The present invention relates to a system for a ship, comprising a plurality of sensors measuring a plurality of data sets including control variables being controllable by the operator of the ship, which data sets each defines a state of the ship at specific sea conditions.

It is well known from US 2004/0193338 that properly adjustment of control variables of a ship such as trim and engine load, may reduce the fuel consumption of the ship significantly. However, it is necessary to adjust such settings on a continuous basis by observing or detecting subtle changes in engine load and attitudes of the ship caused by ever-changing weather and hydro-graphic conditions in order to keep an optimized fuel consumption. Hence, information about the state of the ship at sea is important to minimize the fuel consumption. Container ships and ferries are large ships where such settings are typically changed on an hourly or daily basis rather than by adjustments within minutes and seconds, which for example is the case for small vessels with outboard engines. Based on experience, a captain of a large ship has an idea about the optimum settings of trim and engine power under specific sea conditions, but it is often the case that two captains having experience with the same ship possess very differing opinions as to the optimum settings of the ship in a given situation. Furthermore, humans are not able to capture small changes as for example changes in the weather and sea conditions that affect the optimum set- tings of a ship. Thus even very experienced captains will have very high- variance and non-consistent opinions as to the optimum settings of a ship.

In the field of ship design there is a tradition for using the mathematical equations of hydrodynamics in order to establish a model of the motion of ships in water. However, the equations of hydrodynamics are rather complex and do not have a general solution, so they are only of use in computational hydrodynamics based on numerical approaches or of use under special operating conditions where the equations may be simplified, which implies that the accuracy of the model generally is deteriorated. In both cases the use of the models is cumbersome for many practical purposes due to required processing, and each distinct ship design must be modelled individually. Furthermore, hydro- dynamic equations are solved for a specific sea state and a specific load- ing of the ship, and the theoretical calculations so seldom fit with the actual sea state and the actual conditions of the ship. Therefore, use of hy- drodynamic modelling during actual operation of the ship in order to determine the optimum setting in terms of fuel consumption is impractical because the result is inaccurate and too difficult to obtain dynamically during operation.

It is well know that a ship moving through water sets up a bow wave emanating from the bow of the ship and which creates the majority of the propulsion resistance of the ship because the bow wave carries energy away from the ship at the expense of kinetic energy. A major goal of hull design is therefore to reduce the size of the bow wave in order to improve the ship's fuel economy.

WO 2007/017908 discloses a method for optimizing the onboard consumption of energy on ships by creating a computer model based on hydrodynamic models in order to optimize fuel consumption. The computer model is based on equations describing core components and the structural design of the ship. The computer model of the ship optimizes the operational cost of the ship during operation by receiving signals from a network of sensors and simulating the operation according to the signals received from the sensors. US 2004/0193338 discloses an apparatus for small vessels with outboard engines for selecting the optimum trim of a vessel. The apparatus includes a measuring device for measuring trim and fuel consumption, a generator for generating a statistical model on data received from the measuring device and a trim angle controller for directly selecting a optimum trim of the vessel based on the statistical model. The statistical model is based on data collected for short time periods during the current operation of the ship and a new statistical model based on new measurements each time the vessel is operated.

From US2007/0284475 it is known to record and pre-program a preferred or optimum setting of a ship in the memory of a control device in order to reload the stored settings for later use. The factors used to define the preferred settings for obtaining the optimum or desired performance include engine load, hull inclination, sea conditions, wind ve- locity and wind direction.

US 4,872,118 discloses a system including sensors mounted fore and aft of the ship for monitoring the value of the draft of the ship with the object of monitoring the stability of the ship in order to load the ship with cargo up to the limit determined by the required minimum sta- bility.

In light of the above it is the object of the present invention to provide a system for minimizing the bow wave of a ship during operation at actual sea conditions.

This is achieved by a system according to the opening para- graph wherein said data sets comprise a plurality of system variables affecting the state of the ship, the system employs said measured data sets for generating a statistical model of the measured states of the ship, and wherein a optimum setting of the control variables providing the highest fuel efficiency for a current state of the ship is determined by op- timizing the statistical model with respect to the control variables.

