Probability-Based Statistical Quality Control for Sensor-Driven Process Monitoring and Anomaly Detection
Abstract
Statistical quality control (SQC) provides an essential analytical basis for monitoring process stability, detecting assignable causes, and improving product-quality performance in data-rich production systems. This study develops and evaluates a probability-based monitoring framework that integrates individuals and moving-range control charts, exponentially weighted moving average (EWMA) monitoring, cumulative sum (CUSUM) monitoring, process capability analysis, fitted probability distributions, and anomaly-detection logic. The framework is empirically examined using public, production-quality data comprising 2,103,841 minute-level process-sensor observations and 29,184 hourly laboratory quality measurements collected from a roasting production process. The process-sensor data were aligned with each quality measurement by aggregating the preceding 60-minute operating window, resulting in a complete hourly monitoring dataset with no missing values. Results show that the Weibull distribution provides the lowest information criterion for the quality response, indicating that a simple normal assumption is insufficient for all monitoring decisions. The Phase-I individuals chart identified 79 quality observations outside three-sigma probability limits, while the moving-range chart flagged 532 abrupt changes. Memory-type charts were more sensitive to sustained drift, with the exponentially weighted moving average (EWMA) chart flagging 7,642 observations and the cumulative sum (CUSUM) chart flagging 21,129 observations. Process capability analysis using a declared benchmark interval yielded a process capability index (Cp) of 0.432 and a process capability index adjusted for centering (Cpk) of 0.380, indicating that variability reduction and centering improvement are needed before the process can be considered capable. The strongest empirical relationship was between the third chamber temperature and product quality, with a negative correlation of −0.810. The study contributes a reproducible, interpretable, and probability-based statistical quality control (SQC) framework that can support quality engineers in distinguishing common-cause variation from abnormal process behavior.