JP7460886B2 - Machine Learning Equipment - Google Patents

Description

This relates to a machine learning device that learns the radio wave propagation conditions between wireless devices and other wireless devices.

BACKGROUND ART Systems have conventionally existed in which equipment for heating, ventilation, air conditioning, etc. (hereinafter referred to as HVAC equipment) is connected to a network and managed remotely. In this system, it is necessary to store network addresses and information about the physical location of each HVAC device in a building in association with each other. This mapping work is often done manually in the building, which takes a lot of time and cost.

In response to this problem, efforts have been made in recent years to provide radio wave transmitters and receivers (wireless devices) in each HVAC device, obtain information on radio wave intensity from the transmitters and receivers, and use this information to estimate the location of each HVAC device. For example, Patent Document 1 (JP Patent Publication No. 2016-208265) shows one method for simulating radio wave propagation conditions.

Although it is possible to simulate the radio wave propagation state using existing methods, there is a need for a simpler or more accurate method that is different from the existing methods.

The machine learning device of the first aspect is a machine learning device that learns the radio wave propagation state between a wireless device and another wireless device. The machine learning device includes an acquisition section and a learning section. The acquisition unit acquires the first information as information for acquiring the state variable. The first information is information regarding between a wireless device and another wireless device. For example, the first information includes information regarding the distance between the wireless device and another wireless device, information regarding an article between the wireless device and the other wireless device, and the like. The learning unit learns the state variables and the radio wave propagation state between the wireless device and another wireless device in association with each other.

Here, the acquisition unit acquires first information relating to the relationship between the wireless device and other wireless devices, and the necessary state variables can be obtained from the first information. Since the learning unit learns by associating the state variables with the radio wave propagation state, this machine learning device makes it possible to obtain the radio wave propagation state between the wireless device and other wireless devices.

The state of radio wave propagation between a wireless device and another wireless device can be expressed by any of the following: the amount of radio wave attenuation between the wireless device and the other wireless device, the minimum transmission radio wave strength that can be received by the other wireless device, the reception radio wave strength of the other wireless device in response to transmission from the wireless device at a specified transmission radio wave strength, etc.

The machine learning device according to the second aspect is the machine learning device according to the first aspect, and the first information includes information on the distance between the wireless device and another wireless device. A distance between a wireless device and another wireless device is at least one of the state variables.

Here, the distance between two wireless devices is used as the state variable, making it possible to obtain a more accurate picture of the radio wave propagation state.

The machine learning device of the third aspect is the machine learning device of the first or second aspect, and the first information includes item information about a specific item located between the wireless device and another wireless device.

Here, the state variable can be obtained from information about the item between the two wireless devices that affects the radio wave propagation state. Thereby, a more accurate radio wave propagation state can be obtained.

The machine learning device according to the fourth aspect is the machine learning device according to the third aspect, and the article information includes information on the number of predetermined articles.

Here, the state variables are obtained from the number of specified items, which is relatively simple information. This makes it easier to obtain the radio wave propagation state.

The machine learning device according to the fifth aspect is the machine learning device according to the third or fourth aspect, and the predetermined article is an air conditioner and/or a beam placed in a space above the ceiling.

Here, the learning unit can obtain relatively accurate radio wave propagation conditions in the narrow space above the ceiling, where it is difficult to obtain accurate results using existing radio wave propagation simulation devices.

The machine learning device of the sixth aspect is a machine learning device of any one of the third aspect to the fifth aspect, in which the item information includes information about the size of a specified item.

Here, information regarding the size of a predetermined item located between a wireless device and another wireless device is acquired as the first information. Thereby, a more accurate radio wave propagation state can be obtained.

The machine learning device according to the seventh aspect is the machine learning device according to any one of the third to sixth aspects, and the article information includes information regarding the position of a predetermined article.

Here, information about the position of a specific item located between the wireless device and another wireless device is acquired as the first information. This makes it possible to obtain a more accurate radio wave propagation state.

