RU2519315C2 - Method for automatic control of heading of ship and intelligent system for realising said method - Google Patents

Abstract

FIELD: physics; control.

SUBSTANCE: group of inventions relates to a method for automatic control of heading of a ship and an intelligent system for automatic control of heading of a ship. The method involves using a neural network model of the control object as a model of the control object. In order to adjust fuzzy logic algorithm parameters to be adjusted, the method involves obtaining and identifying heading data of the ship and control data; determining criteria feature data of movement of the ship based on the identified heading data of the ship and control data using a database of the behaviour of the ship on the heading; selecting a neural network model of the control object based on the determined criteria feature data of movement of the ship; determining data of the fuzzy logic algorithm parameters to be adjusted in accordance with the selected neural network model of the control object. The intelligent system includes a fuzzy logic controller, an emulator for correcting control action, feedforward and feedback between the control object, the fuzzy logic controller and the emulator. The emulator is in form of a neural network emulator. The neural network emulator includes a neural network classifier, a neural network model unit of the control object and an optimisation unit. The neural network classifier includes a database of the behaviour of the ship on the heading.

EFFECT: high-speed control processes and efficiency of filtering random perturbations of the system.

14 cl, 6 dwg, 3 tbl

Description

The group of inventions relates to methods and systems for automatic control of a vessel along the course, namely, to a method and intellectual system for controlling the vessel along the course, combining the advantages of neural network systems and control systems based on fuzzy logic.

All existing automatic control systems (ACS) of the vessel on the course, regardless of the design of individual elements, operate on the principle of deviation, i.e. in the system, the actual and set course values are continuously compared and a control signal is generated. Under the influence of this signal, the steering gear shifts the steering wheel and returns the vessel to the set course. In its general form, an ACS vessel at a course includes means for controlling the vessel (i.e., the steering device of the vessel), a regulator (also called autopilot), a control object (i.e., the vessel), and internal and external feedbacks between these elements.

At present, adaptive self-propelled guns on board the course, implemented on the basis of proportional-differential (PD) or proportional-integral-differential (PID) controllers, are widely used. An example of such a system is a steering system, including an automatic adaptive controller in the form of a PID controller [patent RU 2282884 C2, publ. 08/27/2006]. However, the effectiveness of controlling the ship's heading using adaptive systems based on PD and PID controllers depends mainly on the initially accepted mathematical model of the control law, which by no means can always correspond to the actual navigation conditions, therefore such systems cannot provide the optimal operating mode as with automatic stabilization of the ship on course, and during maneuvering. In addition, such systems have low noise immunity when working in rough seas and do not allow adaptive procedures in real time.

At present, adaptive self-propelled guns are being replaced by intelligent automatic control systems (ISAU) for the course of the ship, which have advanced capabilities to perform more complex tasks in an uncertain environment with incomplete information required to solve the course control tasks of the ship. The ISAU vessel at the rate includes systems that use the mathematical apparatus of the theory of artificial neural networks (NS), based on fuzzy logic, based on expert systems, based on genetic algorithms, or a combination thereof.

For example, a ship autopilot is known, including a neural network controller based on the MATLAB software environment [patent US 6611737 B1, publ. 08/26/2003]. This neural network controller contains an information-analytical unit for predicting and generating the ship's heading, connected by data transfer protocols with a heading information gathering unit and a generalized heading control signal output unit. Here, the neural network controller solves other problems of controlling the ship and contains many other blocks that are not directly related to the course of the ship, so the architecture of the used NS is not optimal in terms of training time, ease of implementation and quality of approximation for applied tasks of controlling the course of the ship along the course.

ISAU is known for a vessel at a heading made on the basis of a neural network controller, including an information-analytical unit for processing data to predict and develop a ship's course, which is connected by data transfer protocols with a data collection unit to collect information about the ship's heading and a processed data output unit for output a generalized signal for controlling the course of a vessel [patent RU 2359308 C2, publ. 06/20/2009]. The neural network controller of this system additionally includes two or more computing systems in the form of independent computing algorithms, each of which is connected to the information-analytical unit via an interface and has the ability to process data transmitted by the information-analytical unit, and a knowledge base connected to the information-analytical block interface and designed to store selection criteria informational-analytical unit of a particular computing system. Each of the computing systems is capable of processing data arrays based on such neural networks as Elman's recurrent neural network, multilayer direct distribution network or neural network based on radial basis functions. Here, the information-analytical unit also includes an intelligent classifier connected by an interface to the data collection unit and the processed data output unit and capable of selecting one or more computing systems for optimal data processing.

Despite the many advantages of the ISAU described above, the vessel is heading with a neural network controller in the direct connection of the system, for example, the ability to predict changes in input information, maintaining operability with partial failure of elements and connections, self-learning, flexibility, associativity, absolute and fast convergence to one of the stable states under the action of an input signal falling into the region of attraction of this state, nevertheless, such systems also have disadvantages determined by m, that neural networks are not universal, ie, To reprogram the neural network controller to solve a new problem, it is necessary to virtually re-create the internal architecture of its neural network. And although such a drawback can be overcome with the help of appropriate training of a neural network, the number of types of neural network models of the dynamics of vessel motion in the ISAU at the heading is limited by the performance of the computing architecture of the whole system.

