Elsevier

Pattern Recognition Letters

Volume 30, Issue 3, 1 February 2009, Pages 321-330
Pattern Recognition Letters

Pattern recognition of guided waves for damage evaluation in bars

https://doi.org/10.1016/j.patrec.2008.10.001Get rights and content

Abstract

Guided waves damage identification in bars with neural networks acquires training data from simulation as a cost-effective measure. These neural networks applied with a novel test inputs dependent iterative training scheme are capable of quantifying damages accurately from experimental inputs. The reliability of the predictions depends on the quality of the measured signals, which can be increased by considering more than one signal obtained from different sensor locations or by changing the properties of the interrogation pulse. A parallel network system to process the inputs from these signals collaboratively is described. The core of the system is a data fusion process that associates overlapping intermediate test results while isolating outliers to narrow the training range for improved generalization in the iterative test inputs dependent training scheme. This robust system of signal processing has achieved accurate average damage quantitative results with errors below 4% and 13% the original size of the training parameter space for damage location and depth, respectively, of artificial laminar defects in bars.

Introduction

Guided waves are ultrasonic waves with wavelengths larger than at least one of the characteristic dimensions of the structure (Rose, 1999). The advent of guided waves in quantitative nondestructive evaluation and structural health monitoring in recent years is due to the high capability of these waves to detect and characterize damages in structures (Raghavan and Cesnik, 2007). Guided waves also have the ability to propagate large distances, interrogating less accessible and obstructed locations of the structures for example, the immersed supports of an offshore oil platform (Brincker et al., 1995) or underground gas pipelines (Kim et al., 2006). In addition, high quality yet readily available transducers and sensors for guided waves allow for the development of cost-saving damage detection tools (Staszewski et al., 2004). However, there remains the challenge to quantitatively identify damages from the raw signals alone, which leads to the implementation of advanced signal processing including neural network pattern recognition.

Pattern recognition with neural networks is an advanced signal processing technique that can serve to evaluate damages in structures. In guided wave ultrasonics, pattern recognition is used to identify features from the measured transient wave response signals with the aim to quantify damages from a large parameter space within the structure. Training data for supervised neural networks is obtained from simulation as a cost-effective and practical approach. Simulations of wave response signals can be derived from modeling the damage based on the reflection and transmission coefficients (Wang and Rose, 2003) or finite-elements (Yang et al., 2006). Training with simulated inputs and testing with experimental inputs have shown promising accuracies in quantifying the damages in bars as demonstrated by Liew and Veidt (2005), and in plates as presented by Su and Lin (2004).

Liew and Veidt (2007a) found that when training was repeated for a multi-layer perceptron backpropagation neural network for the same training data and network architecture resulting in different network weights, the damage characterization test results for recognizing experimental inputs were inconsistent. Large fluctuations were observed in identifying the damage location with approximately 20% of the test trials exhibiting errors more than half the size of the damage parameter space. Only the mean prediction of the test trials was found to give reasonable damage quantification results. The inconsistencies are attributed to subtle differences between simulated and measured inputs due to the complex nature of guided waves propagation in structures. This can include noise, mode coupling, dispersion and additional wave modes, which are all difficult to accurately reproduce in simulation (Graff, 1991).

Literature has shown a broad range of optimization techniques for neural network design. Leung and Lam (2003) have proposed the tuning of the neural network structure and parameters by means of a modified genetic algorithm that applies evolutionary computation to locate the learning performance with good fitness. Lin and Lee (1994), and later Huang and Yi (2003) have adopted fuzzy logic in the transfer functions of the network to determine the optimum number of nodes in the hidden layer. Another approach is the selection of optimal initial network weights, which has been studied by Yam and Chow (2000). The learning environment can be improved by means of heuristic techniques that adjust the learning rates during training and numerical optimization techniques in the training algorithm including conjugate gradient and quasi-Newton methods for faster computation with minimum generalization error. Bishop, 1995, Hagan et al., 2002 are comprehensive references for these techniques. All these methods are focused on improving the convergence rate, training speed and generalization error with respect to the training data.

However, there is no known research on optimization techniques that considers the influence of the test inputs, which can be beneficial when there are subtle differences between training and test data. This is the case in the present study for guided waves damage evaluation where training data is obtained from simulations while test data is measured from experiments. In view of this, Liew and Veidt (2007b) developed a training scheme that is governed by the outcome of intermediate test results, introducing neural networks that are inherently dependent on the test inputs. This method involves progressive narrowing of the training range in a series of conventional neural network training and test processes while maintaining the same training data size where generalization and thus the accuracy of average predictions to experimental test inputs can be increased.

An extension of this study was conducted with the integration of a combined network into the test inputs dependent training scheme to improve the robustness and reliability of the system. Sharkey (1999) provided an excellent reference on combined networks that can be categorized as either an ensemble or a modular system. Well-established ensemble techniques that involve resampling of the training data like bagging (Breiman, 1996) and boosting (Freund and Schapire, 1996) can be used to improve the performance of the neural network. In the modular network structure, task decomposition allows module members to be specialized in processing specific tasks in order to produce a better overall solution compared to a single network. The mixture of experts with a gating network to mediate the outputs by Jordan and Jacobs (1994) is a popular approach in modular networks.

