CN111619751A - Ship rolling motion attitude control system based on multiple learning algorithms - Google Patents

Ship rolling motion attitude control system based on multiple learning algorithms Download PDF

Info

Publication number
CN111619751A
CN111619751A CN202010501120.5A CN202010501120A CN111619751A CN 111619751 A CN111619751 A CN 111619751A CN 202010501120 A CN202010501120 A CN 202010501120A CN 111619751 A CN111619751 A CN 111619751A
Authority
CN
China
Prior art keywords
rolling
neural network
ship
layer
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
CN202010501120.5A
Other languages
Chinese (zh)
Inventor
宋伟伟
谭银朝
张玲珑
巩方超
岳昌华
刘胜
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Weihai Ocean Vocational College
Original Assignee
Weihai Ocean Vocational College
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Weihai Ocean Vocational College filed Critical Weihai Ocean Vocational College
Priority to CN202010501120.5A priority Critical patent/CN111619751A/en
Publication of CN111619751A publication Critical patent/CN111619751A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B63—SHIPS OR OTHER WATERBORNE VESSELS; RELATED EQUIPMENT
    • B63B—SHIPS OR OTHER WATERBORNE VESSELS; EQUIPMENT FOR SHIPPING 
    • B63B39/00—Equipment to decrease pitch, roll, or like unwanted vessel movements; Apparatus for indicating vessel attitude
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Abstract

The invention relates to a ship rolling motion attitude control system based on multiple learning algorithms, which mainly comprises a rolling detection device, a ship rolling motion controller comprising a rolling depth learning neural network identifier and a rolling depth learning neural network sub-controller, a fin servo system and a database storing historical rolling data, wherein the ship rolling motion controller comprises a rolling depth learning neural network identifier and a rolling depth learning neural network sub-controller; the rolling detection device is used for detecting target rolling data generated by the target ship in the motion process; the ship rolling motion controller is used for generating a target fin control signal according to the target rolling data and sending the target fin control signal to the fin servo system to control the motion of the monomer fins so as to control the target ship; historical roll data is determined by a ship roll data classification model; the ship rolling data classification model is constructed on the basis of an unsupervised machine learning algorithm. The invention combines a deep learning neural network method with an unsupervised machine learning method, and effectively improves the control effect of the ship rolling motion attitude.

