KR20140118946A - Occupancy estimation method based on carbon dioxide concentration using dynamic neural network - Google Patents

Abstract

The present invention relates to an occupancy estimation method based on carbon dioxide concentration using a dynamic neural network. An occupancy estimation method based on carbon dioxide concentration using a dynamic neural network includes a measuring step of measuring carbon dioxide concentration included in a room atmosphere; a storing step of storing the carbon dioxide concentration measured by the measuring step in real time; a configuring step of configuring at least one dynamic neural network to estimate the number of remaining occupants in a room based on the carbon dioxide concentration stored by the storing step; a neural network training step of comparing the number of remaining occupants, which is estimated by the dynamic neural network configured by the configuring step, with a reference value previously input for the number of remaining occupants, and comparing an error between the number of remaining occupants and the reference value with a preset threshold value in order to control a weight value and a deviation value of a transfer function for configuring the dynamic neural network to allow the error to be decreased; and an estimating step of estimating the number of remaining occupants through the dynamic neural network controlled by the neural network training step. Thus, the present invention proposes a technique of efficiently recognizing the number of occupants in a building. According to the present invention, as compared with those of a occupancy confirming method using a video camera or a motion recognition sensor according to the related art, the installation and maintenance costs are inexpensive and there is no blind spot so that the reliability is improved.

Description

TECHNICAL FIELD [0001] The present invention relates to a method for estimating a carbon dioxide concentration based on a carbon dioxide concentration using a dynamic neural network,

TECHNICAL FIELD The present invention relates to a carbon dioxide concentration-based estimating method of occupancy numbers using a dynamic neural network.

Currently, about 40% of the world's energy consumption is used to control the interior of the building, and most of the energy used to operate the interior of the building is used in HVAC control systems.

Because of the nature of modern people living in buildings for the most part of the day or for work, energy consumption is essential to maintain indoor air quality comfortably and it is expected that this need will continue to increase in the future.

On the other hand, Korean Patent Registration No. 10-0509332 and Korean Patent Laid-Open No. 10-2011-0009508 disclose various techniques for maintaining indoor air quality.

Recently, in addition to the above-mentioned conventional technologies, new techniques for maintaining indoor air quality have been studied. Among them, development of a ventilation control system according to room occupancy has been made.

That is, in the above-mentioned ventilation amount control system, the control of the indoor air quality depends on how many persons are located in the room, so that a method of grasping the room occupancy in real time has a technically important meaning.

In the past, a method has been developed in which a room is imaged with a video camera to identify the room occupancy, and the presence of a room occupant is confirmed by confirming the room occupancy by image processing or installing a motion recognition sensor such as an infrared sensor in the room.

However, video-based methods have high accuracy, but are expensive and may cause privacy problems. In addition, there is a technical limitation in that the method using the motion sensor can not recognize a person who is stationary and can simply identify the presence or absence of a staff member.

Therefore, in order to maintain the indoor air quality, it is required to develop a technique to more efficiently grasp the room occupancy in the building.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to provide a technique for efficiently grasping a room occupancy in a building.

According to an aspect of the present invention, there is provided a method for estimating a CO 2 concentration based on a CO 2 concentration using a dynamic neural network, comprising the steps of: measuring the concentration of carbon dioxide contained in the air; A storing step of storing the concentration of carbon dioxide measured in the measuring step in real time; A neural network constructing step of constructing at least one dynamic neural network unit to estimate the number of residents in the room based on the concentration value of the carbon dioxide stored in the storing step; Comparing the estimated number of residents from the dynamic neural network unit configured in the neural network unit with a preset reference value for the number of residues and comparing an error between the number of residues and the reference value with a preset threshold value, A neural network learning step of adjusting a weight value and a deflection value of a transfer function constituting the dynamic neural network unit in a direction in which the dynamic neural network unit is reduced; And estimating the number of remaining users through the dynamic neural network unit adjusted in the neural network learning step; . ≪ / RTI >

In addition, in the neural network configuration step, the dynamic neural network unit may include an input layer having at least one input node configured to input a concentration value of carbon dioxide stored in the storage unit; A delay unit for sequentially delaying an input signal received from the input layer using a plurality of taps; A hidden layer having a hidden node connected to the delay unit; An output layer having an output node connected to the hidden layer, the output layer outputting the number of residues according to the concentration value input to the input layer; . ≪ / RTI >

On the other hand, in the neural network learning step, the neural network learning unit may be provided to adjust the weight value and the deflection value of the transfer function by a Levenberg-Marquardt algorithm.

