Evaluating the quality of ultrasonic signals using machine learning : A comparative study on binary classification methods applied on ultrasonic chirp signals

Sammanfattning: Advanced bolt-tightening tools such as automatic nutrunners are commonly used in industrial factories to streamline the assembly process. In safety-critical systems, it is crucialto control the tightening process. Ultrasonic clamp force control has been widely investigated in academia, showing higher accuracy than many other control methods, such astorque control. However, the precision of the method highly depends on the choice ofexcitation signal. To mitigate the effect of external variables on the method’s precision,a new sending signal may be derived for each new tightening case. Related work suggeststhe use of cross-correlation methods to find the optimal sending signal. However, suchcalculations are often time-consuming and need to be done before each new tightening.This study proposes an alternative method for finding the optimal sending signal. Theproposed method uses machine learning to substitute the heavy signal processing usedin other solutions. The study’s experiment was conducted by recording more than 50000chirp response signals, which were used as training data for three different binary classifiers. The classifiers obtained an accuracy of up to 93.4% on a test set. In addition,the linear relation between the quality value and bolt parameters was investigated. Thelinear correlation was weak, as the linear correlation coefficient only reached 0.23 at itshighest.