The Use of a Neural Network in Nondestructive Testing by Donald G. Pratt, Mary Sansalone and Jeannette Lawrence April 25, 1990 Nondestructive testing (NDT) methods are techniques used to obtain information about the properties or the internal condition of an object without damaging the object. Thus NDT methods are extremely valuable in assessing the condition of structures, such as bridges, buildings, and highways. Because of the current emphasis on rehabilitation and renovation of structures, there is a critical need for the development of NDT methods that can be used to evaluate the condition of structures so that effective repair procedures can be undertaken. Typically, NDT methods are used to obtain information about a structure in an indirect way. For example, by measuring the speed of stress (sound) waves as they travel through an object and studying how the waves are reflected within the object, one can determine whether or not flaws exist within the object. Of particular interest to structural engineers is the development of NDT techniques for evaluating reinforced concrete structures. Currently, the practical techniques that can detect cracks in concrete use acoustic impact, infrared thermography, and ground penetrating radar. However, none of these methods possesses all the desired qualities of a crack detection system [1,2], which are reliability under various site conditions, capability for rapid testing of large areas, and ease of use. Recently, a new nondestructive testing technique has been developed for finding cracks in concrete structures. This method was developed at the National Institute of Standards and Technology (NIST, formerly National Bureau of Standards) by Carino and Sansalone and is called Impact-Echo [3]. Ongoing research programs at both NIST and Cornell University are aimed at developing the theoretical basis and practical applications for this new technique. One project carried out at Cornell University has developed an automated impact-echo test system in the lab which will be adapted for field use. Key aspects of this project are the development of hardware and software for a field system. The goal is to develop a field test system that is reliable, rapid, and relatively simple to use. OVERVIEW This article presents a new method for automating and simplifying impact-echo signal analysis and data presentation with an artificial intelligence technique that uses a brain-like neural network. We begin with a brief introduction to the impact-echo method. Next, the application of the neural network to the analysis of impact-echo data obtained from concrete plates containing voids is discussed. Two neural network design approaches are reviewed and a discussion of neural network effectiveness is included in the final section. THE IMPACT-ECHO METHOD In impact-echo testing, a stress pulse is introduced into the concrete by mechanical impact. Hardened steel spheres are used to strike the surface, which produces an impact duration of 20 to 80 microseconds, depending on the diameter of the sphere. Such an impact generates a pulse made up of lower frequency waves (generally less than about 50 kHz) that can penetrate into a heterogeneous material such as concrete. The pulse propagates into the concrete and is reflected by cracks and voids and the boundaries of the structure. A transducer that measures displacements at the surface caused by the reflected waves is placed next to the impact point. The recorded surface displacement waveforms can be analyzed to find the depth to a reflecting surface, such as the bottom surface of the plate or an internal crack. For example, in a solid plate the pulse generated by the impact is multiply reflected between the top and bottom surfaces of the plate setting up a transient resonance condition. Each time the pulse arrives at the top surface it produces a characteristic downward displacement. Thus the waveform is periodic. The round-trip travel path for the pulse is approximately equal to twice the thickness of the plate (2T), and the period is equal to the travel path divided by the wavespeed (C). Since frequency is the inverse of the period, the dominant frequency, f, in the displacement waveform is: f = C / 2T (1) The frequency content of a digitally recorded waveform is obtained using the fast Fourier transform (FFT) technique [3,4]. In the amplitude spectrum obtained from the FFT of the waveform] there is a single large amplitude peak at the frequency corresponding to multiple reflections of the pulse between the top and bottom plate surfaces. The frequency value of this peak, which is called the thickness frequency, and the wavespeed in the plate can be used to calculate the thickness of the plate (or the depth of an internal crack if reflections occur from such an internal defect) using Equation (1) rewritten in the following form: T = C / 2f (2) For a wavespeed of 3450 m/s and a peak frequency value of 3.42 kHz, the calculated thickness of the plate is 0.5 m, which agrees with the actual plate thickness is 0.5 m.1 For a given concrete specimen, wavespeed is essentially constant and so Equation (2) relates the frequency of a point on the amplitude spectrum to the depth of a reflecting surface within the specimen. This relationship can be used to convert the horizontal axis of the amplitude spectrum from frequency to depth. In addition, the spectra can be made non-dimensional for a structure of constant thickness if the horizontal axis is expressed as a percentage of the thickness. The resulting graph is called the reflection spectrum. In one example a