Tool wear in machining is the loss of material from the tool cutting edge due to interaction with the workpiece material. Tool life in machining is widely considered to be difficult to predict due to the large number of influencing variables and the stochastic nature of tool wear. Extensive tool wear experimentation to model tool life is expensive and time consuming for different tool-material combinations, and therefore, infeasible in a production environment. In this presentation, machine learning classification methods for tool life modeling in a production environment using shop floor data is described. The idea is to treat shop floor production parts as tool wear experiments and utilize the tool wear data for modeling tool life. Since tool wear can only be measured at the time of tool change in a production environment, tool life modeling is treated as a classification problem. Tool wear experiments are performed, and shop floor data is simulated as tool worn (class 0) and tool not worn (class 1) using the experimental results. Different machine learning classification methods such as Support Vector Machines (SVM), Logistic regression, and Gaussian Process are used to model tool life using the simulated data and the results are compared. Different approaches to generate synthetic data to augment sparse and unbalanced datasets using domain knowledge and experience are presented. Results show that machine learning classification approach offers a practical solution to tool life modeling in a production environment.
- Learn basics of machine learning classification methods
- Identify potential applications in manufacturing for machine learning
- Recognize tools and challenges in implementing machine learning
R&D StaffOak Ridge National Laboratory