Marine dual fuel engines monitoring in the wild through weakly supervised data analytics

https://doi.org/10.1016/j.engappai.2021.104179Get rights and content

Abstract

Background:

Maritime transportation accounts for around 80% of the world freight movements, remarkably contributing to the global environmental footprint. Dual fuel engines, running on both gaseous and liquid fuels, represent a viable way toward the reduction of emissions at the cost of additional complexity in monitoring activities.

Motivation:

Data-driven methods represent the frontier in research and in maritime industrial applications, and they usually require a large amount of labelled data, i.e., sensor measurements plus the associated engine status usually annotated by human operators, which are costly and seldomly available in the wild. Unlabelled samples, instead, are commonly, cheaply, and readily available.

Hypothesis:

The enabling technology for data-driven methods is the availability of a network of sensors and an automation system able to capture and store the associated stream of data.

Methods:

In this paper, we design and propose multiple alternatives toward the weakly supervised marine dual fuel engines data-driven monitoring. To this aim, we will rely on a Digital Twin of the dual fuel engine or on novelty detection algorithms and we will compare them against state-of-the-art fully supervised approaches.

Results:

Results on data generated from a real-data validated simulator of a marine dual fuel engine demonstrate that the proposed weakly supervised monitoring approaches lead to a negligible loss in accuracy compared to costly and often unfeasible fully supervised ones supporting the validity of the proposal for its application in the wild.

Conclusion:

The main outcome is a guideline for selecting the best data-driven dual fuel engine monitoring method according to the available data.

Introduction

Maritime transportation accounts for around 80% of global freight movements (Mangan, 2017). With very few exceptions, vessels are powered by internal combustion engines, burning conventional fossil fuels, producing a large amounts of undesired greenhouse and non-greenhouse emissions (Reitz et al., 2020). In fact, carbon dioxide, carbon monoxide, sulphur oxides, nitrogen oxides, methane, and particulate matter (including black carbon) negatively affect the climate, the environment, and the public health (Manisalidis et al., 2020).

The use of alternative fuels including natural gas, methanol, and biofuels has been proposed as a viable way toward the improvement of the environmental sustainability of the maritime transportation (Hansson et al., 2019, Zuo et al., 2020a, Jiaqiang et al., 2019). In particular, the use of Liquefied Natural Gas (LNG) as a fuel proved to be the most viable solution due to the lower LNG fuel price levels compared to other fossil fuels (Bae and Kim, 2017), the rapid development of the global LNG infrastructure (Thomson et al., 2015), as well as the clean nature of lean combustion, which leads to the reduction of the nitrogen oxides due to the low carbon to hydrogen ratio whilst almost eliminating most of the particulate matter and sulphur oxide emissions (Hansson et al., 2019, Louis, 2001, Zuo et al., 2020b, Qian et al., 2020).

The economic and environmental benefits of using LNG led the marine engine manufacturers to the development of Dual Fuel (DF) versions of both two-stroke and four-stroke Diesel Engines (DEs) (Pavlenko et al., 2020, Zhong et al., 2020) as well as retrofitting kits for converting existing diesel engines (DE) to DF engines1,2 Moreover, a pure natural gas fuelled vessel requires at least 48% more storage capacity (Boretti, 2019). Nevertheless, an obvious drawback of DF engines is the additional technological complexity of the engine fuel system as well as the monitoring, control and safety systems (Boretti, 2019, Wang et al., 2016) to keep their performance always at a desired level. In fact, the degradation of marine engines performance decreases their operational efficiency, leading to a higher fuel consumption and consequently to an increase in greenhouse emissions. For this reason, the implementation of efficient and effective monitoring strategies is of paramount importance to ensure availability, reliability, cost, and environmental sustainability (Gratsos et al., 2009, Lloyd and Cackette, 2001, Xu et al., 2002).

Marine engine manufacturers already provide turnkey monitoring solutions for their DEs. For example, MaK DICARE (CAT, 2019) remote engine monitoring system provides condition-specific maintenance suggestions comparing in real time the engine condition to the desired state and suggesting maintenance actions. Another example is MAN Computer Controlled Surveillance (MAN, 2019), a diagnostic tool for monitoring and storing DEs performance data, and trends dedicated to assisting users in evaluating the machinery status and performance. These systems are based on the knowledge of permissible operating engine parameters and actions are triggered when the monitored parameters exceed their boundaries. The final decision on the actions to undertake is usually left to the operators experience and knowledge, and this needs to be addressed for enabling highly automated or autonomous systems operation. In fact, operators need to be carefully trained, their decisions are biased by their experience, too many degrees of freedom are left to their judgement, their ability to exploit the automation data is limited, and the monitoring process may be stressful and time consuming.

