Ship control motion forecasting method based on expert experience embedded neural network

Abstract

The invention relates to a ship control motion forecasting method based on an expert experience embedded neural network. The ship control motion forecasting method comprises a motion forecasting input factor sensitivity weight distribution method based on an analytic hierarchy process and construction and training of an expert experience embedded neural network model. According to the method, sufficient data mining and learning can be carried out on a large amount of currently accumulated constraint model test data, large-scale and small-scale self-propelled model test data and real ship navigation test data, so that wide and universal forecasting of ship control motion is realized; the problem that a traditional forecasting method is insufficient in forecasting capacity under the complex environment and the special complex working condition of a real ship is solved, and high-precision motion forecasting of the ship under the full working condition is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of ship maneuvering, and particularly relates to a ship maneuvering motion prediction method based on an expert experience-embedded neural network. Background Art

[0002] Currently, the commonly used ship maneuvering motion prediction is based on captive model tests to obtain the hydrodynamic coefficients related to ship motion, establish a six-degree-of-freedom ship motion model, and then perform motion prediction. The equation forms of all maneuvering motion models have their applicable limitations, and the prediction accuracy for complex environments and complex maneuvering conditions is not satisfactory. The six-degree-of-freedom ship motion model consists of six nonlinear differential equations. For small-angle motions, these hydrodynamic forces are approximately linear functions of the motion parameters. For large-angle motions or spatial maneuvers, the expansion must be extended to the second or third order or higher, and the cross-influence cannot be ignored. Therefore, any form of motion equation is limited by the navigation environmental conditions or the flow field pattern of the maneuver, and cannot perform extensive and general predictions. For example, the results of the conventional maneuvering motion model in predicting the ship motion in a large maneuvering state are significantly inaccurate, and it is difficult to obtain satisfactory results in predicting the motion maneuver in the near-surface wave and ice surface environments using the underwater standard maneuvering mathematical model.

[0003] Compared with the technical shortcomings of the modeling and simulation prediction based on captive model tests, the method of directly predicting using free self-propelled tests is an effective way. The full-scale ship test is the ultimate standard for testing the ship maneuvering design scheme and its performance. However, the full-scale ship sea test and the free self-propelled tests on lakes and ship basins all consume a large amount of manpower and financial resources, and there is also the problem of a long test cycle. In addition, there are still many technical and funding problems in the production of models and the use of a large number of high-precision sensors.

[0004] Therefore, it is urgent to design a ship maneuvering motion prediction method based on an expert experience-embedded neural network to solve the problems existing in the above-mentioned prior art. Summary of the Invention

[0005] In view of this, the present invention provides a ship maneuvering motion prediction method based on an expert experience-embedded neural network. The purpose is to construct an expert experience-embedded neural network model (Experience-Informed Neural Network, abbreviated as EINN), which can fully mine and learn a large amount of accumulated captive model test data, self-propelled model test data of different scales, and full-scale ship navigation test data, so as to achieve extensive and general prediction of ship maneuvering motion, solve the problem of insufficient prediction ability of traditional prediction methods in complex shipboard environments and special complex working conditions, and achieve high-precision motion prediction for all ship working conditions.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A ship maneuvering motion prediction method based on an expert experience embedded neural network, the ship maneuvering motion prediction method includes a motion prediction input factor sensitivity weight distribution method based on the analytic hierarchy process and the construction and training of an expert experience embedded neural network model;

[0008] Among them, the motion prediction input factor sensitivity weight distribution method based on the analytic hierarchy process includes the following steps:

[0009] Step 101, based on the operating condition type and the marine environment, construct a motion prediction input element set and an output element set;

[0010] Step 102, solve the motion prediction input element sensitivity weight factor based on the analytic hierarchy process;

[0011] Step 103, perform normalization processing on the motion prediction input element set;

[0012] Among them, the construction and training of the expert experience embedded neural network model includes the following steps:

[0013] Step 201, construct a ship maneuvering motion prediction model based on the long short-term memory neural network;

[0014] Step 202, construct a correction function based on expert experience and train the network model.

