- International Symposium on NDT Contribution to the Infrastructure Safety Systems, 1999 NOV 22-26 Torres, published by UFSM, Santa Maria, RS, Brazil

TABLE OF CONTENTS

Abstract The ultrasound inspection method Materials and methods Mathematical procedures - neural network Experimental results Concluding remarks

Abstract

Non-destructive Testing (NDT) is becoming more and more important in the field of structural integrity of equipment and industrial units. The applied physical principles are, fundamentally, ultrasound, magnetism, X-rays, etc. However , the efficiency of the diagnosis depends fundamentally on the experience of the engineer who carries out the tests.
Among all methods commonly used in NDT, the ultrasound is the most widely known and particularly useful, besides X-rays, for evaluation of metallic materials welding. The application of X-rays in industrial units is more limited for involving the need of heavier equipment, besides the problems of human exposition to the rays.
The main objective of this work was to develop a software that makes the diagnosis less dependent on the experience of the professional who carries out the evaluation and, at the same time, less susceptible to errors in applications of metallic welding evaluation.
With this purpose were used several mathematical procedures and methods and the more significant results were obtained using Neural Networks techniques.
Extensive laboratory experiments were carried out in order to inspect a great number of specimen and collect samples for verifying robustness of the proposed approach.
The results of the work prove the efficiency of the method used to identify the presence of faults. They indicate also the possibility of classifying metallic welding into classes.
The mathematical tools developed in this work for welding analysis in metallic plates, automating the diagnosis can be extended to other welding geometry and involving different metallic alloys.

Keywords: welding, non-destructive testing, ultrasound and neural networks

The ultrasound inspection method

Commonly, the inspection based on ultrasound testing detects discontinuities by monitoring the reflection of the sound transmitted through the material by a gauge accomplished to the piece we want to be inspected. They have also a monitor to visualize the intensity of the reflected energy and to estimate where the interfaces are in the piece. By analyzing these signals, the inspector tries to determine the existence or not of discontinuities inside the material.
Compared to other non-destructive testing techniques, the ultrasound method has some advantages, as described bellow:
• large penetration, which allows the detection of deep discontinuities;
• high sensibility, allowing the detection of 0.5 mm;
• high precision for internal discontinuities;
• it is only necessary to have access to one of the surfaces;
• existing equipment provides immediate indication of the signals;
• the necessary equipment is portable, making the method perfect for field testing;
• it does not have effects on people around the tests.

However, it also presents some disadvantages:
• inspection carried out manually requires great attention;
• signal evaluation demands qualification and experience from the inspector;
• irregularly shaped pieces, with very rough, very thick, small, or non homogeneous surfaces are difficult to inspect and may lead to false diagnosis due to spurious reflections;
• discontinuities very close to the external surface may not be detected, as they can be mistaken for the initial peak associated to the equipment emission;
• equipment calibration requires pattern pieces for reference.
The purpose of this work is to minimize the two first disadvantages mentioned above, automating the diagnosis By developing a procedure that provides automatic diagnosis for the detection of metallic materials welding faults, combining ultrasound techniques and neural networks.

Materials and methods

Ten welded carbon steel plates with geometry as shown in Fig. 1 were inspected. Each plate was previously prepared containing samples of the most frequent types of welding faults and a perfect welding. After the preparation of the pieces, all of them were radiographed in order to verify the presence of faults.

Fig 1: Geometry of the plates (dimensions in mm).


The six different faults considered here are shown in Fig. 2. The inspections were carried out by looking for the position on the surface of the piece where the reflected ultrasound echo was maximized on the equipment screen. For each one of the samples, four inspections were performed by different technicians trying to eliminate the influence of slightly different positions on the surface of the piece chosen for capturing the signal. After that, all the radiofrequency signals were captured by the ultrasound equipment and sent, in ASCII format, to a notebook, running a software developed for this job, through a RS-232 port. A schematic representation of the site for the tests is shown in Fig.3.
Sets of files, containing vectors, whose elements are signal amplitudes, were then formed.
We have used an A-scan screen type that is, basically, an amplitude x time plot, where the horizontal axis is the elapsed time and the vertical deflections are the ultrasound echo amplitudes. The utilization of this kind of screen is well suited for detection of discontinuities, thickness measurement, sonic speed and attenuation of the ultrasonic signal.

Fig 2: Types of welding faults considered: (a) lack of fusion at chamfer, (b) porosity, (c) slag inclusion, (d) crack in the root, (e) lack of penetration, (f) lack of fusion in the root

Fig 3: Site for tests.

The amount of energy reflected, as well as the elapsed times since the transmission of the initial pulse and the echoes reception, are measured by the equipment. In our case, metallic material welding, porosity and lack of fusion are faults which has be behavior of metal-gas interfaces and can be more easily detected. Slag inclusions and other discontinuities, which are metal-solid interfaces, may be identified by partial reflection.
The more common way of generating and receiving sonic waves is by using piezoelectric transducers. We used 45⁰, 60⁰ e 70⁰ angular probes. We used transducers with nominal frequencies of 2MHz, 4MHz and 5 MHz. The best results were obtained with the 700 angular probe working at 5 MHz.

