·Home ·Table of Contents ·Civil Engineering

EIH Model and its application on concrete flaw detection Yang XIANG (College of power and environment engineering, Wuhan Transportation University, Wuhan 430063) S. K. TSO (Department of manufacturing engineering and engineering management, City university of Hong Kong) Xi zhi SHI (State key Lab of vibration, shock & noise, Shanghai Jiaotong university, Shanghai 200030) Contact

Abstract

Concrete is a typical heterogeneous material. In this paper, a difficult problem, i.e. automatic flaw type detection of concrete structure is studied. EIH model for feature extraction of impact-echo signal is set up. A back-propagation neural network classifier is used for classification of extracted features. The results showed that this feature extraction method can resolves flaw type detection of concrete structure well.

1 Introduction

Nondestructive test of heterogeneous materials is an important and difficult work. Concrete is a typical heterogeneous material. Its flaw detection is difficult, especially for detection of small or shallow flaws in concrete structure. Impact-echo test is a simple, suitable for field operation testing method (see Fig.1). Waveform of impact-echo signals in time domain contains the entire information reflected from flaws and the boundaries of the concrete structure. How to extract flaw information for pattern recognition in the waveform is an important task for automatic interpretation. In earlier researches, people determine whether there are flaws or not from waveforms in time domain [1~5]. The difficulty of this method is the overlap of multiple reflected waves. It also can not determine the type, shape and size. Later, people use spectrum analysis to determine the existence of flaw, the depth of the flaw. But it also can not determine the type, shape and size [6,7]. In these researches, only part of the flaw information is used. In addition, during an inspection of bridge deck or highways, signals of hundreds of impact-echo test points will be recorded. Thus an effectively automatic interpretation methods is needed [8,9]. According to the research situation mentioned above, a feature extraction method by using EIH model and a classification method by using a back-propagation neural network are studied in this paper.

Keywords: nondestructive test; concrete; feature extraction; neural networks-based classifier; impact-echo method

Fig 1: Diagram of impact-echo testing system

2 EIH model

Impact-echo signal is actual a kind of "sound". Auditory system of human beings is a kind of perfect sound identification system. So many researchers devoted themselves to the study of auditory model. The potential advantages that can be gained by utilizing auditory models depend on how accurate the models are in mimicking human performance. And building such accurate models depend on the amount of knowledge we have about the auditory system. Based on the achievements of psychophysical and physiological studies of the auditory system, O. Ghitza designed an auditory model [10~12]. The model comprises two stages, one that models the pre-auditory-nerve section of the auditory periphery. The pre-auditory-nerve section has been modeled in considerable detail, guided by the physiological structure of the auditory periphery. In contrast, the post auditory-nerve section is represented in a heuristic manner since little is known, so far, about the operation of auditory functions associated with that part of the auditory periphery. The model consists of 190 cochlear channels, distributed from 200 HZ to 7000 HZ according to the frequency-position relation suggested by [11]. Each channel comprises Goldstein's nonlinear model of human cochlear [12], followed by an array of five level-crossing detectors that simulate the auditory nerve fibers innervating on inner hair cell, at this junction of the model, the sound signal is represented as a 950-dimensional point process that simulates the firing activity of the auditory nerve fibers. As there isn't enough knowledge about the auditory nerve firing activity at present. Therefore, he processes the simulated auditory nerve firing patterns according to principles that are motivated by observed properties of actual auditory nerve response. A representation of this kind is described as an ensemble interval histogram (EIH). In this paper, only the latter stage i.e. EIH model is discussed because the identification of impact-echo signal is easier than sound identification. The EIH model is schematically illustrated in Fig.2.

Fig 2: Diagram of EIH model (a) EIH model    (b) Level-crossing detector

In the EIH model, the ensemble of nerve fibers innervating a single inner hair cell (IHC) is simulated with an array of level-crossing detectors. Each level-crossing detector is equivalent to a fiber of specific threshold. A neural firing is simulated as the positive-going level crossing. The thresholds are distributed across a range of positive levels to account for the half-wave rectification nature of the IHC receptor potential. The value assigned to level j of every filter is a random Gaussian variable, with a mean, L_j, and a standard deviation. The mean values, are uniformly distributed on a log scale over the amplitude range of the multiband pass nonlinear filter (MBPNL) output. The randomness in the values of the jth level across the cochlear channels simulates the fact that diameters and synapse-connection sizes of fibers innervating the same side of different IHC's along the cochlear partition have a certain amount of intrinsic variability. If the magnitude of the input signal is low, only one level will be crossed. However, for large signal magnitudes, several levels will be activated. In figure 2, three levels i.e. L₁, L₂, L₃ are crossed by input signal X(t). There are two positive half-wave in figure 2, thus the total number of level crossing is six.

