Generative adversarial network-based semi-supervised learning for real-time risk warning of process industries

Highlights

• Deep learning methods are developed for building real-time risk warning systems.

• GAN-based semi-supervised learning requires scarce labeled data.

• Semi-supervised model incorporates numerous unlabeled samples into evaluations.

• CNN architectures handle multi-dimensional HAZOP data to enhance warning accuracy.

• Semi-supervised model has better performance for industrial data training.

Abstract

Due to the non-cognition of real-time data, rare loss-based risk warning methods can effectively respond to unexpected emergencies. Machine learning has powerful data processing capabilities and real-time computing functions and thus is suitable for offsetting the shortcomings of traditional risk methods. Risk analysis can be easily employed to perform risk-based data classification for a set of process data. However, the risk analysis process is too complicated to label risk levels for all processes, which is hard to satisfy the requirements of the amount of data for supervised learning. Therefore, the present paper focuses on developing semi-supervised learning methods for the construction of real-time risk-based early warning systems. By using fuzzy HAZOP, we estimate the risk of systems quantitatively based on the process data. With the consideration of scarce labeled data and numerous unlabeled information, we develop the generative adversarial network (GAN)-based semi-supervised learning method to identify the process risk timely. Besides, deep network architecture integrated with the convolutional neural network (CNN) is used for the codification of multi-dimensional process data to enhance the generalization of warning models. Finally, the effectiveness of the proposed method is evaluated through a comparative study with different algorithms on a case of multizone circulating reactor (MZCR).

Introduction

Industrial risk management often depends on two strategies: decreasing the frequency of hazardous events, or reducing the consequence of accidents. The former strategy mostly applies quantitative risk methods to identify critical sub-events and preform risk-based designs (Khan & Abbasi, 1998). Another strategy often reduces the loss of accidents through strengthening device redundancies, designing precursor warnings, and implementing personnel emergencies. Present industries enormously rely on failure data to monitor operating performance of processes, and thus the required improvement or change can only be identified after an incident has occurred (Wang, Khan, Ahmed & Imtiaz, 2016). The United States Center for Chemical Process Safety (CCPS) suggests that: “Facilities should monitor the real-time performance of management system activities rather than wait for accidents to happen” (CCPS, 2007). Investigations of industrial accidents reveal that almost in all cases, a variety of risk symptoms were observable before the crashes but were unfortunately ignored by managers and regulators (Zheng, Chen, Xue & Xue, 2017). As a result, risk-based early warning is a vital topic for improving system security in both academics and practices.

Risk warning system integrates both risk analysis and process alarm for the assessment of process deviations to convey impending hazard information in real-time (Hashemi, 2016). Most of the related researchers have focused on the application in financial risk, earthquake prediction, and flood forecasting (Oliveira, Sa, Lopes, Ferreira & Pais, 2015; Pappenberger, Cloke, Parker, Wetterhall & Richardson, 2015; Sevim, Oztekin, Bali, Gumus & Guresen, 2014). Often, risk warning procedures include three steps: analyzing deviation, establishing identification models, and estimating risk. For instance, Chang, Khan and Ahmed (2011) proposed a risk-based alarm management system with the consideration of hazard probability, hazard impact, and process safety time. Salzano, Agreda, Carluccio and Fabbrocino (2009) constructed an early warning system by predicting the fragility of industrial equipment. Chen, Zhou, Zhang, Du and Zhou (2015) presented a multi-index fuzzy method based on the measured data to accomplish risk early warning. Recently, loss function provides a specific representation of the dynamic relationship between industrial process deviation and system risk. Based on the loss function, Hashemi, Ahmed and Khan (2015) proposed an operational risk-based design approach for early warning systems. Wang et al. (2016) estimated the continuously updated probabilities of undesirable events by using dynamic loss functions and detecting multiple vital variables. Also, loss function has been integrated with the Bayesian theory to construct risk-based online warning systems under uncertainties (Ali & Riaz, 2019; Economou, Stephenson, Rougier, Neal & Mylne, 2016). Since process plants handle hazardous materials in daily operations, the risk warning system is critical to monitor the state of a process in real-time to identify any unsafe conditions before deviation leads to a more severe event (Wang, Khan & Ahmed, 2015). Risk analysis is a systematic and scientific method to predict the risk of possible accidents in industrial systems. However, traditional loss-based risk warning technology requires accident probabilities, which dynamic feature generally takes a month or year as a period, so that it cannot cope with emergencies. Besides, it is difficult to judge situations and make decisions due to the uncertainty, imprecision, and inconsistency of risk symptoms. With the rapid development of timeliness and complexity in process industries, traditional risk analysis technology is increasingly challenging to meet the needs of current business to reduce the accidents. In order to provide risk decision-making in the operational stage in case of the failure of critical protection layers, intelligent risk assessment and online early warning systems need to be developed and deployed (Dai, Wang, Khan & Zhao, 2016).

