Identification of the Metal Cutting Process By Pattern Recognition Analysis of Acoustic Emission and Force Signals.

Diagnostic monitoring of machining processes is a challenging task in the development of automated machine tools as part of the unattended factory of the future. In this dissertation, the feasibility of sensing tool failure and chip breakage during turning has been investigated using pattern recognition analysis of acoustic emission and force signals. Linear discriminant functions (LDF's) were employed for identification of patterns of acoustic emission power spectra, and three dimensional cutting forces representing various states of the machining process. Acoustic emissions detected with a wide b and piezoelectric sensor attached to a tool holder assembly have been found to be highly sensitive to the frictional and fracture mechanisms associated with tool flank, and the breakage of tool and chips, respectively. After an analysis of variance study, the sensitivity was found to be maximum for AE spectral components with 20 KHz b and width within the 100 to 1000 KHz range when normalized with respect to the total power. The best features for LDF's were then determined by selecting the 20 most discriminant spectral components using a statistical class-mean distance measure. The weights of LDF's for classes of sharp tool, worn tool, tool breakage, and chip noise were obtained using the least mean-squared error design criterion based on data obtained off-line. The training and testing of LDF's using the leave-one-out error estimation method yielded high classification performance on data obtained under fixed cutting conditions. To determine the influence of varying cutting conditions on AE spectra, a two-level fractional factorial experiment was conducted yielding a bilinear functional relationship of input variables (cutting speed, feedrate, depth of cut, material hardness, tool grade type, tool wear state) with the AE signal output. This relationship was then used to filter out the effect of controlled variables resulting in an emphasis of the effect of disturbance inputs (tool wear, fracture). The linear discriminant functions based on the filtered output resulted in a performance of over 90% successful recognition of all types of tool failure. A similar filtering and pattern recognition approach was also applied to the normalized force components in the tangential, feed, and radial directions, improving tool wear detection substantially.PhDMechanical engineeringUniversity of Michiganhttp://deepblue.lib.umich.edu/bitstream/2027.42/161652/1/8801312.pd