RU2821616C1 - Self-training system for intelligent agents of ship engine control - Google Patents

Abstract

FIELD: physics.

SUBSTANCE: ship engine intelligent control agents self-training system comprises engine units and assemblies simulation unit based on fuzzy neural network, intelligent agent unit, training unit, fuzzy classifier formation unit, control signal generating unit, control pulse parameters matching unit, fuzzy knowledge base, ship engine convergent network, connected in a certain manner.

EFFECT: providing development of own solutions for control of ship engine taking into account internal and external disturbing effects of environment.

1 cl, 4 dwg

Description

The invention relates to systems for intelligent control of ship engines with the ability to digitally calculate or process data using fuzzy logic, in particular, for the procedure for collecting, accumulating, transmitting and centralizing diagnostic and functional parameters of ship engines.

At the heart of the new generation of engines are intelligent agents that allow subsystems to communicate with the propellers to the engine room and engine control room, as well as to the bridge or even remote objects. An intelligent agent is an agent that acts in an intelligent manner, that perceives its environment, autonomously takes actions to achieve goals, and can improve its performance through learning or knowledge acquisition. Intelligent agents effectively integrate control of starting air valves, start and reverse sequences, control functions and others. These highly integrated systems make it possible to monitor and increase visibility of ship installations, as well as provide real-time engine health monitoring, support intelligent optimization and facilitate decision-making. All this enables ship operators and engine suppliers to operate onboard systems more efficiently. The invention solves problems associated with the intellectualization of an automatic control system for ship power plants.

Today, in such areas as economics, medicine, and the automotive industry, diagnostic and expert systems using “artificial intelligence” based on neuro-fuzzy networks are widely used.

Known KNOWLEDGE MANAGEMENT SYSTEM FOR RESOLVING SITUATIONS under patent RU 2480826 dated 04/27/2013

The invention relates to knowledge management systems for resolving situations (KMS PC) and is intended to support the resolution of problem situations associated with the unsatisfactory quality of specific objects. The technical result is to improve the characteristics of the processed information, as well as to improve the quality of visualization of situations. The knowledge management system for resolving situations contains a knowledge creation block, a knowledge organization block, to the input of which the output of the knowledge creation block is connected, a knowledge localization block, to the input of which the output of the knowledge organization block is connected, a knowledge positioning block, to the input of which the output of the knowledge localization block is connected, and a knowledge reuse block, the inputs of which are connected to the outputs of knowledge localization and positioning blocks. A system integration block, the inputs and outputs of which are connected, respectively, to the inputs and outputs of the blocks of knowledge creation, knowledge organization, knowledge localization, knowledge positioning, knowledge reuse, and a system visualization block, the input of which is connected to the output of the system integration block.

The disadvantages of this technical solution are the lack of an independent connection between the knowledge management system and the environment.

Known INTELLIGENT INFORMATION SYSTEM OF SELECTION "OPTIMEL" under patent RU 2564641 dated 10.10.2015

The invention relates to computer systems that use knowledge-based models, namely to systems that synthesize intelligent solutions in the form of selecting the desired knowledge from a given field of knowledge. The technical result is to provide accelerated access to generated knowledge for the user when solving selection problems. The intelligent information selection system "Optimel" consists of a sequentially connected machine-readable medium, which contains a task distribution block, a knowledge base block, a knowledge input block, a dictionary database block, a working part of the dialogue block, and a computer, through the technical capabilities of which the user has access to intellectual information system "Optimel" and its results.

The disadvantage of implementing this technical solution is the need for manual replenishment of the database by an expert.

There is a known METHOD FOR SEARCHING METHODS FOR RESOLVING TECHNICAL CONTRADICTIONS AND A SYSTEM BASED ON A LEARNING NEURAL NETWORK FOR ITS IMPLEMENTATION under patent RU 2707917 dated 12/02/2019, where the system contains a module for receiving a system user request, a module for recognizing the interface language with a user request, a module for creating an algorithm for generating a request code , module for generating the request code, module for generating the database structure corresponding to the request code, module for generating priorities for training the neural network, module for additional training of the neural network using parameters 1-n corresponding to the request code, neural network, module for generating the response code, module for translating the response code into language interface, module for issuing a response to the user.

