By Roberto Venturini - - PowerPoint PPT Presentation

1 By Roberto Venturini - https://www.flickr.com/photos/robven/1953413479, CC BY 2.0, https://commons.wikimedia.org/w/index.php?curid=57831577

Gerrit-Jan de Bruin Supervision by Jasper van Vliet M.Sc. en dr. Johan Westerhuis Efficient compliance monitoring: Comparison of both airborne and landside sniffing and spectrometric methods to provide direct control on the sulfur emission of ships.

Contents  Introduction  Aim  Analytical techniques  Statistical techniques  Classification with linear boundary  Classification using Z-score  EM algorithm  Outlook Efficient compliance monitoring 3

The emission of SO 𝟑 over time. SO ₂ emissions in the Netherlands 250 Total 60 000 premature deaths, Corbett Transport 200 Shipping 2 year loss, CAFE 150 SO ₂ (kton) 100 50 0 1990 1995 2000 2005 2010 2013 2014 2015 Efficient compliance monitoring 4 Introduction

Maximum allowed FSC 5 Within SECA Global 4 3 FSC [% (m/m)] 2 1 0 01-2010 01-2012 01-2014 01-2016 01-2018 01-2020 $ 40 000 dayˉ¹ Left: Image courtesy of D.J. Oostwoud Wijdenes and National Geographic Society. Efficient compliance monitoring 5 Introduction

Fuel Sulfur Content  FSC = weight of sulphur weight of fuel 16 64.066 ×𝑁 S × ׬ SO 2 − SO 2 bg 𝑒𝑢  FSC = 12 𝑁 C 0.87 ×׬ CO 2 − CO 2 bg 𝑒𝑢 44 × Τ ׬ SO 2 − SO 2 bg 𝑒𝑢  FSC = 0.232 ׬ CO 2 − CO 2 bg 𝑒𝑢 Image courtesy of ILT. Efficient compliance monitoring 6 Introduction

Aim  Compare different techniques and operators for future use for the inspectorate.  Explore the measurements performed so far by all inspectorates in Northern Europe.  What are the compliance rates?  What are the type I and type II errors? I.e. how sure are we that a ship is (non-)compliant? Efficient compliance monitoring 7 Introduction

Image courtesy: ILT Efficient compliance monitoring 8 Introduction

Efficient compliance monitoring 9 Image courtesy: ILT Introduction

TNO/ ILT sniffer Image courtesy: ILT Efficient compliance monitoring 10 Analytical instrument

N = 8049 TNO, 1661 MUMM, 1390 ILT, 743 BSH, 3564 Explicit, 327 DFDS-Maersk, 10 Denmark, 354 Efficient compliance monitoring 15 Campaigns

What fraction is non-compliant? 800 8000 Cumulative count 600 6000 Count 400 4000 200 2000 0 0 < -0.1 0 0.1 0.2 > 0.3 FSC (% m/m) Efficient compliance monitoring 16 Campaigns

What fraction is non-compliant? 800 8000 Cumulative count 600 6000 Count 400 4000 200 2000 0 0 < -0.1 0 0.1 0.2 > 0.3 FSC (% m/m) Efficient compliance monitoring 17 Campaigns

What fraction is non-compliant? Classification N = 19 Accuracy = 47% 7 6 True value 4 2 Efficient compliance monitoring 18 Campaigns

Intermezzo – type I and type II errors Classification N = 19 Accuracy = 47% 7 6 Type 1 True value 4 2 Type 2 Efficient compliance monitoring 19 Intermezzo

Intermezzo – type I and type II errors Classification N = 19 Accuracy = 47% 7 6 Type 1: wrongly accusing True value 4 2 Type 2: overlooking non-compliance Efficient compliance monitoring 20 Intermezzo

Intermezzo – type I and type II errors  What do we want? Low type-I error Equal type-I and High type-I error High type-II error type-II errors Low type-II error Court Efficient compliance monitoring 21 Intermezzo

Intermezzo – type I and type II errors  What do we want? Low type-I error Equal type-I and High type-I error High type-II error type-II errors Low type-II error Court Preselection Efficient compliance monitoring 22 Intermezzo

Intermezzo – type I and type II errors  What do we want? Low type-I error Equal type-I and High type-I error High type-II error type-II errors Low type-II error Court Climate modeling Preselection Efficient compliance monitoring 23 Intermezzo

Intermezzo – type I and type II errors  What do we want? Type I error Type II error Efficient compliance monitoring 24 Intermezzo

What fraction is non-compliant? 800 8000 Cumulative count 600 6000 Count Low type II error Low type I error 400 4000 200 2000 0 0 < -0.1 0 0.1 0.2 > 0.3 FSC (% m/m) Efficient compliance monitoring 25 Campaigns