[0015] Further, the specific steps of step 101 are as follows:

[0016] Step 1011, construct a motion prediction input element set X in , X in =[X′ qt X′ en X′ con ; where X′ qt is the set after processing the ship motion response input element set X qt according to the operating condition type; X′ en is the set after processing the marine environment input element set X en according to the operating condition type; X′ con is the set after processing the ship maneuvering command element set X con according to the operating condition type;

[0017] Step 1012, construct a motion prediction output element set X out , Xout is a set obtained by processing the set X' of output elements of ship motion responses according to the type of operating conditions. out

[0018] Furthermore, the set X of input elements of ship motion responses q t includes: ship speed, attitude angle, angular velocity; the ship speed includes longitudinal ship speed, lateral ship speed, and vertical ship speed; the attitude angle includes course angle, trim angle, and roll angle; the angular velocity includes yaw angular velocity, pitch angular velocity, and roll angular velocity;

[0019] Processing the set X of input elements of ship motion responses according to the type of operating conditions qt is specifically as follows:

[0020] where is the bitwise multiplication operation of vectors, T qt is the operating condition screening vector, T qt is composed of 1 or 0. If a certain operating condition is irrelevant to some elements in X qt , then the corresponding elements in T qt are 0, and the remaining elements are 1.

[0021] Furthermore, the set X of input elements of the marine environment en includes seawater density and sea state elements; the sea state elements include wave height and wave direction angle;

[0022] Processing the set X of input elements of the marine environment according to the type of operating conditions en is specifically as follows:

[0023] where is the bitwise multiplication operation of vectors, T en is the operating condition screening vector, T en is composed of 1 or 0. If a certain operating condition is irrelevant to some elements in X en , then the corresponding elements in T en are 0, and the remaining elements are 1.

[0024] Furthermore, the set X of ship maneuvering instruction elements con includes rudder angle, rotational speed, buoyancy adjustment injection and drainage volume, and trim balance transfer volume; the rudder angle includes the steering rudder and the elevator;

[0025] Processing the set X of ship maneuvering instruction elements according to the type of operating conditions con is specifically as follows:

[0026] where ​is the vector bitwise multiplication operation, T con is the operating condition screening vector, T con It is composed of 1 or 0. If the operating condition is con Some elements in are irrelevant, then T con The corresponding position element is 0, and the other position elements are 1.

[0027] Furthermore, the ship motion response output element set X′ out The data include ship speed, position, attitude angle, and angular velocity; the ship speed includes longitudinal ship speed, transverse ship speed, and vertical ship speed; the position includes longitudinal position, transverse position, and vertical position; the attitude angle includes heading angle, pitch angle, and heel angle; the angular velocity includes bow angular velocity, pitch angular velocity, and roll angular velocity;

[0028] According to the operating condition type, the output element set X′ of the ship motion response is out To process, specifically:

[0029] in is the vector bitwise multiplication operation, T out is the operating condition screening vector, T out It is composed of 1 or 0. If the operating condition is out Some elements in are irrelevant, then T out The corresponding position element is 0, and the other position elements are 1.

[0030] Furthermore, the step 102 specifically includes the following steps:

[0031] Step 1021: Determine the motion prediction input element set X according to the operating conditions. in The sensitivity relationship between the elements in is used to construct the sensitivity weight judgment matrix A, which is as follows:

[0032]

[0033] Where n represents the motion forecast input element set X in There are n elements in total; a ij Represents the motion forecast input element set X in The sensitivity quantization value of the sensitivity comparison between the i-th element and the j-th element in ;

[0034] Step 1022: Obtain the motion prediction input element set X according to the sensitivity weight judgment matrix A. in The sensitivity weight factor of each element ω, the element ω in ω (i) The calculation method is:

[0035]

[0036] In the formula, n represents the set X of input elements for motion prediction in There are n elements in total, indicating a certain element in the sensitivity weight judgment matrix A in the k-th row;

[0037] Step 1023: Conduct a consistency test on the sensitivity weight judgment matrix A, solve the maximum eigenvalue λ of the sensitivity weight judgment matrix A, solve the consistency index CI of the weight matrix A = (λ - n) / (n - 1), where n is the dimension of the sensitivity weight judgment matrix A, solve the consistency ratio CR, CR = CI / RI, where RI is obtained by querying the consistency index table and corresponding values are obtained according to the dimension of the sensitivity weight judgment matrix A. If RI ≤ 0.1, the consistency test is passed; otherwise, readjust the sensitivity quantization numerical elements in the sensitivity weight judgment matrix A until the consistency test is passed.

