Monitoring bivariate autocorrelated process using a deep learning-based control chart: A case study on the car manufacturing industry

Abstract

• Monitoring the bivariate autocorrelated process using machine learning based control charts. • Combination of machine and deep learning models in SPC. • ARL and EARL computation through Monte Carlo simulations. • Monitoring the wheel alignment using the Dürr AG’s ’ x-wheel ’ device. Due to the high frequency of sensor data collection in modern industries, consecutive observations are potentially autocorrelated. Failing to address this issue adequately when designing control charts can lead to numerous false alarms, thereby compromising the efficiency of the monitoring technique. On the other hand, thanks to the evolution of measurement equipment, nowadays more than one quality characteristics are often measured simultaneously. The Hotelling T 2 chart is the most used approach to detect abnormal (or OOC) conditions of processes with multiple quality characteristics. The OOC condition could be due to factors such as shifts in process mean, variance–covariance matrix, outliers, or trends. Conventional statistical control charts have weaknesses in the detection of complex OOC scenarios, while the occurrence of such scenarios is particularly prevalent in today’s modern industries. This paper presents a novel approach for developing control charts using machine learning techniques, specifically LSTM, to monitor bivariate autocorrelated processes. Utilizing a deep memory information structure and novel input features can efficiently capture the complex interdependencies and temporal patterns in autocorrelated processes, leading to suboptimal performance detecting process deviations. The proposed methodology is evaluated using Monte Carlo simulations, demonstrating improved performance compared to traditional control charts. A new practical usage is discovered for monitoring wheel alignment in a car manufacturing assembly using the ‘x-wheel’ device data, specifically designed to examine essential wheel angles. This real-life example demonstrates the feasibility of implementing the proposed method on production lines to oversee a correlated process with two variables.