Meanwhile, the above-mentioned shortcomings are deprived by the ISAU vessel on the course, using fuzzy logic controllers, the main advantages of which are the increased speed of control processes and the high efficiency of filtering random disturbances in processing information. Currently, civilian vessels are already using serial-produced ISAU vessels at the heading, based on fuzzy logic, for example, autopilot models Navipilot AD II from Sperry Marine and NAVpilot 500 from Furuno.

As a prototype, one of the last course-designed vessels developed by ISAU was selected, using the course-control method for the vessel using the fuzzy logic algorithm [application CN 102183957 A, publ. September 14, 2011]. The following operations are performed in this system: setting the course value of the vessel, determining the course deviation as the difference between the actual course value obtained from the feedback data and the set course values and determining the degree of change in the course deviation; setting the input data of course deviation and deviation of the degree of change of course deviation, generating a new deviation of the course and deviation of the degree of change of course deviation, transferring a new course deviation and deviation of the degree of change of course deviation, serving as two input variables for fuzzy variable management, obtaining a fuzzy control rule, required to control the course in accordance with the rules of fuzzy control, the output of the control signal in the form of data of the angle of masonry pen p A and transfer these to the steering apparatus; obtaining a sequence of time data of the actual heading values, masonry angles of the rudder and the degree of change in the course deviation in the form of an input vector for training a network structure based on the support vector method, dynamic identification of the inverse model of the ship’s course heading and the generation of a compensation signal in direct connection; as well as tracking compliance with a given course in accordance with a given control accuracy under the control of the controller on fuzzy logic using a block that implements the support vector method. However, in spite of the increased speed of control processes and the high efficiency of filtering random disturbances inherent in such systems, this system is significantly limited in its adaptability to specific swimming conditions, so that fuzzy sets are set here initially, i.e. they do not change, but are only used to make a decision by the system, and the trained network structures used are of little use for tuning the controller to fuzzy logic under changing swimming conditions, because are selected using the support vector method, i.e. using a linear classifier. Because of this, such a system is rather difficult to adapt to changes in swimming conditions, which in most cases is a non-linear task.

The technical task of the present invention is the creation of such an intelligent ship control system along the course and the corresponding method, which would simultaneously provide both high speed control processes and high filtering efficiency of random disturbances inherent in fuzzy logic systems, as well as the ability to predict changes in input data, the ability to save performance with partial failure of elements and connections, self-learning and flexibility inherent in systems with neural networks, for quick adaptation with high accuracy to any swimming conditions and, in particular, to sharp changes in swimming conditions.

To solve the technical problem, in accordance with one aspect of the present invention, there is provided a method for automatically controlling a vessel in accordance with a course, according to which: a course value of a vessel as a control object is set, control error data is determined from the feedback from the control object as a deviation between the actual and predetermined values course; determine the control effect in the form of data on the angle of masonry pen rudder according to the error control using the fuzzy logic algorithm with a set of fuzzy rules; identify the model of the control object based on the network structure according to the feedback data; and adjust the control action in accordance with the identified model of the control object. What is new is that they use the model of the control object in the form of a neural network model of the control object, and the adjustment of the control action is carried out by setting up customizable parameters of the fuzzy logic algorithm.

In the feedback data for processing using the fuzzy logic algorithm, it is preferable to include at least control data and actual heading data. In such data, the value of the actual linear speed of the vessel may be additionally included.

The feedback data for processing using the fuzzy logic algorithm can include data selected from the group: value of the rudder angle of the rudder, value of the yaw angle of the vessel, rate of change of the yaw angle of the vessel, number of rudder shifts and / or rate of change of the rudder feather masonry.

To configure the fuzzy logic algorithm's tunable parameters, the following operations can be performed: receive and identify vessel movement data at heading and control data; identify the vessel’s heading at the heading and the data of the control actions and determine the data of the criteria for the motion of the ship according to the identified heading information and the data of the control actions using the knowledge base of the ship’s course choose a neural network model of the control object on the basis of certain data of the criterial signs of the movement of the vessel; determine the data of the custom parameters of the fuzzy logic algorithm in accordance with the selected neural network model of the control object.

Here, the identification of the received heading data and the data of control actions can be carried out by analyzing the spectrum of these data using the discrete Fourier transform.

At the same time, when analyzing the spectrum of the data on the ship's heading and heading data, two to eight harmonics of the spectrum can be used.

It is preferable to create and train a knowledge base of vessel behavior along the course using a neural network with radial-base activation functions.

The neural network model of the control object can be selected from several neural network control models from the group: direct distribution neural networks with Levenberg-Marquardt learning algorithm, direct distribution neural networks with Moller learning algorithm, Elman recurring neural networks with Levenberg-Marquardt learning algorithm and / or recurrent neural network Elman with Moller's learning algorithm. Here, it is preferable to use a direct distribution neural network with the Levenberg-Marquardt learning algorithm.

The data of the adjustable parameters of the fuzzy logic algorithm can be determined using the genetic algorithm.

The method can be used the algorithm of fuzzy logic in the form of a set of linguistic conditional operators that determine the specific situation of the vessel's control according to the course, which are preliminarily determined on the basis of expert judgments on two fuzzy variables, which use data of control errors and data of the rate of change of control errors.