The size of the bow wave is a function of the system variables, which comprises factors such as the ship's velocity trough water, the draft and sea and weather conditions and it is therefore affected by adjusting the control variables of the ship such as trim and engine load. Hence, the optimum settings of the control variables affect the state of the ship at sea and hence minimize the bow wave and the fuel consumption. By modelling the fuel efficiency and optimizing with respect to the control variables, which are controllable by the operator of the ship, the bow wave may be minimized by adjusting the ship to the optimum control variables calculated by the system. This provides a dynamical optimization and measurement of the fuel efficiency independently of the design of the ship. Hence, the system is relatively easy to adapt to different ships and the system may learn the particulars of the ship automatically during actual operation of the ship. Moreover, having learned the particulars of a given ship and having generated a statistical model of the ship for operation under actual sea conditions, the models may be fitted manually in order to take into account sudden changes of the ship such as sudden changes of fouling of the hull after being in dock. The ef- feet of fouling may automatically be included in the modelling by including cleaning of the hull as an input parameter to the model. Additionally, the system may also be used to obtain a better utilization of the actual propulsion since the attitude of the propeller with respect to the sea as well as the speed and/or trim of the ship with respect to the current sea condition and the general pitch of the ship, may be optimized by the system.

In a very exact model of a ship as system, said statistical model comprises a first component, which is a model of changes per time unit in non-observable system variables derived by the system, said changes being a function of the current state of said non-observable system variables and said control variables, and a second component which is a statistical regression model wherein said system variables are a function of said non-observable system variables and said control variables. This approach requires a relatively large amount of calculations and process- ing but provides a very precise model of the ship as whole system.

In a practical embodiment of the invention, said statistical model based on the plurality of data sets is a statistical regression model of the fuel efficiency of the ship, which statistical regression model is based on a statistical regression model of the velocity of the ship and a statistical regression model of the fuel consumption of the ship. This is a simplified model, which provides for a simpler implementation of the system, however this approach may show to be sufficient in many practical cases.

Preferably at least one of said plurality of sensors measures the fluctuations of the waves around the ship. This makes it possible to take into account changes in the pattern of the waves and the actual sea state around the ship, which significantly affects the state of the ship and the optimum control variables, and hence the modelling of the ship is improved, which may result in a more precise calculation of the optimum settings of the ship, which again eventually may affect the size of the bow wave and the overall fuel consumption. Having a sensor on each side of the ship makes it possible to measure the blocking effect the ship poses to sea currents, which makes the modelling even more precise. If the output of the sensor measuring the fluctuations of the waves around the ship further comprises the draft of the ship, it is possible to calculate and take into account the variance of the draft of the ship. The draft and the fluctuations of the waves may also be measured by employing pressure sensors mounted on the hull of the ship. By employing a directional sensor measuring the draft of the ship and the fluctuations of the waves around the ship, the amount of data to be processed in the system is limited compared to a alternative wherein the entire wave pattern around the ship is measured by a radar, which typically already may be installed on new ships. In a preferred embodiment, the sensor used to measure the draft and fluctuations of the waves around the ship is a radar or a sonar, which is mountable for instance on the gunwale on the side of the ship and being able to measure directionally towards the sea close to the ship. In such a preferred embodiment, said sensor employs a sampling frequency between 4 and 20 Hz, which has shown to be adequate to obtain a sufficient sampling of the changes in the waves around the ship.

In a practical embodiment, said control variable being controlla- ble by the operator of the ship comprises a trim angle of the ship and a measure of the engine load of the ship.

A simple and practical way to integrate the system according to the invention on new ships is to display the optimum trim and engine load for the current state of the ship as an advice to the operator of the ship. This is a simple way of utilizing the calculated optimum settings and it makes it easy to adjust the settings of the ship. Alternatively, the system may be used for directly controlling the settings of the ship. Directly controlling all the control variables or only some and keeping the rest as an advice to the operator of the ship depending on the individual particulars of the ship and the personnel operating the ship is an evident possibility.

In practical embodiments, the optimum engine load is stated in terms of engine RPM or propeller pitch. However, in general the engine load may be stated in many different ways depending on the given ship and hence engine load may be stated in many different ways such as a percentage.

In order to take into account the changes in the state of the ship with appropriate intervals and keep the energy consuming adjustments of the control variables at a minimum, the optimum trim and engine load for the current state of the ship is updated every 1 to 15 minutes based on measurements of recent changes in the state based on a time window of a corresponding length.

