge structure using soft-computing technique i.e. back propagation neural network. But thorough experimental validation is required in above papers [76-77].The Modal frequency parameters such as natural frequencies and mode shapes are fed as input to BPNN for evaluation of damage intensities of joints in truss bridge structure. Just-Agosto et al. [78] have developed a fault detection technique using the Bayesian probabilistic neural network. A combination of vibration and thermal parameters sent to neural network as input data to detect fault in sandwich composite. A fault diagnosis and condition monitoring technique for internal combustion engine using Discrete Wavelet Transform (DWT) and neural network has been presented by Wu and Liu [79]. They have combined the DWT technique with feature selection of the energy spectrum of diagnosis of faults in rotating engines.
A faults detection methodology for structures using modal parameters and statistical neural network model has been presented by Bakhary et al. [80]. They have considered the effect of uncertainties in developing ANN model, by applying Rosenblueth‟s point estimate method verified by Monte Carlo simulation, the statistics of the stiffness parameters are estimated. But thorough experimental verification is needed in the above paper. A neural network based methodology for condition monitoring of axial flow fan blades has been presented by Oberholster and Heyns [81]. They have developed a methodology in two stages, in the first stage. The Neural networks are trained on features extracted from on-line blade vibration signals measured on an experimental test structure.
In the second stage, neural networks trained on Numerical Frequency Response Function (NFRF) features obtained from a Finite Element Model (FEM) of the test structure. They have concluded that numerical approach is more preferable to the experimental approach where it is less costly to construct, update and test an FEM than to test an experimental or operational structure by means of damage simulation and suggested methodology can handle the online damage classification using sensor for the test structures. But in above paper deviation of results is not discussed. Yeung and Smith [82] have investigated a damage detection methodology, using pattern recognition of the vibration signature and unsupervised Probabilistic Resource Allocating Network (PRAN). They have found that sensitivity of the neural networks can be adjusted so that a satisfactory rate of damage detection can be achieved even in the presence of noisy signals.
A health monitoring of the cantilever beam containing transverse surface crack using neural network techniques has been developed by Suresh et al. [83]. They have calculated modal frequency parameters for different crack locations and depths using analytical
method and these modal parameters are used to train the neural network to detect the damage severity and intensity. A comparative study on the performance neural networks, such as multi-layer perception network, radial basis function network is done by authors.
The authors have found that radial basis function network performance is better than multi-layer perception network. Kao and Hung [84] have presented structural condition monitoring technique using a supervised learning type of Neural System Identification Networks (NSINs).They first identified the undamaged and damaged states of a structural system using NSINs then trained NSINs has used to generate free vibration responses with the same initial condition or impulsive force of structures. In [83-84] it is reported that thorough experimental verification is needed for the authentication of results.