A Deep Learning Framework for Acoustic Condition Monitoring in Machining Nickel Superalloy Turbine Blades
Name
placzek-lplaczek-meng-meche-2025-thesis.pdf
Description
Thesis PDF
Size
11.12 MB
Format
Adobe PDF
Checksum (MD5)
a9c18ff66e94ae190f0375cc16ad24e6
Author(s)
Placzek, Luke Alexander
Advisor(s)
Hardt, David E.
Date Issued
September 2025
Publisher
Massachusetts Institute of Technology
Abstract
The machining of gas turbine blade cooling holes presents persistent manufacturing challenges due to the difficulty of working with single-crystal nickel superalloys, ceramic coatings, and intricate geometries. In the machining of the cooling holes on GE Vernova's 7HA.03 stage 1 turbine blades, over 60 carbide tools are typically consumed per blade, and machining interruptions due to tool breakage occur on average every 30 minutes, leading to over 5 hours of additional process time per blade. While modern CNC machines provide machine state logs through systems like Vimana, due to the size of tools, the RoboDrills in use are only able to detect tool breakage after hole completion, making it impossible to determine when or where within a hole the tool fails. This work was done in conjunction with fellow M.Eng. student Alexander Brush [1], who worked on the statistical modeling of current tool breakage and modification of process parameters to reduce tool breakage. This thesis builds a data pipeline for collecting audio at a 96kHz sample rate, visualize it, and combine it with the machine data and breakage statistics to segment the audio into appropriate files which can be used to train Machine Learning (ML) models to detect tool failures. To test this framework, multiple classification models were trained on 32 hours of segmented audio data synchronized with Vimana machine logs. The data pipeline automatically segments continuous audio recordings into labeled training examples based on machine state, with augmentation methods to mitigate the severe class imbalance. Additionally, a Statistical Process Control (SPC) system was implemented to track tool breakage rates in real-time, using Nelson rules to detect significant changes in process performance. This enables engineers to identify when process modifications impact tool life and correlate changes with specific machines, rows of holes, or operational parameters. The combination of acoustic monitoring for precise breakage detection and SPC for trend analysis provides engineers with actionable insights for process improvement. While the current model accuracy requires improvement before production deployment, this work establishes the infrastructure and methodology for enhanced tool condition monitoring in high-value manufacturing.
MIT Department
Massachusetts Institute of Technology. Department of Mechanical Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