In a preferred embodiment said system variables used to learn the parameters of the statistical model and finding the optimum setting of the control variables in order to minimize the fuel conspumtion, comprises the following factors: mean wind speed, mean response of a wind direction detection, mean rudder angle, standard deviation of rudder angle, standard deviation of trim angle, mean draft, standard deviation of draft, mean trim angle, mean engine load, velocity of the ship through water, fuel consumption. By using these basic factors, experiments have show that a very exact modelling of the ship is obtained. By including mean engine load, which may be obtained from propeller pitch or engine rpm, a relation between engine load and speed is obtained. Evidently, many other factors such as detection of slamming and calorific value and density of the fuel, which also affect the state of the ship, may be included in the data sets.

In an especially useful embodiment, the optimum trim and engine load for the current state of the ship is determined by using nonlin- ear constrained optimization in order to the setting that optimize the fuel efficiency of the ship, while ensuring that the resulting ship speed is at least a user-specified minimum speed. This makes it possible to optimize the control variables of the ship with respect to special practical requirements such as a minimum speed. It is possible to generate the parameters of said statistical regression models by applying machine learning to said data sets. This may for example be achieved by employing an artificial neural network or a general linear model. Machine learning relates to design and devel- opment of algorithms and techniques that allow computers to learn the behaviour of the surrounding world by extracting rules and patterns from massive data sets.

In the following, the invention will be described in further detail by means of examples of embodiments with reference to the schematic drawings, in which

Fig. 1 is a schematic model of the system according to the invention,

Fig. 2 illustrates the terminology of a ship for describing the state of a ship at sea, Fig. 3 shows the relation between trim and propulsion resistance for a ship,

Fig. 4 shows the relation between speed and fuel efficiency of a ship,

Fig. 5 is a state machine describing the principle of a user inter- face according to the invention,

Fig. 6 is an example of a user interface according to the invention.

Fig. 7 is a table showing an example of a measured data set during operation. In general a ship may be considered as a dynamic system where changes in certain factors affecting the ship take place continuously. The factors may be considered as system variables and they define a state of the system and may be denoted as X(t), which is a vector comprising several elements X(t,l), X(t,2)...X(t,N) each representing the system variables such as draft, wind speed and the motion of the ship at a specific time defining a state of the ship.

Some of the system variables are directly observable and some such as the state of the sea are non-observable. The observable system variables such as relative wind speed, speed of the ship through water and fuel consumption may be referred to as a vector function Y(t) and the non-observable system variables may be referred to as a vector function S(t). Hence, at any time in time there is a vector of observable and non-observable system variables. Changes in X(t) per unit of time depend on the current state of the system as well as certain control variables of the system. The control variables of the system comprises factors such as trim, propeller rpm and draft and may also be denoted as a vector function U(t). This relationship may in general be formulated as a statistical differential equa- tion for the non-observable system variable dS/dt = f(S(t), U(t);Pf) + noise, (1) where f is an unknown function defined by parameters Pf and the noise is a random process such as white noise or Gaussian noise also defined by parameters. The observable system variables are functions of the non- observable system variable

Y(t) = g(S(t), U(t);Pg) + noise. (2)

For example the speed of the ship through water, which is an element in Y(t) depends on the non-observable shape of the ship con- tained in S(t) as well as the propeller rpm contained in U(t). In order to utilize these relations to minimize the bow wave and hence the fuel efficiency of the ship, it is necessary to perform three steps in the system: collection of data, learning and optimization of the system. In general the step of collecting data collects samples of the observable system variables Y(t) and the control variables U(t).

The learning process applies machine learning algorithms such as Simulated Annealing and Unscented Kalman filters to the collected data in order to derive the unknown parameters Pf, Pg and the corresponding noise parameters. Optimization is done by collecting recent ob- servations of Y(t) and U(t) and using these to infer the current non- observable variables S(t). Finally, an optimum setting of the control variables U(t) is chosen as the U(t) that will lead to the highest fuel efficiency in the near future, by taking possible constraints such as minimum speed of the ship into account. This exploits that for each U(t) it is possible to predict what will happen by using the quations (1) and (2). Since fuel efficiency is the ships speed divided by fuel consumption, this can be calculated directly from Y(t), since this comprises both speed and fuel consumption. Any non-linear constrained algorithm such as exhaus- tive search could be used for this.