In addition, the connection relationship between the input node and the hidden node can be defined by a hyperbolic tangent sigmoid function.

In addition, the connection relationship between the hidden node and the output node can be defined by a linear function.

As described above, according to the present invention, the sensor unit measures the concentration of carbon dioxide in the room, and the measured concentration value of the carbon dioxide estimates a residual occupant located in the room through the dynamic neural network.

That is, the present invention is less costly to install and maintain than existing methods of estimating the number of occupants using the video camera or the motion recognition sensor, and it is meaningful not only that there is no blind spot but also the reliability is improved.

In addition, since the present invention can accurately estimate the room size, it is possible to minimize the power energy used for ventilation and the cooling / heating energy loss generated when the ventilator is vented to the outside.

FIG. 1 is a block diagram illustrating a method for estimating a CO 2 concentration based on a CO 2 concentration using a dynamic neural network according to the present invention.
2 shows a dynamic neural network according to the present invention.
FIG. 3 shows a first graph measuring carbon dioxide concentration and residence time by time.
FIG. 4 shows a second graph measuring carbon dioxide concentration and residence time by time.
FIG. 5 shows a third graph measuring carbon dioxide concentration and residence time by time.
FIG. 6 shows a result of estimating the occupancy of a room using the ventilation model.
FIG. 7 shows a result of estimating the staff length using the static neural network.
8 shows the result of estimating the occupancy of the room according to the present invention.
FIG. 9 shows another result of estimating occupancy in accordance with the present invention.
FIG. 10 is a block diagram illustrating a method for estimating a CO 2 concentration based on a CO 2 concentration using a dynamic neural network according to the present invention.

The preferred embodiments of the present invention will be described in more detail with reference to the accompanying drawings, in which the technical parts already known will be omitted or compressed for simplicity of explanation.

Before explaining the carbon dioxide concentration-based losing room estimation method using the dynamic neural network according to the present invention, the construction of the present invention will be described first. The carbon dioxide concentration-based loseliness estimation system 10 using the dynamic neural network according to the present invention includes a sensor unit 100, a signal processing unit (not shown), a storage unit 200, a dynamic neural network unit 300, and a neural network learning unit 400 ), Which will be described with reference to Figs. 1 to 9. Fig.

First, the sensor unit 100 is a device for measuring the concentration of a specific component contained in the air in the room. In the present invention, a sensor for measuring the concentration of carbon dioxide contained in the air in the building is used.

The signal processing unit detects a signal of the sensor unit 100 and converts the signal into digital. The sensor unit 100 detects analog signals generated when the concentration of carbon dioxide is measured, amplifies, filters and converts the analog signals into digital data .

The signal processing unit (not shown) detects sensor signals measured by the sensor unit 100, amplifies and filters the signals, converts the sensor signals into digital data, and transmits the digital data to the storage unit 200.

The storage unit 200 is a storage medium for storing concentration values of carbon dioxide measured from the sensor unit 100. The storage unit 200 receives and stores the converted digital data from the signal processing unit in real time.

Meanwhile, the occupancy estimation system according to the present invention includes at least one dynamic neural network unit 300 for estimating the number of residents residing in the room through the concentration value of the stored carbon dioxide from the storage unit 200.

In one embodiment, the dynamic neural network unit 300 includes an input layer 310, a delay unit 320, a hidden layer 330, and an output layer 340, as shown in FIG. The input layer 310, the hidden layer 330, and the output layer 340 included in the dynamic neural network unit 300 are interconnected by the nodes constituting each layer, and the relationship can be defined as a transfer function.