[0038] Furthermore, step 103 specifically includes the following steps:

[0039] Step 1031: Normalize each element X in in the set X of input elements for motion prediction in (i), and the normalization method is as follows:

[0040]

[0041] where and are respectively the upper limit value and the lower limit value of the value range of the element X in (i);

[0042] Step 1032: Construct a set of input elements for motion prediction that combines the sensitivity weight factor The processing method is as follows: where is the bitwise multiplication operation of vectors, that is:

[0043]

[0044] Furthermore, step 201 specifically includes the following steps:

[0045] Step 2011: Define the network structure, including the number of network layers and the number of cell units in each layer. The network includes an input layer, a long short-term memory neural network hidden layer, and an output layer, and the number of cell units included in each long short-term memory neural network hidden layer;

[0046] Step 2012: Process the discard flow of the data of the set of input elements for motion prediction that combines the sensitivity weight factor The processing method is as follows: where f tRepresents the discarded flow processing output of the long short-term memory neural network cell unit at the current moment, W f Represents the weight matrix for discarded flow processing, b f Represents the bias matrix for discarded flow processing, H t-1 Represents the output of the long short-term memory neural network cell unit at the previous moment. That is, at the current moment t, the output at the previous moment t - 1 and the input at the current moment Are simultaneously passed to the sigmod function for weight judgment to determine the degree of data discard;

[0047] Step 2013, the set of input elements for motion prediction integrating the sensitivity weight factor Data update flow processing, the processing method is as follows: Where i t And Respectively represent the input flow and the processing output of the update flow of the long short-term memory neural network cell unit at the current moment. tanh() is the activation function, W i And W c Respectively represent the weight matrices for input flow and update flow processing, b t And b c Respectively represent the bias matrices for input flow and update flow processing, Where Is the element-wise multiplication operation, c t Represents the state of the updated cell unit at the current moment;

[0048] Step 2014, the set of input for motion prediction integrating the sensitivity weight factor at the current moment t Data output flow processing, the processing method is as follows: Where W o Represents the weight matrix for output flow processing, b o Represents the bias matrix for output flow processing, o t Represents the output value of the cell unit output flow processing, H t Represents the output of the long short-term memory neural network cell unit at the current moment. The outputs of each long short-term memory neural network cell unit are weighted and summed through a fully connected layer to obtain the corresponding Prediction result at the current moment

[0049] Furthermore, step 202 specifically includes the following content:

[0050] 2021, for the training samples sampled from the results of the restraint mode test or the results of the small-scale self-propelled model, according to expert experience, correct the X in the training samples out Make corrections, Where Denote the teacher samples corrected based on expert experience, that is, the corresponding true output, and f() represents the expert experience correction function established according to different working condition types in combination with the actual ship navigation test data;

[0051] In 2022, the root mean square error function was used to evaluate the prediction results obtained by training the ship motion prediction model, as follows:

[0052]

[0053] In the formula, i represents the i-th element in the set, and m represents the total number of elements in the set;

[0054] The smaller this value is, the closer the network prediction result is to the true value, and the more mature the network evolution is. According to the calculation result of this value, backpropagation is fed back to the long short-term memory neural network, and the network parameters of the long short-term memory neural network are continuously corrected through iteration, so that is as close as possible to the true output of the actual ship until the calculation accuracy meets the error requirement and the network training is mature.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] The present invention proposes a ship maneuvering motion prediction method based on an expert experience embedded neural network, which has intuitiveness. By directly establishing the corresponding relationship between the ship motion input and output, it avoids modeling the complex internal action relationship of ship motion and the dynamic characteristics of maneuvering motion; at the same time, it has scalability. After the training is mature, the prediction method can be extended to the maneuvering motion prediction of various types of ships; it also has high timeliness and does not rely on traditional ship motion modeling methods (which require captive model tests and a large number of CFD simulation calculations, consuming a lot of manpower, financial resources and a long test cycle). Using this method, by reasonably adjusting the training parameters, a matching motion prediction model can be quickly generated to meet the requirements of rapid design iteration; it makes full and effective use of a large amount of valuable actual ship data (including the ship motion response under environmental actions, the expert experience of crew manual operation, etc.), improves the problem of insufficient prediction ability of traditional prediction methods under complex actual ship environments and special complex working conditions, and realizes a breakthrough in the full-condition high-precision motion prediction technology.

