Data-driven process monitoring and fault analysis of reformer units in hydrogen plants: Industrial application and perspectives

Fault Detection and Identification in an Acid Gas Removal Unit Using Deep Autoencoders

An acid gas removal unit (AGRU) in a natural gas processing plant is designed specifically to remove acidic components, such as carbon dioxide (CO₂) and hydrogen sulfide (H₂S), from the natural gas. The occurrence of faults, such as foaming, and to a lesser extent, damaged trays and fouling, in AGRUs is a commonly encountered problem; however, they are the least studied in the open literature. Hence, in this paper, shallow and deep sparse autoencoders with SoftMax layers are investigated to facilitate early detection of these three faults before any significant financial loss. The dynamic behavior of process variables in AGRUs in the presence of fault conditions was simulated using Aspen HYSYS Dynamics. The simulated data were used to compare five closely related fault diagnostic models, i.e., a model based on principal component analysis, a shallow sparse autoencoder without fine-tuning, a shallow sparse autoencoder with fine-tuning, a deep sparse autoencoder without fine-tuning, and a deep sparse autoencoder with fine-tuning. All models could distinguish reasonably well between the different fault conditions. The deep sparse autoencoder with fine-tuning was best able to do so with very high accuracy. Visualization of the autoencoder features yielded further insight into the performance of the models, as well as the dynamic behavior of the AGRU. Foaming was relatively difficult to distinguish from normal operating conditions. The features obtained from the fine-tuned deep autoencoder in particular can be used to construct bivariate scatter plots as a basis for automatic monitoring of the process.

Improved Process Monitoring Scheme Using Multi-Scale Independent Component Analysis

The task of fault detection is crucial in modern chemical industries for improved product quality and process safety. In this regard, data-driven fault detection (FD) strategy based on independent component analysis (ICA) has gained attention since it improves monitoring by capturing non-gaussian features in the process data. However, presence of measurement noise in the process data degrades performance of the FD strategy since the noise masks important information. To enhance the monitoring under noisy environment, wavelet-based multi-scale filtering is integrated with the ICA model to yield a novel multi-scale Independent component analysis (MSICA) FD strategy. One of the challenges in multi-scale ICA modeling is to choose the optimum decomposition depth. A novel scheme based on ICA model parameter estimation at each depth is proposed in this paper to achieve this. The effectiveness of the proposed MSICA-based FD strategy is illustrated through three case studies, namely: dynamic multi-variate process, quadruple tank process and distillation column process. In each case study, the performance of the MSICA FD strategy is assessed for different noise levels by comparing it with the conventional FD strategies. The results indicate that the proposed MSICA FD strategy can enhance performance for higher levels of noise in the data since multi-scale wavelet-based filtering is able to de-noise and capture efficient information from noisy process data.

Backstepping Methodology to Troubleshoot Plant-Wide Batch Processes in Data-Rich Industrial Environments

Troubleshooting batch processes at a plant-wide level requires first finding the unit causing the fault, and then understanding why the fault occurs in that unit. Whereas in the literature case studies discussing the latter issue abound, little attention has been given so far to the former, which is complex for several reasons: the processing units are often operated in a non-sequential way, with unusual series-parallel arrangements; holding vessels may be required to compensate for lack of production capacity, and reacting phenomena can occur in these vessels; and the evidence of batch abnormality may be available only from the end unit and at the end of the production cycle. We propose a structured methodology to assist the troubleshooting of plant-wide batch processes in data-rich environments where multivariate statistical techniques can be exploited. Namely, we first analyze the last unit wherein the fault manifests itself, and we then step back across the units through the process flow diagram (according to the manufacturing recipe) until the fault cannot be detected by the available field sensors any more. That enables us to isolate the unit wherefrom the fault originates. Interrogation of multivariate statistical models for that unit coupled to engineering judgement allow identifying the most likely root cause of the fault. We apply the proposed methodology to troubleshoot a complex industrial batch process that manufactures a specialty chemical, where productivity was originally limited by unexplained variability of the final product quality. Correction of the fault allowed for a significant increase in productivity.

Integrated Diagnostic Framework for Process and Sensor Faults in Chemical Industry

This study considers the problem of distinguishing between process and sensor faults in nonlinear chemical processes. An integrated fault diagnosis framework is proposed to distinguish chemical process sensor faults from process faults. The key idea of the framework is to embed the cycle temporal algorithm into the dynamic kernel principal component analysis to improve the fault detection speed and accuracy. It is combined with the fault diagnosis method based on the reconstruction-based contribution graph to diagnose the fault variables and then distinguish the two fault types according to their characteristics. Finally, the integrated fault diagnosis framework is applied to the Tennessee Eastman process and acid gas absorption process, and its effectiveness is proved.