The above-mentioned model may be denoted as a state space model, which is a description of discrete states used as a model of a system. However, a considerably more simple case of the general model is described in more detail below with reference to a preferred embodi- ment. First it is assumed that certain observable factors are very close to their non-observable counterparts such that the elements in S(t) may be considers known. For example, the observed draft is very close to the true draft. Second it is assumed that no changes will occur in S for the near future say about 5 minutes. Then the differential equation (1) may be disregarded and S(t) assumed constant for a time interval. With these two simplifying assumptions, as regression model is obtained

Y(t) = g(S(t), U(t); Pg) + noise, where Y(t) only contains the speed of the ship trough water denoted V(t) and the fuel consumption denoted F(t), while S(t) is the observed factors such as draft, wind speed, wind direction and so on. U(t) still contains the control variables of the ship such as trim and propulsion rpm. The parameters Pg and the noise parameters may be learned from observations of Y(t), S(t) and U(t) collected over several time periods. In this embodiment the optimization requires no prediction as the effect on Y(t) is obtained directly from a statistical regression model.

Fig. 1 shows a system 1 according to the invention. The system 1 is adapted to be installed on larger ships such as container ships and ferries, where proper adjustment of control variables may have a significant influence on size of the bow wave and hence the fuel consumption and where the adjustment of the control variables is quite often carried out on an hourly or daily basis. The control variables used for minimizing the bow wave of such ships depend on the individual ship design and typically include trim, draft and engine load in terms of engine rpm and propeller pitch. The system 1 of Fig. 1 comprises a plurality of sensors 2, which during operation of the ship, measure a plurality of factors that affect and define a state of the ship. The factors comprise both non- controllable external factors such as weather and sea conditions, and control variables such as rudder angle, trim and engine load, which are controllable by the operator of the ship and these factors are considered as system variables.

The system 1 is based on data driven modelling, i.e. modelling of the ship at sea conditions based on empirical data by means of machine learning methods. Once the sensors 2 have collected enough data, an actual learning of the system 1 can take place. The learning basically consists of learning parameters of certain statistical mathematical models that relates the states and the settings of the ship.

For measuring factors such as wind speed, rudder angle, propeller pitch, and engine rpm, the system 1 typically employs sensors 2 al- ready installed on most ships. Other sensors 2 such as an electronic inclinometer and a microwave radar for measuring draft and the fluctuation of the waves travelling around the ship are added when installing the system 1. In Fig. 1 each sensor 2 is connected to a client 3 so that the system 1 comprises a numbers of subsystems running in parallel. Raw data measured by the sensors 2 are pre-processed in the clients 3 in order to provide a central server 4 with data that is prepared for learning and modelling. The pre-processing in relation to each sensor 2 depend on the measured data and the format of the raw data. For some sensors 2, the raw data are simply provided with a time stamp and for others the measured data may be used to calculate a derived data type required in the central server 4.

The communication between the sensors 2 and the clients 3 is done individually according to the protocols implemented in each sensor 2. When the required pre-processing of the data received from the sen- sors 2 has been carried out, the data are forwarded as a data message to the central server 4. The pre-processing of the data may further comprise error detection and corresponding correction and handling of any detected errors before the data is forwarded to the central server 4. The central server 4 is a central hub for the system 1 and handles data communication between the individual components of the system. Data messages handled in the central server 4 comprise a name, a value and a time stamp and are stored in a database 5. The central server 4 and an advice generator 7 further process the captured data in a manner to be described in more detail below.

The system 1 further comprises a surveillance system 6 keeping track of which processes and programs are responsible for the data in the central server 4. The processes and programs may be one of the clients 3, the advice generator or any other program that sends data mes- sages to the central server 4. By monitoring the age of the data messages in the central server 4, the surveillance system 6 can determine if the programs and processes are working correctly. If the age of a specific data message handled in the central server 4 exceeds a predefined limit, the surveillance system 6 may restart the process or program re- sponsible for the specific type of data message.

The system 1 has a user interface 8 for presenting the optimum setting of the control variables generated by the advice generator 7 to the operating personnel on the bridge of the ship. The relatively simple principle of the user interface is to be described later in more details with relation to Figs. 5 and 6.

