Machine learning approach in non-intrusive monitoring of tool wear evolution in massive CFRP automatic drilling processes in the aircraft industry Articles uri icon

publication date

  • October 2022

start page

  • 622

end page

  • 639

volume

  • 65

International Standard Serial Number (ISSN)

  • 0278-6125

Electronic International Standard Serial Number (EISSN)

  • 1878-6642

abstract

  • This research presents an analysis of real production data of an automatic drilling industrial system and emphasizes its ability as a process control indicator in terms of tool wear. In particular, the study is framed in Carbon-fiber-reinforced polymer composites (CFRPs) drilling operations carried out at Airbus facilities. The industrial process data were directly collected from the manufacturing plant in Getafe (in the Madrid-Spain region) and come from three different sources: spindle power consumption signals, obtained from the internal instrumentation of the machine, cutting tools wear analysis, and hole quality inspection. The main goal is to use different machining features such as tool accumulated cutting time, together with signal features to feed Machine Learning (ML) algorithms to predict tool wear. To address the inherent variability of complex production systems, it has been proposed a specific methodology that is applicable to control machining operations. The approach includes data collection, data pre-processing, and the application of Linear Regression, k-Nearest Neighbors, and Random Forest ML algorithms. As an outcome to be predicted, a novel qualitative scale of the general condition of the drill is proposed. The predictive models show promising results bearing in mind the quality and quantity of the available data – up to 3500 holes drilled with 8 diamond-coated tungsten carbide tools under different work conditions (number of layers, thickness, and others). The relevance of the benchmarks defined as representative features of the spindle power consumption as well as other machining-related parameters and their relationship with tool wear has been discussed. The Random Forest model gets the best results, being the most interesting variables the accumulated cutting time and the maximum spindle power consumption, and the most irrelevant, the number of parts to be drilled.

subjects

  • Aeronautics
  • Biology and Biomedicine
  • Industrial Engineering
  • Materials science and engineering
  • Mechanical Engineering

keywords

  • digitization; machine learning; predictive maintenance; spindle power consumption; tool wear monitoring