Proceedings of ICASI - 2004 International Conference on Advances in Structural Integrity July 14-17, 2004, Indian Institute of Science, Bangalore,

India

ICASI/XX-XXX

Time domain structural health monitoring with magnetostrictive patches using five stage hierarchical neural networks.
Ghosh D. P. (a) and Gopalakrishnan S. (b)
(a) Graduate student (b) Assoc. professor, Dept. of Aerospace Engineering, Indian Institute of Science, Bangalore 560012 ABSTRACT
An integrated method for damage detection of composite laminates is presented in this paper using time domain data obtained from magnetostrictive sensors and actuators and artificial neural networks (ANN) identification with five stage hierarchical neural network (HNN). Magnetostrictive actuators are activated through an actuation coil, which vibrates the composite laminate. The presence of delamination is sensed through a sensing coil as open circuit voltage. The ANN is applied to establish the mapping relationship between structural damage status and sensor open circuit voltage. A fivestage hierarchy of networks is used for the identification procedure. The results of delamination damage detection for composite laminate show that the method developed in this paper can be applied to structural damage detection and health monitoring for various industrial structures. To demonstrate this approach, numerical simulations are carried out on a composite cantilever beam to identify size and location of delamination using the sensor data for a known actuation for a certain combination of sensor and actuator locations. Keywords: Magnetostrictive, SHM, ANN, Hierarchical Neural Network, Inverse Problem.

1. INTRODUCTION
Composites have revolutionized structural construction. With the availability of smart materials and the feasibility of embedding them into or bonding them to composite structures, smart structural concepts are emerging to be attractive for potential high performance structural applications. 1 Authors4 had developed a finite element formulation for inbuilt magnetostrictive patches for performing numerical simulations. The mathematical relationships between structural response (sensor open circuit voltage) and structural damage status are very complex. ANN has particular advantage in establishing accurate mapping relationships between sensor data and physical parameters of structural damage. Hung and Kao2 and Yun and Bahng 3 reported their researches on structural damage detection using ANN, and their results showed that ANN is a highly effective tool for identifying structural damage. By using the hierarchical scheme, a complicated large-scale system is decomposed into a set of lower

order subsystems and a coordination process, and thus becomes tractable. Here five stages of hierarchy are considered for structural health monitoring case.

2. FORWARD ANALYSIS
Application of magnetic field causes strain in the magnetostrictive material (Terfenol-D) and the stress, changes magnetic flux density of that material. 4 Using this phenomenon, the time domain sensor response due to actuation in the actuator can be computed through finite element formulation. In the forward analysis for a given actuation history and a given damage condition, sensor response history can be obtained for a particular structure. Detail procedure for forward analysis can be obtained from 5.

3. INVERSE ANALYSIS
Damage identification is the geometric inverse problem, where actuation and response of structure are known but structure is unknown. Artificial neural networks (ANNs) can provide non-linear parameterized

mapping between the response and damage of the structure with unknown function relationship.

3.1. Hierarchical Neural Network

With the common three-layer neural network architectures, networks lack internal structure; as a consequence, it is very difficult to discern characteristics of the knowledge acquired by a network in order to evaluate its reliability and applicability. Alternative neural-network architecture is presented, based on a hierarchical organization. By using the hierarchical scheme, a complicated large-scale system is decomposed into a set of lower order subsystems and a coordination process, and thus becomes tractable. A five-stage hierarchical neural network is designed by combining a multilayer perceptron for first stage and mixture-of-experts in the subsequent stages. The second stage mixture-of-experts, ensembler network, learns to minimize the overtraining errors. The third stage mixture-of-experts, validation network, learns to minimize the validation errors. And the fourth stage mixture-of-experts, expert networks, learns to minimize the error of network due to loss of information for input space dimension reduction. And Finally fifth stage committee machine choose the appropriate expert network from all expert networks. The whole procedure is discussed as follows.

otherwise it is overtrained. Every trained neural network is tested through the test data set, which gives the testing errors as a measure of generalization of the neural networks. These testing errors with the weightages of the neural networks give the weighted average of testing error, as a measure of testing error of the ensemble network. Next issue in the neural network is the generalization of the network using the testing error of ensemble network.

3.3. Validation Network

As, every neural network within the ensemble network is trained with same training sample; these neural networks need testing for generalization of the ensemble network. For testing of network, sample data set is separated into two groups, one for training and other for testing (validation). After training of neural networks, with the training sample, the neural networks simulate the testing sample and get the testing error. These testing errors with their weightages give the measure of generalization of the ensemble network. However, every network in the ensemble network is trained and tested through the same set of training data set and testing data set respectively. To get a more general input-output mapping from a set of sample data set, different division of training sample and testing sample is essential. This can be done through a systematic manner using validation network. Validation Network consists of some number of ensemble networks. These ensemble networks are trained and tested with same sample data set with different partition for training data set and testing data set. However, every validation network is train for a fixed set of sample data set.