In this case, linguistic conditional operators can be formed into a matrix using the Gaussian function and the sigmoid membership function.

In accordance with another aspect of the present invention, there is provided an intelligent automatic course control system for a ship, comprising: a ship as a control object; a fuzzy logic controller for determining the control action on the control object by a control error, as a deviation between the actual and set course values set in the direct circuit of the system; an emulator for adjusting the control action by identifying the model of the control object based on the network structure according to feedback data; and direct and feedback between the control object, the controller on fuzzy logic and the emulator, included in the direct and reverse circuits of the system, respectively. What is new is that the emulator is installed in the reverse circuit of the system and is made in the form of a neural network emulator, in which a neural network control model is used as a model of the control object, and which is made with the possibility of adjusting the control action by adjusting the controller's adjustable parameters to fuzzy logic.

The vessel can be equipped with means for laying the rudder pen, means for setting the heading value and means for determining the vessel's motion data along the heading, while: one input of the controller on fuzzy logic is connected to the outputs of the means for setting the heading value and means for determining the vessel motion data for the heading for receiving error data management; the other input of the controller on fuzzy logic is connected to the output of the neural network emulator to obtain data on customizable parameters of the controller on fuzzy logic; the controller output on fuzzy logic is connected to the input of the masonry tool steering wheel for transmitting control data; one input of the neural network emulator is connected to the output of the means for determining the vessel’s motion at the heading for receiving the ship’s motion data at the heading; the other input of the neural network emulator is connected to the output of the controller on fuzzy logic to obtain control data.

A neural network emulator can include a neural network classifier, a block of a neural network model of a control object and an optimization block, while: the input of a neural network classifier is an input of a neural network emulator for receiving vessel movement data along the course and data of control actions from a control object and controller on fuzzy logic, respectively; the output of the neural network classifier is connected to the input of the block of the neural network model of the control object for transmitting data of the criterion characteristics of the vessel’s movement, determined by the neural network classifier according to the vessel’s heading and heading data; the output of the block of the neural network model of the control object is connected to the input of the optimization block for transmitting data of the parameters of the neural network corresponding to the neural network model of the control object loaded into the block of the neural network model of the control object based on the criterion characteristics of the vessel; the output of the optimization block is the output of a neural network emulator for transmitting to one of the inputs of the controller on fuzzy logic the data of the controller's adjustable parameters on the fuzzy logic determined by the optimization unit according to the loaded neural network model of the control object.

Preferably, when the neural network classifier contains a course knowledge base of a ship’s course for determining the criteria for a ship’s motion according to the course of the ship’s course and control data.

The invention is further explained in more detail with reference to the accompanying drawings, in which:

FIG. 1 is a functional block diagram of an intelligent automatic control system (ISAU) of a vessel at the heading of an embodiment of the present invention;

FIG. 2 is a graph of the root mean square error (RMS) of the control versus the type of neural network used;

FIG. 3 is a graph of the root mean square error (RMS) of training neural networks on navigation conditions for a trawler;

FIG. 4 is a graph of changes in control accuracy during training of neural networks for various types of vessels;

FIG. 5 is a graph of the dependence of the integral component on the number of harmonics for a coastal passenger passenger vessel for various sailing conditions;

FIG. 6 is a graph illustrating the adaptation of a trawler-type vessel at a 10 deg course with a sharp change in wind speed and wave height using an ISAU vessel at the heading shown in FIG. one.

The present invention is illustrated by the example of the implementation of the intelligent automatic control system (ISAU) of the vessel at the rate.

As shown in FIG. 1, an ISAU vessel at a heading rate in accordance with an embodiment of the present invention generally includes a vessel as a control object (OS) 1, a fuzzy logic controller (RNL) 2, and a neural network emulator (NSE) 3.

OS 1 in particular is equipped with a means for laying the rudder pen for laying it at an angle α (for example, a steering machine), means for setting the heading value K _З (for example, heading unit) and means for determining the vessel's movement data along the course (for example, gyrocompass, lag), such as the actual value of the heading K and the value of the linear speed of the vessel, which are included in the data vector of the actual course of the vessel V _y .

RNL 2 serves to generate a control action (α) on the masonry tool of the rudder in accordance with the adopted control algorithm based on fuzzy logic and specified control quality based on input data in the form of a control error e, and RNL 2 parameters configured with using NSE 3. Here, the control error e is defined as the deviation between the actual heading value K received from the means for determining the vessel’s movement data at the heading and the set heading value K _З received from the heading means but.

NSE 3 is used to configure customizable parameters of RNL 2 by determining the vector of customizable parameters V _r based on the data vector of control actions V _α received from the output of RNL 2 and the vector of data of the actual heading of the vessel V _y received from means for determining the data on the movement of the vessel along the course.

In this example, NSE 3 includes the following blocks: neural network classifier (NSC) 4, neural network model of the control object (NSM OU) 5 and optimization block (BO) 6. The device and principle of operation of blocks 3-6 will be explained below.

The ISAU-proposed vessel at the rate provides for the consistent implementation of the following main modes: data preparation, training of neural networks NSE 3, functioning.