Description

Ship rolling motion attitude control system based on multiple learning algorithms
Technical Field
The invention relates to the field of ship motion attitude control, in particular to a ship rolling motion attitude control system based on multiple learning algorithms.
Background
Deep learning can automatically learn characteristics which represent deeper essence of data from a large amount of data, and has large data generalization capability.
The ship is often interfered by environmental factors such as sea waves and the like during navigation, and inevitably swings, so that great potential safety hazards are caused to the offshore operation of the ship particularly under severe sea conditions.
A complex problem is that of ships sailing at sea. The ship inevitably swings and tilts under the action of random sea waves, so that the normal navigation of the ship on the sea surface is influenced, the navigation performance and the manipulation performance of the ship are influenced to a certain extent, and the disturbance of the sea waves also influences the motion state of objects on the ship. Particularly under severe sea conditions, for passenger ships, civil ships and small fishery ships, because the displacement is low (generally between 1000t to 3000 t), the small ships are greatly influenced by sea waves, large-amplitude swinging motion can occur under the action of the sea waves, and the maximum inclination angle can even reach more than 15 degrees. In these cases, the motion attitude of the hull is controlled mainly by using PID control, adaptive control, robust control or a single neural network method, but the above method has poor stability in controlling the motion of the hull.
Disclosure of Invention
The invention aims to provide a ship rolling motion attitude control system based on multiple learning algorithms, which combines a deep learning neural network method with an unsupervised machine learning method and effectively improves the ship rolling motion attitude control effect.
In order to achieve the purpose, the invention provides the following scheme:
a ship rolling motion attitude control system based on multiple learning algorithms comprises: the system comprises a rolling detection device, a ship rolling motion controller, a fin servo system, a single fin and a database;
the rolling detection device is used for detecting target rolling data generated by a target ship in a motion process and sending the target rolling data to the ship rolling motion controller; the target rolling data is a rolling angle of the target ship;
the database is connected with the ship rolling motion controller; the ship rolling motion controller is used for generating a target fin control signal according to the target rolling data and the data in the database and sending the target fin control signal to the fin servo system; the ship rolling motion controller comprises a rolling deep learning neural network identifier and a rolling deep learning neural network sub-controller;
the fin servo system is used for controlling the motion of the monomer fin according to the target fin control signal so as to realize the control of a target ship;
historical roll data and historical fin control signals are stored in the database; the historical rolling data is obtained by inputting original historical rolling data into a ship rolling data classification model, and the historical fin control signal is obtained by inputting an original historical fin control signal into a ship rolling data classification model; the ship rolling data classification model is constructed on the basis of an unsupervised machine learning algorithm.
Optionally, the rolling deep learning neural network identifier is configured to calculate rolling prediction data according to the historical rolling data and the historical fin control signal;
the rolling deep learning neural network sub-controller is used for generating a target fin control signal according to the rolling prediction data, the historical rolling data, the historical fin control signal and the target rolling data.
Optionally, the rolling deep learning neural network identifier and the rolling deep learning neural network sub-controller both adopt a feed-forward back propagation neural network structure based on a deep learning neural network.
Optionally, the rolling deep learning neural network identifier adopts a two-layer BP neural network structure; the input layer of the layer 1 BP neural network structure comprises 4 nodes, the hidden layer comprises 10 nodes, and the output layer comprises 1 node; the input layer of the layer 2 BP neural network structure includes 4 nodes, the hidden layer includes 9 nodes, and the output layer includes 1 node.
Optionally, the rolling deep learning neural network sub-controller adopts a two-layer BP neural network structure; the input layer of the layer 1 BP neural network structure comprises 6 nodes, the hidden layer comprises 11 nodes, and the output layer comprises 2 nodes; the input layer of the layer 2 BP neural network structure includes 6 nodes, the hidden layer includes 7 nodes, and the output layer includes 1 node.
Optionally, the performance index function of the rolling deep learning neural network identifier is a mean square error of an actual rolling angle and an output rolling angle, and the performance index function of the rolling deep learning neural network sub-controller is a mean square error of an actual rolling angle.
Optionally, the rolling deep learning neural network identifier and the rolling deep learning neural network sub-controller both adopt an improved random gradient descent method or an improved negative gradient steepest descent method to correct the network weight.
Optionally, the rolling deep learning neural network identifier and the rolling deep learning neural network sub-controller both adopt a batch normalization algorithm to optimize the network weight.