frequency peak at 3.42 kHz appears as a peak at a depth of 100%, indicating reflection from the bottom of the plate. In another example, a reflection spectrum obtained from an impact-echo test on a 0.4 m thick plate containing a 0.4 m diameter void located 0.3 m below the top surface of the plate. Reflection from the void produces a dominant peak at about 75% of the plate thickness. In the impact-echo method, tests are carried out at selected points on the structure, the location of which depends on the geometry of the structure and the type and size of flaw one is trying to locate. In a typical filed application, tests would be carried out at many individual points. Automating the interpretation of reflection spectra is necessary for a rapid and easy to use field test system. We used an artificial neural network as a way of training the computer to recognize the key features of reflection spectra. INTERPRETING IMPACT-ECHO DATA A commercial neural network simulation package called BrainMaker, produced by California Scientific Software, was chosen to interpret the results of impact-echo tests. This product allows the user to adjust the various network parameters, such as the number of neurons in each layer, the format of the inputs and outputs, the neuron transfer function, etc. The software has a proprietary back propagation algorithm that uses integer math and runs at 500,000 connections per second. Creating and training a network is done in a graphical interface, with pull-down menus and dialog boxes for use with the keypad or a mouse. The program is very easy to use and comes with extensive documentation that provides an excellent introduction to neural networks, both in theory and application. Reflection spectra are the inputs to the neural network. In the first design approach, two outputs were used which represented 1) the probability of a flaw and 2) the depth of the flaw. This design proved too difficult; an analysis is presented in the next section. The final network design used 11 output neurons: one is the probability that a flaw exists and ten others are for the approximate depth of the flaw. The ten depth outputs give the flaw depth within each 10% increment of the structure's thickness. The absence of a flaw shows up on a reflection spectrum as a single peak at 100% of the structure thickness, and so a flaw probability of 0% is associated with a flaw depth of 100%. A reflection spectrum and the corresponding network output for a solid 0.4 m thick slab shows a low flaw probability and a high probability at 100% of the slab's thickness. A reflection spectrum and neural network output obtained from a test on a 0.4 m thick slab containing a 0.2 m void at a depth of 0.2 m shows a high flaw probability coupled with a high probability at 50%, indicating a flaw between 40% and 50% of the thickness of the slab. Thus the network is capable of detecting the presence of a flaw and resolving the flaw depth to within 10% of the thickness of the structure. In order for the network to learn to interpret reflection spectra correctly, the training set must include a wide range of flaw conditions. Each member of the training set includes the reflection spectrum obtained at a particular test point and the target output for this point. The target output is the flaw probability and the depth of the flaw, both of which must be accurately known. Some of this data is acquired from impact-echo tests on laboratory specimens containing simulated voids. However, it is impractical to construct laboratory specimens for every case one would like to use in training a network. So, the results obtained from numerical simulations of impact-echo tests on structures containing voids [5] are also used. Numerical simulations provide a fast and inexpensive way to generate a variety of data for the training set, compared with using laboratory specimens. The network used in the examples described above was trained with data from laboratory specimens and numerical simulations. The system used to do impact-echo testing in the laboratory includes data acquisition hardware with 12-bit resolution installed in a portable 80386-based computer operating at 25Mhz. The displacement transducer uses a small conical piezoelectric element attached to a large brass backing. This transducer has a broadband output that provides a very faithful response to displacement. The sensitivity is on the order of 2 X 10^8 volts per meter. Stress pulses are introduced into the structure using mechanical impact, either by dropping hardened steel spheres or using a spring-loaded impactor. The sampling and triggering parameters for the data acquisition card are under software control, and are set so that the data is taken automatically when an impact is produced. All the signal analysis is done in software, including the FFT amplitude spectrum computation and the neural network simulation. These two algorithms account for the majority of the processing time. A supervisory program is being developed with the capacity to gather test data for training new networks, run tests using previously trained networks, and display the reflection spectrum and network output. At the present stage of development, a single test takes about two seconds from the time the impact is produced to the point at which the output is displayed on the screen. THE NEURAL NETWORK DESIGN This application was designed using the BrainMaker simulator from California Scientific Software. The training algorithm is the back propagation algorithm and the sigmoid transfer function is selected. The learning rate, which