RU2815322C2 - Method of locating fire source using multilayer recurrent perceptron - Google Patents

Abstract

FIELD: fire-fighting equipment.

SUBSTANCE: present invention relates to methods of ensuring fire safety in premises of fire hazardous facilities. Technical result is achieved by determining the location of the fire source using a multilayer recurrent perceptron, at that, to measure fire factors, multi-parameter fire sensors of a fire alarm system are used, which are placed in a room using a genetic algorithm, all measurements of temperature, optical density, concentration of carbon monoxide of air and coordinates of location of multiparameter fire sensors are entered into multilayer recurrent perceptron, wherein the multilayer recurrent perceptron is trained by the Levenberg–Marquardt method and the multilayer recurrent perceptron is validated using fire simulation, and then using the trained neural network, the location of the fire source is determined.

EFFECT: reducing the error of determining the location of the fire source with reliability to the zone of the fire source and reducing the time spent by the equipment on detecting the fire source.

1 cl, 5 dwg

Description

The present invention relates to methods for ensuring fire safety in the premises of fire-hazardous objects, which serve to determine the location of the fire with accuracy up to the fire zone and can be used in fire safety information support systems for faster and more efficient localization and elimination of the fire.

The “Fire Hazard Monitoring System” is known (RU No. 26430 U1, A62C 37/00, A62C 37/10, publ. 12/10/2002), which includes sensors interconnected by loops connected to the control panel, characterized in that it additionally contains a central device and group devices, sensors are made in the form of analog sensors, each of which contains an addressable module, while the sensors are connected to each other and to group devices by ring loops so that each loop is equipped with a separate group device, and each sensor is equipped with a cutter for short-circuited sections of the corresponding loop , while the central device is connected to group devices with a control panel.

Also known is the “Method for automatic detection of the initial stages of fires in the premises of fire-hazardous objects containing heat-generating equipment” (RU No. 2646204, G08B 17/00, G08B 17/12, publ. 03/01/2018), which consists in cyclically determining the values of controlled parameters of fire hazards , comparing the results obtained with the specified maximum permissible values and signaling a fire if these values are reached, characterized in that in order to detect a fire at the initial stage of its development, the entire set of physical parameters controlled at the facility are combined into groups, each of which characterizes the processes of occurrence and the development of a fire in a specific room, while each monitored parameter in the group is assigned a serial number and, at each control cycle, numerical series are formed from the obtained parameter values in accordance with their serial numbers, which are stored and, in subsequent control cycles, compared with the newly formed numerical series, which, in case of a discrepancy, is also remembered for subsequent comparison, while the memorization of new numerical series is carried out during the initial period of training and self-tuning or periodically during the time during which changes in the monitored parameters reach the maximum possible values, but do not exceed the maximum permissible values, and after the end of the period training and self-tuning, when a numerical series appears that differs from those stored, they signal the occurrence of the initial stage of a fire in the room.

Also known is the “Fire detection method and intelligent control station for implementing the method” (RU No. 2344859, A62C 3/00, G08B 17/00, publ. 01/27/2009), which consists of constantly identifying fire hazard factors in the environment of the controlled object, transforming the identified factors into an array of digitized data, comparing this data array with an array of a priori data, classifying the obtained comparison results in accordance with extrema and generating, depending on the hazard class, a control signal. The intelligent control station contains a housing with an aspiration device, a processor, a control device, an actuator assembly and a fire extinguishing channel arranged in series, wherein the aspiration device consists of an inlet pipeline, a temperature sensor, a fan, a coarse dust filter and a fine dust filter located in series. , sensor module, exhaust pipe, and the processor contains a block of analog-to-digital converters, a block of functional measurements and correlations, a control and programming block

As a significant drawback of these methods, it is necessary to indicate the need to perform a large complex of preliminary work on the analysis of each controlled object, the lack of optimization of determining the coordinates of the location of the fire and optimization of the location of sensors, and the lack of automatic adaptation of the system to the parameters of the object.

The closest technical solution of the claimed invention is the procedure for measuring fire parameters with multi-criteria and fire sensors of the fire alarm system, presented in the description of the patent for the utility model “Ship device for determining the source of fire with a multi-criteria fire detector using a neural classifier” (No. 198734, G08B 19/00, publ. 01/27/2009) designed to transmit a signal about the type of fire source to the microcontroller of an analog-to-digital converter.

The disadvantage of multi-criteria fire detectors is that they produce a low-informative discrete signal obtained using threshold processing, which does not allow the source to be localized with sufficient accuracy and the type of fire source to be determined.