That is, a unit node constituting each layer has a multi-input 1 output structure for receiving a plurality of input variables (X1, X2, ..., Xn) and calculating one output variable (Y) The relationship can be defined by the transfer function f (X) as described in Equation (1) below.

[Equation 1]

X, which is a variable introduced into the transfer function in Equation (1), is a sum of products of input variables (X1, X2 ... Xn) and weight values (W1, W2 ... Wn) introduced into each unit node And the bias value (b, bias).

&Quot; (2) "

As a result, the transfer function that defines the relationship among the unit nodes constituting the dynamic neural network unit 300 according to the present invention can be expressed by the following equation (3).

&Quot; (3) "

(Where A _k is the output variable, W _{i, j} is the weight, P _i is the input variable, B _j is the deflection value)

As can be seen from Equations 1 to 3, the transfer function that defines the relationship between the nodes constituting the dynamic neural network unit 300 is a function having two parameters of the weight value W and the deflection value b , Which may be a linear or non-linear function, as the case may be.

The input layer 310 constituting the dynamic neural network unit 300 includes at least one input node configured to input a concentration value of carbon dioxide stored in the storage unit 200. [

The delay unit 320 sequentially delays the input signal received from the input layer 310 using a plurality of taps. As shown in FIG. 2, the delay unit 320 includes a plurality of tabs to delay the input value input from the input node of the input layer 310 in a stepwise manner.

For example, the input value input from the first input node 311 to the first tap 321 moves to the second tab 322 through the nth tab, which are the lower tabs, A new input value is transmitted from the first input node 311 to the tap 321. [

The hidden layer 330 has a hidden node connected to each tap constituting the delay unit 320 and the output layer 340 has an output node 341 connected to each hidden node. In order to measure the indoor carbon dioxide concentration value, two carbon dioxide measurement sensors are used in the embodiment of the present invention. The input layer 310 is composed of two input nodes, and the output layer 340 is connected to the input layer 310 so that the number of the output nodes 341 is preferably one.

In the present invention, a process of training a dynamic neural network unit 300 in which an input variable such as a carbon dioxide concentration value is inputted in real time is preceded. In the dynamic neural network unit 300, a dynamic neural network unit 300 It can be explained that it is a process of learning so as to accurately estimate the remaining personnel according to the concentration value measurement.

In the present invention, the neural network learning unit 400 compares the estimated number of residents and the number of residues estimated in the dynamic neural network unit 300 with a preset reference value, compares the error of the number of residues and the reference value with a predetermined threshold value, The dynamic neural network unit 300 is learned by adjusting a weight value and a deflection value of a transfer function constituting the dynamic neural network unit 300 in a direction in which the error of the dynamic neural network unit 300 is reduced.

That is, the neural network learning unit 400 sets the weight and deflection values, which are parameters constituting the dynamic neural network unit 300, as an arbitrary initial value, and applies the input parameter of the reference value to the dynamic neural network unit 300. The neural network learning unit 400 applies a reference value input variable to the dynamic neural network unit 300 and compares the output variable with the output variable of the reference value to adjust the weight value and the deflection value so that the error of each output variable is minimized By repeating this, the optimum weight value and deflection value are calculated.

In this case, the neural network learning unit 400 compares the error of each output variable with a preset threshold value in the storage unit 200, and outputs the weighted value in a direction in which the error of the output variable of the dynamic neural network unit 300 decreases And the deflection value.

Herein, the input parameter of the reference value is the indoor carbon dioxide concentration value measured for a predetermined time, and the output parameter can be set as the number of occupants remaining in the measurement object space of the building.

The neural network learning unit 400 may be configured to perform a Levenberg-Marquardt algorithm such that the error between the output variable output from the dynamic neural network unit 300 and the output variable of the reference value is minimized in the process of training the dynamic neural network unit 300. [ To adjust the weighting and deflection values of the transfer function.