State-of-the-art methods try to overcome the limitations of exploiting the human in the loop for monitoring activities by exploiting instead numerical methods (Kowalski et al., 2017, Cipollini et al., 2018a, Cipollini et al., 2018b). For this purpose, a gold-standard solution is to compare the engine behaviour in dynamic conditions with the normal (expected) behaviour provided by an accurate Digital Twin Grimmelius et al. (2007). This solution enables the identification of unexpected behaviour and to establish trends in temporal performance variation. Numerical methods play a central role in developing an accurate Digital Twin of the engine for the prediction of key performance parameters. In particular, engine modelling has been performed by employing commercial or custom-made tools based on first principles and thermo-physical processes fundamentals. A number of engine models of varying complexity are reported in the pertinent literature (Xiang et al., 2019, Baldi et al., 2015, Reitz and Rutland, 1995). Detailed modelling approaches (of the zero-dimensional to three-dimensional type) result in computationally demanding simulations and consequently are unsuitable for real-time engine monitoring applications (Stoumpos et al., 2018, Stoumpos et al., 2020).

To develop models suitable for effective monitoring in real operational conditions, two main alternatives exist. The first one is to exploit approximate but computationally efficient first-principle models of the mean value type (Geertsma et al., 2017, Geertsma et al., 2018) or the combined mean value/0D type (Baldi et al., 2015), whilst the second one is to exploit the historical data acquired by the modern automation systems to build accurate data-driven models. The first approach results in faster but often inaccurate predictions, which limits their effectiveness for engine monitoring in real time operation (Geertsma et al., 2018). The second approach represents the frontier in both research and industrial application and is highly dependent on the availability of an adequate amount of historical data (Talaat et al., 2018, Cipollini et al., 2018b). Their enabling technology is then the availability of a network of sensors and an automation system able to capture and store the associated stream of data, which are nowadays readily available. This data, which is often called unlabelled as there is usually no associated annotation about the status of the engines, are commonly, cheaply, and readily available (Munim et al., 2020). In fact, these annotations are costly and rarely available in the wild since they require the supervision of the engine performed by an human operator (Nixon et al., 2018). In some cases, this labelling activity requires to reduce the operation or to eventually stop the vessel or to maintain the engine. Consequently, unlabelled data is available in large quantities for a large period of time with a very high frequency and just a very small amount of these data are actually labelled during planned maintenance (every few years) or during exceptional disruptions (few time in the life of a vessel).

In this respect, this study aims at designing and proposing multiple alternatives for the reduction of the use of labelled data toward a weekly supervised monitoring for marine DF engines targeting to reduce as much as possible, the necessity of labelled data at least to a realistic level which is realistic to retrieve in the wild. Furthermore, this study proves that the preceding proposal does not compromise the modelling accuracy below a level that prevents their use in real operations.

The approach employed in this study includes the following three steps: (a) the Fully Supervised Performance estimation; (b) the Fully Supervised Health Status Estimation, and; (c) the Weakly Supervised Health Status Estimation. The Fully Supervised Performance Estimation step includes the design of a Digital Twin, exploiting state-of-the-art supervised data-driven methods for enabling the prediction of the engine performance and emissions parameters based on the control variables (e.g. engine load and engine speed), in healthy engine conditions. This step actually does not employ labelled data; instead it employs the acquired data from engine operation under healthy conditions. The Fully Supervised Health Status Estimation step focuses on developing models capable of classifying the status of the engines as healthy or faulty and it is accomplished by employing two approaches. The first one employs the Digital Twin developed in the first step to estimate the deviation (drift) of the parameters of the actual engine operation (based on the acquired data) from the respective Digital Twin predicted parameters. The second one exploits state-of-the-art supervised data-driven methods to classify the status of the investigated engine based on the control and performance parameters. This step requires labelled data with the engine under healthy and faulty conditions. The Weakly Supervised Health Status Estimation step focuses on reducing the amount of labelled data required to build the models developed in the second step by employing two approaches. The first one focuses on the estimation of the engine performance parameters variation from the respective parameters calculated by the Digital Twin by employing a limited amount of labelled data for tuning the drift detection model. The second one, instead, will exploit state-of-the-art unsupervised data-driven methods to detect abnormal conditions (anomalies) of the investigated engine by employing as input the considered control and performance parameters. The weakly supervised health status estimation step employs the models trained just with data acquired under the engine healthy conditions from the engine monitoring system. These models are subsequently fine tuned with a very small amount of labelled data. Fig. 1 depicts our proposal with a simple graphical representation.

Based on the preceding methodology, this study contributes to the better understanding of the effects of using, in multiple methods, labelled and unlabelled data as well as to quantify the methods accuracy deterioration in cases when the available labelled data is limited. Furthermore, it is demonstrated whether it is possible to monitor the investigated engine status by employing a weakly supervised method with a realistic amount of data.

