RU198734U1 - Marine device for determining the source of ignition by a multicriteria fire detector using a neural classifier - Google Patents

Abstract

A ship-based device for determining the source of fire by a multicriteria fire detector using a neural classifier (MFI) is designed to transmit a signal about the type of fire source to the microcontroller of an analog-to-digital converter. for ship premises. To build a neural classifier, a programming language, a Sequential classification model and a learning library were used The path of training a neural network is provided through a three-layer neural network with 12 neurons in two hidden layers, a heterogeneous set of activation functions - sigmoid for the first layer and ReLU for the second. The configuration of the third layer of the neural network with predictive functions forms classes of fire sources, containing three neurons in the output layer. The technical result is an increase in the reliability of identifying the fire source, which makes it possible to quickly and more efficiently localize and extinguish a fire in the premises of a ship or ship. 2 h. p. f-ly, 3 ill.

Description

The utility model relates to fire detection equipment in the premises of a ship, used to identify the source of ignition.

A known system for detecting fire hazard and fire in a submarine compartment (RU 2598782, G08B 17/00, G08B 25/14, 22.10.2014), which carries out comprehensive control and analysis of the physical parameters of monitored objects and factors affecting the occurrence and development of a fire hazardous situation ...

Also known is the "Fire hazard control system" (RU No. 26430 U1, A62C 37/00, A62C 37/10, 04.07.2002), which refers to fire safety systems and is designed to prevent a fire, mainly at various objects of increased fire and explosion hazard, in particular on ships and ships.

The disadvantage of these systems is that they only warn of a fire. This information is not enough to make a decision on the organization of fire containment and extinguishing processes, and therefore it is necessary to conduct additional reconnaissance.

The proposed utility model can be used in modern fire alarm systems (SPS) with multicriteria fire detectors (MPI), therefore the closest technical solution (prototype) to the claimed one is the MPI type "Bark ME" of the complex of technical means of fire automatic equipment "Gamma-01 "Produced by OOO NPO Fire Automation Service (www.npo.pas). The specified MPI monitors the change in the temperature of the air in the room, the degree of smoke and the percentage of carbon monoxide.

The disadvantage of the MPI used in the ATP is that they are not very informative and do not allow determining the type of ignition source.

Various types of combustible loads can ignite in the ship's premises: fuel, oils, rags, insulation of cable lines of the electric power system, etc. At the same time, based on the statistical significance, for a particular room, the number of probable fire sources is small and in most cases ranges from 3 to 5. Therefore, for ship premises for various purposes, it is possible to determine in advance many possible fire sources. Let us agree to call such sources of fires typical sources of ignition.

The utility model is aimed at achieving a technical result, which consists in increasing the reliability of identifying the source of a fire, which makes it possible to quickly and more efficiently localize and extinguish a fire in the premises of a ship or ship.

To achieve the specified technical result, a shipboard device for determining the source of ignition of the MPI is proposed, which contains channels for measuring fire parameters, such as measuring temperature, optical density, carbon monoxide and flame, the outputs of which are connected to the inputs of the microcontroller of an analog-to-digital converter, while the device is equipped with a memristor of a neural classifier, the inputs of which are connected to the outputs of the channels for measuring the parameters of the fire, and the output is connected to the additional control input of the microcontroller of the analog-to-digital converter.

Private signs are:

- in the neural classifier, the neural network learning path is provided through a three-layer neural network with 12 neurons in two hidden layers, a heterogeneous set of activation functions - sigmoid for the first layer and ReLU - for the second.

- the configuration of the third layer of the neural network with predictive functions forms the classes of sources of fire, containing three neurons in the output layer.

This proposal is based on the analysis of the test results of Bark ME (Expert) combined addressable analogue fire detectors, as well as an infrared flame detection camera under conditions of exposure to test fires (according to GOST 53325-2012).

When performing typical fire tests, a program for visualizing the operation of fire detectors was used, which displayed the dynamics of changes in monitored parameters and recorded their relative values when the multicriteria fire detection algorithm was triggered. To recognize the type of ignition source, a neural network was used, to the inputs of which the signals of the triggered MPI were fed.

This solution makes it possible to provide an increase in the information content of a fire alarm system with an FDI, thereby quickly and efficiently localizing and extinguishing a fire.

The essence of the utility model is illustrated by diagrams.

The figure 1 shows a diagram of a ship's device for determining the source of ignition by a multicriteria fire detector using a neural classifier.

Figure 2 shows a diagram of a routed analog-to-digital converter based on a memristor with a neural classifier.

Figure 3 shows a block diagram of a neural classifier.