3 Wavelength distribution of different flaws

For using the EIH model, experiments are carried out on three concrete slabs which contain man-made flaws. One slab is solid, another slab contains a void of f50mm, the other slab contains delamination at mid depth. The slabs are 1 meter long and 1 meter wide and 0.2 meter high (see table 1). Signals are collected by impact-echo testing system. As shown in Fig.1, the experiment system consists of one broadband accelerate transducer, one charge amplifier, a

Specimen	Sizes of slabs (m3)	Types of flaws	Sizes of flaws (mm)
1	1×1×0.2	None	__
2	1×1×0.2	Void	F50
3	1×1×0.2	Delamination	200×200
Table 1: Specification of specimens

Fig 3: Time series of three types of signal (a) solid slab    (b) slab with void    (c) slab with delamination

digital oscilloscope and a computer. The piezoelectric accelerate transducer B&K 8309 (the frequency band is 0~50 kHz ), the charge amplifier YE5858, the digital oscilloscope HP infinium 54815A are chosen. Hardened bearing balls are used as impact sources. Such a source produces a well-defined and mathematically simple input which in turn generates waves with characteristics that facilitate signal interpretation. The diameters of the bearing ball are 12.3 and 14.288mm. In order to reduce the influence of exciting force on amplitude of signals, the normalization and centralization of signals are carried out before feature extraction. The arriving point of R wave is set as original point. The length of each signal is 800ms. After denoising, the signal is used for feature extraction.

In this paper, the normalized and centralized signal is expressed as X(n). The zero-crossing wavelength is defined as the lasting time between two successive changes of amplitude from negative to positive. The length of time is counting by the number of sampling interval. Then four levels which are increasing progressively on a log scale over the amplitude range of [0,1] are set. In this paper, the mean values are assigned as 0.045, 0.122, 0.331, 0.9. As similar as the zero-crossing wavelength, the level-crossing wavelength is defined as the lasting time between the two successive changes of amplitude from lower than this level to over this level. An estimate of the interval probability density function of a given level can be obtained by computing a histogram of the intervals from the point process data produced by the level-crossing detector. Only intervals between successive upward-going level crossing are considered. Function S(a_i) is used to express the total number of zero-crossing and level-crossing relate to wavelength a_i thus the probability density function of wavelength is:

Here, the lower limit of summation is selected as 50 because the upper limit of frequency of the signal is 50 kHz. The upper limit of the summation is selected as 500 because the lower limit of the frequency of the analyzing signal is 5 kHz.

Fig 4: Wavelength distribution curves of different flaws (a) solid slab (b) slab with void (c) slab with delamination

In this paper, by counting the total number of every level-crossing detector for a specific wavelength, the probability density function of the wavelength can be determined. The probability density functions of wavelength for different flaws are illustrated in fig. 4. Fig. 4 (a), (b), (c) are the wavelength distribution curve of different types of flaws. It can be seen that the concentrated regions are different for different flaws.

4 Feature extraction and classification

A feature extraction process can efficiently provide class separation with a small number of features. After the process, each impact-echo signal can be expressed as a feature vector. In fact, non-destructive flaw detection of concrete structure is a kind of pattern recognition. The concept of pattern recognition may be expressed in terms of the partition of feature space (or mapping from feature space to decision space). Suppose that N features are to be measured from each input pattern. Each set of N features can be considered as a vector X, called a feature (measurement) vector, or a point in the N-dimensional feature space W.

The problem of classification is to assign each possible vector or point in the feature space to a proper pattern class. This can be interpreted as a partition of the feature space into mutually exclusive regions and each region will correspond to a particular pattern class.

According to the wavelength distribution curves, we divided the wavelength into 45 region from 50 to 500, count up the number of wave in each region, a feature with 45 dimensions have been got. Then it can be used for classification of different flaws.

In classification, the training set contains 55 samples collected from solid slab, 70 samples collected from slab with delamination, 70 samples collected from slab with void. The total number of training sample is 195. The test set contains 89 samples collected from solid slab, 118 samples collected from slab with delamination, 118 samples collected from slab with void. The total number of testing set is 325.

