·Table of Contents
·Reliability and Validation 2
The Artificial Intelligence in Service of Ultrasonic Inspection ReliabilityF. Bettayeb
Scientific Research Center on Welding And Non Destructive Testing,
C.S.C, Route de Dely Brahim, BP: 64, Chèraga, Algiers.
Tel/fax: (213-2) 361850.
H. Benbartaoui , B.Rouaroua
University of Science and Technology USTHB, El Alia BP:32, Algiers.
Usually, during the examination procedure the ultrasonic operator is alone, depending only on his own observation of the signal features, such us echo shape, amplitude level, defect position within the joint geometry, rotation of the transducer around the defect location, etc. These parameters to which, for some of them no figure can be put, are combined by the operator in a non explicit way to lead to the diagnosis. In the aim of automation, this interpretation practice found a natural extension in the exploitation of artificial intelligence computer tools.
Usually automated data interpretation will be done using statistical or artificial intelligent techniques. Among these tools, the artificial neural networks have been verified to be very efficient in the area of pattern recognition, since the current study formalism is dealing with a signal shape recognition combined with heuristics rules. The system drawn here by the artificial neural networks method, is based on an automatic classification and learning algorithm of A-scan data from defects, thanks to an automatic apprenticeship stage.
For accuracy and effectiveness needs, our classification scheme has been accompanied by an experiment ultrasonic inspector with a qualification of a level 2 certified by the IAEA agency.
Neural networks are nowadays predominantly applied in classification tasks. The neural network impose strong requirements on the data and the inspection, however when these are accomplished then good automatic classification systems can be developed. The architecture of a back propagation network is used due to the fact that it is the most robust and common network. Networks used for classification have commonly as much input neurones as there are features and as much output neurones as there are classes to be separated  . In the back propagation algorithm which is based on a gradient descent-method, each neurone of a layer is connected to each neurone in the previous and in the next layer.
In the case of ultrasonic system the inputs are represented by the shape envelope of the A-scan defects and the outputs are 2 classes about planar and volumetric defects.
3.1. Ultrasonic data features for defect recognition.
A detailed knowledge of the interaction between ultrasonic waves and defects, the propagation medium and the conditions in which ultrasonic investigations are curried out, are all elements of basic importance in defect recognition. The strategy is to extract some parameters, enabling the featuring of the pulse echo envelope reflected from a defect, on A-scan images obtained using a single transducer, relative both to the maximum reflection transducer position and to the transducer-defect distance variation. Some of the recognition rules for the most important defects are drawn in the table below.
|Defect Type||Echo form||Echo form after transducer movement from the defect location|
|Mvt // to the joint||Mvt ^ to the joint||Around the joint|
|Porosity||With jagged aspect||The echo disappears rapidly||The echo declines quickly||The echo is nearly constant.|
|PorosityGroup||Set of small echoes with stiff fronts||The global echo persists, with mobility of the sub echoes position.||Same conditions as // movement.||Same conditions as the other movements.|
|Slag Inclusion||has an irregular front||The echo persists but its shape changes||The echo go away rapidly||The echo decreases gradually|
|ImperfectPenetration||The echo is high with stiff front.||The echo behaviour subsists||As for // movement||The echo disappears|
|Lack of fusion||As for lack of penetration||Same conditions.The only||Same conditions differentiation is||Same conditions. about the defect position.|
|Cracks||High echoes superposition, with stiff front||The echo persists||The echo persists||The echo reduces and raises without going out|
|Table 1: "Defect recognition rules from manual testing"|
Neural network classifier
Choosing the architecture of a neural network for a particular problem usually requires some prior knowledge of the problem's complexity. The network topology, however directly affects the 2 most important factors of neural network training. Both theoretical studies and simulations, show that larger than necessary, networks tend to overfill the training data and thus have poor generalisation, while too small a network will have difficulty learning the training samples . Currently, there are no formal methods to customise or select the right network structure. Here the Ann classifier is trained to represent some decision between the 2 classes to be recognised. The system architecture is drawn in the following figure.
|Fig 1: "Network architecture"|
The input layer consists of the envelop shape of the defect signal called the defect map acquired from a numerical oscilloscope on 1024 samples in a vectorial representation. And the output layer is a Boolean representation of planar or volumetric defect decision represented with 2 classes.
5.1. Learning process
The algorithm starts with random initial weights and learns with different maps, respectively. All of the connections in the network are adaptive and are trained with back propagation algorithm. Since the back propagation algorithm requires differentiability along the signal path of the network, we adopt as a transfer function at each unit the familiar 'Sigmoid' function as follows:
So, when x is presented to the network and propagated forward, the state xj in each layer is:
After learning each layer will have different internal representation.
In order to evaluate the weights, they have to be trained from a random set of initial weights, using the training back propagation algorithm, with a minimising process of an objective function. This has be done by the choice of the conventional mean squared error function (MSE), between the actual output xj and the desired one xd.
Minimising an objective function is usually performed by a gradient descent procedure, which requires the computation of the gradient of the objective function with respect to connection weights, and back propagation is just an efficient way to compute this gradient .The modification of weights is performed as the ensuing algorithm.
In this paper we have worked on artificial flaws from steel plates, and on some natural defects from welds, for which the obtained signals are sampled on 1024 points. And for each defect location, we have taken several signals dealing with different transducer positions. The learning is performed on about 90 patterns. Future work concerns the search of an optimal configuration of the global network in which a sub net of each defect family will be included, so that to obtain a more categorical apprenticeship stage.
|© AIPnD , created by NDT.net|||Home| |Top||