A ship ice load prediction method based on physical information neural network

Abstract

The present invention relates to the field of ship ice load, and in particular to a ship ice load prediction method based on physical information neural network, which realizes the intelligent prediction of ship ice load. First, the ice load force of a ship sailing in the broken ice area and the number and relative collision speed of ice particles at the current moment are calculated by CFD-DEM coupling, which are used as training and verification data for the network. Then, the momentum conservation equation between the momentum of ice particles and the ship ice load is constructed as a physical constraint and added to the loss function of the network, and finally a fully connected network structure is established. The present invention establishes a neural network data set, and uses the physical equations related to floating ice and ship ice load to constrain the neural network, and combines the neural network with the physical equations, so as to calculate and predict the resistance of the subsequent ice surface during real-time icebreaking; when the present invention is combined with the display unit of the hull, it can be adjusted according to the actual ship speed required for icebreaking.

Description

Technical Field

[0001] The invention relates to the field of ship ice load, and in particular to a ship ice load prediction method based on physical information neural network. Background Art

[0002] Polar navigation is an important research branch of ship and ocean engineering. Currently, researchers have established a certain research foundation on the interaction mechanism between ships and ice through numerical simulation and experiments. Especially in the predictive measurement modeling process, the actual collision calculation between the ice surface and the ship is completed through algorithm calculation, which is generally based on neural network steady and transient laminar flow simulation at low Reynolds number.

[0003] However, there are still deficiencies in accurately simulating the complexity and variability of icebreaking, especially for the ice particle factors in the icebreaking process, as well as the intelligent forecasting and prediction in the icebreaking process. That is, when breaking ice on large plates, the different hardness in the early detection process may cause damage to the hull due to other reasons. Therefore, it is also necessary to make predictions for one second and beyond in the icebreaking process to determine whether the ice at the front of the hull can be broken stably. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a ship ice load prediction method based on physical information neural network.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In a first aspect of the present invention, the present invention provides a ship ice load prediction method based on physical information neural network, comprising the following steps:

[0007] A. Use computational fluid dynamics methods to obtain the data set required for physical information neural network;

[0008] (1) Constructing control model based on CFD-DEM method;

[0009] The continuity equation and Navier-Stokes equation are selected as the basic equations of the fluid phase. The Reynolds average method is used to establish the fluid phase control equation. The shear stress transport model is used as the turbulence model to close the control equation. The fluid volume method is used to capture the free surface waves generated during the ship's navigation.

[0010] The contact model of discrete element particles is the Hertz-Mindlin model, and the nonlinear increase of the contact force is affected by the displacement. The contact between the two particles is the discrete phase basic equation, and the control model is obtained after calculation;

[0011] (2) Establish a physical model of ships and floating ice;

[0012] The hull model uses the Kriso Container Ship model with a scale ratio of 31.6; Icebreaking model: Broken ice is regarded as a composite particle composed of several spherical particles;

[0013] (3) CFD-DEM coupled simulation;

[0014] The longitudinal and transverse lengths of the computational domain were set to 6.0 times the length of the ship, and the vertical height was set to 4.5 times the length of the ship. At the same time, in the boundary condition settings, the inlet, upper boundary, and lower boundary were set as the velocity inlet, the outlet boundary was set as the pressure outlet, and the lateral boundary was set as the symmetry surface. The hull surface was configured to have an anti-slip wall condition. In the simulation calculation, the Euler method was used to fix the ship, and the fluid rushed from the front velocity inlet to the ship model at a set speed. At the same time, an ejector was set at the velocity inlet to spray crushed ice toward the ship model in a particle flow. The computational domain was discretized using a structured hexahedral grid, and local grid refinement was performed in the free surface, ship model, hull Kelvin wave and other areas to ensure the accuracy of the flow field calculation. A six-layer prismatic grid was generated near the hull surface to simulate the boundary layer flow.