Data preparation mode

The regime for preparing vessel course data consists of three related processes: generating source data vectors, identifying source data vectors, and creating working data arrays.

The formation of the initial data vectors V _α and V _y takes place according to the following scheme: ship heading data is collected using the input interface into the source vectors, each of which is a collection of ship heading data, yaw rate, rudder masonry, rudder speed rudder feather, average ship speed.

The identification of the initial yaw characteristics data vectors of the vessel consists in assigning identification marks P _{j to them} . Identification marks correspond to the unique properties of these vectors. The process of identifying the vectors V _α and V _y consists in performing the corresponding calculations on their elements in the NSC 4. Here, the physical principles of identifying the source data vectors are analyzing the data spectrum obtained using the discrete Fourier transform and measuring the statistical characteristics of the signal from OA 1.

The process of creating working arrays of data on the ship's heading consists in establishing a correspondence to each knowledge base formed in long-term memory, which is provided in NSC 4, the initial data vector of the actual heading of the ship Yj of the identification mark vector Pj. As a result of this, a working array of data on the behavior of the vessel at the rate {Y, P} j is formed, after which further operations are performed on them, such as training and / or processing, depending on the operating mode of the ISAU vessel at the course. The array is a knowledge base created so that in the future, when the same or similar statistical characteristics of the behavior of the vessel at the heading appear, according to the vector of identification marks P _j, select a ready-made (i.e. trained) NS as a neural network model of the vessel in block 5 to the HCM DU obtained via SFN 4 vector criterial signs V _a vessel for further defining movement via BO 6 parameters, as a vector V _r adjustable controller parameters for adjusting RNL 2 which correspond to selected th neural network model of the ship.

Training mode

In the training mode, the weighting coefficients of the connection matrices of the NS of a certain architecture, formed in advance with the activation function, are changed. The purpose of the training is to configure the NSM OS 5 neural network of weighting coefficients w _{jq of} its matrix of connections W in the form of a vector of neural network (input) parameters between input X _j and output Y _j data vectors of NS for a defined set of all possible combinations of label vector P _j , which is achieved by searching for the best characteristics of the used NS architecture: the data training algorithm, the algorithm for processing the vessel’s course data, the number of NS layers, the number of neurons in each NS layer.

After training the NS, any input vector X _j belonging to the set of training pairs, but not coinciding with its values, will correspond to the output vector Y _j uniquely. Thus, trained NSs are ready to process heading data.

Functioning mode

When the vessel moves, the ISAU vessel along the course collects the data of the vectors V _α and V _y , in particular from signals such as the yaw angle of the vessel, the rate of change of this angle, the number of rudders, the rate of change of the rudder feather masonry. Then there is a spectral analysis of these vectors V _α and V _y with the selection of several harmonics for each of the above signals, plus the value of the linear velocity of the vessel received from the means for determining the vessel's motion data at the heading is added to the data of the vector V _y , whereby the motion data is obtained the vessel at the heading under current external influences F. The collected data in the form of vectors V _α and V _y are fed to the input of NSC 4, which according to the data of these vectors determines which of the neural networks available in the NSM OS 5 is more just fits this data, after which it loads the most suitable NS into the OSM OU 5 as a neural network model of the vessel. According to the selected neural network model of the vessel, BO 6 loads the optimal RNL 2 parameters corresponding to this model in the form of a vector V _r . T.O. The main task of BO 6 is to configure RNL 2 according to the neural network model of the vessel NSM OU 5, selected using NSC 4, with current external influences F, i.e. under specific operational and weather conditions, the vessel is moving on course.

Further, the operating principle and principles of synthesis of individual elements of the ISAU vessel at the course will be explained in more detail using examples.

RNL 2 is functionally a set of linguistic conditional operators or fuzzy associations that define specific situations of course control of a vessel. Here in RNL 2 the following parameters are used: K - output signal (actual ship heading); To _Z - signal at the input (given course of the vessel); e - deviation between the given and the actual vessel at the heading; e? - the rate of change of error e (the rate of departure of the vessel off course); u - controller output.

Two fuzzy variables e and e? are presented in the form of nine fuzzy subsets: from zero ZE to a high positive value of PL and a high negative value, as presented in table 1.

Table 1

Matrix of fuzzy ship heading control associations

The matrix of fuzzy associations (rules) for the ISAU vessel presented in Table 1 is compiled based on expert judgment, where each group of elements in the matrix gives one fuzzy association (rule) that indicates how the control variable u should be changed for the observed values of the input fuzzy variables e and e ?.

As an example, we give an interpretation of the rule (PL, ZE, PL) in natural language. If the error e between the set and actual values of the course K _W and K is positive and large, and the rate of change of the error e? close to zero, the control signal (control action) to the steering machine should be maximum.

Here, for fuzzy subsets of the matrix of fuzzy associations of two fuzzy variables e and e? the Gauss function and the sigmoid membership function are selected.

A fuzzy subset of fuzzy variables ZE (values close to zero) are chosen narrower than others. This improves the accuracy of control near to the value of the course to the _W and increase the robustness of the course ISAU ship.

