Data-driven Vessel Performance Monitoring ErhvervsPhD projekt af Benjamin Pjedsted Pedersen FORCE Technology / DTU Poul Andersen Maritime Engineering, Department of Mechanical Engineering, DTU Jan Larsen Department of Informatics and Mathematical Modeling, DTU Peter Sinding FORCE Technology Skibsteknisk Selskab d. 12. november 2012 1
Outline 1. Fuel/Power Performance Monitoring 2. Traditional performance monitoring 3. New approach 4. Artificial Neural Network model or Gaussian Processes 5. Input data Measured or Noon report data 6. Training of the ANN and GPR 7. Predictions Results 8. Post-processing trend and event detections 9. Conclusion Skibsteknisk Selskab d. 12. november 2012 2
Ship Propulsion Performance - Factors Engine Efficiency Identification of Ship Performance factors Propeller Efficiency Hull Efficiency Wind Fuel Power Thrust Resistance Poor Maintenance Propeller Fouling Waves Shallow Water Draught & Trim Water Temp/ Density Hull Fouling Skibsteknisk Selskab d. 12. november 2012 3
Ship Propulsion Performance Expressions of propulsion power/fuel performance 1. Two identical conditions compared at two different times δec = 2. Comparison of measured and theoretical power/fuel consumption EC K EC = EC EC EC EC t 0 observed theoretical 0 (1) (2) Fuel Consumption % 160 150 140 130 120 110 100 90 80 Fuel Consumption due to Fouling of Hull & Propeller Hull Cleaning Docking 17-11-2003 17-01-2004 18-03-2004 18-05-2004 18-07-2004 17-09-2004 17-11-2004 17-01-2005 Skibsteknisk Selskab d. 12. november 2012 4
Ship Propulsion Performance Expressions of propulsion performance comparing the predicted energy consumption E C with the observed or measured energi consumption EC relative prediction errors ω (1) Skibsteknisk Selskab d. 12. november 2012 5
Traditional Performance Evaluation Calculate theoretical propulsion power Empirical hull resistance, propeller and hull efficiency Resistance, self propulsion and open water model test as support Wind resistance estimation Compare theoretical propulsion power with the measured power Skibsteknisk Selskab d. 12. november 2012 6
New approach non-linear prediction Regression models combined with automatic data sampling Regression of ship states (specific draught, speed, wind, power, etc.) Prediction of the power for new states and compare with the actual power Measured samples of 10 min time series Speed through the water Heading Relative wind speed and direction Air temperature GPS cog, sog, etc. Propeller shaft power Propeller revolutions Hindcast data Wind speed Wind direction Sign. Wave height Mean wave period Skibsteknisk Selskab d. 12. november 2012 7
Artificial Neural Network (ANN) Multivariate non-linear regression y z M ( x) = j= 0 d = g w i= 0 w (1) j x i ji (2) z kj kj (9) (10) Where: x is the array of the measured input data. y is the output, in this case the propulsion power w (1) weights of the first hidden layer g nonlinear activation function tangent hyperbolic The optimum number of hidden units and/or layers have to be determined by trial and error. Skibsteknisk Selskab d. 12. november 2012 8
Gaussian Process Regression (GPR) Multivariate Gaussian distribution over functions Non-parametric model all data have to used for every regression f ( x) GP( m( x), k( x, x') ) y = N( µ, Σ) or (9) Where: x is the array of the measured input data. y is the output, in this case the propulsion power m(x) or μ is the mean function E[f(x))] k(x,x ) or Σ is the covariance function E[(f(x) m(x))((f(x )-m(x ))] The covariance function (Squared Exponential) is optimized by a Automated Relevance Method (ARD), where the length-scale is the controlling parameter for each input variable. GPR is a fast method for smaller data sets but the complexity increases with the number of inputs O(n 3 ) and is not appropriate for very larger data sets. GPR indicates the relevance of the each of the inputs variables as length-scale the smaller the more relevant GPR also finds the predictive variances Skibsteknisk Selskab d. 12. november 2012 9
Input data 1. Measured data from Torm Marie two month four sperate data sets 2. Noon report data from Torm Marie two years 3. Noon report data from five containerships for a period of up to 10 years Skibsteknisk Selskab d. 12. november 2012 10
Input data from Torm Marie 1. Input based on measured data, with the measured propeller power used as output. Hindcast has been evaluated for every 10 min interval. 2. Input based on noon reports, with specific fuel oil consumption (ton/hr) used as output. Data Data source M NR Speed through water Measured onboard X Relative wind velocity Measured onboard X Relative wind direction Measured onboard X Air temperature Measured onboard X Propulsion power Measured onboard X Logged mean speed Noon report X Sea water temperature Noon report X X Air temperature Noon report X X Arrival draught fore Noon report X Arrival draught aft Noon report X Report time, UTC Noon report X True wind speed Hindcast X X True wind direction Hindcast X X Significant wave height Hindcast X X True wave direction Hindcast X X Wave period Hindcast X X Relative wind speed Derived from hindcasts X X Relative wind direction Derived from hindcasts X X Skibsteknisk Selskab d. 12. november 2012 11
Input data Torm Marie measurements Number of Mean Number of Start date Trim Data set valid noon draught, U Samples End date Ta-Tf min -U max P min -P max reports T m M N [m] [m] [knots] [kw] 1 236 2 109 3 301 4 555 09-02-2008 14-02-2008 22-03-2008 27-03-2008 30-01-2008 06-02-2008 01-03-2008 11-03-2008 3 7.4 2.4 14.2-16.2 7573-11283 4 7.85 2.7 13.6-15.1 7750-9248 7 12.15 0 13.4-16.0 9 13.0 0 13.0-15.9 8138-11216 9741-12096 Skibsteknisk Selskab d. 12. november 2012 12