[0015] B. Constrain the neural network using the physical equations related to ice floes and ship ice loads;

[0016] Since there is momentum conservation between the ship and the ice, the momentum conservation equation between the ice particle momentum and the ship ice load is constructed as the loss function of the network to physically constrain the network, and the following loss function is constructed:

[0017] Loss total =Loss function +Loss data

[0018] Where: Loss function The loss in is based on the energy conservation equation between the ship and the ice. data is the predicted value of ice load f; the two can be expressed as:

[0019]

[0020]

[0021] Where n is the number of data points in the ship-ice collision dataset, F n is the predicted value of the network; t n is a time series; N represents the total number of samples, and the average of N samples is taken as the loss;

[0022] C. Predict ice loads through time series using physical information neural network methods;

[0023] The ice load force of ships sailing in the broken ice area is calculated by CFD-DEM coupling, and the number of ice particles, relative collision velocity and frequency at the corresponding time are collected; based on the ice load force on the ship, the number of ice particles at the corresponding time, and the relative collision velocity and frequency of ice particles, the physical information neural network method is used to predict the ice load through time series to obtain the time series data of ice load.

[0024] As a preferred technical solution of the present invention, the computational fluid dynamics method in step A is to establish a ship ice collision model using a CFD-DEM method for simulating ship navigation in an ice-breaking area;

[0025] The ship ice collision model includes an ice particle model for ship icebreaking simulation. The ice particles are considered as composite particles composed of spherical particles with a length or width of 20 cm, a thickness of 4 cm, and a density of 910.0 kg / m 3 , Poisson’s ratio is 0.3, Young’s modulus is 7.09 Pa, and the crushed ice enters the flow field through the ejector.

[0026] As a preferred technical solution of the present invention, the data category of the data set in step A includes: the number of ice particles at the current moment, the relative collision speed between the ice particles and the ship at the current moment, the relative collision frequency between the ice particles and the ship at the current moment, and the ship ice load force calculated based on the data category of the data set.

[0027] As a preferred technical solution of the present invention, 70% of the data set in step A is used as a training set, and 30% is used as a validation set.

[0028] As a preferred technical solution of the present invention, the physical equation in step B is a momentum conservation equation between the momentum of ice particles and the ice load of the ship.

[0029] As a preferred technical solution of the present invention, the structure of the physical information neural network in step C is a fully connected neural network, the input layer is a time series, and the output layer is the value of the ice load force in the current time series, the number of ice particle collisions, and the relative speed between the ice particles and the ship.

[0030] In a second aspect of the present invention, the present invention provides a ship ice load prediction system based on a physical information neural network, the system comprising:

[0031] The acquisition module is used to collect the ship status and ice load data in the current state;

[0032] A data module, used to carry PINNs and a ship ice load prediction module, the ship ice load prediction module includes the steps of the method described in steps A to C, and outputs the ship ice load and its predicted value in the current state;

[0033] An image module is used to display the ship ice load and its predicted value obtained by the data module and display them on an image device;

[0034] The exception information unit is used to collect exception information and distribute it to the person in charge after visual processing on the front end.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention establishes a neural network data set, and utilizes physical equations related to floating ice and ship ice load to constrain the neural network, and combines the neural network with the physical equations, thereby calculating and predicting the resistance of the subsequent ice surface during real-time icebreaking. When the present invention is combined with a display unit of the hull, it can complete adjustment according to the actual ship speed required for icebreaking. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0038] Figure 1 The ice particle model of the present invention has a length of 20 cm and a thickness of 4 cm;

[0039] Figure 2 The model and calculation domain diagram of the physical model of the ship and floating ice of the present invention

[0040] Figure 3 It is a structural diagram of the physical information neural network of the present invention;

[0041] Figure 4 is a schematic diagram of ice load in Example 1;

[0042] Figure 5 is a schematic diagram of the average number of ice collisions in Example 1;

[0043] Figure 6 Schematic diagram of ice load image display in Example 1. DETAILED DESCRIPTION

[0044] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0045] Example 1

[0046] like Figure 1-3 As shown, the present invention provides a ship ice load prediction method based on physical information neural network, comprising the following steps:

[0047] A. Use computational fluid dynamics methods to obtain the data set required for physical information neural network;

[0048] (1) Constructing control model based on CFD-DEM method;

[0049] The continuity equation and Navier-Stokes equation are selected as the basic equations of the fluid phase. The Reynolds average method is used to establish the fluid phase control equation. The shear stress transport model is used as the turbulence model to close the control equation. The fluid volume method is used to capture the free surface waves generated during the ship's navigation.