Optionally, the ship rolling data classification model is a classification model constructed after the rolling data is classified by using a partition-based clustering method fuzzy K-mean method in an unsupervised learning algorithm.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention provides a ship rolling motion attitude control system based on multiple learning algorithms, which comprises a rolling detection device, a ship rolling motion controller, a fin servo system, a single fin and a database, wherein the ship rolling motion controller comprises a rolling deep learning neural network identifier and a rolling deep learning neural network sub-controller; the rolling detection device is used for detecting target rolling data generated by the target ship in the motion process; the ship rolling motion controller is used for generating a target fin control signal according to the target rolling data; the fin servo system is used for controlling the motion of the single fins according to the target fin control signal so as to realize the control of the target ship; the database is connected with the ship rolling motion controller; historical roll data and historical fin control signals are stored in the database; historical rolling data and historical fin control signals are determined by a ship rolling data classification model; the ship rolling data classification model is constructed on the basis of an unsupervised machine learning algorithm. The invention combines a deep learning neural network method with an unsupervised machine learning method, and effectively improves the control effect of the ship rolling motion attitude.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.
FIG. 1 is a block diagram of a ship rolling motion attitude control system based on various learning algorithms according to the present invention;
FIG. 2 is a schematic structural diagram of a ship rolling motion attitude control system based on various learning algorithms according to the present invention;
FIG. 3 is a network structure diagram of a rolling deep learning neural network identifier NNI of the ship rolling motion attitude control system based on multiple learning algorithms;
FIG. 4 is a network structure diagram of a rolling deep learning neural network sub-controller NNC of the ship rolling motion attitude control system based on various learning algorithms;
FIG. 5 is a schematic flow chart of the Batch Normalization (BN) algorithm of the present invention; FIG. 5(a) is a schematic flow diagram of the upper portion of the Batch Normalization (BN) algorithm of the present invention; fig. 5(b) is a schematic flow chart of a lower part of the Batch Normalization (BN) algorithm of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention aims to provide a ship rolling motion attitude control system based on multiple learning algorithms, which combines a deep learning neural network method with an unsupervised machine learning method and effectively improves the ship rolling motion attitude control effect.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
The invention provides a ship rolling motion attitude control system based on multiple learning algorithms, which comprises the steps of firstly, constructing a ship rolling data classification model by using an unsupervised machine learning method; and then training experimental data of ship rolling motion by using a deep belief network DBN (direct binary digit network), wherein each layer in the deep belief network is a limited Boltzmann machine, when unsupervised training is carried out, firstly training a first layer (training according to a standard RBM), then taking a pre-trained hidden node of the first layer as an input node of a second layer, training the second layer, and finally training the whole network by using a BP (back propagation) algorithm. The RBM represents the data as a probabilistic model through learning and can also be used to generate new data once the model is trained or converged to a steady state through unsupervised learning, thus giving an automatic learning result of the rolling motion of the vessel. Compared with the traditional PID control, the self-adaptive control, the robust control and the single neural network method, the algorithm is improved in the aspects of automatic learning capacity, shortened training time, stability of a control system and the like.
As shown in fig. 1, the invention aims to provide a ship rolling motion attitude control system based on multiple learning algorithms, which comprises: the device comprises a rolling detection device, a ship rolling motion controller, a fin servo system, a single fin and a database.
The rolling detection device is used for detecting target rolling data generated by a target ship in a motion process and sending the target rolling data to the ship rolling motion controller; the target roll data is a roll angle of the target vessel.
The database is connected with the ship rolling motion controller; the ship rolling motion controller is used for generating a target fin control signal according to the target rolling data and the data in the database and sending the target fin control signal to the fin servo system; the ship rolling motion controller comprises a rolling deep learning neural network identifier and a rolling deep learning neural network sub-controller. The target fin control signal is a control signal of a fin angle of a single fin (stabilizer).
The fin servo system is used for controlling the motion of the single fin according to the target fin control signal so as to control a target ship through the single fin (stabilizer fin).
Historical roll data and historical fin control signals are stored in the database; the historical rolling data is obtained by inputting original historical rolling data into a ship rolling data classification model, and the historical fin control signal is obtained by inputting an original historical fin control signal into a ship rolling data classification model; the ship rolling data classification model is constructed on the basis of an unsupervised machine learning algorithm.