Based on a priori heuristic considerations, it was accepted that continuous fuzzy subsets in each of the sets overlap by about 20-30% percent. If the overlap is too large, differences between the values corresponding to different subsets are lost. With too little overlap, there is a tendency to “two-digit” control, which leads to a deterioration in the quality of the process (ambiguity of the solution, large overshoot, etc.). In real conditions, the overlap allows smoothing the transition from one control action to another in the course of the ISAU vessel operation at the heading.

The calculation of the functioning parameters of RNL 2 can be represented as the following algorithm:

- calculate (measure) the output value of the process y (t) (i.e., the dependence of K on time);

- calculate the error e (t) and the rate of change of the error e? (t) for time t;

- conduct the formation of the corresponding fuzzy subsets by quantizing the quantities e (t) and e? (t);

- calculate on the basis of the formed membership functions the degree of membership of the error and the derivative of the error M (e _i ) and M (e? _i );

, вычисляемыми по степеням принадлежности посылок M(e_i) и M(e?_i) с помощью нечеткой логической операции «И»;- activate from the matrix of rules of conclusion with degrees of affiliation

calculated by the degrees of membership of the premises M (e _i ) and M (e? _i ) using the fuzzy logical operation "AND";

до величины

методом кодирования по минимуму корреляции;- “cut off” the value

up to

coding method to minimize correlation;

путем выполнения нечеткой логической операции «ИЛИ» между значениями

- всех активизированных заключений;- calculate the actual output degree of ownership

by performing a fuzzy logical OR operation between values

- all activated conclusions;

- determine the output of the fuzzy controller u by a discrete analogue of the centroid method.

Those. here a simplified algorithm of fuzzy inference can be applied according to the following form of writing predicate rules: P _i : if e is a _i1 and e? is a _i2 , then c = z _i , where z _i are real numbers, and a _ij are fuzzy numbers from the Gauss functions and sigmoid functions.

When implementing RNL 2, as an algorithm for finding the optimal values of custom parameters, i.e. parameters of predicate rules, using BO 6, it is preferable to use a genetic algorithm in BO 6, which avoids stopping the algorithm at a local extremum and does not depend on the dimension of the problem. Evaluation of the above parameters of RNL 2 should be carried out according to the value of the objective function, which, as one of the options, can reflect the minimum standard error (RMS).

Below is a brief example of a genetic algorithm implemented using BO 6 and allowing one to determine the data of the vector of tunable parameters V _r based on the data of the vector of parameters of the neural network W obtained using the NSM OU 5.

Step 10. Randomly generating 54 adjustable RNL 2 parameters with 50% uniform distribution relative to the starting value for ten possible settings.

Operation 20. Determination of the management effectiveness criterion for each RNL 2 tuning option on the neural network model of the vessel according to the data of the vector W.

Operation 30. Strengthening the criteria for RNL tuning options that exceed the average criterion with β> 1; weakening the criteria for the settings of RNL 2, which have a value below the average criterion, using 0 <β <1, using the following formulas:

, где n=720;

where n = 720;

;K _av =

;

=K_i (K_i - K_cp)β, если K_i>K_cp;

= K _i (K _i - K _cp ) β if K _i > K _cp ;

=K_i (K_i · K_cp)β, если K_i<K_cp.

= K _i (K _i · K _cp ) β if K _i <K _cp .

Operation 40. Randomly choosing depending on the size of the criteria of ten RNL tuning options.

не станут равными между собой или параметры РНЛ 2 для всех десяти вариантов не станут одинаковыми.Operation 50. Randomly mixing among themselves the corresponding parameters of the selected RNL tuning options 2. Repeat operations 20-40 until all the criteria

will not become equal or RNL 2 parameters for all ten options will not be the same.

Operation 60. Carrying out the mutation process, which consists in a stochastic change in part of the chromosomes. Moreover, each gene of the line that undergoes a mutation, with some low probability, changes to another gene. In binary coding during mutation, the values of individual bits are reversed.

Operation 70. After receiving a quasi-optimal solution (exit operations 20-40), the solution is thinned by the random search method using the adaptation procedure of the distribution of the search step:

- changing the direction of the search, taking into account the history of the search for dependencies

,

,

- количество параметров настройки, k - шаг итерации,

- оптимальное значение вектора,

, 0<B₂<B₁<1 - коэффициенты учета предыстории;where i =

is the number of settings, k is the iteration step,

is the optimal value of the vector,

, 0 <B ₂ <B ₁ <1 - coefficients of accounting history;

- change in the value of the random step (the procedure allows us to reduce the value of the random step (variance) as we approach the optimum point).

Depending on the information obtained during the search at each iteration, the value x _{0 is} determined that minimizes the objective function:

,

,

where random settings are generated taking into account the probability

P (x) = exp [- (xx ₀ ) / 2σ ² ], where

,

,

where 0 <g <1, g is a parameter that determines the decrease in variance at an unsuccessful step.

- отношение текущей дисперсии к первоначальной.Thus, the search is localized in the vicinity of the optimal point. The condition for the end of the stage is

- the ratio of the current variance to the original.

The data of the vector of parameters of the neural network W as input for the above-described genetic algorithm is obtained from the NSM OU 5 block according to the neural network model of the vessel, selected on the basis of the vector of the vessel motion criteria vector V _A , determined using the NSC 4.