For the influence of environment noise and the characteristic of impact-echo signal, the flaw identification of concrete structure is a very complex pattern recognition problem. The mapping from feature space to decision space is a nonlinear process. It is suitable for using neural networks to carry out classification. In this paper, a back-propagation neural network classifier was used.

This classifier is a three-layer network, consisting of an input layer, a hidden layer and an output layer. The number of neurons in the hidden layer is 25. The third layer is out put layer, its output value is determined by the value of the class it represented. The weights were initialized randomly in the range (-0.1,0.1). The activation functions of the neuron in the hidden and output layer are the sigmoid function, whereas the input units are linear, A conjugate gradient method is used for fast convergence of the supervised learning algorithm. The input pattern is normalized prior to presentation to the neural network and this normalized input pattern is fed to the neural network. By using this back-propagation neural network, the results of classification as shown in table 2.

True conditions		Classified conditions			Average detection rate
		Solid	Void	Delamination
Training set	Solid	96.36%	0.00%	3.64%	94.36%
	Void	7.14%	92.86%	0.00%
	Delamination	1.43%	4.29%	94.29%
Testing set	Solid	91.01%	2.25%	6.74%	90.15%
	Void	2.54%	88.98%	8.47%
	Delamination	2.54%	6.78%	90.68%
Table 2: Classification results of different flaws

5 Conclusion

An EIH model based on the achievements of psychophysical and physiological studies of auditory system is used in this paper. Feature extraction of impact-echo signals is carried out by using this EIH model. A back-propagation neural network is used for classification of these features. The results of classification test demonstrated that the feature extraction method and the neural network are suitable for flaw type detection and the automatic interpretation of impact-echo signal.

6 Acknowledgements

This work was supported by the project of the Ministry of transportation of China (95-06-02-04), the science fund of Hubei province (99j079), the project of city university (9360030) and NSFC-69772001.

References

1. Carino N. J., Sansalone M., Pulse-echo method for flaw detection in concrete, Technical Note 1199, National Bureau of Standards, July, 1984.
2. Sansalone M., Carino N. J., Detecting honeycombing, the depth of surface-opening cracks, and ungrouted ducts using the impact-echo method, ACI Concrete International, April, 1988, pp38-46.
3. Carino N. J., Sansalone M., Hsu N. N., Flaw detection in concrete by frequency spectrum analysis of impact-echo waveforms, International Advances in Nondestructive Testing, 12th Edition, Gordon & Breach Science Publishers, New York, 1986, pp117-146.
4. Lin Y., Sansalone M., Detecting flaws in concrete beams and columns using the impact-echo method, Materials Journal of the American Concrete Institute, July-August 1992, pp394- 405.
5. Yasunori Sakata and Masayasu Ohtsu, Crack evaluation in concrete members based on ultrasonic spectroscopy, ACI Materials Journal, V.92, No.6, November-December 1995, pp686-699.
6. Jiunn-Ming Lin and Mary Sansalone, Impact-echo studies of interfacial bond quality in concrete: Part I -Effects of unbonded fraction of area, ACI Materials Journal, V.93, No.3, May-June 1996, pp223-232.
7. Jiunn-Ming Lin, Mary Sansalone, and Randall Poston, Impact-echo studies of interfacial bond quality in concrete: Part II -Effects of bond tensile strength, ACI Materials Journal, V. 93, No.4 July-August 1996, pp318-326.
8. Pratt D., Sansalone M., Impact-echo signal interpretation using artificial intelligence, ACI Materials Journal, 89(2), pp.178-187
9. Valenzuela M., Sansalone M., Krumhansl C., Streett W. B., Use of sound for the interpretation of impact-echo signals, In Proceedings, 4^th conference on Auditory display, Palo alto, CA, 1997, pp.47-56
10. Oded Ghitza, Auditory models and human performance in tasks related to speech coding and speech recognition, IEEE Transactions on speech and audio processing, January 1994, Vol.2, No.1, Part II, pp.115-132.
11. D. Greenwood, A cochlear frequency-position function for several species - 29 years later, J. Acoust. Soc. America, Vol. 87, no. 6, 1990, pp2592-2605
12. J. L. Goldstein, Modeling rapid waveform compression on the basiar membrane as a multiple - bandpass - nonlinearity filtering, Hearing Res., Vol.49, 1990, pp39 -60

© AIPnD , created by NDT.net

|Home|    |Top|