The system works as follows. The request is entered through the request receiving module in the form of interface text. The next, sequential module recognizes the interface language with the user's request, in the following modules an algorithm is created for generating the request code and the request code. Then the signal goes to the database structure generation module, and the next module generates priorities for training the neural network. Additional training of the neural network using the parameters corresponding to the request code is carried out using the previous module. Next, the neural network searches for methods for resolving technical contradictions using the principles of morphological analysis and fuzzy logic. The method of morphological analysis is based on combinatorics. Its essence lies in the idea of obtaining a detailed description of all existing and possible (admissible) technical systems of the class under study, followed by a search in this set for a description of the technical system that most fully corresponds to the goal. To do this, a group of basic technical and/or social and/or economic characteristics is identified in the product or object of interest. For each characteristic, alternative options are selected, that is, possible versions of the object. By combining them, many different solutions are obtained, including those of practical interest. Previously, the neural network forms stable relationships between the characteristics of the object corresponding to the request code and its technical and/or social and/or economic characteristics in the form of layering of pseudo-planes, learns from the resulting relationships and searches for a solution to overcome a particular technical contradiction or technical problem and clarifies it in various ways filters, the direction of their solution, then the neural network establishes the relationship between the subject under study and its mention in patent databases and each set is formed in its own information space, after which, using the morphological design method, a set of matrices is formed, each of which is formed by ordered searches for exhaustive and finite a list of technical and/or social and/or economic characteristics taken into account when applying technical solutions with establishing the type of dependency between them, after which, based on the recorded relationships, they work with clear and fuzzy logic, create a ranking system, which is the basis for training the neural network and the basis of material for generating a detailed response to the request. Next, a response code is generated, which is translated into the interface language, and the output module provides a response to the user.

The disadvantage of this system is the use of enumeration of options available in the knowledge base, instead of a solution formed from the current state of the environment.

The TECHNICAL PROBLEM of the proposed invention is to create a self-learning system for intelligent agents for controlling a ship's engine, ensuring the development of its own solutions, adequate to the environment, for controlling a ship's engine.

The TECHNICAL RESULT lies in the fact that the constructed intelligent engine control system, through the use of self-learning of an intelligent agent, forms its own solution based on internal and external disturbing influences of the environment.

The technical result is achieved by using a set of distinctive features of the proposed system and the interconnection of its blocks and nodes.

To solve the set technical problem and achieve a technical result, a self-learning system for intelligent ship engine control agents is proposed.

THE ESSENCE OF THE INVENTION is as follows.

This system allows for monitoring, analysis, diagnosis of the technical condition and management of the entire complex dynamic system, consisting of main and auxiliary elements of the control system, elements of load support systems, elements of fire safety systems,

The block structure of the system provides a procedure for collecting, accumulating, transferring and centralizing diagnostic and functional parameters of the elements of the control system. In real operation conditions of the vessel, in any mode provided for the type of vessel being operated (sailing, maneuvering, movement in ice, etc.), information is continuously collected from all sensors installed on the ship's power plant, as well as from sensors reflecting external operating conditions (outside temperature, humidity, wind speed and direction, sea waves, etc.).

All information from the sensors is supplied to the engine components and assemblies modeling unit, where it is processed and sorted according to their belonging to the subsystems and components of the ship's power plant and synchronously, through several parallel channels, transmitted to the intelligent module.

Unlike a conventional agent, which reacts to external influences by searching through the rules available in the knowledge base, which are not replenished independently, an intelligent agent contains blocks that allow it to independently develop new knowledge in the course of its work.

The overall marine engine control architecture includes three parts: 1) plant/subsystem models that can simulate the characteristics of the propulsion system and its drives, 2) individual distributed control system algorithms, 3) a multi-agent knowledge base that can support the exploration and application of advanced intelligent control strategy . The various models and controls that make up it can be easily replaced because their interfaces will be clearly defined by the interdependency map. The purpose of this intelligent platform is to create an integrated modeling and simulation environment that integrates various interdependent plant and subsystems of a marine engine, in addition, this platform can monitor the states of major components and systems and based on this develop advanced intelligent control strategies.

Information about the state and changes in the environment, about the current operating mode of the engine and other key parameters of the ship's power plant and external environment is distributed into several streams using the interface. The knowledge base is designed to accumulate information about engine control. The main element of the self-learning system is a fuzzy classifier, the task of which is to generate a control signal for an intelligent agent.