The vessel’s neural network models used in the NSM OS 5 were obtained using the IS-2005 simulator of the Engineering Center for Information and Control Systems CJSC. Using this simulator, six different ship models were modeled (coastal passenger vessel (PSPP), trawler, transport refrigerator (TR), car-passenger vessel (APP), tanker, supertanker (ST)), the characteristics of which are given in Table 2.

table 2

Vessel characteristics

The following abbreviations are used in table 2: VFSh - fixed pitch propeller; VRS - adjustable pitch propeller.

For each of the vessels, simulation of the following sailing conditions was carried out at a speed of 8.0-15.6 knots: No. 1 - wind speed of 1-2 m / s, wave height of 3% security 0.25 m; No. 2 - wind speed 2-5 m / s, wave height 0.85 m; No. 3 - wind speed 5-8 m / s, wave height 1.25 m; No. 4 - wind speed of 8-10 m / s, wave height of 1.25. For each of the trajectories of the vessel’s movement, for various operational factors, 79 neural networks were trained that differ in type, architecture, and learning algorithm.

Based on the results of the assessment of the standard deviation, eight NSs were selected, which make it possible to obtain high-quality neural network models of vessels for all studied operating conditions (see Fig. 2). In FIG. 2, the following notation of the selected NS is used: I-III - direct distribution neural networks with the Levenberg-Marquardt learning algorithm; IV - direct distribution neural network with the Moller learning algorithm (a combination of the conjugate gradient method with the quasi-Newtonian approach); V-VII - Elman recurrent neural networks with the Levenberg-Marquardt learning algorithm, VIII - Elman recurrent neural networks with the Moller learning algorithm. For all types of NS 100, 500 and 1000 training cycles were carried out.

Analysis of the simulation results showed that it is preferable to use a direct distribution neural network with the Levenberg-Marquardt learning algorithm as the vessel’s neural network model.

A computer simulation was also carried out to determine the dependence of the standard deviation of training of neural networks on the above navigation conditions for various vessels, the characteristics of which are presented above in table 2. Analysis of the simulation results showed that the standard deviation of training of neural networks is practically unchanged. For example, as shown in FIG. 3, the standard deviation of training neural networks for a trawler at a speed of 8, 12, and 15 knots varies slightly. This once again confirms the possibility of using the theory of neural networks as a neural network model of vessel movement at the heading for the NSM OS 5. The analysis of the simulation results showed that due to the worsening conditions of navigation (increase in wind speed from 1-2 to 8-10 m / s , increase in wave height from 0.25 m to 1.25 m) for TR (15 knots), tanker (15 knots), SSPP (8 knots), APP (15 knots), ST (15, 6 knots) RMS values NA increase, but remain generally insignificant.

Additionally, computer simulation was carried out to determine the dependence of the standard deviations of the training of neural networks on the types of vessels, the characteristics of which are presented in table 2. The analysis of the simulation results showed that the standard deviations of the training of the NS, simulating the movement of vessels along the course, have no characteristic dependencies.

T.O. different vessels in different sailing conditions require different neural network models simulating the course of the vessel.

To determine the sufficient number of training cycles for neural networks simulating the course of the vessel, an additional study was conducted. Neural networks, from the above, were selected the best for a particular vessel in the indicated navigation conditions. For example, in FIG. Figure 4 shows a graph of the decrease in the standard deviation of the training for the NS for six vessels, the characteristics of which are given in table 2.1, under sailing conditions No. 1, the analysis of which shows that for all six vessels a significant decrease in the standard deviation of training occurs up to 400 training cycles, then there is a slight adaptation. Similarly, it was determined that the number of training cycles is significant 450 (for vessels under conditions No. 2), 350 (for vessels under conditions No. 3) and 300 (for vessels under conditions No. 4). Also, analysis of the simulation results for various vessels in different sailing conditions showed that the significant number of training cycles is 300 (for vessels of the tanker and trawler type), 350 (for vessels of the APP and PSPP type) and 450 (for vessels of the ST and TR type). Thus, the analysis of the studies showed that for various vessels and different navigation conditions 450 cycles of training of any type of neural network are sufficient.

The vector data of the criteria for the motion of the vessel V _A as input to select and load the optimal neural network model of the vessel from several models, for example, the above models, into the NSM block 5 OU 5, obtained from the NSC 4 according to the feedback in the form of a vector of actual ship heading data V _y and the data vector of the control actions V _α .

To identify the course behavior of the vessel along the course (especially the effect of changes in the laying of the rudder feathers, weather factors, etc.), NSC 4 uses a spectral analysis of the course of the vessel along the course (vector V _y ) with the corresponding curve of commands for changing the position of the rudder (vector V _α ). For spectral analysis, the discrete Fourier transform was used.

To study and identify a sufficient number of harmonics for spectral analysis, an algorithm was developed for the special NSC 4 module to restore the ship's motion curve to eight harmonics, and determine the corresponding integral component from other harmonics. For each model of the vessel, after carrying out a spectral analysis and obtaining eight maximum harmonics, the integral criterion I of the possibility of using fewer harmonics to identify the neural network model of the vessel in the NSC 4 knowledge base is calculated. Here, the main signal Y ₁ consists of the sum of eight components, and the compared signal Y ₂ consists of the sum of the investigated number of components:

;

;

;

;

,

,

; ω - наименьшая частота из заданных гармоник; N - количество введенных гармоник. В таблице 3 приведены значения интегральных составляющих относительно количества гармоник для всех типов судов, приведенных в таблице 2.Where

; ω is the lowest frequency of the given harmonics; N is the number of introduced harmonics. Table 3 shows the values of the integral components with respect to the number of harmonics for all types of vessels, shown in table 2.