The communication between the sensors 2 and the clients 3 and the subsequent pre-processing performed in the clients 3 before sending data messages to the central server 4, depends in particular on the communication interface and the input from the individual sensors 2. In a practical embodiment, the measurements from two microwave radars are received through a serial RS485 interface and the number of measurements received at the corresponding clients 3 may vary from less than 1 to about 70 measurements per second. The pre-processing is a down sampling to reduce the rate at which measurements are sent to the central server 4. Down sampling may for instance be carried out by dividing a second into N equally long intervals, where N in a preferred embodiment is 3. At the end of each interval the median of all measurements in the interval is sent to the central server 4 and the original measurements in the interval are discarded. In general, analogue values may be pulled once every second from the individual sensors 2 by the corresponding clients 3. This is for instance the case in a practical embodiment wherein the measured values of the rudder angle and the trim are read as voltages in the range of -10 V to 10 V, and the fuel consumption is read in a voltage range from 0.25 V to 2.5 V.

In a preferred embodiment, at least a part of the system 1 is based on the NMEA 0183 specification for communication between marine electronic devices already installed on the ship. However, parts of the system 1 may, depending on the actual implementation, be based on different types of networks including standard wireless local area networks.

The ship's draft is defined as the distance from the keel to the waterline WL, i.e. the vertical line that is placed halfway between the aft and forward perpendiculars as shown in Fig. 2. The draft is measured by employing two microwave radars placed on either side of the ship so that any changes in heeling angle, i.e. the angle of rotation around the longitudinal axis of the ship, affect each radar equally but with opposite signs and thereby false draft measurements from changes in heeling an- gles are cancelled out. However, changes in the pitching angle, i.e. the angle of rotation around an axis through the centre of the ship perpendicular to the longitudinal axis and parallel to the water, may lead to false draft measurements. Therefore, this measurement is corrected by the trim angle, which is identical to the pitching angle. It would also be possible to obtain a similar correction by providing a third microwave radar at a different point along the length of the ship, but this would be more cumbersome and expensive than the preferred solution with two radars and a inclinometer, which is in all cases required to measure the trim angle. The raw data obtained from the two radars may be used in many conceivable ways to obtain information about the variance and fluctuations of the waves around the ship. However, the system does not necessarily require a radar or radars to obtain raw data sets about the entire wave pattern around the ship. Experiments have shown that a sufficient measure of the draft and the variance of the waves around the ship may be obtained by using a radar or radars only measuring a limited area of the waves close to the ship. Such sensors for measuring the fluctuations of the waves and the draft are typically mounted on the gunwale of the ship, but in general the position depends on the design of the individual ship. A sensor operating as a relative level measurement unit may be able to obtain such measurements. Such sensors are operated by transmitting an ultrasonic or electromagnetic pulse from the sensor to the surface being monitored, from which surface the pulse is reflected back to the sensor. The output of the sensor is typically achieved by dividing the time of flight of the pulse with two, and correct with the temperature and convert the signal to a raw data measure corresponding to the measured distance. A practical embodiment of the invention uses an ABM300-XXXR microwave radar from ABM Sensor Technology Inc. sold under the name "Smart" Radar Explosion Proof Meas- urement Sensors.

A key factor affecting the state of the ship is the propulsion resistance, which is the resistance from the sea and the air that the ship is moving through to the ship's forward motion. Hence, the propulsion resistance is highly dependent on factors such as wind speed and waves in the sea. One of the sensors of the system 1 is a special wind direction detector adapted to measure the wind direction as a suitable measure. The wind direction detector comprises a number of cells (typically 4 or more), where each cell responds maximally to wind from a certain direction and to wind from other direction. Furthermore, the sea current is measured indirectly by measuring the difference between the ship heading and the GPS heading. The system 2 also calculates other derived factors based on sensor readings. The variance of the rudder angle is used as an indication of the sea state including the sea current and the variance of the draft signals measured by the two radars, is used as an indi- cation of the roughness of the sea and the ship motion. The variance of the trim angle is used as an indication of the current sea state and ship motion.