controls the amount adjustment to the weights, is set to a nominal value of 1 (0 prevents training; 4 is the absolute maximum). The training tolerance, which specifies how close the output must be to the training pattern to be considered correct, is set to 0.1 (90% accuracy within the possible output range). Three layers are used. The first layer is the input layer which reads in the data to be analyzed. The second or "hidden" layer processes the information from the first layer and sends it to the third, or output layer, which produces the result. In order to use a back propagation network, a training file is needed which consists of sets of input and output pairs. Each pair of input data and known output results is called a fact. This application's training file consists of 59 facts. Each fact has 150 inputs and 11 outputs, hence there are 150 input neurons and 11 output neurons. Each input neuron is assigned a vertical slice of the reflection spectrum. The value presented to each input neuron represents the amplitude at a particular frequency range which is 1/150 of the waveform's total frequency range. One of the 11 outputs correspond to the probability or certainty of a flaw, and 10 others the range of flaw depth. For training the appropriate flaw depth is set to 1 with all the others set to 0. The appropriate flaw depth is the known state of the test specimen. To train the network, the program presents the facts one at time and computes the actual network output for that fact. The actual output is compared to the known result and the difference is used to make adjustments to the network connections. Facts for which the network's output is not within the training tolerance are considered bad, and statistics are displayed as such on the screen. The inputs, outputs, and hiddens can be displayed as numbers, symbols, pictures or thermometers. While training, the network is shown all of the facts, over and over until it learns everything to the performance level specified. The first design used only two output neurons: one for the probability of a flaw and the other represented the depth of the flaw directly by its numeric output value. Although this network trained quickly (86 runs in 15 minutes on a 25 MHz 386), it did not test well. It was observed that the output was sensitive to the amplitude of the inputs rather than the features. It did not pass the test on laboratory samples within the required accuracy. Upon consideration, it was thought that the network was experiencing difficulty in the way a person might. Imagine trying to judge the exact length of lines on a wall from quite a distance away with nothing to compare them to. This is a difficult task. But if asked what the relative length of two lines is (e.g., Is the first line half the length of the second?), it becomes an easy task. This concept sparked an idea for a new design. The new design allowed the neural network to answer "yes" or "no" to questions like "Is there a flaw at a depth of 10 - 20%?", rather than ask it to come up with a precise number. The second design used 11 output neurons instead of 2. By adding more output neurons which represent the flaw depth in increments, it is easier for the network to train. With multiple outputs (each of which represents the probability of a flaw existing within a particular range of the total depth), the network picks one of many instead of using one neuron to indicate the depth directly. Distributing the output has also been found by California Scientific Software to be a good design technique. This scheme also permits the detection situations where the network is unable to make an accurate classification after it's trained. In some cases, the output conditions may not make sense. For example, when the network says that the flaw depth may be at 10% AND it may be at 50% (which is indicated by both neurons being partially turned on), it means the network is having trouble interpreting the input. If the first network were to encounter such an ambiguous case, the single output would indicate some depth and it would be hard to interpret the difficulty it was having. Still, after increasing the number of output neurons, the network had difficulty passing the test on laboratory samples. After training, histogram diagrams were examined. The histogram shows that the neuron connections are tending to bunch up toward the negative end of the weight values. This is often a bad sign that the network is making major changes to the weights without being effective (the number correct is only 47 out of 54 at this point). Sometimes a network eventually trains and tests out well when this happens, but this one did not. It was found that 10 hidden layer neurons was too few. The problem was alleviated by increasing the number of hidden neurons to 20. It had taken 169 iterations to train but now with 20 hidden neurons the new network trained in 72 iterations, and it got all of the testing facts correct. ADVANTAGES OF THE NEURAL NETWORK The ability of the neural network to learn the key features of input patterns makes it a useful tool for interpreting impact-echo reflection spectra. The relative ease with which a network can be defined, trained, and used makes the technique attractive for developmental work where the system is likely to undergo many revisions before a final system is produced. Once the design change to 11 outputs was conceived, implementation was accomplished in a few hours. The network output is a set of probabilities that provides a simple way to measure the certainty of the result. For example, if the