Image-based Measurement of Material Roughness using Machine Learning Techniques

• Published in: Procedia CIRP 2020, Vol.95, p.377-382
• Main Authors: Giusti, Alessandro, Dotta, Matteo, Maradia, Umang, Boccadoro, Marco, Gambardella, Luca M., Nasciuti, Adriano
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Electrical discharge machining (EDM); Surface analysis
• ISSN: 2212-8271; 2212-8271
• DOI: 10.1016/j.procir.2020.02.292

In-situ metrology is essential for closed-loop control of machining processes to achieve zero-defect manufacturing: in this context, using inexpensive industrial cameras integrated in machine tools is a widespread solution for dimensional measurements, but has not yet been adopted for measuring surface characteristics such as roughness. This task is challenging because surface appearance can be complex and difficult to model by standard machine vision algorithms. Optical methods are preferred because they can be easily integrated in the machining process. This is useful for EDM machines, since if the measured surface roughness does not match the requirements after the process has concluded and the workpiece removed from the machine, there is no way of correcting the error: in fact, exact repositioning of workpiece and electrode, and recreating the microscopic gap conditions which are required to resume machining, is in many cases impossible. Motivated by these requirements, this paper presents a machine-integrated, inexpensive optical measurement system, measuring Ra values comparable to results originating from contact profilometers, and with the potential to deliver additional information on surface topography. In fact, characterizing surface roughness just by the Ra value is often insufficient in practice: different surface morphologies, can have the same Ra value but different topographies, which may lead to different optical, haptic or functional properties of the surfaces. The proposed approach relies on a Convolutional Neural Network (CNN) that given as input a small square picture representing a small portion of the surface, returns the Ra value of that part of the surface. The CNN is first trained using a collection of many training instances, where each training instance is a pair composed by one input patch and the corresponding expected output, i.e. the true Ra value of the surface visible in the patch. Once trained, the CNN is deployed in the EDM machine and used to predict the Ra for new surfaces that were not part of the training. The paper reports extensive qualitative and quantitative experimental results for a range of different roughness values (0.2 < Ra [um] < 2.0).