[0057] Other features and advantages of the present invention will be described in the subsequent description, and, in part, will be obvious from the description, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained by the structure pointed out in the description and the drawings. Brief Description of the Drawings

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following briefly introduces the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0059] Figure 1 Fig. shows the flowchart of the ship maneuvering motion prediction method based on the expert experience embedded neural network in the embodiment of the present invention. Detailed implementation manners

[0060] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0061] An embodiment of the present invention proposes a ship maneuvering motion prediction method based on an expert experience embedded neural network. As shown in the attached Figure 1 The ship maneuvering motion prediction method includes a motion prediction input factor sensitivity weight allocation method based on the Analytic Hierarchy Process (AHP) and the construction and training of an expert experience embedded neural network model (EINN).

[0062] The motion prediction input factor sensitivity weight allocation method based on the Analytic Hierarchy Process (AHP) includes the following steps:

[0063] Step 101: Based on the operating condition type and the marine environment, construct a motion prediction input element set and an output element set;

[0064] Step 102: Solve the motion prediction input element sensitivity weight factors based on the Analytic Hierarchy Process;

[0065] Step 103: Normalize the motion prediction input element set;

[0066] The construction and training of the expert experience embedded neural network model (EINN) includes the following steps:

[0067] Step 201: Construct a ship maneuvering motion prediction model based on the Long Short-Term Memory neural network (abbreviated as LSTM, an improved form of the Recurrent Neural Network RNN);

[0068] Step 202, construction of a correction function based on expert experience and training of the network model.

[0069] The specific steps of step 101 are as follows:

[0070] Step 1011, construct a set X of input elements for motion prediction based on the type of operating conditions and the marine environment in , X in = [X′ qt X′ en X′ con ; where X′ qt is the set obtained by processing the set X of input elements for ship motion response according to the type of operating conditions qt ; X′ en is the set obtained by processing the set X of input elements for the marine environment according to the type of operating conditions en ; X′ con is the set obtained by processing the set X of ship maneuvering instruction elements according to the type of operating conditions con .

[0071] Step 1012, construct a set X of output elements for motion prediction based on the type of operating conditions and the marine environment out , X out is the set obtained by processing the set X′ of output elements for ship motion response according to the type of operating conditions out .

[0072] The set X of input elements for ship motion response qt includes: ship speed, attitude angle, angular velocity; the ship speed includes longitudinal ship speed, lateral ship speed, and vertical ship speed; the attitude angle includes course angle, trim angle, and roll angle; the angular velocity includes yaw angular velocity, pitch angular velocity, and roll angular velocity;

[0073] The processing of the set X of input elements for ship motion response according to the type of operating conditions qt is specifically as follows:

[0074] where is the bitwise multiplication operation of vectors, T qt is the operating condition screening vector, T qt is composed of 1 or 0. If a certain operating condition is irrelevant to some elements in X qt , then the corresponding elements in T qt are 0, and the remaining elements are 1.

[0075] The set X of input elements for the marine environment en includes seawater density and sea state elements; the sea state elements include wave height and wave direction angle;

[0076] According to the type of operating conditions, the input element set X of the marine environment en is processed as follows:

[0077] where is the bitwise multiplication operation of vectors, and T en is the operating condition screening vector, T en consists of 1 or 0. If a certain operating condition is irrelevant to some elements in X en , then for T en the elements at the corresponding positions are 0, and the elements at the remaining positions are 1.

[0078] The ship's maneuvering instruction element set X con includes rudder angle, rotational speed, buoyancy adjustment injection displacement, and longitudinal trim balance transfer water volume; the rudder angle includes the steering rudder and the elevator;

[0079] According to the type of operating conditions, the ship's maneuvering instruction element set X con is processed as follows:

[0080] where is the bitwise multiplication operation of vectors, and T con is the operating condition screening vector, T con consists of 1 or 0. If a certain operating condition is irrelevant to some elements in X con , then for T con the elements at the corresponding positions are 0, and the elements at the remaining positions are 1.