The essence of the invention is illustrated by graphic materials that show:

Fig. 1 - self-learning system of an intelligent agent FIG. 2 - marine engine control diagram FIG. 3 - diagram of the intelligent agent FIG. 4 - converged ship network architecture

THE SYSTEM CONSISTS of a modeling block 1 of engine components and assemblies based on a fuzzy neural network, an intelligent agent block 2, a training block 3, a fuzzy classifier generation block 4, a control signal generation block 5, a control pulse parameters matching block 6, a fuzzy knowledge base 7, convergent network of the ship engine 8. The output of the modeling block 1 is connected to the intelligent agent block 2, and the input of the modeling block 1 is connected to the ship engine parameter sensors and elements of the ship's unified convergent network 8, and the output of the intelligent agent 2 is connected to the training block 3 and to the input of the fuzzy base knowledge 7, where for self-training of the neural network they use the possibility of fast learning based on deep trust networks and a limited Boltzmann machine, while the location of the output of the training block 1 provides the possibility of receiving a signal to the fuzzy classifier generation block 4, which is interconnected with the fuzzy database block 7, as well as the receipt of a signal to the control signal generating unit 5, which is interconnected with the control pulse parameters matching unit 6, installed with the ability to receive information from the sensors of the ship engine 8.

The Intelligent Engine Control System OPERATES AS follows.

The intelligent agent block receives 2 current information from the modeling block of engine components and assemblies 1, reflecting the parameters of their operation. Intelligent agent 2 analyzes data on the state of parameters affecting engine operation, after which it generates control signals that are most appropriate in the current situation, which are fed through the interface to the knowledge base 7 for subsequent verification of their adequacy and efficiency. At the same time, the signal is sent to the fuzzy classifier generation block 4.

Training the fuzzy classifier 4 consists of finding the vector K that minimizes the distance between the experimental data and the results of logical inference.

The fuzzy classifier is formed from several layers D, each of which has a width of N _i neurons.

Self-learning of the network consists of finding the correct weights that can be used to correctly classify input examples.

In this case, fast learning based on deep belief networks (DBN) and restricted Boltzmann machine (RBM) is used to self-train the neural network.

The method is based on changing the artificial neural network model from discriminant to generative. A discriminant model is a model that models the classification characteristics of a network. The objective function that needs to be minimized is the error between the required classification targets and the resulting targets. On the other hand, a generative model is a model that models the ability of a network to generate input data. The goal is to minimize the error between the data generated by the model and the original data.

The generative model is able to regenerate the original data given the states of the hidden units, which is its representation of the real world data. This model is called a deep trust network (DBN). The DBN model allows the network to generate visible activations based on the states of its hidden units, which represents the network's belief.

To obtain the states of the hidden blocks corresponding to the visible data, a restricted Boltzmann machine is used between every two successive layers of the network.

To optimize a given configuration of visible and hidden blocks, the energy of the model must be minimized.

The function of pre-training is to apply the data to the input layer of the two-layer RBM and obtain the hidden activations as a result, and then regenerate the visible activation model, and finally generate the hidden activation model. In this way, the model's weights can be updated for given input data.

After training the current layer, its weights are frozen and the hidden layer's activations are used as visible inputs to the next layer, and the same training algorithm is applied. The resulting deep learning network weights are used to initialize the fine-tuning stage.

The fine-tuning phase is a conventional backpropagation algorithm. For classification tasks, a layer is added on top of the network, the width of which is equal to the number of targets or classes. Each neuron in this layer is activated for each class label, and the rest are deactivated. Backpropagation starts with the weights obtained in the pre-training phase. Top layer activations are obtained for each example of the training set or batch of examples obtained in the forward path, and then the error signal between the obtained activations and the required targets is passed back to the network to adjust the weights.

The self-learning process consists in the fact that during engine operation, an intelligent agent periodically polls the values of sensors and controllers at the input and position signals of control systems at the output. It then compares the current value of the controls for this rule with the values in the knowledge base. Next, the knowledge base section “required position of controls” is adjusted taking into account new experience gained during operation and in accordance with which ship engine control impulses are formed.

The ultimate goal of the self-learning process is to analyze the characteristics of marine engine installations and systems, configure an algorithm for optimizing its operating parameters, and ultimately develop an intelligent automation and operation management strategy.

Claims (1)

A self-learning system for intelligent agents for controlling a ship engine, including a block for modeling engine components and assemblies based on a fuzzy neural network, an intelligent agent block, a training block, a fuzzy classifier generation block, a control signal generation block, a block for matching control pulse parameters, a fuzzy knowledge base, a convergent network ship engine, while the output of the modeling block is connected to the intelligent agent block, and the input of the modeling block is connected to the parameter sensors of the ship engine and elements of the ship's unified convergent network, the output of the intelligent agent is connected to the training block and to the input of the fuzzy knowledge base, where for self-learning of the neural network use the possibility of fast learning based on deep trust networks and a limited Boltzmann machine, the location of the output of the training block provides the possibility of receiving a signal to the fuzzy classifier generation block, which is interconnected with the fuzzy database block, as well as receiving a signal to the control signal generation block, which is interconnected with a block for matching control pulse parameters, installed with the ability to receive information from ship engine sensors.