flaw probability is 55%, the network is suggesting uncertainty in the data, compared with an output of 98%, which shows close correlation with members of the training set. The neural network provides an automated method of determining flaws in concrete without destroying the structure. Testing of the neural network revealed a success rate of about 90% with laboratory concrete samples. Success is difficult to precisely determine for several reasons. One difficulty occurs when the sensor is placed near the edge of a flaw. The network output may be vague or confusing. The edge of a flaw can cause reflections from many levels in the concrete. In this case, the network output could be taken in the context of the results of tests of nearby areas to determine that it was in fact an edge which caused the confusing output. This decision could be automated by another neural network which looked at the results of several tested proximal areas at once. Other approaches for finding flaws range from the drilling of core samples to the use of radar. The first method is destructive, time-consuming and only permits checking a small percentage of the area. The second require expensive equipment and isn't effective when there's steel reinforcement. These approaches experience the same problem when the sensor is not placed directly over the flaw. They also have other problems of not being capable of rapidly testing large areas, reliable under various site conditions or easy to use. A neural network is better because it uses a non-destructive technique, the system can be built from off-the-shelf parts, its speed enables quicker interpretation of results, its flexibility lends it to use as a developmental tool, and the results will be consistent. CONCLUSION A new method for automatic interpretation of nondestructive test data has been presented. The use of an artificial neural network provided a quick and accurate means of interpreting the results of impact-echo tests obtained from concrete structures. On-going work is focusing on developing a rugged field test instrument based on the impact-echo laboratory test system. When this objective is realized, a tool will be available for rapid and reliable detection of cracks in concrete structures. To date, the impact-echo testing technique has been used in trail field studies for detecting voids in a concrete ice-skating rink [6] and in reinforced concrete slabs [7]. Once a rapid field instrument is developed, the method can be used routinely for nondestructive testing of plate-like structures such as slabs, pavements and walls. For these applications, it is expected that a neural network will be used to automate signal processing. A Canadian mining company is currently negotiating with Cornell University for a system that will help them determine if the structure of a decommissioned mine is safe enough to recommission the mine. Acknowledgements: Research sponsored by grants from the Strategic Highway Research Program, Project C-204 and from the National Science Foundation (PYI Award). BrainMaker neural network simulation software ($195) was provided by California Scientific Software, 10141 Evening Star Drive #6, Grass Valley, CA 95945-9051. (916) 477-7481. -------------------- Footnotes: 1. The frequency resolution in the amplitude spectrum and thus the accuracy of plate thickness or crack depth predictions will depend on the sampling rate and duration of the recorded waveform. References: 1. Manning, D.G. and Holt, F.B., "Detecting Deterioration in Asphalt-Covered Bridge Decks," Transportation Research Record 899, 1983, pp. 10-20. 2. Knorr, R.E., Buba, J.M., and Kogut, G.P., "Bridge Rehabilitation Programming by Using Infrared Techniques," Transportation Research Record 899, 1983, pp. 32-34. 3. Sansalone, M. and Carino, N.J., "Impact-Echo: A Method for Flaw Detection in Concrete Using Transient Stress Waves," NBSIR 86-3452, NTIS PB #87-104444/AS, Springfield, Virginia, September, 1986, 222 pp. 4. Carino, N.J., Sansalone, M., and Hsu, N.N., "Flaw Detection in Concrete by Frequency Analysis of Impact-Echo Waveforms," in International Advances in Nondestructive Testing, Vol. 12, ed. W. McGonnagle, Gordon and Breach Science Publishers, 1986, pp. 117-146. 5. Sansalone, M., and Carino, N.J., "Transient Impact Response of Plates Containing Flaws," in Journal of Research of the National Bureau of Standards, Vol. 92, No. 6, Nov-Dec 1987, pp. 369-381. 6. Sansalone, M., and Carino, N.J., "Laboratory and Field Studies of the Impact-Echo Method for Flaw Detection in Concrete," Nondestructive Testing of Concrete, SP-112, American Concrete Institute, Detroit, 1988, pp. 1-20. 7. Sansalone, M. and Carino, N.J., "Detecting Delaminations in Concrete Slabs with and without Overlays Using the Impact-Echo Method," ACI Materials Journal, V. 85, No. 2, Mar.-Apr. 1989, pp. 175-184. 8. Stanley, J., "Introduction to Neural Networks," (c) California Scientific Software, Sierra Madre, California, January, 1989 About the authors: Donald G. Pratt is a doctoral student in Civil Engineering at Cornell University. Mary Sansalone received a Ph.D. in structural engineering from Cornell University, where she is an assistant professor. Prior to joining the faculty at Cornell, she was a research engineer with the National Institute of Standards and Technology. Mr. Pratt and Dr. Sansalone may be reached at Cornell University, Hollister Hall, Ithaca, NY 14853. Jeannette (Stanley) Lawrence is a technical writer specializing on the subject of neural networks. She may be reached at California Scientific Software, Grass Valley, CA.