[0081] The ship's motion response output element set X' out includes ship speed, position, attitude angle, and angular velocity; the ship speed includes longitudinal ship speed, lateral ship speed, and vertical ship speed; the position includes longitudinal position, lateral position, and vertical position; the attitude angle includes course angle, longitudinal trim angle, and transverse tilt angle; the angular velocity includes yaw angular velocity, pitch angular velocity, and roll angular velocity;

[0082] According to the type of operating conditions, the ship's motion response output element set X' out is processed as follows:

[0083] where is the bitwise multiplication operation of vectors, and T out is the operating condition screening vector, T out consists of 1 or 0. If a certain operating condition is irrelevant to some elements in X' out , then for T out the elements at the corresponding positions are 0, and the elements at the remaining positions are 1.

[0084] Step 102 specifically includes the following steps:

[0085] Step 1021: According to the operating conditions, judge the sensitivity relationship among the elements in the motion prediction input element set X in and construct a sensitivity weight judgment matrix A as follows:

[0086]

[0087] In the formula, n represents that there are n elements in the motion prediction input element set X in ; a ij represents the sensitivity quantization value of the comparison of the sensitivity between the i-th element and the j-th element in the motion prediction input element set X in ;

[0088] If two elements are compared and their sensitivity degrees are the same, the sensitivity quantization value a ij takes 1; if two elements are compared and one element is slightly more sensitive than the other under this operating condition, the sensitivity quantization value a ij takes 3; if two elements are compared and one element is significantly more sensitive than the other under this operating condition, the sensitivity quantization value a ij takes 5; if two elements are compared and one element is strongly more sensitive than the other under this operating condition, the sensitivity quantization value a ij takes 7; if two elements are compared and one element is extremely more sensitive than the other under this operating condition, the sensitivity quantization value a ij takes 9; if two elements are compared and one element is between the above adjacent judgments in terms of sensitivity under this operating condition, the sensitivity quantization value a ij takes 2, 4, 6 or 8.

[0089] In addition, if the sensitivity quantization value of the comparison of the sensitivity between the i-th element and the j-th element is a ij , then the sensitivity quantization value of the comparison of the sensitivity between the j-th element and the i-th element is represented by 1 / a ij .

[0090] Step 1022: According to the sensitivity weight judgment matrix A, obtain the sensitivity weight factors ω of the elements in the motion prediction input element set X in , and the calculation method of the element ω (i) in ω is as follows:

[0091]

[0092] In the formula, n represents that there are n elements in the motion prediction input element set X in , represents a certain element in the sensitivity weight judgment matrix A in the i-th row.

[0093] Step 1023: Conduct a consistency test on the sensitivity weight judgment matrix A, solve for the maximum eigenvalue λ of the sensitivity weight judgment matrix A, solve for the consistency index CI of the weight matrix A, where CI = (λ - n) / (n - 1), n is the dimension of the sensitivity weight judgment matrix A, solve for the consistency ratio CR, and CR = CI / RI. Here, RI is obtained by querying the consistency index lookup table according to the dimension of the sensitivity weight judgment matrix A to find the corresponding value. If RI ≤ 0.1, the consistency test is passed; otherwise, readjust the sensitivity quantization numerical elements in the sensitivity weight judgment matrix A until the consistency test is passed.

[0094] The specific steps of step 103 are as follows:

[0095] Step 1031: Normalize each element X in in the motion prediction input element set X in (i), and the normalization method is as follows:

[0096]

[0097] where and are respectively the upper limit value and the lower limit value of the value range of the element X in (i);

[0098] Step 1032: Construct a motion prediction input element set that integrates the sensitivity weight factor The processing method is as follows: where is the bitwise multiplication operation of vectors, that is:

[0099]

[0100] The specific steps of step 201 are as follows:

[0101] Step 2011: Define the network structure, including the number of network layers and the number of cell units in each layer. The network includes an input layer, a long short-term memory neural network (LSTM) hidden layer, and an output layer, and the number of cell units included in each LSTM hidden layer;

[0102] Step 2012: Process the discard flow of the data in the motion prediction input element set that integrates the sensitivity weight factor The processing method is as follows: where f t represents the output of the discard flow processing of the LSTM cell unit at the current moment, W f represents the weight matrix of the discard flow processing, b f represents the bias matrix of the discard flow processing, and Ht-1 represents the output of the long short - term memory neural network (LSTM) cell unit at the previous moment. That is, at the current moment \(t\), the output at the previous moment \(t - 1\) and the input at the current moment are simultaneously passed to the sigmod function for weight judgment to determine the degree of data discard;

