Advances in Process Safety and Protection of Cyber-Physical Systems

Safety has remained the foremost concern in process systems engineering for decades. The growing complexity of processes, coupled with the integration of the Internet into industries, has led to an increase in unintentional man-made disasters with severe consequences. Thankfully, numerous innovative technologies are now being implemented in actual industries. These include mathematical model-based approaches, machine-learning algorithms, and data-driven methodologies. The advent of industrial Internet technology has interconnected traditional processing industries with external networks, essentially transforming them into Cyber-Physical Systems (CPS), where the cyber component becomes integral to industrial processes. While this integration brings advantages, it also introduces significant safety risks. This Special Issue compiles five high-quality papers highlighting the latest advancements in Process Safety and the Protection of Cyber-Physical Systems. Among these, three papers delve into process fault diagnosis across industries, while two papers concentrate on estimating source terms for environmental protection in industrial parks in the event of accidents.

Monitoring and diagnosing faults in industrial processes have grown increasingly crucial to ensure their smooth operation. The primary objective of fault diagnosis is to identify any anomalies within these processes. As industrial systems become more intricate, especially with the integration of traditional processes into Cyber-Physical Systems via the Internet, establishing precise mathematical models and gathering ample empirical data becomes more challenging. Consequently, approaches reliant on first-principle models and expert knowledge lack widespread credibility. Techniques like multivariate statistical process monitoring (MSPM) [1], including principal component analysis (PCA) [2], canonical variate analysis (CVA) [3], nonlinear PCA (NLPCA) [4], kernel PCA (KPCA) [5], dynamic PCA (DPCA) [6], and deep PCA (DePCA) [7], fall under the category of data-driven methods. These techniques have demonstrated exceptional efficacy in practical fault diagnosis scenarios. They do not necessitate intricate mathematical models or empirical knowledge; instead, they detect faults by analyzing data changes. This Special Issue gathers three papers that represent the latest advancements in data-driven fault diagnosis techniques for industrial processes.

Shang et al. delved into the detection of valve stiction within control systems, which is a prevalent cause of oscillation faults in process control. This issue can significantly degrade control performance and lead to the instability of the system. Their work introduces a novel approach to identifying valve stiction. This method involves preprocessing and reconstructing data from the controller output and the controlled process variable. Initially, dynamic slow feature analysis (DSFA) is employed to extract slow features (SFs). Subsequently, it measures the Hurst exponent of the slowest SF to gauge its long-term correlation, culminating in the definition of a new valve detection index used to pinpoint valve stiction. Through simulation studies and real case analyses, the efficacy of this method has been demonstrated.

Jin et al. investigated roof-caving incidents in coal mines, which is a domain where knowledge and standardization are notably lacking. To address this, their study introduces an ontology-based semantic modeling approach for roof fall accidents aimed at facilitating the sharing and reuse of crucial knowledge for informed decision making in this context. Utilizing the ontology modeling tool Protégé, they constructed the ontology framework. This method incorporates ontology-driven deep information mining and semantic reasoning, employing semantic rules derived from expert insights and data fusion technology to comprehensively assess potential risks in mines. Through the analysis of roof fall cases, the effectiveness and practicality of this approach were affirmed, validating its utility in addressing such incidents.

Li et al. conducted fault detection for nonlinear and dynamic processes. Since most industrial systems today are nonlinear and dynamic, traditional fault detection techniques can hardly extract both nonlinear and dynamic features simultaneously. Li et al. proposed a novel nonlinear dynamic process monitoring method, namely, the “canonical variate kernel analysis” (CVKA). The new CVKA method was applied to a TE process case study, proving the excellent performance of CVKA compared to other common approaches in dynamic nonlinear process monitoring for TE-like processes. The reason behind the outperformance of CVKA is that it combines CVA in sequence for dynamic feature extraction with a kernel principal component analysis for nonlinear features from CVA residual space.

The leakage of harmful and/or hazardous gases poses a significant threat to the surrounding environment and people’s lives, leading to potential explosions, poisoning, and other critical issues. Source Term Estimation (STE) serves the purpose of swiftly localizing the source location and estimating pertinent information to minimize property damage and safeguard individuals once an accident occurs [8]. STE holds paramount importance in environmental protection within the chemical industry. Various methods for Source Term Estimation (STE) have emerged, relying on static air pollution monitoring networks (AQMNs) [9,10] and mobile sensing devices like UAVs or monitoring vehicles [11,12]. Most of these methods utilize an atmospheric transport and dispersion (ATD) model to predict the concentrations of leaked hazardous gases from possible sources. These predictions are then compared with multi-sensor data, leveraging deviations to pinpoint the real sources of the leak. However, their applicability in real-time scenarios is hindered by high computational costs and associated time constraints. To address this, several machine-learning methods have been proposed for STE, aiming to alleviate the computational load offline. This concept involves training a machine-learning-based STE model using data generated by the ATD model and then employing the pre-trained model for real-time STE with multi-sensor data. Nonetheless, the following key challenge remains: trained data must encompass all potential leakage scenarios, which is a feat often challenging in practical application. This Special Issue compiles two papers that tackle the challenges associated with Source Term Estimation in chemical industries.

Liang et al. explored a data-driven approach utilizing multiple sensors for Source Term Estimation (STE) within chemical industry settings dealing with hazardous gas leaks. Their novel contribution focused on circumventing the need to generate all conceivable leakage scenarios. This approach centered on constructing a data-driven STE model using historical multi-sensor observations that encompassed specific independent hazardous gas leakage scenarios (IHGLSs) within a targeted chemical industry park. Subsequently, this established STE model was employed to analyze online process data from multiple sensors, enabling real-time STE for the chemical industry park. To illustrate this method, it was applied to conduct STE for hazardous gas leak scenarios where a Gaussian plume model could aptly describe atmospheric transport and dispersion.

Liu et al. investigated the fusion of the grey wolf optimizer algorithm with a refined diffusion model for Source Term Estimation (STE). Their study involved the application of the following four optimization algorithms: the Gray Wolf Optimizer (GWO), Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Ant Colony Optimization (ACO) within the context of STE. Performance comparisons revealed that GWO outperformed the others in terms of computational efficiency and adaptive convergence. Building upon this finding, the study combined GWO with a modified Gaussian diffusion model featuring a surface correction factor to estimate the emission source term. Simulation outcomes showcased that this adjusted Gaussian plume model significantly enhanced the accuracy of STE, indicating its potential for use in emergency warnings and safety monitoring situations.