In the dynamic neural network unit 300 of the present invention, the connection relationships between the input nodes 311 and 312 and the hidden nodes 331, 332, and 333 constituting the hidden layer 330 The connection relationship between the hidden nodes 331, 332 and 333 and the output node 341 constituting the output layer 340 is defined by the hyperbolic tangent sigmoid function as shown in Equation (4) May be defined by the same linear function, but are not necessarily limited thereto.

&Quot; (4) "

&Quot; (5) "

The method for estimating the carbon dioxide concentration based on the carbon dioxide concentration using the dynamic neural network according to the present invention will be described with reference to FIG. 1 to FIG.

1. Measurement step < S101 >

In this step, the concentration of carbon dioxide contained in the indoor air is measured. That is, the sensor unit 100 is installed at a predetermined position to measure the concentration of carbon dioxide contained in the air in the building, and the concentration value of carbon dioxide is measured in real time. At this time, the signal processing unit detects analog signals generated when the sensor unit 100 measures the concentration of carbon dioxide, amplifies and filters the digital signals, converts the analog signals into digital data, and transmits the digital signals to the storage unit 200.

2. Storage step < S102 >

In this step, the storage unit 200 receives the concentration value of the carbon dioxide converted into the digital data in step S101 from the signal processing unit and stores it in real time.

3. Neural network configuration step <S103>

In this step, at least one dynamic neural network unit 300 is configured to estimate the number of residents in the room based on the concentration value of carbon dioxide stored in the storage unit 200 in step S102.

Here, the dynamic neural network unit 300 is a neural network that learns the change of the carbon dioxide concentration value over time, and includes an input layer, a delay unit, a hidden layer, and an output layer, as described above, and thus a detailed description thereof will be omitted .

4. Neural network learning step <S104>

In this step, the neural network learning unit 400 compares a reference value (actual number of occupant numbers) previously input with respect to the estimated number of remaining persons and the remaining number of persons from the dynamic neural network unit 300 in step S103, The threshold value and the bias value of the transfer function constituting the dynamic neural network unit 300 are adjusted in a direction in which the error of the reference value is reduced.

In other words, the neural network learning unit 400 compares the error of each output variable with a preset threshold value in the storage unit 200, and outputs the output of the reference value and the output variable of the dynamic neural network unit 300, And the deflection value. In this step, the input parameter of the reference value is the carbon dioxide concentration value of the room measured for a predetermined time, and the output variable can be set as the actual remaining number in a specific area of the building.

In this step, the neural network learning unit 400 uses the Lebenberg mathematical algorithm to minimize the error between the output variable output from the dynamic neural network unit 300 and the output variable of the reference value, And normalized it with Bayesian rules.

5. Estimation Step < S105 >

In this step, an arbitrary dynamic neural network unit with the smallest error between the output variable output through the dynamic neural network unit 300 and the output variable of the reference value among the dynamic neural network units learned in step S104 is selected.

Accordingly, the user can obtain the output variable value close to the actual occupant number by installing the sensor unit on the specific space for estimating the occupant, and applying the measured carbon dioxide concentration value to the input parameter of the selected dynamic neural network unit.

<Examples>

In order to verify the estimation of the residential capacity of the dynamic neural network, residential personnel and carbon dioxide concentration were measured in the actual building from February 21 to 25, 2012. Table 1 below shows the measurement date and time of the data for three days used in the present embodiment. Since the aspect ratio of the space was large, the measurement was performed using two devices in consideration of spatial distribution. The error of the measuring instrument is 3% or 50ppm of the measured value, the resolution is 1ppm, and the data is every minute. The reference value of the room occupant was obtained by counting the number of people entering and leaving by installing a video camera at the entrance.

Data # 1 Data # 2 Data # 3 Start 2/21 07:30 2/23 08:30 2/24 08:30 End 2/22 05:00 2/24 05:00 2/25 05:00

In the present embodiment, the generation and training of the dynamic neural network for estimating the occupancy are performed using the Neural Network Toolbox of Matlab. The input values of the neural network are two concentrations of carbon dioxide measured from the sensors, and the output values are set as the occupancy numbers. At this time, since the natural ventilation amount is not given in advance, the influence of the change in the ventilation amount is not considered.

