Incorporation of process-specific structure in statistical process monitoring: A review

A Hybrid Data-Based and Model-Based Approach to Process Monitoring and Control in Sheet Metal Forming

“…Industrial Process Monitoring (IPM) is an activity that involves tracking of machinery, systems, and processes in real-time to achieve production consistency, safety, and economic viability [7]. The focus of IPM has evolved through the years, initially almost entirely dedicated to improving fault detection performance while, more recently, diagnosis and root-cause analysis is gaining importance in which the set of variables causing a fault are identified and isolated [16]. Moving forward, the emphasis is in the direction of predictive analytics where the evolution of operational risks is assessed, which aids better planning and shutdown operations, minimizes production losses, and secures equipment/operator safety [7].…”

“…Reis [16] presents a review on the prevalent process monitoring approaches highlighting recent trends within IPM and the variety of approaches emerging within applied statistics, engineering, and machine learning communities. The review points out that classical monitoring approaches, which incorporate a generic probabilistic model structure representing normal operating conditions (NOC), such as Shewhart [22], Exponentially Weighted Moving Average (EWMA), and Cumulative Sum (CUMSUM) [23], are unsuitable for root cause analysis as these methods do not contain process-specific structure information other than the parameter estimates of the NOC model obtained from process data [7].…”

“…As a consequence, monitoring approaches have evolved to multivariate approaches and further to high-dimensional methodologies such as Principal Component Analysis (PCA) and Partial Least Square (PLS) [7]. However, these approaches are non-causal and, hence, would lead to ambiguous results when used for fault diagnosis purposes [16]. A good example is the smearing-out effect in PCA-based process monitoring [24][25][26] in which, for a given change in the monitoring statistic, the method is not able to distinguish between the set of variables causing a fault and the variables that are a consequence of the fault.…”

“…As a response to this limitation, subsequent approaches aimed at incorporating more of the causal information into process NOC models, i.e., the knowledge regarding causes and effects of a change in state are embedded into the underlying model structure to effectively guide fault diagnosis and troubleshooting activities. A common perception among process monitoring research [7,16,27] is that the inclusion of process-specific information into the monitoring methods shall bring extra sensitivity toward process fluctuations leading to finer fault detection/diagnosis. Though this will benefit the fault detection capability, fault diagnosis/troubleshooting activities will reap significantly greater benefits from the added sensitivity since these activities are known to consume a considerably larger portion of the process downtime [7].…”

“…In this context, several data-driven [7][8][9][10][11], model-based [12][13][14], and statistical [15] approaches have been proposed to support Processes 2020, 8, 89 2 of 11 the identification of the underlying root cause of a fault. However, most of these approaches lack features necessary to completely diagnose and isolate a fault [16]. Similarly, numerous control approaches have been developed to regulate increasingly complex manufacturing processes [17,18].…”