3.2. Ensemble Network

Often artificial neural networks are prone to overtraining, where network trains the computational and experimental noises. And there is no direct rule to draw the line between well training and overtraining for a set of training examples and network architecture. One of the indirect ways to get a measure of overtraining of the network is Ensemble Network. In ensemble network, a number of neural networks are train with the same training samples but with different initial condition, learning rate, training algorithm, network architecture and training sequence (for sequential learning). In training phase, each network trains and generates training error for the training samples. On the basis of these training errors, weightages of the trained neural network is determined, where less training error gives more weightage of the neural network. In the execution phase, these weightages are used to get weighted average of all neural networks output as the output of ensemble network. From the distribution of the output of different neural networks and their corresponding weightage one measure of overtraining can be computed. If the outputs of different neural network are close (at least for those has more weightage) then networks are well trained

3.4. Dimensional Reduction

In time domain structural health monitoring, time histories of the sensor outputs (open circuit voltages) are the original input space for the input-output mapping, which is very high dimensional. So, it is not possible to train the network taking full dimension of the original input space (sensor output). To address this issue, different dimension reduction procedures are available in the literature, which can reduce the dimension of the original input space keeping main features of the high dimensional data. But for structural health monitoring, suitable dimension reduction procedure is not available, which will reduce the dimension of the input space preserving the signature of damage from the high dimensional sensor output data. To overcome this problem a number of reductions procedure is taken to increase the chance of preserving the damage signature in the reduced input data sets. In

the next section this issue is discussed in a systematic manner.

3.5. Expert
Expert consists of a number of validation networks, which are train and tested through same output of a sample data set but with different lower dimensional input of the data set. These different input data sets come from different dimension reduction procedure of the original input set, which is high dimensional sensor output. So, every expert is a mapping from sensor output to the damage properties of the structure and performance of the expert depends on the performances of its validation networks, which is trained through dimensionally reduced input sample set.

wise location of the delamination is difficult using general type of experts. Ones the span wise location is determined using general type of experts, these conditional experts are trained taking samples from within that location.

3.8. Training and Testing of HNN

In training phase, every multilayer perceptrons (MLP) are trained and tested with their corresponding training and testing samples. These training will give their training and testing error. Although network architectures and learning algorithms are same, due to different initial condition, learning rate, momentum rate and learning sequence the trained MLPs will be different. Depending upon the training error every trained MLP is associated with their weightage. In the ensemble network these weightages and their testing errors are considered to calculate average testing error. These testing errors are for every ensembler network. All ensemble networks within one validation network calculate are weighted average testing error. This is the training and testing phase of the hierarchical network. Training and testing phase is limited within ANN and ensembler network.

3.6. Committee Machine

But some times, due to this reduction procedure the mapping looses the output uniqueness, which is essential for the training of neural network. It is perhaps impossible to combine simplicity and accuracy in a single model of ANN. Single multi-layered perceptron (MLP) uses a black box approach to globally fit a single function into the data, thereby losing insight into the problem. This problem was studied6 by partitioning the input and output space into a piecewise set of subspaces, with each subspace having its own expert. Size of these subsets are such that the output uniqueness within a subset is preserved as well as sufficient number of sample data is available to train and test the network. Then for every subset one expert is trained and tested taking sample data from these subsets. Similar to the division of output space, input space can be divided in different subsets on the basis of sensor, actuator and actuation combination. But as the input space subdivision is known a priory, committee machine for input space subdivision is not required.

3.9. Execution of HNN

In identification phase, execution samples (time domain sensor output) are feed into every expert to get their opinion and confidence. From this mutual information the location and size of the delamination is obtained with the level of confidence. Every expert pass on these execution samples to their subordinates, validation networks after assigned dimension reduction. These validation networks also pass on the sample to ensembler network. Similarly ensembler network gives these samples to their subordinate MLP (ANN) network. These MLP execute these samples and give their opinion. As different MLP are trained differently, their opinion will also be different. These opinion and their corresponding training weightage will create opinion and variance for the ensembler network. After getting all the ensembler level information, information on validation network will be created by fusing this ensembler level information. In the execution phase of validation network, the opinion and variance of ensembler network is used. Opinion of validation network is created by the opinion of ensembler network and their corresponding weighted average testing error. Similarly variance of validation network is created from variance of ensembler network with their weighted average testing error. If this opinion is with in the range of inherited expert and variance is with in their accepted limit, this validation network is considered as

3.7. Conditional Expert

Always some dimension of output space is difficult to train than other dimension of output space. In SHM of composite laminate, depth wise training is difficult than span wise training. To solve this kind of problem some conditional experts are trained. If the dimension of output subset of some expert is less than the original dimension of output space, the value for the missing dimension is required a priory to execute the network. These experts are called Conditional Expert. Conditional experts are used to get finer location for some dimensions in the output space, when locations for remaining dimensions are already known by other type of experts. For structural health monitoring problem in composite laminate, identification of layer

active validation network. As the dimension reduction algorithm is different between different validation networks, one weightage for dimension reduction is used to give more weightage for better algorithm. This operation will be done for all experts. For execution in the expert level, information are fused from subordinate, validation networks. Active experts are determined depending upon the maximum number of active validation network within that expert. Then weighted opinion and weighted variance among active validation network are calculated, as the opinion and variance of the hierarchical network (HNN). If more than one active expert are available, committee machine calculate the opinion and variance of the HNN considering variance of all active experts.