[0050]

[0051] Where: ρ is the fluid density; U i (i=1, 2, 3) are the fluid velocities in the x, y, and z directions, and P represents the fluid pressure.

[0052] The contact model of discrete element particles is the Hertz-Mindlin model, and the nonlinear increase of the contact force is affected by the displacement. The basic equation of the contact between two particles, that is, the discrete phase, is as follows:

[0053] F contact =F s +F t

[0054] F s =-K n d n -N n V n

[0055]

[0056] Where: F s The magnitude of the normal component of the contact force, F t is the magnitude of the tangential component of the contact force, K n is the normal spring stiffness, d n is the normal overlap distance between particles, N n is the normal damping, V n is the relative normal velocity between particles, C fs is the static friction coefficient, d t is the tangential overlap distance between particles, V t is the relative tangential velocity between particles.

[0057]

[0058]

[0059] Where: E eq , R eq, M eq , G eq Respectively represent equivalent elastic modulus, equivalent radius, equivalent mass, and equivalent shear modulus; N n damp and N t damp are the normal damping coefficient and the tangential damping coefficient.

[0060]

[0061] Where: N n damp 、N t damp are the forward restitution coefficient and the tangential restitution coefficient, respectively.

[0062] (2) Establish a physical model of ships and floating ice;

[0063] The hull model adopts the Kriso Container Ship model with a scale ratio of 31.6. The other parameters are shown in Table 1.

[0064]

[0065] Table 1 Ship model parameters

[0066] Ice breaking model: The broken ice is regarded as a composite particle composed of several spherical particles, and the length (width) and thickness of the ice particles are 20 cm and 4 cm respectively. The density of the model ice is 910.0 kg / m 3 , Poisson's ratio is 0.3. The Young's modulus of the model ice is 7.09 Pa. The properties of the crushed ice are defined in the Lagrangian phase, and the crushed ice is defined to enter the flow field through the ejector; Figure 2 shown.

[0067] (3) CFD-DEM coupled simulation;

[0068] In order to accurately simulate navigation in the icebreaking area, the longitudinal and transverse lengths of the computational domain are set to 6.0 times the ship length, and the vertical height is set to 4.5 times the ship length. At the same time, in the boundary condition setting, the inlet, upper boundary and lower boundary are set as the velocity inlet, the outlet boundary is set as the pressure outlet, and the lateral boundary is set as the symmetry surface. The hull surface is configured to have an anti-slip wall condition. In the simulation calculation, the Euler method is used to fix the ship, and the fluid rushes from the front velocity inlet to the ship model at a set speed. At the same time, an ejector is set at the velocity inlet to spray crushed ice to the ship model with a certain particle flow. The computational domain is discretized using a structured hexahedral grid, and local grid refinement is performed in areas such as the free surface, ship model, and hull Kelvin wave to ensure the accuracy of the flow field calculation; a six-layer prismatic grid is generated near the hull surface to simulate the boundary layer flow.

[0069] B. Constrain the neural network using the physical equations related to ice floes and ship ice loads;

[0070] Since there is momentum conservation between the ship and the ice, the momentum conservation equation between the ice particle momentum and the ship ice load is constructed as the loss function of the network to physically constrain the network, and the following loss function is constructed:

[0071] Loss total =Loss function +Loss data

[0072] Where: Loss function The loss in is based on the energy conservation equation between the ship and the ice. data is the predicted value of ice load f. Both can be expressed as:

[0073]

[0074] Where n is the number of data points in the ship-ice collision dataset, F n is the predicted value of the network; t n is a time series; dt represents the differential sign; N represents the total number of samples, and the average of N samples is taken as the loss.

