Knowledge Creation via Data Analytics in a High Pressure Die Casting Operation Public
Downloadable Contentopen in viewer
Big Data is a term typically associated with large internet entities such as Facebook, eBay, and Google where every click, search, and upload builds an actionable dataset used to target advertisements and enhance the user experience. The Data Science realm classifies data as Big Data by the three V’s: volume, velocity and variety. The high-pressure die-casting (HPDC) process is a commonly employed method of producing large volumes of cast components particularly in aluminum alloys. The automated pieces of equipment are interfaced such that any signal passed from one machine to another, or sensor input, is data which may be relevant to the output of the process. For each part cast, it is possible to record the input parameters: melt temperature, lock up tonnage, cycle time, plunger velocities, cavity fill time, intensification pressure, dwell time, and spray time to name a few. Outputs parameters are generally lacking across all parts, rather they are measured on an audit basis. Is die casting process data truly big data? In the scale of the internet giants and major banking and credit firms, no. However, in some respects to the three V’s, it is. Certainly, in a die casting facility with multiple machines running production the velocity of data generation is high on the input side. Die casting data resides in spreadsheets, databases, images, and shift notes. Thus, variety of data is a consideration. Generating great volume, unfortunately, is often a challenge. It is posited that the same tools can be used to gain knowledge into the HPDC process as in other truly Big Data environments. Production HPDC process data, generously donated by FCA Kokomo Casting Plant, covering one year of production across 12 die casting machines and 20 die cavities was used to assess the applicability of machine learning algorithms such as Random Forest, Support Vector Machine, and XGBoost for the prediction of casting quality and performance metrics. The challenges which arise from the characteristics of materials processing data, using HPDC as the exemplar, have been identified. The proper data preparation methods for machine learning have been described. Predictive modeling of part quality and mechanical properties of die-cast engine blocks has been performed with an emphasis on model evaluation and cross-validation.
- Defense date
- Date created
- Resource type
- Rights statement
Permanent link to this page: https://digital.wpi.edu/show/3484zk755