Many processing techniques have been developed to clearly reveal and provide quantitative information concerning defects in infrared thermography. One such approach employs artificial neural networks, which have recently obtained considerable success in a variety of applications. In this research, classical transient heat conduction modeling and artificial neural network approaches will be used to detect defects, to characterize their depth and size and to estimate the thermal contrast signal.

Theoretically, solving the transient heat conduction equation with the appropriate boundary conditions allows the description of subsurface defects and anomalies of materials and complex structures. The solution of direct problems in infrared thermography, whereby the temperature distribution is computed for known defect geometry and properties, provide useful information to interpret experimental data, to optimize the heat perturbing and scanning configuration, to determine the limits of applicability of the thermographic approach and, especially in this study, to train desired neural networks [1].

Except for certain simple geometries, it is extremely difficult to obtain an analytical solution of the problem. Numerical computation methods such as the finite difference method and the finite elements method are generally suited for the majority of practical problems [2]. In this study, a three-dimensional heat conduction problem with appropriate initial and boundary conditions applied in thermal non-destructive evaluation has been solved using the finite difference method. This simulation is done on a MasPar computer which contains 2K processors, using the MPL programming language (a parallel version of the C programing language). Also, the results obtained by the finite difference method for a sample without defects is compared with its one-dimensional analytical solution.

Neural networks are known for their high processing speed, high classification accuracy, low sensitivity to noise, and easy thresholding capability to yield a binary image used, for instance, in automated detection [3]. In this study, various interconnected feedforward multilayered neural networks have been designed and evaluated in order to reveal and estimate the parameters of defect in the modeled sample. By using the entire temperature versus time curve or thermal contrast curve as the input to the neural networks and through suitable training, the networks have been made to respond in the desired way to the various features of the thermal curves.

Figure 1. Two possible architectures for flaw detector
and estimator network in infrared thermography.

For example, a two-step architecture, as shown in Figure 1, will be defined in a first step to classify the thermal curves corresponding to each pixel of a sequence of infrared images as a flow detector network, and in a second step to proceed to the spatial analysis as estimation network. The training will be done either with experimental (real or simulated defects), or with theoretical data. Theoretical data seems to be encouraging for two major reasons: first a large training set is required, which is difficult to obtain with real experiments, and second, simulated cases are not representative enough.

Comparisons will be made between the neural network results and both the model and the actual experimental results in terms of specimen parameters, and output signal (thermal contrast) estimation.

This method is believed to offer advantages in terms of speed and efficiency as compared to current inverse techniques (analytical or based on empirical relationships).

References

1. Xavier Maldague, "Nondestructive Evaluation of Materials by Infrared Thermography", Springerverlag London Limited, 1993.
2. V. Vavilov, X. Maldague, B. Dufort, Robitaille and J. Picard "Thermal nondestructive testing of carbon epoxy composites: detailed analysis and data processing" NDT & E International Volume 26 Number 2 1993.
   
3. D. R. Prabhu and W. P. Winfree, "Neural Network Based Processing of Thermal NDE Data for Corrosion Detection", QNDE, vol. 12, 1993.