Table 3

The values of the integral components

After computer simulation on the IS-2005 simulator with six ship models under four different weather conditions and discrete Fourier transform, label vectors were obtained that characterize the amplitudes of the eight fundamental harmonics when the ship heading, yaw rate, rudder angle and average linear speed (65 elements in total for each vector). An analysis of the data showed a pronounced dependence of the integral component of the error I on the number of harmonics in modeling the movement of the vessel. If for large vessels it is possible to limit ourselves to two or three harmonics, then for vessels of small displacement eight harmonics are required. Also, from the analysis of the obtained results, it was concluded that the spectral characteristics of the models of ships change with changing weather conditions, which is especially typical for small tonnage vessels such as PSPP and trawler. For example, in FIG. 5 is a graph of the dependence of the value of the integral component I on the number of harmonics for the AMS for various swimming conditions (No. 1-4, described above). Studies for all considered ship models showed that the error in modeling various types of ships depends on the number of harmonics involved. If in small-tonnage vessels an almost linear dependence is observed, then in more inertial large-tonnage vessels it is non-linear, therefore it is possible to use fewer harmonics.

For the classification of ship motion curves by criteria, such as spectral characteristics, it is preferable to use neural networks with radial basis activation functions (RBF).

96 examples were used to teach RBF NS. Data was read from files, scaled and entered into an array.

Next, clustering was performed according to the “k-means” method of each class into six clusters using the “kmeans” function, that is, in each cluster there were 24 objects. The vectors of the centroids of the clusters were recorded in a special array. The next step was to determine the deviations of the Gaussian function for each centroid of the clusters, for which the most accurate and effective method of “k-nearest neighbors” was used. The bell radius of each core was calculated as the average distance to k = 3 (it was experimentally established that the network error is the smallest at k = 3) of the nearest neighboring centroids and recorded in an array. The result of such training are weights - vectors of centroids and threshold layer levels - deviations from the array. As a result, the responses of the Gauss functions were obtained for each of the six types of ships. The result of the experiments and training of NSC 4 was a knowledge base for 24 neural network models of the movement of various vessels under various sailing conditions.

Functional testing of methods and algorithms synthesized according to the principles described above was carried out in two stages. The first stage consisted of computer simulation to fill the NSC knowledge base with 4 ship models under various sailing conditions. The second stage was to conduct simulation using an IS-2005 simulator for a model of a ship of a certain type, at different speeds and changing sailing conditions in order to verify the adaptation process of RNL 2.

Computer simulation was carried out for four conditions of external influences, indicated above as No. 1-4. An analysis of the results showed that NSE 3 requires additional tuning in bad weather conditions, which confirms the validity of including NSC 4 with the knowledge base of NS with the best identification of the vessel's heading along the course under various sailing conditions (in particular, Nos. 1–4). The adaptation of RNL 2 carried out by NSE 3 by adjusting the parameters of the latter led to new values of the tuned parameters of RNL 2.

After a series of modeling experiments, the NSC 4 knowledge base was stored in the NSC 4 knowledge base, for example, 12 neural network models of the trawler were saved for different weather conditions and different speeds. As an example in FIG. Figure 6 shows the adaptation process for a vessel of a trawler type with a sharp change in weather, in particular from navigation conditions No. 1 to navigation conditions No. 4, due to the selection of the optimal ones from the number of saved neural network models. Analysis of the adaptation data shows that the root-mean-square integral criterion for the deviation of the vessel from the course decreased from 2.43 (in the first three minutes) to 0.16 (in the last 3 minutes), and this criterion was optimized under the condition that the steering machine operates no more than 2 / 3 of the study period, which significantly exceeds the corresponding indicators of known analogues (for example, according to technical solutions RU 2359308 C2 and CN 102183957 A, described above).

When adapting RNL 2, other target functions are possible, allowing to minimize the load on the steering machine, subject to the conditions of high-quality keeping the ship on course. For example, if you reduce the requirements of the operating time to 1/2, then the load on the steering machine will decrease, the operating time of the steering machine will decrease, but the quality of keeping the vessel on course will worsen, while the integral criterion in this case will be 0.31, which still allows significantly exceed the characteristics of known analogues.

It should be understood that the above examples are used only to illustrate the feasibility of the present invention and a number of its advantages, and these examples do not limit the scope of legal protection presented in the claims, while a specialist in this field is relatively easy to implement other variants of the invention without departure from the essence of the invention within the scope of legal protection.