The first graph of FIG. 3 shows the carbon dioxide concentration and the residence time of Data # 1 over time, the second graph of FIG. 4 shows the carbon dioxide concentration and the residence capacity of Data # 2 over time, Shows the carbon dioxide concentration and longevity of the data # 3.

In the dynamic neural network of the present embodiment, the number of tapped delay lines, which are a plurality of tabs constituting the delay unit, is set to 1 to 90. 3, the carbon dioxide concentration values respectively measured from the first sensor and the second sensor are input to the first input node and the second input node of the input layer constituting the dynamic neural network unit as input variables And the number of persons entering and leaving the same time zone was set as an output variable at the output node of the output layer.

In this case, the hyperbolic tangent sigmoid function as shown in Equation (4) is used as a transfer function for calculating the input layer and hidden layer of the dynamic neural network, and the transfer function for calculating the output layer includes a linear function Was used.

In the present embodiment, the dynamic neural network unit is trained by utilizing the first day data (Data # 1). The Levenberg-Marquardt algorithm was used to optimize the weighting and deflection values of the dynamic neural network during training and it was normalized by Bayesian regulation.

In addition, the dynamic neural network unit trained with the carbon dioxide concentration value and the residence capacity measured in the second and third day data (Data # 2, 3) was tested. Here, the training results were evaluated using root-mean-square (RMS), and the training of the dynamic neural network was repeated eight times to select the neural network having the lowest RMS value.

FIG. 8 is a result of estimating the output variable, the capacity factor, by applying the carbon dioxide concentration value of the second day data (Data # 2) to the dynamic neural network unit trained by the first day data (Data # 1) The result of estimating the output variable as the input variable is the carbon dioxide concentration value of the third day data (Data # 3) in the dynamic neural network unit trained by the data (Data # 1). The solid lines in Figs. 8 and 9 are the actual occupant seats, and the dashed lines represent the occupant seats estimated by the dynamic neural network unit.

&Lt; Comparative Example 1 &

A single space full mixed ventilation model, one of the conventional techniques for estimating occupancy, was used to estimate occupancy. Such a ventilation model is a model for a change in the concentration of carbon dioxide in a single space, and can be expressed by Equation (6) below assuming the same air volume at the inlet and outlet and perfect mixture of the indoor air.

&Quot; (6) &quot;

(Where V is the actual volume, C is the carbon dioxide concentration, t is the time, Cin is the concentration of the introduced carbon dioxide, Cout is the concentration of carbon dioxide removed outdoors, Q is the amount of ventilation, G is the amount of carbon dioxide generated per hour)

If the initial concentration is C0 when the occupant and air volume are constant, the indoor carbon dioxide concentration is derived as shown in Equation (7) below.

&Quot; (7) &quot;

(Where t is the indoor carbon dioxide concentration, V is the actual volume, t is the time, Cin is the incoming carbon dioxide concentration, Q is the ventilation amount, G is the amount of carbon dioxide generated per hour,

FIG. 6 shows a result of estimating the occupancy number by applying the carbon dioxide concentration value measured in the second day data (Data # 2) to the ventilation model. As shown in FIG. 6, it can be seen that the estimation method of the occupancy number estimated through the ventilation model has a large error and variation in comparison with the actual occupation number.

In other words, the estimation method based on the ventilation model is inconvenient to estimate only the carbon dioxide concentration value or the concentration gradient because the indoor pollution concentration changes exponentially according to the amount of pollution generation and the amount of ventilation.

&Lt; Comparative Example 2 &

A static neural network model, which is one of the conventional techniques for estimating the room occupancy, was used to estimate occupancy. For each input node of the input layer constituting the static neural network, the present carbon dioxide concentration value and the concentration change value are input as input variables, and the occupancy number is estimated through the output node of the output layer.