With respect to the radars, the pre-processing takes into account the effects of trim, the angle of the mounting of the radar with re- spect to the ship and the position of the radar relative to the midship post and the keel of the ship. The true trim angle of the ship and the trim angle of the inclinometer is measured during installation when the ship is stationary at port so that the latter can be calibrated. Before the actual learning process of the system 1 other factors may be derived and calculated from the raw data measured by the sensors 2 and added to the data sets. For instance the standard deviation of the trim angle is calculated to capture the motion of the ship, the standard deviation of the radar measurements is calculated to capture the power in the waves passing close to the ship. Further, the standard deviation of the angle of the rudder or rudders of the ship is calculated to capture the amount of work done by the rudder, which is an important contributing factor to the propulsion resistance of the ship. This factor is necessary as a complement to the mean of the rudder angle since the latter may be well near zero taken over a period of for instance 3 minutes, even when the rudder due to surface currents is working hard to maintain the course of the ship and therefore the rudder angle mean may be misleading if used alone.

Variables defining the state of the ship having the system 1 in- stalled are sampled in discrete time with a sampling frequency varying from 1 Hz for most signals and preferably about 6 Hz for the microwave radar, which empirically has shown to be sufficient to measure the fluctuation and the variation of the waves travelling close to the ship. The speed and fuel consumption of the ship do not depend directly on instan- taneous fluctuations in the factors defining the state of the ship, but rather on long-term quantities. Therefore, the samples are divided into non-overlapping time windows of for instance 3 to 5 minutes within which mean and variance of each variable are used as appropriate to express the current state of the ship. The table of Fig. 7 illustrates the samples taken during an actual operation of a system according to the invention. In this embodiment, 180 samples taken over a period of 3 minutes make up a time window, which may be used for learning and/or analysis/optimization with respect to the state measured over the time window. The time window could be different, say 15 minutes, but not too much more, since then the optimization would then not take into account sudden changes in ship states.

Techniques known in the art of machine learning such as whitening and de-correlation of the data dimensions, which is standard tech- niques in the machine learning field, is applied before the actual learning process of the system 1. This is done using singular value decomposition and corresponds to a reorientation of the coordinate system so that data in the multidimensional state variables is uncorrelated with all other variables and has unit variance. This is done to facilitate the learning process i.e. the machine learning of the system.

As described above the basic factors to define the state of a ship comprise the mean wind speed, mean response from a wind direction sensor, mean rudder angle, standard deviation of draft measured by a first microwave radar, standard deviation of draft measured by a second microwave radar, standard deviation of rudder angle, standard deviation of trim angle, mean draft, means trim angle, mean engine load, velocity of the ship through water and fuel consumption. The details of such variables may vary with the particulars for the individual ship such as the number of rudders. However, other and more complex state vari- ables such as mean difference between GPS and ships heading, the ship ^' s velocity over ground measured by a GPS, detection of main frequency in the measured radar wave spectra and other derived state variables, may be employed in more advanced embodiments. However, due to sea currents it is normally preferred to used the ship's velocity through water as a basic measurement for the modelling of the ship.

In a preferred embodiment, the density and calorific value of the fuel is taken into account since these factors may vary from bunkering to bunkering. The makes it possible to optimize the setting of the ship more precisely with respect to metre per energy unit and instead of the actual fuel consumption measured in litre per second. Furthermore, additional specific sensors installed on the ship may measure various other factors. Another conceivable measure is detection of wave slamming, which may be detected by sensors positioned on the hull. Such measurements may advantageously be utilized in the further processing of the collected data and subsequently in the modelling and optimization done by the system 1.

The purpose of dynamically modelling the ship under actual sea condition is to minimize the bow wave in order to optimize the fuel effi- ciency. To this end it is appreciated that the optimum setting of the control variables for the current state of the ship may minimize the bow wave and hence optimize the fuel efficiency. The fuel efficiency E of a ship may be measured in metre per litre and may be calculated by measuring the speed through water denoted V measured in metre per second and the fuel consumption, F, measured in litre per second. The fuel efficiency E of a ship in operation depends on many factors but only a few of these are as mentioned above controllable by the captain of the ship. The control variables being controllable by the operator of the ship typically include draft, trim and engine load of the ship and may be re- ferred to jointly as a variable L) comprising the control variables of the ship. The engine load of a ship may in general terms be controlled through either the propeller pitch and/or engine rpm depending on the individual ship. However, this distinction is not important in relation to a general description, but when the system 1 according to the invention is adapted and implemented on a particular ship, pitch and/or rpm is used as appropriate.