[0103] Step 2013, the set of input elements for motion prediction integrating the sensitivity weight factor Processing of the update flow direction of data, and the processing method is as follows: where \(i\) t and respectively represent the input flow direction and the processing output of the update flow direction of the long short - term memory neural network (LSTM) cell unit at the current moment. \(tanh()\) is the activation function, \(W\) i and \(W\) c respectively represent the weight matrices for processing the input flow direction and the update flow direction, \(b\) t and \(b\) c respectively represent the bias matrices for processing the input flow direction and the update flow direction, where is the element - wise multiplication operation, \(c\) t represents the state of the updated cell unit at the current moment;

[0104] Step 2014, the set of input for motion prediction integrating the sensitivity weight factor at the current moment \(t\) Processing of the output flow direction of data, and the processing method is as follows: where \(W\) o represents the weight matrix for processing the output flow direction, \(b\) o represents the bias matrix for processing the output flow direction, \(o\) t represents the output value of the processing of the output flow direction of the cell unit, \(H\) t represents the output of the long short - term memory neural network (LSTM) cell unit at the current moment. The outputs of each long short - term memory neural network (LSTM) cell unit are weighted and summed through a fully - connected layer to obtain the corresponding prediction result at the current moment

[0105] The specific content of step 202 is as follows:

[0106] 2021. For the training samples sampled from the results of the restraint - mode test or the results of small - scale self - propelled model, according to expert experience, correct \(X\) in the training samples out where where represents the teacher sample corrected based on expert experience, that is, the corresponding true output, and \(f()\) represents the expert - experience correction function established according to different working - condition types and combined with the actual ship - trial data;

[0107] In 2022, the root mean square error function was used to evaluate the prediction results obtained by training the ship motion prediction model, as follows:

[0108]

[0109] Where represents the i-th element in the set, and m represents the total number of elements in the set;

[0110] The smaller this value is, the closer the network prediction result is to the true value, and the more mature the network evolution is. According to the calculation result of this value, backpropagation is fed back to the long short-term memory neural network (LSTM), and the network parameters of the LSTM are continuously corrected through iteration, so that is as close as possible to the actual output of the real ship until the calculation accuracy meets the error requirements and the network training is mature.

[0111] The ship maneuvering motion prediction method based on the expert experience embedded neural network proposed by the present invention has intuitiveness. By directly establishing the correspondence between the ship motion input and output, it avoids the modeling of the complex internal action relationship of the ship motion and the dynamic characteristics of the maneuvering motion; at the same time, it has scalability. After the training is mature, the prediction method can be extended to the maneuvering motion prediction of various types of ships; it also has high timeliness and does not rely on the traditional ship motion modeling method (which requires captive model tests and a large number of CFD simulation calculations, consuming a large amount of manpower, financial resources, and a long test cycle). Using this method, by reasonably adjusting the training parameters, a matching motion prediction model can be quickly generated to meet the requirements of rapid design iteration; it makes full and effective use of a large amount of valuable real ship data (including the ship motion response under environmental actions, the expert experience of crew manual operation, etc.), improves the problem of insufficient prediction ability of the traditional prediction method under complex real ship environments and special complex working conditions, and realizes a breakthrough in the full working condition high-precision motion prediction technology.

[0112] Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A ship maneuvering motion prediction method based on an expert experience embedded neural network, characterized in that, The ship maneuvering motion prediction method includes a sensitivity weight assignment method for motion prediction input factors based on the analytic hierarchy process and the construction and training of an expert experience embedded neural network model; Among them, the sensitivity weight assignment method for motion prediction input factors based on the analytic hierarchy process includes the following steps: Step 101, based on the operating condition type and marine environment, construct a set of motion prediction input elements and a set of output elements; Step 102, solve the sensitivity weight factor of the motion prediction input elements based on the analytic hierarchy process; Step 103, perform normalization processing on the set of motion prediction input elements; Among them, the construction and training of the expert experience embedded neural network model includes the following steps: Step 201, construct a ship maneuvering motion prediction model based on the long short-term memory neural network; Step 202, construct a correction function based on expert experience and train the network model.