In addition, the rolling detection device is also connected with the database; the rolling detection device sends target rolling data to the database, and the ship rolling motion controller sends target fin control signals to the database so as to update data stored in the database.
Preferably, the ship rolling data classification model is a classification model constructed by classifying the rolling data by a partition-based clustering method fuzzy K-mean method in an unsupervised learning algorithm.
Wherein the roll deep learning neural network identifier is configured to calculate roll prediction data based on the historical roll data and the historical fin control signal; the rolling deep learning neural network sub-controller is used for generating a target fin control signal according to the rolling prediction data, the historical rolling data, the historical fin control signal and the target rolling data.
The rolling deep learning neural network identifier and the rolling deep learning neural network sub-controller both adopt a feedforward back propagation neural network structure based on a deep learning neural network DLNN.
The rolling deep learning neural network identifier adopts a two-layer BP neural network structure; the input layer of the layer 1 BP neural network structure comprises 4 nodes, the hidden layer comprises 10 nodes, and the output layer comprises 1 node; the input layer of the layer 2 BP neural network structure includes 4 nodes, the hidden layer includes 9 nodes, and the output layer includes 1 node.
The rolling deep learning neural network sub-controller adopts a two-layer BP neural network structure; the input layer of the layer 1 BP neural network structure comprises 6 nodes, the hidden layer comprises 11 nodes, and the output layer comprises 2 nodes; the input layer of the layer 2 BP neural network structure includes 6 nodes, the hidden layer includes 7 nodes, and the output layer includes 1 node.
The performance index function of the rolling deep learning neural network identifier is the mean square error of an actual rolling angle and an output rolling angle, and the performance index function of the rolling deep learning neural network sub-controller is the mean square error of the actual rolling angle.
The rolling deep learning neural network identifier and the rolling deep learning neural network sub-controller both adopt an improved random gradient descent method SGD or an improved negative gradient steepest descent method RMSProp to correct the network weights (described in detail below).
The rolling deep learning neural network identifier and the rolling deep learning neural network sub-controller adopt a Batch Normalization (BN) algorithm to optimize network weights.
According to the technical scheme, the ship rolling motion attitude control system based on multiple learning algorithms provided by the invention adopts a ship rolling motion controller based on a deep learning neural network DLNN, selects a specific ship type and a single fin type and size, can effectively simulate a non-dynamic system, and utilizes a clustering method fuzzy K-mean method based on division in an unsupervised learning method to classify rolling data and construct a classification model. And training the experimental data of the ship rolling motion by using a deep learning neural network DLNN on the basis, and providing an automatic learning result of the ship rolling motion attitude. Compared with the traditional PID control, the self-adaptive control, the robust control and the single neural network method, the system has the advantages that the automatic learning capability, the training time shortening, the control stability and the like are improved.
The ship rolling motion attitude control system based on various learning algorithms described in fig. 1 is described below with reference to specific embodiments.
As shown in fig. 2, the ship rolling motion attitude control system based on multiple learning algorithms has the following action sequence:
(1) obtaining signals output by a measurement system
Figure RE-GDA0002595822810000072
k represents the current time of day and k represents the current time of day,
Figure RE-GDA0002595822810000073
indicating the roll angle at the current time.
(2) Signal for calculating prediction output of NNI sequence of rolling deep learning neural network identifier
Figure RE-GDA0002595822810000074
k +1 represents the next time instant,
Figure RE-GDA0002595822810000075
indicating the predicted roll angle at the next time.
(3) According to
Figure RE-GDA0002595822810000076
And
Figure RE-GDA0002595822810000077
and (4) training a rolling deep learning neural network sub-controller NNC by adopting a deep learning neural network.
Figure RE-GDA0002595822810000078
Indicating the target roll angle at the next time.
(4) According to
Figure RE-GDA0002595822810000079
And
Figure RE-GDA00025958228100000710
and (4) training a rolling deep learning neural network identifier NNI by adopting a deep learning neural network.
(5) The control signal α (k +1) output by the NNC is calculated. α (k +1) represents a control signal of the fin angle at the next time.
The ship rolling motion attitude control system can correct the weights and thresholds of the sub-controllers and the identifier according to the information accumulated by online observation under the condition of perturbation of system parameters or existence of modeling errors, effectively control the system and ensure the operation performance of the ship rolling motion attitude control system.
As shown in fig. 3, the roll deep learning neural network identifier NNI utilizes a two-layer feed-forward back propagation neural network structure, the input layer of the layer 1 BP neural network structure includes 4 nodes, the hidden layer includes 10 nodes, and the output layer includes 1 node. The input layer of the layer 2 BP neural network structure includes 4 nodes, the hidden layer includes 9 nodes, and the output layer includes 1 node. Wherein:
the input layer of the layer 1 BP neural network is composed of:
Figure RE-GDA0002595822810000071
where i denotes the ith input neuron of the layer 1 BP neural network.
The hidden layer of the layer 1 BP neural network is formed as follows:
Figure RE-GDA0002595822810000081
wherein j represents the jth hidden layer neuron of the layer 1 BP neural network;
Figure RE-GDA0002595822810000082
connecting weights between nodes of an input layer and a hidden layer in a layer 1 BP neural network;
Figure RE-GDA0002595822810000083
threshold values of nodes of the hidden layer;
Figure RE-GDA0002595822810000084
is the output of the hidden layer; gi() Which is the excitation function of the hidden layer, α is the fin angle,
Figure RE-GDA00025958228100000811
is a roll angle.
The output layer of the layer 1 BP neural network is composed of:
Figure RE-GDA0002595822810000085
wherein the content of the first and second substances,
Figure RE-GDA00025958228100000812
is the output of the layer 1 BP neural network;
Figure RE-GDA0002595822810000086
connecting weights between nodes of a hidden layer and an output layer of a BP neural network of a layer 1;
Figure RE-GDA0002595822810000087
is the node threshold value of the output layer of the BP neural network of the 1 st layer; f. ofi() Is as followsAnd taking the excitation function of the output layer of the 1-layer BP neural network as a sigmoid function.
The input layer of the layer 2 BP neural network is composed of:
Figure RE-GDA0002595822810000088
where m represents the mth input neuron of the layer 2 BP neural network.
The hidden layer of the layer 2 BP neural network is formed as follows:
Figure RE-GDA0002595822810000089
wherein the content of the first and second substances,
Figure RE-GDA00025958228100000810
connecting weights between nodes of an input layer and hidden layers in a layer 2 BP neural network;
Figure RE-GDA0002595822810000091
a threshold value of each node of the hidden layer;
Figure RE-GDA0002595822810000092
is the output of the hidden layer;
Figure RE-GDA0002595822810000093
is the excitation function of the hidden layer.
The output layer of the layer 2 BP neural network is composed of:
Figure RE-GDA0002595822810000094
wherein the content of the first and second substances,
Figure RE-GDA00025958228100000910
is the output of the layer 2 BP neural network;
Figure RE-GDA0002595822810000095
implicit to layer 2 BP neural networkThe connection weight between each node of the layer and the output layer;
Figure RE-GDA0002595822810000096
a node threshold value of a layer 2 BP neural network output layer; f. ofi() And taking the excitation function of the output layer of the BP neural network of the layer 2 as a sigmoid function.
Selecting a performance index function:
Figure RE-GDA0002595822810000097
i.e. the mean square error of the actual roll angle and the predicted roll angle.
The algorithm for correcting the network weight by adopting the improved random gradient descent method (learning rate attenuation minimum batch gradient descent training method) is as follows:
given dataset X1 ═ X(1),x(2),……,x(n)K training period, b learning rate decay minimum, α initial learning rate (step size)0。
Randomly sampling 10 pieces of data X1 ═ { X ═(1),x(2),x(3),x(4),x(5),x(6),x(7),x(8),x(9),x(10)} ＝{1.35,0.92,0.62,0.42,0.29,0.2,0.1,0.07,0.05,0.03}。
The 10 pieces of rolling data are taken from the experimental data of a ship course/rolling semi-physical simulation system.
The average loss value gradient calculation formula of the layer 2 neural network sampling data is as follows:
Figure RE-GDA0002595822810000098
Figure RE-GDA0002595822810000099
Figure RE-GDA0002595822810000101
Figure RE-GDA0002595822810000102
the calculation formula of the average loss value gradient of the layer 1 neural network sampling data is as follows:
Figure RE-GDA0002595822810000103
Figure RE-GDA0002595822810000104
Figure RE-GDA0002595822810000105
Figure RE-GDA0002595822810000106
α calculation formula of attenuation learning ratei＝(1-i/k)α0+-i/kb；for i＜k；k＞＝100。
The modification formula of the network weight is as follows: w is ai＝wi-αΔw。
As shown in figure 4 of the drawings,
Figure RE-GDA0002595822810000107
in the form of historical roll angle data,
Figure RE-GDA0002595822810000109
for the predicted roll data obtained by the roll deep learning neural network identifier, α (k-1), α (k) are historical fin control signals,
Figure RE-GDA0002595822810000108
for a target roll angle of a roll deep learning neural network controller, an NNC (neural network sub-controller) utilizes a two-layer feedforward back propagation neural network structure, an input layer of a 1 st layer BP (back propagation) neural network structure comprises 6 nodes, a hidden layer comprises 11 nodes, and an output layer comprises 2 nodes; the input layer of the layer 2 BP neural network structure comprises 6 nodesThe hidden layer comprises 7 nodes, and the output layer comprises 1 node; wherein:
the input layer of the layer 1 BP neural network is composed of:
Figure RE-GDA0002595822810000111
the hidden layer of the layer 1 BP neural network is formed as follows:
Figure RE-GDA0002595822810000112
wherein x represents the x input neuron of the NNC layer 1 BP neural network, and y represents the y neuron in the hidden layer of the NNC layer 1 BP neural network;
Figure RE-GDA0002595822810000113
connecting weights between nodes of an input layer and a hidden layer in a layer 1 BP neural network;
Figure RE-GDA0002595822810000114
a threshold value of each node of the hidden layer;
Figure RE-GDA0002595822810000115
is the output of the hidden layer; gi() Is the excitation function of the hidden layer.
The output layer of the layer 1 BP neural network is composed of:
Figure RE-GDA0002595822810000116
wherein the content of the first and second substances,
Figure RE-GDA00025958228100001110
and α (k +1) is the output of the layer 1 BP neural network;
Figure RE-GDA0002595822810000117