Claims (14)

1. A method of automatically controlling a ship at a heading, including the following operations, sets the heading value of the ship as a control object, determines data of control errors as a deviation between the actual and set course values according to feedback from the control object, determines the control action in the form of pen laying angle data the rudder of the vessel according to the control error using the fuzzy logic algorithm with a set of fuzzy rules, identify the model of the control object based on the network structure according to direct communication, and adjust the control action in accordance with the identified model of the control object, while the adjustment of the control effect is carried out by setting up custom parameters of the fuzzy logic algorithm, characterized in that a neural network model of the control object is used as a model of the control object, while for setting up custom the fuzzy logic algorithm parameters perform the following operations receive and identify the vessel movement data at the heading and control data actions, identify the vessel’s heading at the heading and the data of the control actions and determine the data of the criteria for the ship’s motion using the identified heading and the data of the control actions using the knowledge base of the ship’s course behavior, select a neural network model of the control object based on certain criteria signs of movement of the vessel, determine the data of the adjustable parameters of the fuzzy logic algorithm in accordance with the selected neural network model about EKTA control. 2. The method according to claim 1, characterized in that the feedback data for processing using the fuzzy logic algorithm include at least data of the control actions and data of the actual value of the course. 3. The method according to claim 2, characterized in that the value of the actual linear speed of the vessel is additionally included in the feedback data for processing using the fuzzy logic algorithm. 4. The method according to claim 2, characterized in that the feedback data for processing using the fuzzy logic algorithm includes data that is selected from the group: value of the rudder angle, value of the yaw angle of the vessel, rate of change of the yaw angle of the vessel, number of rudders and / or rate of change of masonry feather steering. 5. The method according to claim 1, characterized in that the identification of the received data on the vessel's heading along the course and the data of the control actions is carried out by analyzing the spectrum of these data using a discrete Fourier transform. 6. The method according to claim 5, characterized in that when analyzing the spectrum of the data on the vessel’s heading and heading data, two to eight harmonics of the spectrum are used. 7. The method according to claim 5, characterized in that the course knowledge base of the vessel’s behavior is created and trained using a neural network with radial basis activation functions. 8. The method according to claim 1, characterized in that the neural network model of the control object is selected from several neural network control models from the group: direct distribution neural networks with the Levenberg-Marquardt learning algorithm, direct distribution neural networks with the Moller training algorithm, Elman recurrent neural networks with Levenberg-Marquardt learning algorithm and / or Elman's recurrent Neural network with Moller learning algorithm. 9. The method according to claim 8, characterized in that as a neural network model of the control object, a direct distribution neural network with the Levenberg-Marquardt learning algorithm is used. 10. The method according to claim 1, characterized in that the data of the custom parameters of the fuzzy logic algorithm is determined using the genetic algorithm. 11. The method according to claim 1, characterized in that they use a fuzzy logic algorithm in the form of a set of linguistic conditional statements that determine the specific course of the vessel’s course control, which are preliminarily determined based on expert judgment on two fuzzy variables, which use these control errors and control error rate of change data. 12. The method according to claim 11, characterized in that the linguistic conditional operators are formed into a matrix using the Gaussian function and the sigmoid membership function. 13. Intelligent system of automatic control of the vessel along the course, including the vessel as a control object, a fuzzy logic controller for determining the control effect on the control object by a control error, as a deviation between the actual and set course values, installed in the direct circuit of the system, an emulator for adjusting the control impacts by identifying the model of the control object based on the network structure according to feedback data, direct and feedback between the control object, a fuzzy logic emulator and an emulator included in the forward and reverse circuits of the system, respectively, characterized in that the emulator is installed in the reverse circuit of the system and is made in the form of a neural network emulator, in which a neural network control model is used as a model of the control object, and which is executed with the ability to configure custom controller parameters on fuzzy logic according to feedback data, where the neural network emulator includes a neural network classifier, a block of a neural network model of the control object I and the optimization unit, while the neural network classifier contains a knowledge base of the ship’s course behavior to determine the criteria for the ship’s movement according to the course of the ship and the data of control actions, the input of the neural network classifier is the input of the network simulator to receive the data of the ship’s course and control data effects from the control object and the controller on fuzzy logic, respectively, the output of the neural network classifier is connected to the input of the block of the neural network model of the object and the control for transmitting data of the criterion characteristics of the vessel’s movement, determined by the neural network classifier according to the vessel’s heading and heading data, the output of the neural network model block of the control object is connected to the input of the optimization block to transmit the data of the neural network parameters corresponding to the neural network model of the control object loaded in the block of the neural network model of the control object on the basis of the criterion signs of the vessel’s motion, the output of the optimization block is the output of the neuro etevogo emulator to transmit at one of the inputs of the fuzzy logic controller to the data controller configurable parameters for the fuzzy logic block certain optimization of a neural network model of the control object loaded therein. 14. The system according to item 13, wherein the vessel is equipped with a means of laying the rudder, means for setting the heading value and means for determining the vessel’s movement along the course, with one fuzzy logic input connected to the outputs of the means for setting the heading value and means for determining vessel motion data for heading to receive control error data, another fuzzy logic input of the controller is connected to the output of a neural network emulator to receive fuzzy logic controller settings, output is a fuzzy logic controller is connected to the input of the rudder feather masonry tool for transmitting control data, one input of the neural network emulator is connected to the output of the vessel motion detection data for the course to receive the vessel movement data along the course, the other input of the neural network emulator is connected to the controller output on the fuzzy logic to obtain control data.

Images