[0075] C. Predict ice loads through time series using physical information neural network methods;

[0076] The ice load force of ships sailing in the broken ice area is calculated by CFD-DEM coupling, and the number of ice particles, relative collision speed and frequency at the corresponding time are collected.

[0077] Detailed data categories include: the number of ice particles at the current moment, the relative collision speed between ice particles and the ship at the current moment, and the relative collision frequency between ice particles and the ship at the current moment. Among them, the number of ice particles and the relative collision speed are measured using the image recognition method, and the relative collision frequency is measured using the interval sampling method. The momentum of a single ice particle acting on the ship can be calculated using the relative collision speed and the mass of the ice particles, and then multiplied by the number and frequency to obtain the ship's ice load.

[0078] For the above samples, 70% of the sample points are used as the training set of the physical information neural network, and the remaining 30% are used as the validation set. The 70% number not only meets the large number of sample points required for training, but also ensures that the data will not be overly redundant, resulting in overfitting of the training. The 30% data set ensures that the data volume is rich enough to exclude special results, so that there is a reasonable reference standard to define whether the training is successful when outputting the results.

[0079] Based on these data, a physical information neural network method is used to predict ice load through time series. In this framework, the time series is used as input, and the value of the ice load force, the number of ice particle collisions, and the relative speed between the ice particles and the ship of the current time series are output through a fully connected network. The data is segmented according to the input and output of the neural network, and training begins. An automatic learning rate adjustment strategy is adopted. If the loss value does not decrease in 10 iterations, the learning rate is adjusted by increasing or decreasing it by 20%. The entire training process contains 200,000 iterations. At the beginning of the training, the loss decreases slowly. In the training interval of 30,000 to 80,000 steps, the loss value decreases rapidly. At the end of convergence, the loss value shows oscillation. This is because the collision between ice particles and ships is a probabilistic event. Each small batch of data cannot accurately represent the distribution of ice loads during the entire voyage, resulting in oscillations in the loss value at the end of training. Finally, the trained network is used to predict the ship's ice load in the next second. The prediction error is about 14%, as shown in Figure 2. Figure 3 As shown in the process cycle, the following is a specific calculation example:

[0080]

[0081]

[0082] Wherein, m is the total mass of ice particles, which is the product of the number n and the individual mass, v is the relative velocity between the ice particles and the hull, and F is the ice load force. When the actual simulation is performed according to the calculation process of the present invention, after inputting the corresponding data, the following can be obtained: Figure 4-5 The data shown, such as Figure 6 As shown, the image directly output by the method can directly show the ice load state under the current state.

[0083] In another embodiment, the present invention provides a ship ice load prediction system based on a physical information neural network, the system comprising:

[0084] The acquisition module is used to collect the ship status and ice load data in the current state;

[0085] A data module, used to carry PINNs and a ship ice load prediction module, the ship ice load prediction module includes the steps of the method described in steps A to C, and outputs the ship ice load and its predicted value in the current state;

[0086] An image module is used to display the ship ice load and its predicted value obtained by the data module and display them on an image device;

[0087] The exception information unit is used to collect exception information and distribute it to the person in charge after visual processing on the front end.