connecting weights between nodes of a hidden layer and an output layer of a BP neural network of a layer 1;
Figure RE-GDA0002595822810000118
is the node threshold value of the output layer of the BP neural network of the 1 st layer; f. ofi() And taking the excitation function of the output layer of the BP neural network of the layer 1 as a sigmoid function.
The input layer of the layer 2 BP neural network is composed of:
Figure RE-GDA0002595822810000119
h denotes the h-th neuron of the NNC layer 2 BP neural network.
The hidden layer of the layer 2 BP neural network is formed as follows:
Figure RE-GDA0002595822810000121
wherein the content of the first and second substances,
Figure RE-GDA0002595822810000122
connecting weights between nodes of an input layer and hidden layers in a layer 2 BP neural network;
Figure RE-GDA0002595822810000123
a threshold value of each node of the hidden layer;
Figure RE-GDA0002595822810000124
is the output of the hidden layer; gi() Is the excitation function of the hidden layer.
The output layer of the layer 2 BP neural network is composed of:
Figure RE-GDA0002595822810000125
wherein u (k) is the output of the layer 2 BP neural network;
Figure RE-GDA0002595822810000126
connecting weights between nodes of a hidden layer and an output layer of a BP neural network of a layer 2;
Figure RE-GDA0002595822810000127
a node threshold value of a layer 2 BP neural network output layer; f. ofi() And taking the excitation function of the output layer of the BP neural network of the layer 2 as a sigmoid function.
The selected performance index function is:
Figure RE-GDA0002595822810000128
the modified negative gradient steepest descent method RMSProp is adopted to correct the network weight, and an inertia term ξ which converges to a minimum value is introduced to be 10-7And an attenuation factor hyperparameter β, where β is 0.7 based on empirical values.
The calculation formula of the output layer connecting node weight of the layer 2 neural network is as follows:
Figure RE-GDA0002595822810000129
the calculation formula of the weight of the hidden layer connection node of the layer 2 neural network is as follows:
Figure RE-GDA00025958228100001210
the calculation formula of the output layer connecting node weight of the 1 st layer neural network is as follows:
Figure RE-GDA0002595822810000131
the calculation formula of the hidden layer connection node weight of the layer 1 neural network is as follows:
Figure RE-GDA0002595822810000132
where α is the learning rate.
Accumulating the square of the current gradient in cache, with the aim of eliminating the sign in the gradient, then:
cachei＝β.cachei+(1-β)(Δωi)2；
the correction formula of the network weight is as follows:
Figure RE-GDA0002595822810000133
as shown in fig. 5, the basic idea of the Batch Normalization (BN) algorithm adopted by the present invention is: sampling a small batch of data, and then carrying out normalization processing on the output of the batch of data in each layer of the neural network, namely subtracting the average value of the output value of a certain neuron in a certain layer and dividing the average value by the standard deviation of the output value. Since the mean and variance of the data in the batch cannot replace the mean and variance of all the data, the mean and variance of the whole data are replaced by the mean and variance in the operation, and attenuation factors are introduced to perform attenuation accumulation on the mean and variance, so that the training time of the depth model can be greatly accelerated.
The output values after batch normalization are:
Figure RE-GDA0002595822810000134
wherein the mean of the neuron outputs is:
Figure RE-GDA0002595822810000135
the standard deviation of the neuron output is:
Figure RE-GDA0002595822810000141
and introducing an attenuation factor decade to carry out attenuation accumulation on the mean value and the variance as follows:
μ＝decay·μ+(1-decay)μsample
the innovation points of the invention are as follows:
1. a ship rolling motion control system based on a deep learning neural network DLNN is provided.
2. And classifying the rolling data by using a partition-based clustering method fuzzy K-mean method in unsupervised machine learning to construct a classification model.
3. A neural network identifier and a neural network controller network structure based on the deep learning neural network DLNN are constructed.
4. The experimental data of the ship rolling motion are used for training, and the automatic learning result of the ship rolling motion attitude is given, so that the attitude forecast of the ship rolling motion in the effective time in the future is realized.
The invention provides a ship rolling motion attitude control system based on multiple learning algorithms, which comprises the following steps: the device comprises a rolling detection device, a ship rolling motion controller based on a deep learning neural network and a fin servo system. Compared with the prior art, the ship rolling motion controller based on the deep learning neural network is designed, and the number of the neurons with the activation function is increased by increasing the number of the neurons of the hidden layer of the deep learning neural network, so that the system has strong learning capability. Grouping a large number of parameters by adopting a method of pre-training and fine-tuning, pre-training hidden nodes of each layer unsupervised layer by layer, and training the whole network by utilizing a BP algorithm after pre-training of each layer is finished. The technical scheme effectively saves training overhead and ensures the excellence of the performance of the control system while utilizing the degrees of freedom provided by a large number of parameters of the model. Therefore, the stability of the ship roll angle control is effectively improved by jointly applying various learning algorithms.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other.
The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.

