Machining process monitoring using multivariate latent variable methods

This thesis outlines a multivariate latent variable approach to Tool Condition Monitoring (TCM) for machine tools. The multivariate approach provides similar levels of predictive ability to other commonly used modeling approaches applied in the field of machining but has the added benefit of providing the user with detailed process insight. This insight can be used to highlight process shifts and provide underlying root cause information which can aid in taking corrective action and in making process improvement decisions.;This thesis deals with the following four topics: (1) The Projection to Latent Structure-Discriminant Analysis (PLS-DA) is used to integrate data from multiple sensors for both individual and multiple machines. The use of a new feature-augmented model is proposed to provide the user with a simple interface to monitor quality across a complex cell involving multiple machines with different process states. Experiments were conducted on three different milling machines under different operating conditions: sharp, worn and chipped tools as well as conditions which generate chatter. The results show that the proposed technique can successfully differentiate between the different conditions tested across a group of similar machines. (2) The same sensory system was used to build a new monitoring model for the surface roughness Ra of the machined parts. A Projection to Latent Structure (PLS) model was used to integrate sensory data from individual and multiple machines and to relate this information to the surface roughness of the machined parts. The sensory signals were further analyzed and monitoring features including non-linear terms were extracted. The model was tested under different conditions with the results showing that the proposed technique can be used for on-line monitoring of surface roughness. Based on the sensory system configuration, PLS was able to successfully detect new process conditions and shifts; both those modeled and unmodeled. In addition, a comparison between PLS and the current modeling approaches like artificial neural nets and multiple regression analysis was carried out for surface roughness modeling. (3) Jaeckle and MacGregor [23] introduced a technique to estimate the required operating conditions in order to achieve a specific process yield with a desired set of characteristics. This dissertation provides a detailed study of the application of such a technique for estimating the cutting parameters of a machining process. In this work the null space solutions were considered in a new way to support machine tool and operating parameter selection decisions based on achieving a desired quality level while maximizing productivity. This approach was demonstrated on a side end-milling process using the PLS model developed for the surface roughness Ra estimation with the Material Removal Rate (MMR) used as the productivity measure. The model inversion solutions were validated experimentally and the results show very good agreement between the surface finish achieved for the estimated cutting parameters. (4) The multivariate latent concepts (PCA, PLS) were applied on two different industrial case studies using data from the General Motors engine plant in St. Catherines, Ontario. The industrial case studies were performed to gain insight into the implementation issues that can be experienced as well as highlight the value of the insight provided by the multivariate latent variable approach.