Machine learning has powerful data processing capabilities and real-time computing functions, which are suitable for offsetting the shortcomings of traditional risk methods. Different from model-based approaches, machine learning focuses on driving potential information to process multidimensional and time-varying data features from process variables. Typically, support vector machine (SVM), artificial neural network (ANN), and hidden Markov model (HMM) provide considerable performance for predictive warning (Dabrowski, Beyers & Villiers, 2016; Xu, Yang & Wang, 2017; Yang, Li, Ji & Xu, 2001). As can be seen, machine learning has unique potentialities in developing real-time early warning systems. More recently, the computational complexity of machine learning methods caused by a large amount of process data can be improved by deep learning techniques (LeCun, Bengio & Hinton, 2015). Zhang and Zhao (2017) proposed an extensible deep belief network (DBN) based fault diagnosis model for fault classification and early warning. Wu and Zhao (2018) developed a convolutional neural network (CNN) for chemical process fault warning. Zheng et al. (2017) integrated significant accident warning features and proposed Pythagorean-type fuzzy deep denoising autoencoder (PFDDAE) technology to achieve high accuracy for risk classification and accident warning. However, in terms of the early warning system, a critical problem is the lack of labeled faulty-case data. Numerous unbalanced data contains a small amount of accident data and a large amount of fault-free data, which cannot be directly trained by machine learning. Risk analysis techniques can be employed to identify risks and perform risk-based data classification from a set of process data. However, the risk analysis process is too complicated to label corresponding risk levels for thousands of process data from plants. Therefore, due to the unbalanced information and scarce labeled data, numerous deep supervised learning methods are difficult to be exploited for warning system modeling in the real application.

To address the above problems, a semi-supervised deep learning model was first proposed for soft sensing (Yao & Ge, 2018). Traditional semi-supervised learning methods including generative models, mixture models and graph-based methods, utilize unlabeled data to either modify or reprioritize hypotheses obtained from labeled data alone, and thus the representation of past models is constrained by the high computational complexity affected by the generated model (Zhu, 2008). As a result, a generative adversarial network (GAN) was proposed to overcome the shortcoming of approximate calculation for thorny probabilities and to provide a super presentation for semi-supervised learning (Goodfellow, Abadie, Mirza, Xu & Farley, 2014). However, the emphasis of most GAN-based application is placed on the generator rather than the discriminator. That is, the purpose of GAN-based applications is to guide the generator to generate data or pictures that are close to the real data (Douzas & Bacao, 2018; Wang & Liu, 2020; Lian, Jia, Zareapoor, Zheng & Luo, 2019; Chen, Lv & Wang, 2019). By contrast, we focus on the data analysis capability of GAN with a small amount of labeled data. Several related works can be found in the field of bearing fault diagnosis, wind turbine fault detection, and gear reliability classification (Guo, Li, Song, Wang & Chen, 2019; Li, He, Li & Chen, 2019; Liu, Qu, Hong & Zhang, 2019). GAN could be trained in an enormous precision with limited labeled data and a large amount of unlabeled data. It means that a large number of unlabeled process data from distributed control systems (DCSs) can be exploited reasonably.

Present paper focuses on the construction of real-time risk warning systems by using GAN. The significant contribution of present works is twofold: first, based on the process data, a complete fuzzy HAZOP-based risk analysis procedure is proposed to perform hazard identification, risk analysis, and decision making; second, GAN is developed to address the problem of the scarce labeled data for real-time risk assessment. Besides, deep network architecture integrated with CNN is exploited to codify multi-dimensional data and enhance the generalization of warning models. Proposed real-time risk warning framework not only can deeply extract the feature information from scarce labeled data but also can apply extra numerous unlabeled process samples to improve the model performance as well.

The remaining papers are organized as follows. In Section 2, the mechanism of risk warning methodology including risk-based pretreatment process and GAN-based semi-supervised learning is presented. A case study of a multizone circulating reactor is manifested in Section 3. Finally, conclusions are made in Section 4.