Claims (9)

1. A ship rolling motion attitude control system based on multiple learning algorithms is characterized by comprising the following steps: the system comprises a rolling detection device, a ship rolling motion controller, a fin servo system, a single fin and a database;
the rolling detection device is used for detecting target rolling data generated by a target ship in a motion process and sending the target rolling data to the ship rolling motion controller; the target rolling data is a rolling angle of the target ship;
the database is connected with the ship rolling motion controller; the ship rolling motion controller is used for generating a target fin control signal according to the target rolling data and the data in the database and sending the target fin control signal to the fin servo system; the ship rolling motion controller comprises a rolling deep learning neural network identifier and a rolling deep learning neural network sub-controller;
the fin servo system is used for controlling the motion of the monomer fin according to the target fin control signal so as to realize the control of a target ship;
historical roll data and historical fin control signals are stored in the database; the historical rolling data is obtained by inputting original historical rolling data into a ship rolling data classification model, and the historical fin control signal is obtained by inputting an original historical fin control signal into a ship rolling data classification model; the ship rolling data classification model is constructed on the basis of an unsupervised machine learning algorithm.
2. The system according to claim 1, wherein the rolling deep learning neural network identifier is configured to calculate rolling prediction data according to the historical rolling data and the historical fin control signals;
the rolling deep learning neural network sub-controller is used for generating a target fin control signal according to the rolling prediction data, the historical rolling data, the historical fin control signal and the target rolling data.
3. The system of claim 1, wherein the rolling deep learning neural network identifier and the rolling deep learning neural network sub-controller both adopt a feedforward back propagation neural network structure based on a deep learning neural network.
4. The ship rolling motion attitude control system based on multiple learning algorithms according to claim 1, characterized in that the rolling deep learning neural network identifier adopts a two-layer BP neural network structure; the input layer of the layer 1 BP neural network structure comprises 4 nodes, the hidden layer comprises 10 nodes, and the output layer comprises 1 node; the input layer of the layer 2 BP neural network structure includes 4 nodes, the hidden layer includes 9 nodes, and the output layer includes 1 node.
5. The ship rolling motion attitude control system based on multiple learning algorithms according to claim 1, characterized in that the rolling deep learning neural network sub-controller adopts a two-layer BP neural network structure; the input layer of the layer 1 BP neural network structure comprises 6 nodes, the hidden layer comprises 11 nodes, and the output layer comprises 2 nodes; the input layer of the layer 2 BP neural network structure includes 6 nodes, the hidden layer includes 7 nodes, and the output layer includes 1 node.
6. The system as claimed in claim 1, wherein the performance index function of the rolling deep learning neural network identifier is the mean square error of the actual rolling angle and the output rolling angle, and the performance index function of the rolling deep learning neural network sub-controller is the mean square error of the actual rolling angle.
7. The system of claim 1, wherein the rolling deep learning neural network identifier and the rolling deep learning neural network sub-controller both adopt an improved random gradient descent method or an improved negative gradient steepest descent method to modify the network weights.
8. The system of claim 1, wherein the rolling deep learning neural network identifier and the rolling deep learning neural network sub-controller optimize network weights using a batch normalization algorithm.
9. The system for controlling ship rolling motion attitude based on multiple learning algorithms according to claim 1, characterized in that the ship rolling data classification model is a classification model constructed by classifying rolling data by a partition-based clustering method fuzzy K-mean method in an unsupervised learning algorithm.
CN202010501120.5A 2020-06-04 2020-06-04 Ship rolling motion attitude control system based on multiple learning algorithms Withdrawn CN111619751A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010501120.5A CN111619751A (en) 2020-06-04 2020-06-04 Ship rolling motion attitude control system based on multiple learning algorithms

