Harnessing Real-Time Sensor Data for Fault Diagnosis in the Continuous Production of Copper Magnet Wires

Headnote

Abstract

The production of defective products leads to substantial material and financial losses. It is critical to detect manufacturing defects as early as possible to reduce waste. To do so, many manufacturers have harnessed technology by installing sensors to monitor their production processes in real time. Nonetheless, a significant portion of the data collected by the sensors remains unanalyzed due to processing and analytical constraints, resulting in missed opportunities for early fault diagnosis and quality improvement. This research leverages real-time sensor data from Rea Magnet Wire, one of the world's largest manufacturers of magnet and nonferrous wire products, to identify process variables responsible for defects in the continuous manufacturing of copper wires. Due to the continuous nature of the manufacturing process, data alignment was performed prior to model training. Three machine learning models (Random Forest, XGBoost, and Logistic Regression) were implemented to predict faults in copper magnet wire production using 21 process variables obtained from real-time sensor data from their Lafayette plant. We observe that interactions between process variables are critical for fault diagnosis, given that nonlinear methods have better performance. Furthermore, we study the effect of process variables on quality measures using cumulative local effects.

Keywords

Data Analysis, fault diagnosis, quality, continuous manufacturing

(ProQuest: ... denotes formulae omitted.)

1 Introduction

In various manufacturing systems, an output variable (e.g., quality measure) is influenced by input variables (e.g., process parameters) [1]. Technological advances in manufacturing have enabled the integration of sensors into production lines, facilitating real-time data collection during continuous manufacturing processes [2]. However, dozens or even hundreds of variables can be generated every second or minute continuously over time from these systems. The substantial volume of data gives rise to the curse of dimensionality, where the sample size required to accurately estimate relationships between variables increases exponentially with the number of features [3]. Furthermore, time series data collected in continuous manufacturing presents unique challenges, such as temporal misalignment and nonlinear trends [4].

This study focuses on Rea Magnet Wire Company, Inc., one of the world's largest magnet and non-ferrous wire product manufacturers. Founded in 1933, Rea produces insulated copper, aluminum, and brass magnet wire used in the manufacture of motors, transformers, and coils, as well as various specialty wire products [5]. For this research,...