2. The ship maneuvering motion prediction method based on an embedded neural network of expert experience according to claim 1, wherein The specific steps of the above-mentioned step 101 are as follows: Step 1011, construct a set X of motion prediction input elements based on the types of operating conditions and the marine environment in , X in = [X′ qt X′ en X′ con ; where X′ qt is the set obtained by processing the set X of ship motion response input elements according to the types of operating conditions qt ; X′ en is the set obtained by processing the set X of marine environment input elements according to the types of operating conditions en ; X′ con is the set obtained by processing the set X of ship maneuvering instruction elements according to the types of operating conditions con . Step 1012, construct a set X of motion prediction output elements based on the type of operating conditions and the marine environment out , X out is a set obtained by processing the set X' of ship motion response output elements out according to the type of operating conditions.

3. The ship maneuvering motion prediction method based on an embedded neural network of expert experience according to claim 2, wherein, The set of input elements X for ship motion response qt includes: ship speed, attitude angle, angular velocity; the ship speed includes longitudinal ship speed, lateral ship speed, and vertical ship speed; the attitude angle includes course angle, trim angle, and roll angle; the angular velocity includes yaw angular velocity, pitch angular velocity, and roll angular velocity; According to the type of operating conditions, the input element set X of the ship motion response qt is processed as follows: Among them is the bitwise multiplication operation of vectors, T qt is the operation condition screening vector, T qt is composed of 1 or 0. If the operation condition is irrelevant to some elements in X qt then the elements at the corresponding positions in T qt are 0, and the elements at the remaining positions are 1.

4. The ship maneuvering motion prediction method based on an expert experience embedded neural network according to claim 3, characterized in that, The set X of input elements of the marine environment en includes seawater density and sea state elements; the sea state elements include wave height and wave direction angle; According to the type of operating conditions, the input element set X of the marine environment en is processed, specifically as follows: Among them is the bitwise multiplication operation of vectors, T en is the vector for screening operating conditions, T en is composed of 1 or 0. If the operating condition is irrelevant to some elements in X en then the elements at the corresponding positions in T en are 0, and the elements at the remaining positions are 1.

5. The ship maneuvering motion prediction method based on an expert experience-embedded neural network according to claim 4, wherein The set X of ship operation instruction elements con includes rudder angle, rotational speed, buoyancy adjustment injection displacement, and longitudinal trim balance displacement; the rudder angle includes the rudder and the elevator; According to the type of operating conditions, process the set X of ship maneuvering instruction elements con The specific processing method is as follows: wherein is the bitwise multiplication operation of vectors, T con is the vector for screening operating conditions, T con is composed of 1 or 0. If the operating condition is irrelevant to some elements in X con then for the corresponding position elements in T con are 0 and the elements in the remaining positions are 1.

6. The ship maneuvering motion prediction method based on an expert experience embedded neural network according to claim 5, characterized in that, The set X' of the output elements of the ship motion response out includes ship speed, position, attitude angle, and angular velocity; the ship speed includes longitudinal ship speed, lateral ship speed, and vertical ship speed; the position includes longitudinal position, lateral position, and vertical position; the attitude angle includes course angle, trim angle, and roll angle; the angular velocity includes yaw angular velocity, pitch angular velocity, and roll angular velocity; According to the type of operating conditions, process the set X' of output elements of ship motion response out The specific processing method is as follows: Among them is the bitwise multiplication operation of vectors, T out is the operation condition screening vector, T out is composed of 1 or 0. If the operation condition is irrelevant to some elements in X′ out then the elements at the corresponding positions of T out are 0, and the elements at the remaining positions are 1.