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010501120.5A CN111619751A (en) 2020-06-04 2020-06-04 Ship rolling motion attitude control system based on multiple learning algorithms

Publications (1)

Publication Number Publication Date
CN111619751A true CN111619751A (en) 2020-09-04

Family

ID=72269120

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010501120.5A Withdrawn CN111619751A (en) 2020-06-04 2020-06-04 Ship rolling motion attitude control system based on multiple learning algorithms

Country Status (1)

Country Link
CN (1) CN111619751A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112800075A (en) * 2021-02-01 2021-05-14 上海海事大学 Ship operation forecast database updating method based on real ship six-degree-of-freedom attitude data

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112800075A (en) * 2021-02-01 2021-05-14 上海海事大学 Ship operation forecast database updating method based on real ship six-degree-of-freedom attitude data
CN112800075B (en) * 2021-02-01 2023-09-29 上海海事大学 Ship manipulation prediction database updating method based on six-degree-of-freedom attitude data of real ship

Similar Documents

Publication Publication Date Title
KR102492318B1 (en) Model training method and apparatus, and data recognizing method
CN108764540B (en) Water supply network pressure prediction method based on parallel LSTM series DNN
Zhang et al. Convergence of gradient method with momentum for two-layer feedforward neural networks
CN111176122B (en) Underwater robot parameter self-adaptive backstepping control method based on double BP neural network Q learning technology
CN111506970A (en) Ship hydrodynamic performance evaluation method
CN113238666B (en) Prediction method of ship motion attitude of GRU (generic routing framework) optimized based on sparrow search algorithm
Pillai et al. Extreme learning ANFIS for control applications
Suryo et al. Improved time series prediction using LSTM neural network for smart agriculture application
CN109165421A (en) Marine shafting bearing negative charge values prediction technique based on genetic algorithm optimization BP neural network
Praczyk et al. Decision system for a team of autonomous underwater vehicles—preliminary report
CN111619751A (en) Ship rolling motion attitude control system based on multiple learning algorithms
Bondarenko et al. Neurons vs weights pruning in artificial neural networks
CN113420508B (en) Unit combination calculation method based on LSTM
CN107273971B (en) Feed-forward neural network structure self-organization method based on neuron significance
JPWO2022024211A5 (en)
CN113626177A (en) CPU resource load condition prediction method based on LSTM
Shrestha et al. Adaptive delay learning in SpikeProp based on delay convergence analysis
Sun et al. A fast algorithm for finding global minima of error functions in layered neural networks
JP6974955B2 (en) Crossbar structure and optimization problem solution search system
CN113836818A (en) Ocean current motion prediction algorithm based on BP neural network prediction model
Sheel et al. Accelerated learning in MLP using adaptive learning rate with momentum coefficient
CN113807005A (en) Bearing residual life prediction method based on improved FPA-DBN
Nagatani et al. Restricted Boltzmann machine associative memory
Ghorrati A New Adaptive Learning algorithm to train Feed-Forward Multi-layer Neural Networks, Applied on Function Approximation Problem
Zhang et al. Short-term tidal level forecasting based on self-adapting PSO-BP neural network model

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WW01 Invention patent application withdrawn after publication

Application publication date: 20200904

WW01 Invention patent application withdrawn after publication