excitations with different frequencies (50, 500. 5000 Hz) are exerted on the dynamic model of the composite laminate, and the vibration responses of 550 different cases are numerically simulated for each frequency. These 550 cases include the intact laminate, laminates with delamination damage at different layers and of 50 different delamination sizes (10 mm to 500 mm) at each layer. For structural health monitoring Hierarchical Neural Network (HNN) is used. Vibration responses of higher dimension (500 time steps) for a given delamination (open circuit voltage in magnetostrictive sensor), is preprocessed for dimension reduction in the input space of the neural network. Here for simplicity first fifteen optimum values and their location of the sensor open circuit voltage and their time integrals are taken as the input space of the four type of validation neural network. Every validation network consists of four ensemble networks. These ensemble networks are trained and tested by same sample data set but with different random partitioning between training and testing data sets. So every ensemble network is trained and tested by a fixed set of training and testing data sets respectively. In order to identify the delamination length at each layer, one BP neural network with 15 inputs and 2 outputs are designed. One hidden layer of node strength 10 is taken as the net architecture. Every expert is trained by the sample data within their expertise location. These samples are for the delamination in the corresponding location. Thus there are thirty experts for each actuation frequency to predict the size and location of the delamination. A number of delamination identification is performed using proposed Hierarchical neural network and shown in the Figure-3. It is shown that depth identification is difficult than span wise identification.

4. NUMERICAL EXAMPLES
In this paper a numerical study on 12 layered beam containing two patches, one acting as an actuator and the other as a sensor has been presented.

4.1 Forward Analysis

Numerical simulation is carried out by considering a unidirectional laminated composite beam (Figure-1). Elastic modulus of composite is assumed 181(E1) and 10.3(E2) GPa. Poison ratio, density and shear modulus of composite are taken as 0.0, 1.6 gm/c.c. and 28(G12) GPa respectively. Elastic modulus (Em), poison ratio (m), shear modulus (Gm) and density (m) of magnetostrictive material are assumed 30 GPa, 0.0, 23 GPa and 9.25 gm/c.c. respectively. Magnetomechanical coupling coefficient is 15E-09 m/amp. Direct transient dynamic analysis has been done with 500 time steps to calculate open circuit voltage of the sensor in each time steps. Relative permeability (r) is assumed as 10 for magnetostrictive material. Number of coil turn in sensor (Ns) and actuator (N) is assumed 1000. Actuation current at actuator (I) is taken as 1.0 Amp at three different frequencies. 4.1.1. Result and Discussion Numerical results have been simulated for a fixed position of sensor and actuator combination, for different locations of the delamination. Figure-2 shows open circuit voltages in the sensor when the delamination is between 4th and 5th layer with input frequency 5kHz.

5. CONCLUSIONS
The study demonstrates the use of vibration response using magnetostrictive sensor and actuator of a structure for health information of the structure. This study is successful in classifying and identifying structural damage location and severity using the designed hierarchical neural network (HNN). The results show that HNN is a powerful tool for establishing the mapping relationships between open circuit voltages and the structural damage status, and demonstrate the ability of HNN for structural damage detection.

4.2. Inverse Problem

In this study, sinusoidal actuation current is considered in the actuation coil. Three sin wave

REFERENCES
1. Loewy, R.G. Recent developments in smart structures with aeronautical applications, Smart Materials and Structures, 6, 1997, pp. R11-R42. S.L. Hung and C.Y. Kao, Structural damage detection using the optimal weights of the approximating artificial neural networks. Earthquake Eng. Struct Dyn. 31 (2002), pp. 217-234. C.B. Yun and E.Y. Bahng, Substructural identification using neural networks. Comp. Struct. 77 (2000), pp. 41-52 Butler, J.L, Application manual for the design of Terfenol-D magnetostrictive 5.

2.

6.

3. 4.

transducers, Edge Technologies Inc., Ames Iowa, 1988. Ghosh D. P. and Gopalakrishnan S.; Structural health monitoring in a composite beam using magnetostrictive material through a new FE formulation. International Conference on Smart Materials, Structures and Systems; Dec12-14, 2002, Indian Institute of Science, Bangalore, India. Ghosh D. P. and Gopalakrishnan S.; Identification of delamination size and location of composite laminate from time domain data of magnetostrictive sensor and actuator using artificial neural network. Proceeding of the SEC 2003, December 12-14, structural engineering convention an international meet.

FIGURES AND TABLES

Figure 1: Laminated Beam with Actuator, Sensor and Delamination.

Figure 2: Actuation frequency 5000 Hz

Figure 3: Identification using HNN

Figure 4

Figure 5