ҧ Z-score N = 5552 (69%)  𝐼 0 : The ship has a FSC of 0.1 wt. % or less.  𝐼 1 : The ship has a higher FSC than 0.1 wt. %. 𝑦−𝜈 0  𝑨 = 𝑡 𝑦 Τ 𝑜  Z-score can be calculated to p-value with a significance level Efficient compliance monitoring 26 Campaigns

Z-score with 𝜷 = 𝟏. 𝟏𝟔 9 % 91 % Efficient compliance monitoring 27 Campaigns

Z-score with 𝜷 = 𝟏. 𝟏𝟔 Classification N = 19 Accuracy = 68% 11 2 True value 4 2 Efficient compliance monitoring 28 Campaigns

Another approach Efficient compliance monitoring 30 Another approach

What fraction is non-compliant?  How many port state controls should take place?  How reliable are climate modellings assuming 100% compliance?  What is the catch rate? Efficient compliance monitoring 31 Another approach

EM-algorithm Efficient compliance monitoring 32 Another approach

EM algorithm  Guess initial parameters  Calculate responsibility  Maximize likelihood of all parameters Efficient compliance monitoring 33 Another approach

𝛿 𝑗,0 + 𝛿 𝑗,1 = 1 For each datapoint i 𝛿 𝑗,0 = 1 𝛿 𝑗,0 = 0.5 𝛿 𝑗,0 = 0 𝛿 𝑗,1 = 1 𝛿 𝑗,1 = 0 𝛿 𝑗,1 = 0.5 Efficient compliance monitoring 34 Another approach

EM algorithm  Guess initial parameters  Calculate responsibility  Maximize likelihood of all parameters 𝑜 𝑙 𝜈 𝑙 = 1 ෞ ෍ 𝑦 𝑗 𝑜 𝑙 𝑗∈𝑙 𝑜 𝑙 𝜏 𝑙 = 1 𝑦 𝑗 − 𝜈 𝑙 2 ෞ ෍ 𝑜 𝑙 𝑗∈𝑙 Efficient compliance monitoring 35 Another approach

EM algorithm  Guess initial parameters  Calculate responsibility  Maximize likelihood of all parameters 𝑜 𝑙 𝜈 𝑙 = 1 ෞ ෍ 𝑦 𝑗 𝑜 𝑙 𝑗∈𝑙 𝑜 𝑙 𝜏 𝑙 = 1 𝑦 𝑗 − 𝜈 𝑙 2 ෞ ෍ 𝑜 𝑙 𝑗∈𝑙 Efficient compliance monitoring 36 Another approach

EM algorithm  Guess initial parameters  Calculate responsibility  Maximize likelihood of all parameters Iterate until convergence Efficient compliance monitoring 37 Another approach

EM-algorithm 96 % N = 5552 (69%) 𝜈 1 = 0.06 wt−% 𝜏 1 = 0.04 wt−% 𝜈 2 = −1.1 wt−% 𝜏 2 = 0.8 wt−% 4 % Efficient compliance monitoring 38 Another approach

What fraction is non-compliant? 800 8000 Cumulative count 600 6000 Count 400 4000 200 2000 0 0 < -0.1 0 0.1 0.2 > 0.3 FSC (% m/m) Efficient compliance monitoring 39 Another approach

EM algorithm  Guess initial parameters  Calculate responsibility likelihood prior 2 ) 𝜌 𝑙 ฏ 𝒪(𝑦 𝑗 | ෞ 𝜈 𝑙 , ෞ 𝜏 𝑙 𝛿 𝑗,𝑙 = ෞ 2 ) + 𝜌 2 Lognormal(𝑦 𝑗 − 0.1| ෞ 2 ) 𝜌 1 𝒪(𝑦 𝑗 | ෞ 𝜈 1 , ෞ 𝜏 1 𝜈 2 , ෞ 𝜏 2 evidence  Maximize likelihood Efficient compliance monitoring 40 Another approach

𝑜 𝑙 𝑦 𝑗 𝑜 𝑙 σ 𝑗∈𝑑 𝜈 𝑜𝑑 = 1 𝜈 𝑑 = ෞ ෞ ෍ log 𝑦 𝑗 − 0.1 𝑂 𝑑 𝑂 𝑜𝑑 𝑜 𝑙 𝑦 𝑗 − 𝜈 𝑑 2 𝑗∈𝑜𝑑 σ 𝑗∈𝑑 𝑜 𝑙 𝜏 𝑑 = ෞ 𝜏 𝑜𝑑 = 1 log 𝑦 𝑗 − 0.1 − 𝜈 𝑜𝑑 2 𝑂 𝑑 ෞ ෍ 𝑂 𝑜𝑑 𝑗∈𝑜𝑑 Type your footer here 41

Efficient compliance monitoring 42 Another approach

97 % 3 % Efficient compliance monitoring 43 Another approach

Outlook  Determine the relation between type I and type II errors more precisely.  Better instruments will result in better accuracy.  Better validation makes the introduction of supervised methods possible. Efficient compliance monitoring 44 Outlook