Mathematical procedures - neural network
Neural networks have been trained to carry out complex functions in many applications including pattern classification, systems identification, voice and image processing and controls. In this work we used neural networks to identify the presence of faults in welds and classify them. The resulting can be a powerful tool in quality controlling and to identify systematic problems in welding processes in automated industries.
Initially we tried a network with one layer of neurons, using the architecture shown in Fig.4. We developed training and classification routines.

Fig 4: Detail of the training network with one layer.


The signals to be analyzed were grouped in two different ways as described bellow:
• (i) The signals were divided in three classes, being the first one a set of all the possible faults, the second one samples without discontinuities and the third one a set of root reflections. Root reflections are signals frequently found in practice, when we have a perfect weld, which may lead us to erroneous fault diagnostics.
• (ii) The signals were divided in eight classes: the six faults shown in Fig. 3, the signal without discontinuities and the root reflection signal.

In the case (i) we had the welds classified into approved and imperfect ones. In the case (ii) they were additionally classified according to their fault class identification. To train the neural network we tested four times each one of the eight plates which contained themselves samples of the eight patterns, totaling a set of 256 vectors. The signals obtained from the two additional pieces, which also contained the six fault patterns, were used for testing the neurons, making a set of 144 vectors. This network consisted of a single neuron layer with weights being adapted by the Widrow-Hoff training rule. We had a weight for each one of the elements of the input vectors. A bias signal was also used, as indicated on the upper path in Fig. 4, where b is a scalar.
During the process, we tested different quantizer functions, as linear function, hard limiter function and sigmoid functions (logarithmic and hyperbolic tangent). Similar performance in terms of convergence and processing speed was observed in all cases.
As learning rate, we used a factor of 0.99 of the maximum learning rate, which corresponded also to the stability condition for the algorithm. The desired value for the maximum quadratic error was arbitrarily chosen as a sequence of decreasing values till the value of "1".
Random values were used to initialize elements of the weight vector W=[ w₁ , w₂ , . . . , w_n ]. The final weights obtained were then introduced as initial weights for a new execution. After reaching the specified error, we had the parameters for the trained network.
The next step was to apply the non-training vectors through the network and observe the results, which will be shown in the next section. All redundant information was removed for efficiently classifying the vectors. With this purpose, we transformed the signal from time domain to frequency domain via Fast Fourier Transform. Then we identified the peak frequencies of each sample, keeping only a range of points around the dominant frequency and applying these new vectors to the network. We repeated these procedures enlarging the range around the dominant frequency. We arrived at this methodology after a visual analysis of the frequency spectrum where the peak frequencies were quite similar for all the analyzed signals.
After this procedure, we modified the network architecture, by including a hidden layer with four neurons. The first layer had a non linear quantizer of sigmoid type, and the second layer had a linear quantizer, as shown in Fig.5.

Fig 5: Neural network architecture using a hidden layer.

Experimental results

For case (i), carried out in time domain, 256 samples (32 in class 0, 196 in class 1 and 32 in class 2), one layer neural network, we have a percentage of correct classification of 100%. This indicated that is possible to find the existence of faults.
The results for the case (ii), in time domain, 256 samples (32 in each one of the eight classes), one layer neural network was 89,84% of correct classification
Based on the results obtained, we carried out the classification, using the trained network, of the 48 non-training samples. The results are respectively 27,8% and 14,9%.
Two reasons were considered to explain the low percentage of correct classification for the samples that did not participate in training the neural network. The first one was that the number of training samples was not large enough or not representative for adequately training the network. The second one was that when the number of samples was close to the number of elements in the weight vector, the network converged for any vectors we included for training after a certain number of steps.
Then, we transformed via FFT the 256 training samples to frequency domain and took the module for four range of points around the peak frequency.
Finally, we experimented networks with a hidden layer and normalized input vectors, getting better results: 100% (training samples) and 81.25%(non training samples) for the three classes problem.

Concluding remarks

The results presented here lead to some interesting conclusions. In all situations we tested the samples using the signal with and without a DC level. Similar results were obtained in all cases, meaning that the DC component did not give important class information. Analogous observation was made regarding the phase information. In the cases shown above in frequency domain for the three-class (training and non-training samples) and eight-class (training samples), a correct classification was reached.
The results of present work proves the efficiency of the method used to identify the presence of fault. The results indicate also the possibility of classifying metallic welding into classes.
The mathematical tools developed in this work for welding analysis in metallic plates, automating the diagnosis can be extended to other welding geometry and involving different metallic alloys.

© NDT.net - info@ndt.net

|Top|