The technical result of the claimed invention is to increase the reliability of determining the location of the fire source with accuracy up to the fire source zone, which makes it possible to quickly and efficiently determine the fact of a fire, localize and eliminate the source of the fire in the premises of fire-hazardous objects.

The technical result of the claimed invention is achieved in the proposed method for determining the location of a fire using a multilayer recurrent perceptron (MRP), using the results of measuring fire factors, in which, unlike the prototype, to measure fire factors, multi-parameter fire sensors of the fire alarm system are used, which are placed in indoors using a genetic algorithm, all measurements of temperature, optical density, carbon monoxide concentration in the air and the coordinates of the location of multi-parameter fire sensors are entered into the MRP, training and validation of the MRP is carried out using fire modeling. In this method, instead of the threshold signals of multi-criteria sensors, continuous digital signals of multi-parameter sensors are used, which makes it possible to take into account the correlation dependencies between parameters - fire factors, the introduction of the received digital information from the sensors into the MRP allows the accumulation of data that trains the neural network, and is also used for the optimal placement of multi-parameter sensors genetic algorithm.

To determine the coordinates of the location of the fire, an MRP with three layers is used, containing eight and five neurons in two hidden layers, two neurons in the output layer with a set of activation functions - a hyperbolic tangent and a learning method with backpropagation of the error in time to all layers of the neural network. Activation function - hyperbolic tangent is an everywhere differentiable activation function and is suitable for backpropagation networks. Compared to other common activation functions, this function shows a lower error rate.

Validation of the neural network is carried out by simulating a room with an arrangement of fire sources. Fire simulation is carried out using a Fire Dynamics Simulator (FDS), which creates a digital twin of the fire. A digital fire twin has a number of advantages over a full-scale fire model: safety, low cost, large fire coverage of premises.

The neural network is presented in the form of a multilayer perceptron.

The combination of the essential features listed above increases the reliability of identifying the source of the fire.

The illustrations accompanying the description show:

Figure 1 – general diagram of the interaction of fire safety systems, consisting of: 1 – block for collecting information from multi-parameter sensors, 2 – multi-parameter sensors, 3 – information collection system, 4 – control unit, 5 – information system.

Fig. 2 – block diagram of the MRP, consisting of: 6 – signals from the output of multi-parameter sensors, 7 – delay block, 8 – input layer, 9 – first hidden layer, 10 – feedback of the first hidden layer, 11 – second hidden layer, 12 – feedback of the second hidden layer, 13 – output layer, 14 – feedback of the output layer.

Figure 3 is a graph of MRP training results with input delays, where 15 is a training graph, 16 is a validation graph, 17 is a testing graph.

Figure 4 is a diagram of the location of fire sources in the simulated room, where 18, 19, 20, 21, 22 are fire sources.

Figure 5 is a diagram for determining a fire zone, where 18, 19, 20, 21, 22 are fire sources, 23, 24, 25, 26, 28 are fire sources, with dots, colors corresponding to the numbers of fire sources, showing the results of the neural network according to determining the location of the fire.

A method for determining the location of a fire using a multilayer recurrent perceptron is presented in the general diagram of the interaction of fire safety systems (Fig. 1), which shows a block diagram of a fire protection system, which includes a block for collecting information from multi-parameter sensors 1, consisting of multi-parameter fire sensors 2 and an information collection system 3, control unit 4 and information system 5, which contains a neural network unit.

The neural network block is shown in Figure 2. The neural network is a multilayer recurrent perceptron with input delays. Signals from the output of multi-parameter sensors 6 pass through a delay block 7. The delay block 7 forms at the output at a discrete point in time a measurement vector of q fire factors measured by one multi-parameter sensor, , at the moments of time Input layer 8 determines the vector of signals with delays received from all multi-parameter sensors sensors located in positions p=1,...,m, where , is the number of positions of multi-parameter sensors in the controlled room. The input layer is represented by a vector:

which is supplemented by the coordinate vector of multiparameter sensors:

The number of input layer nodes is determined by the formula:

where is the number of multi-parameter fire protection system sensors, is the number of fire factors measured by each sensor, is the number of time delays at the input of the neural network, is the number of sensor location coordinates, . The first hidden layer 9 MRP has eight fully connected neurons with an activation function

where ν is the output signal of the neuron adder, characterizing its local field, e ≈ 2.71828 is a mathematical constant. The first hidden layer has the feedback of the first hidden layer 10. The second hidden layer 11 of the MRP has five fully connected neurons with an activation function determined by formula (4) and the feedback of the second hidden layer 12. The output layer 13 contains two neurons and the feedback of the output hidden layer 14. The output layer generates a signal Y, which determines the coordinates of the location of the fire.