It must be noted that a large amount of labelled data acquired from marine DE and DF engines pertaining to faulty conditions are not currently available in the literature. To overcome this limitation, this study employs data generated from a validated simulator of a marine DF engine capable of simulating both healthy and faulty conditions (Stoumpos et al., 2020). This study demonstrate that the proposed weakly supervised monitoring approaches lead to a negligible deterioration of the prediction accuracy compared with the costly and often unfeasible fully supervised ones, supporting the validity of the proposal for its application in the wild.

The rest of the paper is organised as follows. Section 2 provides an overview of the state-of-the-art on engine modelling with a particular focus on maritime applications. Section 3 introduces the engine simulator employed to generate the required data. Section 4 describes how the dataset used in this study has been generated by using the simulator described in Section 3. Section 5 presents our proposal. Section 6 tests and demonstrates its validity employing the data described in Section 4. Section 7 summarises the main findings of this study.

Section snippets

Related work

The marine DEs and DF engines design, development, optimisation, and monitoring procedures are nowadays increasingly based on mathematical modelling, numerical simulations, and data-driven models, rather than on experiments and prototyping (Lebedevas et al., 2020). He and Rutland (2004), for example, developed a general DEs simulation tool with a small computer resource footprint for engine design based on Artificial Neural Networks.

Karlsson et al. (2010) and Stewart and Borrelli (2008) point

The considered DF engine

This study investigates the Wärtsilä 9L50DF engine, which is a marine four-stroke, turbocharged and intercooled DF engine (Wärtsilä, 2012) that is employed for ship propulsion or electrical generation, in the latter case as part of a generator set (Wärtsilä, 2012). The engine is capable of operating in two distinct modes: (i) the gas mode running on natural gas and liquid pilot fuel, usually Light Fuel Oil, that is injected in the engine cylinders for initiating the combustion of the premixed

Dataset generation

For the reason described in the introduction and for confidentiality issues, datasets corresponding to the investigated marine DF engine under faulty conditions were not available to the authors, therefore, this study employed the physical model of high fidelity that was developed and validated in previous authors’ studies (Stoumpos et al., 2018), which was briefly described in Section 3.2, to generate it. For this purpose, multiple simulation runs, corresponding to different scenario, were

Methods

In this section, the proposed models are presented, which address the problem described in Section 1 exploiting the datasets described in Section 4.

The problem described in Section 1 can be associated to a conventional ML framework (Shalev-Shwartz and Ben-David, 2014), in which one has to consider an input space XRd and an output space Y and the goal estimating the unknown rule μ:XY which associates an element yY to an element xX. ML techniques estimates μ through a learning algorithm AH:Dn×

Experimental results

In this section, we will present the results and the quality of the methodologies presented in Section 5 for solving the problem described in Section 1 (weekly supervised monitoring for marine DF engines targeting to reduce, as much as possible, the necessity of labelled data at least to a realistic level which is realistic to retrieve in the wild without compromising the recognition accuracy) by means of the data described in Section 4.

Conclusions

Marine dual fuel engines, running on both gaseous and liquid fuels, represent a viable way toward the reduction of emissions since maritime transportation accounts for around 80% of the world freight movements, remarkably contributing to the global environmental footprint. The side effect of this transition is in additional complexity in monitoring activities which are required to keep their performance always at the desired level. In fact, the degradation of marine engines performance

CRediT authorship contribution statement

Andrea Coraddu: Conceptualization, Methodology, Data curation, Writing - original draft, Writing - reviewing and editing. Luca Oneto: Conceptualization, Methodology, Data curation, Writing - original draft, Writing - reviewing and editing, Software, Validation. Davide Ilardi: Data curation, Methodology, Writing - original draft. Sokratis Stoumpos: Data curation, Writing - original draft. Gerasimos Theotokatos: Supervision, Conceptualization, Reviewing, Writing - original draft.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (90)

  • QianY. et al.

    Improvement of combustion performance and emissions in a gasoline direct injection (GDI) engine by modulation of fuel volatility

    Fuel

    (2020)
  • ReitzR.D. et al.

    Development and testing of diesel engine CFD models

    Prog. Energy Combust. Sci.

    (1995)
  • ShaylerP.J. et al.

    The exploitation of neural networks in automotive engine management systems

    Eng. Appl. Artif. Intell.

    (2000)
  • ShinS. et al.

    Deep neural network model with Bayesian hyperparameter optimization for prediction of NOx at transient conditions in a diesel engine

    Eng. Appl. Artif. Intell.

    (2020)
  • StoumposS. et al.

    Marine dual fuel engine modelling and parametric investigation of engine settings effect on performance-emissions trade-offs

    Ocean Eng.

    (2018)
  • SyedJ. et al.

    Artificial neural network modeling of a hydrogen dual fueled diesel engine characteristics: An experiment approach

    Int. J. Hydrogen Energy

    (2017)
  • TalaatM. et al.