A ship's device for detecting a source of ignition by a multicriteria fire detector using a neural classifier (Fig. 1), which includes: a carbon monoxide measurement channel containing an electrochemical gas sensor 1 and an operational amplifier 2; a channel for measuring optical density containing a photodiode 3, a current-voltage converter 4, an IR diode 5, an IR diode current stabilizer 6; a temperature measurement channel containing a thermistor 7 and a resistive divider 8; a flame measurement channel containing an IR receiver 9, a unit consisting of a resistor, a capacitor and a potentiometer 10, a comparator 11; the change signals are fed to the memristor of the neural classifier 12, which provides digital information to the ADC microcontroller 13.

Diagram of a routed analog-to-digital converter based on a memristor with a neural classifier (Fig. 2), consisting of three main blocks: a block of comparators 1, a block-calculator 2 of the required bit width for an individual ADC (IADC) for each of the inputs, as well as the neural network itself 3. The block of comparators is responsible for comparing the analog voltage received at a specific input with the balancing voltage generated by the HC for this input signal. The scheme for calculating the required bit depth sets the number of bits in the generated IADC for each of the input signals for the NS, which will determine the speed and accuracy of obtaining the result. The basis of the entire ADC is the NS itself, which forms the IADC of a given bit depth, which forms the count value based on the information coming from the output of the corresponding comparator.

In the structural diagram of the neural classifier (Fig. 3), using a multi-layer neural network of direct propagation (multilayer perceptron), consisting of: a three-layer neural network with 12 neurons in two hidden layers 1 and 2, a heterogeneous set of activation functions - sigmoid for the first layer and ReLU for the second. The configuration of the third layer 3 solves the problem of the device - at the output layer we obtain the result of determining (each of the 3 neurons of the output layer) the classes of fire sources.

The algorithm of the neural classifier was considered as an example.

Let the values of the factors be determined - xi , where i is the number of the MPI factor. Among the MPI factors, the following are distinguished: flame ( i = 1 ); smoke content,% ( i = 2 ); temperature, ºС ( i = 3 ); carbon monoxide concentration, ppm ( i = 4 ). The output values are yj , where j is the number of the group of ignition sources. Among the sources of ignition are identified: spray or spilled oil, fuel ( j = 1 ); destruction of thermal insulation or ignition of oily rags ( j = 2 ); sparking of switchboard equipment or short circuit in cable routes ( j = 3 ).

To build a neural classifier, the Python programming language , the Sequential classification model and the Keras learning library were used. Let's create two hidden layers 1, 2 and one output layer 3 of the neural network:

model = Sequential ()

model.add (Dense (12, input_dim = 4, activation = 'sigmoid')))

model.add (Dense (12, activation = 'relu')

model.add (Dense (3, activation = 'sigmoid'))

The first numbers transmitted by Dense are the numbers of neurons experimentally optimized by varying the structure of the neural network. Changing the number of hidden layers 1, 2 and the neurons contained in them allows to achieve a better predictive quality of the neural network classification model.

In order to solve the categorization problem, before starting training the model, it is compiled using the compile method, where "binary_crossentropy":

model.compile (loss = "binary_crossentropy", optimizer = "adam",

metrics = ['accuracy'])

The fit () method is used to train the neural network:

model.fit (X, Y, epochs = 1000, batch_size = 10)

Next, the result of training the neural network is evaluated. The evaluate () method returns the values of the loss function and metric for the trained model:

scores = model.evaluate (X, Y)

According to the training results, the accuracy metric took the maximum value of 1, which means that the training was successful.

Then the trained neural network is applied to the resulting data that were not involved in the training. In three experiments, the model was able to verify the cause of the fire with an accuracy of 94.89%, 99.88% and 86.52%.

The use of the claimed utility model in the ATP with DF allows real-time recognition of the type of fire source. The device is able to quickly learn new, previously unknown types of fire sources.

Claims (3)

1. A ship-based device for determining the source of ignition by a multicriteria fire detector containing channels for measuring fire parameters, such as measuring temperature, optical density, carbon monoxide and flame, the outputs of which are connected to the inputs of the microcontroller of an analog-to-digital converter, characterized in that it is equipped with a memristor of a neural classifier , the inputs of which are connected to the outputs of the channels for measuring the parameters of the fire, and the output is connected to the additional control input of the microcontroller of the analog-to-digital converter. 2. The device according to claim 1, characterized in that in the neural classifier the neural network learning path is provided through a three-layer neural network with 12 neurons in two hidden layers, a heterogeneous set of activation functions - sigmoid for the first layer and ReLU - for the second. 3. The device according to claim. 2, characterized in that the configuration of the third layer of the neural network with predictive functions forms the classes of sources of fire, containing three neurons in the output layer.

Images