MRP training is performed using the Levenberg-Marquard method of backpropagation of errors over time in real time. Two types of data are used to train a neural network. The first type of data is obtained by simulating a fire using a Fire Dynamics Simulator (FDS), which creates a digital twin of the fire. A digital fire twin has a number of advantages over a full-scale fire model: safety, low cost, large fire coverage of premises. Data of the second type is obtained as a result of fire bench tests, for example, at NPO Fire Automation Service LLC in Shuya or at the Federal State Unitary Enterprise Krylov State Scientific Center.

The proposed method for determining the location of a fire using a multilayer recurrent perceptron includes the following steps:

1. Creating a digital model - a double of the development of a fire in a room or performing fire bench tests.

2. Determination of a plan for the optimal location of multi-parameter sensors in a controlled room.

3. Obtaining data from multi-parameter sensors placed in optimal positions in the room at a given height.

4. Training of MRP using the method of back propagation in time.

5. Validation of the neural network using new data not used in training.

6. Application of a trained neural network in determining the location of the fire source.

Example 1. Multi-parameter sensors are selected that measure fire factors: temperature, carbon monoxide concentration and smoke concentration. The sensors are installed in optimal positions calculated using a genetic algorithm. Using a fire dynamics simulator (FDS), the results of measuring fire factors in the coordinates of sensor installation are obtained.

A MRP with input delays is constructed and trained in time using the Levenberg-Marquard method. In FIG. Figure 3 shows the following graphs of the dependence of the root mean square error (rms) on the number of the training step: training schedule 15, training verification (validation) schedule 16 and testing schedule 17. The dependence shows that already at the sixteenth training step, MRP training was achieved.

In FIG. Figure 4 shows a simulated room with marked fire sources 18, 19, 20, 21, 22. The dimensions of the room are 7 m long and 5 m wide. Table 1 shows the coordinates of the fire sources, measured from point 23 (see Figure 4).

Table 1. Coordinates of fire sources Source no. Length distance, m Width distance, m 18 4.0 0.9 19 0.6 2.9 20 3.1 4.8 21 4.6 2.4 22 6.4 0.6

With the optimal arrangement of multi-parameter sensors in Fig. Figure 5 shows areas marked using MCI as zones where a fire occurred. Table 2 shows the numbers of fire sources and the corresponding numbers of ellipses.

Table 2 Results of identifying fire sources Fire number 18 19 20 21 22 Ellipse number 24 25 26 27 28

Figure 5 shows the location of fires 18, 19, 20, 21, 22 with marked fire localization zones in the form of a projection onto the plane of the room plan. To determine the ability of a neural network to localize a fire, multiple computer simulations of the fire of various combustible materials (cable, paper, gasoline, alcohol-containing substances and household waste) and determination of the location of the fire were performed. The dots in Figure 5 show the results of determining the location of the fire, and the ellipses show zones 24, 25, 26, 27, 28, which characterize the accuracy of localizing the fire for different fire sources.

The probability of correctly identifying the source of the fire is 92%. To activate automatic fire extinguishing systems in large rooms, this size of fire areas is quite satisfactory.

The neural network is trained in the Matlab R2019a environment. The trained neural network recognizes the location of the fire, starting from the fifth second of the fire, as it tracks the trend of increasing fire factors: temperature, visibility and concentration of carbon monoxide in the air, measured by multi-parameter sensors. In addition, the coordinates of multiparameter sensors from which data are received are transmitted as input parameters to the neural network. Since in the experiments multiparameter sensors are located at the same height, two coordinates were used (location along the length and width of the room). The MRP has 51 elements of the input layer (five samples of three measured parameters are supplied from the output of each of the three sensors and an additional six parameters of the sensor coordinates). The output layer of the MRP has two output parameters, which are the result of determining the coordinates of the fire source along the width and length of the room.

When performing standard fire tests, a program is used to visualize the operation of multi-parameter sensors, which displays the dynamics of changes in the monitored parameters and records their relative values when the multi-parameter fire detection algorithm is triggered. This solution makes it possible to increase the information content of a fire alarm system with multi-parameter sensors, thereby localizing and extinguishing a fire faster and more efficiently.

Claims (1)

A method for determining the location of a fire using a multilayer recurrent perceptron, using the results of measuring fire factors, while to measure fire factors, multi-parameter fire sensors of the fire alarm system are used, which are placed in the room using a genetic algorithm, all measurements of temperature, optical density, carbon monoxide concentration air environment and the location coordinates of multi-parameter fire sensors are entered into a multilayer recurrent perceptron, characterized in that the multilayer recurrent perceptron is trained using the Levenberg-Marquard method and the multilayer recurrent perceptron is validated using fire modeling, then the location of the fire source is determined using a trained neural network.