FIG. 7 shows a result of estimating the staff length by applying the carbon dioxide concentration value measured in the second day data (Data # 2) to the static neural network model. As shown in FIG. 7, it can be seen that the fluctuation range of the estimation value is decreased as compared with the method using the ventilation model, by using the static neural network with the multi-layer perceptron.

In the conventional static neural network model, the current concentration and the concentration change value have to be inputted to each input node of the input layer constituting the static neural network in order to estimate the occupancy, but the dynamic neural network unit 300 according to the present invention has the input Since the delay unit 320 including a plurality of tap delay lines is located between the layer 310 and the hidden layer 330, the input variable transmitted through the input node of the input layer 310 is sequentially delayed, Since all of the n tap delay lines are transmitted to the hidden layer 330, a change with time can be learned.

That is, since the dynamic neural network unit 300 of the present invention derives an output variable in consideration of a change with time in comparison with the number of occupants, which is an output variable estimated through Comparative Examples 1 and 2, 2 is closer to the actual number of residents than the estimated number of residents.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. And the scope of the present invention should be understood as the following claims and their equivalents.

10: Estimation system of CO2 concentration based on dynamic neural network
100:
200:
300: dynamic neural network unit
310: input layer
311: First input node
312: second input node
320:
321: First tab
322: Second tab
323: The third tab
330: Hiding layer
331: first hidden node
332: second hidden node
333: Third hidden node
340: Output layer
341: Output node
400: Neural network learning part

Claims (3)

A measuring step of measuring the concentration of carbon dioxide contained in the indoor air;
A storing step of storing the concentration of carbon dioxide measured in the measuring step in real time;
A neural network constructing step of constructing at least one dynamic neural network unit to estimate the number of residents in the room based on the concentration value of the carbon dioxide stored in the storing step;
Comparing the number of the residual persons estimated by the one or more dynamic neural network units with the reference value inputted in advance for the number of the remaining persons and comparing the error between the residual number of persons and the reference value with a preset threshold value, A neural network learning step of learning the dynamic neural network unit by adjusting a weight value and a deflection value of a transfer function constituting the dynamic neural network unit in a direction; And
Wherein the dynamic neural network having the lowest root-mean-square (RMS) value for the estimated number of residents among the at least one dynamic neural network learned in the neural network learning step finally selects the residual number to estimate, An estimation step of estimating the number of remaining users through a neural network; Lt; / RTI &gt;
In the neural network configuration step, the dynamic neural network unit
An input layer having at least one input node adapted to input a stored concentration value of carbon dioxide;
A delay unit for sequentially delaying an input signal received from the input layer using a plurality of taps to learn a change with time;
A hidden layer having a hidden node connected to each tap constituting the delay unit; And
An output layer having an output node coupled to the hidden layer and outputting a number of residues according to a concentration value input to the input layer; / RTI &gt;
Wherein the input layer, the hidden layer and the output layer are interconnected by at least one unit node so that their relationship can be defined by the transfer function, and the unit node receives a plurality of input variables, However,
The transfer function is a function having the weight value and the deflection value as parameters as shown in the following Equation (1)
Characterized in that the neural network learning step adjusts the weighting and deflection values of the transfer function by a Levenberg-Marquardt algorithm
A Method for Estimating Gap Length Based on Carbon Dioxide Concentration Using Dynamic Neural Network.
[Equation 1]

(Where A _k is the output variable, W _{i, j} is the weight, P _i is the input variable, and B _j is the deflection value) The method according to claim 1,
Characterized in that the connection relationship between the input node and the hidden node is defined by a hyperbolic tangent sigmoid function
A Method for Estimating Gap Length Based on Carbon Dioxide Concentration Using Dynamic Neural Network. 3. The method of claim 2,
Characterized in that the connection relationship between the hidden node and the output node is defined by a linear function
A Method for Estimating Gap Length Based on Carbon Dioxide Concentration Using Dynamic Neural Network.

Images