Input data noon reports Small time density one point about every 24 hour Readings are subject human errors (position N/S E/W) Many values are weather observations (wave height, wind force) Draught and trim (loading computer or departure/arrival readings) The output variable HFO/day tank reading Data available, every ship make noon reports no matter how well it is equipped Noon reports from Torm Marie Date UTC 09-12-2006-05-12-2008 Number of valid samples Mean draught [m] Trim, Ta-Tf [m] Ship speed [knots] Seawater temp [ C] Specific HFO [tons/day] 323 7.35-15.35 0-3.4 9.9-17.5 12-32 1.1-3.9 Skibsteknisk Selskab d. 12. november 2012 13
Container ship data No ship specific information were available only noon report data Ship ID Total period [Years] Total number of noon reports Number of dockings Number of hull cleanings 1 9.25 2337 2 7 2 9 2283 2 2 3 7.75 2268 1 3 4 10.5 2679 2 3 5 10 2564 2 4 Skibsteknisk Selskab d. 12. november 2012 14
Input data hindcasts Wind speed and direction, significant wave height, period and direction, for a given time and place. Restricted information in smaller seas e.g. the Mediterranean, The North Sea For noon report data the hindcasts are an average of noon report period with one hour intervals in order to give more representative value since the speed and fuel consumption is a average value. All informations from NOAA www.noaa.gov Skibsteknisk Selskab d. 12. november 2012 15
Training and cross-validations of ANN and GPR Leave-One-Out for GPR All data except one are trained and tested with the missing one > N trainings 80%/20% for ANN Division into 5 data sets that are trained and tested alternately -> 5 trainings Cross-validation error Skibsteknisk Selskab d. 12. november 2012 16
Variable analysis Evaluation of the most important parameter by different input set combinations and Automatic Relevance Determination ARD Relevant variables: 1. Mean draught 2. Trim 3. Ship speed 4. Relative wind speed 5. Relative wind direction 6. Wave height 7. Relative wave directions 8. Water temperature 9. Air temperature 10. time numeric (Matlab time 1 per day) Skibsteknisk Selskab d. 12. november 2012 17
GPR - Variable relevance Measured data from Torm Marie Skibsteknisk Selskab d. 12. november 2012 18
GPR - Variable relevance Noon report data from Torm Marie Skibsteknisk Selskab d. 12. november 2012 19
GPR - Variable relevance Container ship #1 Input combination ID 17 18 20 21 1 NR.Ulog x x x x 2 NR.Uobs x x 3 NR.Tsw x x x x 4 NR.Tm x x x x 5 NR.Trim x x x x 6 NR.True_wind_speed_m_s x x x x 7 NR.True_relative_wind_direction_deg x x x x 8 HC.Vrel x x 9 HC.gammarel x x 10 HC.mean.Hs x x 11 HC.mean.Tp x x 12 HC.var.Hs x x 13 HC.var.Tp x x 14 HC.var.Td x x 15 HC.var.Ws x x 16 HC.var.gamma x x 17 NR.UTC x x 27 NR.Sea_state_m x x x x 28 NR.True_relative_sea_direction_deg x x x x Skibsteknisk Selskab d. 12. november 2012 20
Prediction errors and variance Torm Marie NR data Container ship #1 Skibsteknisk Selskab d. 12. november 2012 21
Predictions results Comparison of regression models and data sets Skibsteknisk Selskab d. 12. november 2012 22
Results measured dataset comparison Skibsteknisk Selskab d. 12. november 2012 23
Long term trend detection For the container ships only Trend of the relative prediction error ω(t) Linear trend assumed fitted by weighted least square regression n 2 α i ( ωi f ( ti, )) wi = 2 i= 1 σ i E( ) = w α 1 Skibsteknisk Selskab d. 12. november 2012 24
Event detection Prediction of known and unknown events Known: Dockings, hull cleanings Unknown: Propeller damages Skibsteknisk Selskab d. 12. november 2012 25
Conclusion input data based on measurements Prediction of mean power over a 10 minutes period based on measured input data Relative mean error less than 2 % with ANN and GPR method The error is reduce approximately 1% by introducing hindcast as input variables The introduction of hindcast eliminates the need for wind measurements which is a major benefit It is not possible to compare the methods with measured input data and noon report input data due to lack of data of noon reports for in the best case 9 noon reports to one dataset (#4) Because the present input data has a limited range of speed, wind, power and loading conditions, the prediction window is similarly small, but it may be representative of the most common situations. Skibsteknisk Selskab d. 12. november 2012 26
Conclusion input data based on noon reports Prediction of the fuel consumption per hour can be predicted with ~7% accuracy for the Torm Marie data and with ~5% accuracy for the container ship data with GPR ANN with only 5 hidden units is in general the best Time as an variable reduce the error significantly by ~1% GPR and ANN gives equal level of prediction errors Data driven models are dependant on good input data and are thus not able to work from day one. Suggestions for improvement More consistent Noon report speed tests Rating of docking and hull cleanings they are usually not of the same quality Better hind cast data Skibsteknisk Selskab d. 12. november 2012 27
Acknowledgements Many thanks to: The ship-owner Torm The crew on Torm Marie, in particular Chief Engineer Rasmus Hoffman The project has been sponsored by the Danish Industrial PhD program and Danish Centre for Maritime Technology DCMT Skibsteknisk Selskab d. 12. november 2012 28