Myths and misconceptions of data-driven methods: Applications to process safety analysis. (February 2022)

• Record Type: Journal Article
• Authors: Wen, He; Khan, Faisal; Amin, Md. Tanjin; Halim, S. Zohra
• Is Part Of: Computers & chemical engineering. Volume 158(2022)
• Publication Date: 2022-02
• Subjects: Process safety analysis -- Data-driven method -- Data source -- Myth -- Misconception; ADR Accurate diagnosis rate -- AHP Analytic hierarchy process -- AIC Akaike information criterion -- ANN Artificial neural network -- BIC Bayesian information criterion -- BN Bayesian network -- BT Bow tie -- CACE Computers & Chemical Engineering -- CART Classification and regression tree -- CNN Convolutional neural network -- CPT Conditional probability table -- DM Data mining -- DT Decision tree -- ETA Event tree analysis -- FAR False alarm rate -- FDD Fault detection and diagnosis -- FDR Fault detection rate -- FMEA Failure mode and effects analysis -- FT Fuzzy theory -- FTA Fault tree analysis -- HAZOP Hazard and operability studies -- HSE Health, Safety, and Environment -- ICA Independent component analysis -- ISO International Organization for Standardization -- JHM Journal of Hazardous Materials -- JLPPI Journal of Loss Prevention in the Process Industries -- JRR Proceedings of the Institution of Mechanical Engineers, Part O: Journal of Risk and Reliability -- JSR Journal of Safety Research -- KICA Kernel independent component analysis -- KMC K-means clustering -- KNN K-nearest neighbor -- LOPA Layer of protection analysis -- LR Logistic regression -- MAR Missed alarm rate -- ML Machine learning -- MSPM Multivariate statistical process monitoring -- M-ANN Myth of ANN (overuse of ANN) -- M-BN Myth of BN (missing BN's underlying assumption) -- M-CC Myth of correlation coefficient (using correlation coefficient for model verification) -- M-DR Myth of data representation (improper data representation) -- M-EA Myth of error analysis (absence of error analysis) -- M-MSPM Myth of multivariate statistical process monitoring (Absence of model behavior analysis for MSPM) -- NBC Naïve bayes classifier -- PCA Principal component analysis -- PCC Pearson correlation coefficient -- PLS Partial least square -- PN Petri net -- PSEP Process Safety and Environmental Protection -- PSP Process Safety Progress -- QRA Quantitative risk analysis -- RA Risk Analysis -- RESS Reliability Engineering & System Safety -- RF Random forest -- RNN Recurrent neural network -- SGD Stochastic gradient descent -- SS Safety Science -- SVM Support vector machine -- WoS Web of Science Core Collection; Chemical engineering -- Data processing -- Periodicals; 660.0285
• DOI: 10.1016/j.compchemeng.2021.107639
• Languages: English
• ISSNs: 0098-1354

Abstract

Highlights: This study analyzes the five most common myths and misconceptions of data-driven methods observed in recent literature. The study observes data representation, appropriate assumptions, and use of method without detailed understanding of its limitation are key issues. The authors believe this work will help peers to avoid the myths studied here. Abstract: With a rapid realization of process digitization, data-driven methods are being increasingly adopted in process safety analysis. However, the use of data-driven methods contains a varied degree of myths and misconceptions resulting from the application of a method or the data representation that does not follow a proper scientific notion. These myths and misconceptions cause significant errors in terms of results and their interpretability. Hence, the purpose of this study is set to analyze the most common myths and misconceptions of data-driven methods observed in the recent literature. In the current work, we have analyzed 500 public domain articles from 1990 to 2020, published in 10 renowned safety journals. The analysis attempts to address the following questions: (i) What are the key data in process safety analysis? (ii) What are the sources of data? (iii) What does the data-driven method mean? (iv) What are the common myths and misconceptions of data-driven methods? and (v) How frequently such myths and misconceptions are occurring? After analyzing the 500 articles, it is observed that most of the myths are related to improper data representation, missing appropriate assumptions, and blanket use of methods without a detailed understanding of their limitations. The authors believe this work will help peers to avoid the myths studied here, use data-driven methods with scientific rigor, and present findings in a meaningful way.