7. The ship maneuvering motion prediction method based on the expert experience embedded neural network according to claim 6, characterized in that The specific steps of the above-mentioned step 102 are as follows: Step 1021: According to the operating conditions, judge the sensitivity relationship among the elements in the set X of motion prediction input elements, and construct a sensitivity weight judgment matrix A as follows: in Specifically as follows: Wherein, n represents the set X of input elements for motion prediction in There are n elements in total; a ij represents the sensitivity quantization value for comparing the sensitivities of the i-th element and the j-th element in the set X of input elements for motion prediction in of the set X of input elements for motion prediction; Step 1022: Obtain the set X of motion prediction input elements according to the sensitivity weight judgment matrix A in Each element sensitivity weight factor ω, and the element ω in ω (i) is calculated as follows: In the formula, n represents the set X of input elements for motion prediction in There are a total of n elements in it, and k represents a certain element in the sensitivity weight judgment matrix A in the k-th row; Step 1023, conduct a consistency test on the sensitivity weight judgment matrix A, solve the maximum eigenvalue λ of the sensitivity weight judgment matrix A, solve the consistency index CI of the weight matrix A = (λ - n) / (n - 1), where n is the dimension of the sensitivity weight judgment matrix A, solve the consistency ratio CR, CR = CI / RI, where RI is obtained by querying the consistency index table according to the dimension of the sensitivity weight judgment matrix A, and if RI ≤ 0.1, the consistency test is passed, otherwise, readjust the sensitivity quantization numerical elements in the sensitivity weight judgment matrix A until the consistency test is passed.

8. The ship maneuvering motion prediction method based on an expert experience embedded neural network according to claim 7, wherein The specific steps of the above-mentioned step 103 are as follows: Step 1031, for each element X in the set X of motion prediction input elements in perform normalization processing, and the normalization processing method is as follows: in (i) wherein and are the upper limit value and the lower limit value of the value range of element X in (i), respectively; Step 1032, construct a set of input elements for motion prediction that incorporates sensitivity weight factors The processing method is as follows: Among them is the bitwise multiplication operation of vectors, that is:

9. The ship maneuvering motion prediction method based on an expert experience embedded neural network according to claim 8, wherein The specific steps of the above-mentioned step 201 are as follows: Step 2011, clarify the network structure, including the number of network layers and the number of cell units in each layer. The network includes an input layer, a long short-term memory neural network hidden layer, and an output layer, and the number of cell units included in each long short-term memory neural network hidden layer; Step 2012, the set of motion prediction input elements integrated with the sensitivity weight factor Processing of the discard flow of data, the processing method is as follows: where f t represents the output of the discard flow processing of the long short-term memory neural network cell unit at the current moment, W f represents the weight matrix of the discard flow processing, b f represents the bias matrix of the discard flow processing, H t-1 represents the output of the long short-term memory neural network cell unit at the previous moment, that is, at the current moment t, the output at the previous moment t - 1 and the input at the current moment are simultaneously passed to the sigmod function for weight judgment to determine the degree of data discard; Step 2013, the set of motion prediction input elements fused with the sensitivity weight factor Data update flow processing, the processing method is as follows: where i t and respectively represent the input flow and the processing output of the update flow of the long short-term memory neural network cell unit at the current moment, tanh() is the activation function, W i and W c respectively represent the weight matrices of the input flow and the update flow processing, b t and b c respectively represent the bias matrices of the input flow and the update flow processing, where is the element-wise multiplication operation, c t represents the state of the updated cell unit at the current moment; Step 2014, the motion prediction input set of the fusion sensitivity weight factor at the current t moment Output flow processing of data, the processing method is as follows: Among them, W o represents the weight matrix of output flow processing, b o represents the bias matrix of output flow processing, o t represents the output value of the output flow processing of the cell unit, H t represents the output of the long short-term memory neural network cell unit at the current moment. The outputs of each long short-term memory neural network cell unit are weighted and summed through the fully connected layer to obtain the corresponding prediction result at the current moment ​ 10. The ship maneuvering motion prediction method based on an expert experience-embedded neural network according to claim 9, characterized in that, The specific content of the above-mentioned step 202 is as follows: In 2021, for the training samples sampled from the results of the restraint mode test or the small-scale self-propelled model results, according to expert experience, correct X in the training samples out where represents the teacher sample corrected based on expert experience, that is, the corresponding true output, and f() represents the expert experience correction function established by combining the actual ship navigation test data according to different working condition types;​ 2022, use the root mean square error function to evaluate the prediction results obtained by training the ship motion prediction model, specifically as follows: In the formula, i represents the i-th element in the set, and m represents the total number of elements in the set; The smaller this value is, the closer the network prediction result is to the true value, and the more mature the network evolution is. According to the calculation result of this value, backpropagation is fed back to the long short-term memory neural network, and the network parameters of the long short-term memory neural network are continuously corrected through iteration, so that it is as close as possible to the real output of the actual ship until the calculation accuracy meets the error requirement and the network training is mature.

Images