As existing combined network techniques also focus on the training data, a new approach known as a parallel network is introduced in this paper for integration into the test inputs dependent training scheme. Each neural network in parallel processes data acquired from a different sensor location or from changes in the properties of the interrogation pulse excitation. Test results from the neural networks are then merged statistically by data fusion to produce a new training range. The data fusion process involves associating classification overlaps and isolating outliers in order to reduce the size of the damage parameter spaces for subsequent neural network processing in the test inputs dependent training scheme. Klein (1993) pointed out the benefits of data fusion for neural networks processing a multitude of sensors, which includes higher signal-to-noise ratio, better reliability and increased hypothesis discrimination.

The experimental setup for the guided wave ultrasonics application in bars and the concepts behind the simulation program are described in Section 2. Subsequently, the experimental and simulated transient wave responses are compared, leading to the explanation of the feature extraction method. In Section 3, the neural network design is presented followed by a detailed explanation of the test inputs dependent parallel network scheme including the data fusion process. The potential of the system is supported by results and discussion in Section 4, ending with conclusions in Section 5.

Section snippets

Experiment and simulation

Fig. 1 shows the schematics of the experimental setup for exciting and measuring guided waves in a 2 m long 12 mm × 6 mm aluminum bar. A 2 mm × 12 mm × 6 mm longitudinal Pz27 piezoceramic transducer was adhesively bonded to one end of the bar and wired to a SRS DS345 function generator through a Krohn-Hite 7500 amplifier for the excitation of 8-cycle narrowband guided wave pulses with f center frequency. A 4 mm × 12 mm × 6 mm brass backing mass was attached behind the transducer to increase the signal-to-noise

Neural network design

The parallel network system was the combination of conventional neural networks of the same architecture. This was assumed a valid concept as all the neural networks in the system processed wavelet transform inputs derived from damage parameters within the same parameter space. The fundamental architecture selected was a feedforward backpropagation multi-layer perceptron with a single hidden layer governed by the following equation:Ov=m=1MWv,mFu=1UWm,uIu+Bm+Bvwhere I is the input, O is the

Results and discussion

Computations were conducted with MATLAB® running Neural Network Toolbox 4.0.4 in default settings. The same neural network architecture described in Section 3.1 was maintained throughout the respective configurations in the parallel network system for consistency and K = 50 test trials were conducted. n = 1 and imax = 2 were selected to allow adequate range reduction for improved generalization while data fusion removed any present outlying configurations.

The computation time varied broadly among

Conclusions

The test inputs dependent parallel network system is an interesting and novel framework that combines the robust features of data fusion and the improved generalization concept of the WRS technique. The system has been specifically designed for guided waves damage identification in bars for structural health monitoring but is generally applicable for any pattern recognition problems with relatively high uncertainties in predictions from repeated tests due to subtle differences between training

Acknowledgments

The authors appreciate the invaluable advice and feedback provided by Assoc. Prof. David Mee from discussions early in this research. C.K. Liew is also grateful for the support received from an International Postgraduate Research Scholarship (IPRS) awarded by the Australian Department of Education Science and Training (DEST), and a UQ Graduate School Scholarship (UQGSS) from the Unversity of Queensland (UQ).

References (34)

  • K.F. Graff

    Wave Motion in Elastic Solids

    (1991)
  • M.T. Hagan et al.

    Neural Network Design

    (2002)
  • Huang, X.M., Yi, J.K., 2003. A method of constructing fuzzy neural network based on rough set theory. In: Proc. 2nd...
  • M.I. Jordan et al.

    Hierarchical mixtures of experts and the EM algorithm

    Neural Comput.

    (1994)
  • Kim, J.Y., Lee, D.H., Park, K.S., Jo, Y.D., Choi, S.C., Lee, C.H., Song, S.J., Cheong, Y.M., 2006. Long range...
  • L.A. Klein

    Sensor and Data Fusion Concepts and Applications

    (1993)
  • H.F. Leung et al.

    Tuning of the structure and parameters of a neural network using an improved genetic algorithm

    IEEE Trans. Neural Networks

    (2003)
  • Cited by (15)

    • Convolutional Neural Networks for vehicle damage detection

      2022, Machine Learning with Applications
    • Vibration-based diagnostics of epicyclic gearboxes – From classical to soft-computing methods

      2019, Measurement: Journal of the International Measurement Confederation
      Citation Excerpt :

      The necessity for fusing together information obtained via different means was already pointed out e.g. by Sheng (NREL report, pages 33 and 134)[63]. Various algorithms that can perform data or decision fusion exist, including exempli gratia Ensembles of Neural Networks [222], applied in SHM mostly to interpretation of ultrasonic measurements [223,224], but some cases related to vibration analysis can also be pointed out [214,152]. In particular, Zhang et al. tested an ensemble of neuro-fuzzy networks fed with different features in a parallel gearbox health diagnosis [152].

    • On the selection of advanced signal processing techniques for guided wave damage identification using a statistical approach

      2014, Engineering Structures
      Citation Excerpt :

      For determining the severity of damages in one-dimensional waveguides with a limited number of sensors, pattern recognition and optimization are two commonly used approaches. Pattern recognition approach, such as supervised learning [27,28], applies prior experience to make sense of new data in the damage identification. Optimization approach [29–32] minimizes the discrepancy between the numerically predicted structural responses and the measured data by altering the damage parameters of a pre-defined model in order to determine the location and severity of the damage in the structure being tested.

    • Non-destructive evaluation (NDE) of aerospace composites: Structural health monitoring of aerospace structures using guided wave ultrasonics

      2013, Non-Destructive Evaluation (NDE) of Polymer Matrix Composites: Techniques and Applications
    View all citing articles on Scopus
    View full text