Having the system installed on a ship the first step is to collect data from the sensors 2 of the system 1. In order to obtain a solid basis of data for modelling the ship, data is typically collected for a relatively long time period, i.e. in the order of months, which is necessary in order to obtain a proper statistical regression model of the ship.

Since the fuel efficiency E depends on both the state of the ship S and the controllable factors L), a change in S may change the relation between E and L) as illustrated in Figs. 3 and 4. Fig. 3 shows that the re- lation between trim and propulsion resistance depends on the current state of the ship, and Fig. 4 shows that the same principle holds for the relation between speed V and fuel efficiency E.

In order to apply machine learning, the system 1 measures among others fuel efficiency E, controllable factors L), speed V and fuel consumption F over time. The measured data is applied to machine learning methods in order to obtain a statistical mathematical model that has learned the relations between the states and the settings of the ship as accurately as possible. In a preferred embodiment of the invention, the system 1 models the speed V of the system as a function of the current state S of the ship and the controllable factors L). In another model, the system 1 models the fuel consumption F of the ship as a function of the current state S and the controllable variables L). Parameters used in these statistical regression models are derived from the collected data, and it is important to collect enough data so that the relations may be modelled correctly. The models are cases of statistical regression and any suitable model, such as a general linear model or a neural network, may be used to model the relations and obtain the parameters of the statistical models. In general the approach according to the invention is to determine the posterior distribution of the parameters of the model, conditioned on the observations, and to employ this posterior distribution to predict the result on speed and fuel consumption of any setting of control variables. However, if enough data is available, a good approximation is to choose the parameters that maximize the posterior, which is proportional to the product of the model likelihood and the prior distribution of the parameters. The prior may be chosen in a variety of fashions. In a preferred embodiment, the prior is expressed through regu- larization, which in the case of a neural network may be done using the technique of weight decay. By applying a suitable regularization method, it is ensured that the models are not too flexible, which would make them perform badly when used on new data.

The system 1 employs an artificial neural network where regularization is done through weight decay and where the optimal value for the weight decay parameter is found through iteration and evaluation on a separate validation data set. For the purpose of learning the parameters for the two models with any given weight decay parameter, conjugate gradient descent with line search is used.

Once the statistical regression models of speed V and fuel consumption F have been trained enough to model the ship during opera- tion, the optimum settings may be predicted for the current state of the ship by optimizing the relation E = V/F with respect to the settings of L) and the recent measurements defining the current state of the ship. The optimizing algorithm employed by the system for this purpose may be any suitable, non-linear constrained optimizer and it is given the trained statistical regression models, e.g. neural networks, as the function to optimize with respects to the settings L) of the particular ship.

Any suitable algorithm for solving constrained non-linear optimization may be used and since the preferred embodiment only employs two or three dimensions of settings L), the system 1 employs exhaustive search evaluating every point in a fine grid spanning the space of possible value of L).

Choosing optimum values of trim and engine load for the current state of the ship maximizes the fuel efficiency E. The constraints are typically that the speed must exceed a minimum value given for the particular ship and trim and power must be within physically possible values.

Draft may be optimized similarly to trim as described above. However, draft is more constrained by cargo and bunker load so that draft may not be changed as freely as trim and power. For some ships, draft can be controlled by bunkering of fuel oil and the frequency at which bunkering is done.

The advice of the system is presented on a graphical user interface at the bridge of the ship. The information displayed on the graphical user interface is based on a finite state machine as shown in Fig. 5. Each state of the finite state machine corresponds to one particular display image of the graphical user interface that informs the captain on the bridge as to a proper action to take. Fig. 6 shows an embodiment of a design of the graphical user interface in a practical embodiment. The system works by calculating a new state variable S and the corresponding controllable setting L) every 3-5 minutes and the user interface then gives advice to the captain by means of some simple actions such as increase power if it is too low, decrease power if it is too high, increase trim, if it is too low and decrease trim, if it is too high. Threshold values are used to make sure that the system does not require the captain to adjust trim and power too often.

Hence, the advantage of the system is that dynamical measurements of factors affecting the state of the ship is used to provide a statistical model for dynamically optimizing the control variables of the ship during operation and thereby obtaining the most energy efficient setting of the ship.

The invention should not be regarded as being limited to the described embodiments. Several modifications and combinations of the dif- ferent embodiments will be apparent to the person skilled in the art.