شبکه عصبی مصنوعی برای کنترل کیفیت با تست اولتراسونیک در مقاومت در برابر نقطه جوش

An artificial neural network is proposed to solve problems in the interpretation of ultrasonic oscillograms obtained by the pulse echo method. The artificial neural network classifies resistance spot welds in several quality levels through their respective ultrasonic oscillograms. The inputs of the artificial neural network are vectors obtained from each ultrasonic oscillogram with the help of a MATLAB® program. The training of the artificial neural network uses supervised learning mechanism and therefore each input has the respective desired output (target). There are four targets, one for each considered quality level. The available data set is randomly split into a training subset (to update weight values) and a validation subset (to guard against overfitting by means of cross validation). The number of neurons in the hidden layers is selected considering the overfitting phenomenon. This research work has the aim of contributing to the automation of quality control processes in resistance spot welding.

Resistance spot welding (RSW) is extensively used for joining sheet steel in the automotive industry [1] and [2]. The trend to reduce the high number of spot welds per vehicle (3000–4000 [3]) imposes the optimization and fine-tuning of reliable quality control systems [4]. The pulse echo method is an ultrasonic non-destructive testing technique suitable for the quality control in RSW [5] and [6]. This method obtains from each spot weld an ultrasonic oscillogram that allows estimating the quality level of the aforementioned spot weld. Sometimes the ultrasonic oscillogram is hard to interpret by a human expert or the task of interpreting repeatedly oscillograms for a long time gives rise to the drop of the human expert's efficiency [7]. Therefore, the automation of the interpretation of ultrasonic oscillograms would improve the quality control performance in RSW. Artificial neural networks (ANN) are mathematical models that imitate the behaviour of the human nervous system and hence have a parallel, distributed and adaptive processing capable of mapping non-linear and complex systems, in which the regression methods have their limitations [8], [9] and [10]. For this reason, the ANN are extensively used in pattern recognition tasks [11], [12], [13], [14] and [15]. The interpretation on each ultrasonic oscillogram, in order to classify the respective spot weld in a certain quality level, is a pattern recognition problem, so an ANN is proposed to carry out the automation of the interpretation of ultrasonic oscillograms. An ANN, just like a human being, learns by means of training. A supervised learning mechanism is used in the training of the ANN in which a set of input/target pairs is utilized (a target is the desired output respective to a certain input). In the training, the synaptic weights (each link between neurons has a synaptic weight attached to it) are repeatedly adjusted to reduce the error between the experimental outputs and the respective targets until a certain value of error is achieved [16].

The object of this work is achieved because the proposed neural model has shown its effectiveness as a tool to carry out the automation of the classification of resistance spot welds in quality levels through their respective oscillograms obtained by ultrasonic testing: •The proposed ANN produces good results in the classification of input vectors non-used in the training. The ANN proves its adaptability, its robustness and therefore its ability to generalize. •The selection of the significant characteristics of the phenomenon in question in order to make the inputs of the ANN is very important. The way of representing the ultrasonic oscillograms by means of 10-component vectors, that contain the relative heights of the echoes and the distance between consecutive echoes, is appropriate.