[0088] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A ship ice load prediction method based on physical information neural network, characterized in that: The following steps are involved: A. Use computational fluid dynamics methods to obtain the data set required for physical information neural network; (1) Constructing control model based on CFD-DEM method; The continuity equation and Navier-Stokes equation are selected as the basic equations of the fluid phase. The Reynolds average method is used to establish the fluid phase control equation. The shear stress transport model is used as the turbulence model to close the control equation. The fluid volume method is used to capture the free surface waves generated during the ship's navigation. The contact model of discrete element particles is the Hertz-Mindlin model, and the nonlinear increase of the contact force is affected by the displacement. The contact between the two particles is the discrete phase basic equation, and the control model is obtained after calculation; (2) Establish a physical model of ships and floating ice; The hull model uses the Kriso Container Ship model with a scale ratio of 31.6; Icebreaking model: Broken ice is regarded as a composite particle composed of several spherical particles; (3) CFD-DEM coupled simulation; The longitudinal and transverse lengths of the computational domain were set to 6.0 times the length of the ship, and the vertical height was set to 4.5 times the length of the ship. At the same time, in the boundary condition settings, the inlet, upper boundary, and lower boundary were set to the velocity inlet, the outlet boundary was set to the pressure outlet, and the lateral boundary was set to the symmetry surface. The hull surface was configured to have an anti-slip wall condition. In the simulation calculation, the Euler method was used to fix the ship, and the fluid rushed from the front velocity inlet to the ship model at a set speed. At the same time, an ejector was set at the velocity inlet to spray crushed ice toward the ship model in a particle flow. The computational domain is discretized using structured hexahedral meshes, and local mesh refinement is performed on the free surface, ship model, and hull Kelvin wave regions to ensure the accuracy of flow field calculations; a six-layer prismatic mesh is generated near the hull surface to simulate boundary layer flow; B. Constrain the neural network using the physical equations related to ice floes and ship ice loads; Since there is momentum conservation between the ship and the ice, the momentum conservation equation between the ice particle momentum and the ship ice load is constructed as the loss function of the network to physically constrain the network, and the following loss function is constructed: Loss total =Loss function +Loss data ; Where: Loss function The loss in is based on the energy conservation equation between the ship and the ice. data is the predicted value of ice load f; the two can be expressed as: Where n is the number of data points in the ship-ice collision dataset, F n is the predicted value of the network; t n is a time series; N represents the total number of samples, and the average of N samples is taken as the loss; C. Predict ice loads through time series using physical information neural network methods; The ice load force of ships sailing in the broken ice area is calculated by CFD-DEM coupling, and the number of ice particles, relative collision velocity and frequency at the corresponding time are collected; based on the ice load force on the ship, the number of ice particles at the corresponding time, and the relative collision velocity and frequency of ice particles, the physical information neural network method is used to predict the ice load through time series to obtain the time series data of ice load.

2. A ship ice load prediction method based on physical information neural network according to claim 1, characterized in that: The computational fluid dynamics method in step A is to establish a ship ice collision model using a CFD-DEM method for simulating ship navigation in an ice-breaking area; The ship ice collision model includes an ice particle model for ship icebreaking simulation. The ice particles are considered as composite particles composed of spherical particles with a length or width of 20 cm, a thickness of 4 cm, and a density of 910.0 kg / m 3 , Poisson’s ratio is 0.3, Young’s modulus is 7.09 Pa, and the crushed ice enters the flow field through the ejector.

3. A ship ice load prediction method based on physical information neural network according to claim 2, characterized in that: The data categories of the data set in step A include: the number of ice particles at the current moment, the relative collision speed of ice particles and the ship at the current moment, the relative collision frequency of ice particles and the ship at the current moment, and the ship ice load force calculated based on the data category of the data set.

4. The ship ice load prediction method based on physical information neural network according to claim 1 is characterized in that: 70% of the dataset described in step A is used as a training set, and 30% is used as a validation set.

5. The ship ice load prediction method based on physical information neural network according to claim 1 is characterized in that: The physical equation in step B is the momentum conservation equation between the momentum of ice particles and the ice load of the ship.

6. The ship ice load prediction method based on physical information neural network according to claim 1 is characterized in that: The structure of the physical information neural network in step C is a fully connected neural network, the input layer is a time series, and the output layer is the value of the ice load force in the current time series, the number of ice particle collisions, and the relative speed between the ice particles and the ship.

7. A ship ice load prediction system based on physical information neural network for implementing the method of claim 1, characterized in that: System includes: The acquisition module is used to collect the ship status and ice load data in the current state; A data module, used to carry PINNs and a ship ice load prediction module, the ship ice load prediction module includes the steps of the method described in steps A to C, and outputs the ship ice load and its predicted value in the current state; An image module is used to display the ship ice load and its predicted value obtained by the data module and display them on an image device; The exception information unit is used to collect exception information and distribute it to the person in charge after visual processing on the front end.

Images