    A hybrid model of an artificial neural network with thermodynamic model for system diagnosis of electrical power plant gas turbine

    Eng. Appl. Artif. Intell.

    (2018)
  • ThomsonH. et al.

    Natural gas as a marine fuel

    Energy Policy

    (2015)
  • TurksonR.F. et al.

    Artificial neural network applications in the calibration of spark-ignition engines: An overview

    Eng. Sci. Technol. Int. J.

    (2016)
  • WakodeV.R. et al.

    Regression analysis and optimization of diesel engine performance for change in fuel injection pressure and compression ratio

    Appl. Therm. Eng.

    (2017)
  • WangY.S. et al.

    An engine-fault-diagnosis system based on sound intensity analysis and wavelet packet pre-processing neural network

    Eng. Appl. Artif. Intell.

    (2020)
  • XiW. et al.

    A feature extraction and visualization method for fault detection of marine diesel engines

    Measurement

    (2018)
  • XiangL. et al.

    Investigation on gaseous fuels interchangeability with an extended zero-dimensional engine model

    Energy Convers. Manage.

    (2019)
  • XuK. et al.

    Fuzzy assessment of FMEA for engine systems

    Reliab. Eng. Syst. Saf.

    (2002)
  • ZhaiY.J. et al.

    Neural network model-based automotive engine air/fuel ratio control and robustness evaluation

    Eng. Appl. Artif. Intell.

    (2009)
  • ZhongW. et al.

    Experimental study the effect of injection strategies on combustion and emission characteristics in gasoline compression ignition engines using gasoline/hydrogenated catalytic biodiesel blends

    Fuel

    (2020)
  • ZuoQ. et al.

    Effect of critical dual-carrier structure parameters on performance enhancement of a dual-carrier catalytic converter and the gasoline engine system

    Energy Convers. Manage.

    (2020)
  • ZuoQ. et al.

    Effect of different exhaust parameters on NO conversion efficiency enhancement of a dual-carrier catalytic converter in the gasoline engine

    Energy

    (2020)
  • AdamsR.A. et al.

    Sobolev Spaces

    (2003)
  • AhmedR. et al.

    Automotive internal-combustion-engine fault detection and classification using artificial neural network techniques

    IEEE Trans. Veh. Technol.

    (2015)
  • AntonyJ.

    Design of Experiments for Engineers and Scientists

    (2014)
  • BorettiA.

    Advantages and disadvantages of diesel single and dual-fuel engines

    Front. Mech. Eng.

    (2019)
  • BreimanL.

    Random forests

    Mach. Learn.

    (2001)
  • BrzozowskaL. et al.

    An application of artificial neural network to diesel engine modelling

  • CaiC. et al.

    A novel approach for marine diesel engine fault diagnosis

    Cluster Comput.

    (2017)
  • CATC.

    DICARE - system for diesel engine diagnostic and predicitive maintenance

    (2019)
  • Christen, C., Brand, D., 2013. IMO tier 3: gas and dual fuel engines as a clean and efficient solution. In: Conseil...
  • Fernández-DelgadoM. et al.

    Do we need hundreds of classifiers to solve real world classification problems?

    J. Mach. Learn. Res.

    (2014)
  • GT-SUITE Manual

    (2019)
  • GoodfellowI. et al.

    Deep Learning

    (2016)
  • Gratsos, G.A., Psaraftis, H.N., Zachariadis, P., 2009. Life cycle cost of maintaining the effectiveness of a ship’s...
  • Grimmelius, H., Mesbahi, E., Schulten, P., Stapersma, D., 2007. The use of diesel engine simulation models in ship...
  • HansonR.K. et al.

    Survey of rate constants in the n/h/o system

  • HeY. et al.

    Application of artificial neural networks in engine modelling

    Int. J. Engine Res.

    (2004)
  • Herceg, M., Raff, T., Findeisen, R., Allgowe, F., 2006. Nonlinear model predictive control of a turbocharged diesel...
  • Cited by (26)

    • A digital twin-based decision support approach for AGV scheduling

      2024, Engineering Applications of Artificial Intelligence
    • Hierarchical level fault detection and diagnosis of ship engine systems

      2023, Expert Systems with Applications
      Citation Excerpt :

      It operates in harsh environments, and it is not easy to collect fault data and labels under maritime operations, considering the complicated interests among sailors, ship owners, and manufacturers. Especially labeling DF engine data requires skilled experts to determine whether data are normal or abnormal, and its tasks are costly and tedious (Coraddu et al., 2021). Due to the lack of data and labels for various engine conditions, most studies on monitoring and fault detection and diagnosis of marine engines have applied machine/deep learning technology to simulated data, lab-scaled data, or limited operating condition data to predict ship performance and defects.

    View all citing articles on Scopus
    View full text