The machine learning device of the eighth aspect is a machine learning device of any one of the third aspect to the seventh aspect, in which the item information includes information regarding the orientation of a specified item.

Here, information regarding the orientation of a specific object located between the wireless device and another wireless device is acquired as the first information. This makes it possible to obtain a more accurate radio wave propagation state.

The machine learning device according to the ninth aspect is the machine learning device according to the first aspect or the second aspect, and the learning unit learns by associating state variables and radio wave propagation states based on the learning data set. The learning data set consists of the results of actual measurement of the radio wave propagation state and the state variables at the time of the actual measurement of the radio wave propagation state.

Here, actual measurements of radio wave propagation conditions are taken, and the learning unit learns using a learning dataset consisting of the state variables and results at that time. This makes it possible to obtain more accurate information about radio wave propagation conditions.

The machine learning device of the tenth aspect is the machine learning device of any one of the third aspect to the eighth aspect, in which the learning unit learns by associating state variables with radio wave propagation conditions based on a learning dataset. The learning dataset is composed of results of actual measurements of the radio wave propagation conditions and state variables obtained from item information at the time of the actual measurements of the radio wave propagation conditions.

Here, the state of radio wave propagation is actually measured, and the learning section performs learning using a learning data set consisting of the state variables and results obtained at that time. Thereby, a more accurate radio wave propagation state can be obtained.

The machine learning device of the eleventh viewpoint is a machine learning device of the first viewpoint, the second viewpoint, or the ninth viewpoint, and the learning unit adjusts the coefficients of the linear model of the following equation 1 to and the state of radio wave propagation.
Formula 1:

The machine learning device of the twelfth viewpoint is one of the third to eighth viewpoints or the machine learning device of the tenth viewpoint, and the learning unit adjusts the coefficients of the linear model of the following equation 2. By this method, state variables and radio wave propagation states are associated and learned.
Formula 2:

The machine learning device of a thirteenth aspect is the machine learning device of any one of the first aspect to the twelfth aspect, further comprising an output unit and an update unit. The output unit outputs an estimation result of the radio wave propagation state. The update unit updates the learning state of the learning unit by evaluating the difference between the estimation result of the radio wave propagation state and the result of actually measuring the radio wave propagation state.

Here, the learning state of the learning section is improved and a more accurate radio wave propagation state can be obtained.

を更新することによって、学習部の学習状態を更新する。
評価関数１：

The machine learning device according to the fourteenth aspect is the machine learning device according to the eleventh aspect, and further includes an output unit and an update unit. The output unit outputs the estimation result of the radio wave propagation state. The update unit updates the equation 1 in the above equation 1 so that the following evaluation function 1 becomes small.

By updating , the learning state of the learning section is updated.
Evaluation function 1:

Here, the learning state of the learning unit improves, and more accurate radio wave propagation conditions can be obtained.

を更新することによって、学習部の学習状態を更新する。
評価関数２：

The machine learning device according to the fifteenth aspect is the machine learning device according to the twelfth aspect, and further includes an output unit and an update unit. The output unit outputs the estimation result of the radio wave propagation state. The update unit updates the equation 2 above so that the following evaluation function 2 becomes small.

By updating , the learning state of the learning section is updated.
Evaluation function 2:

Here, the learning state of the learning unit improves, and more accurate radio wave propagation conditions can be obtained.

FIG. 2 is a simple vertical cross-sectional view showing a building in which a plurality of BLE modules in which a machine learning device learns radio wave propagation conditions is arranged, and a beam and an air conditioner existing in a space above the ceiling of the room. A plan view of the first floor of the building, including the arrangement of beams and air conditioners in the space above the ceiling. FIG. 1 is a diagram illustrating the configuration of an HVAC management system including a machine learning device.

(1) Overview of HVAC management systems (1-1) Background of the need for HVAC management systems With the increasing demand for energy saving and rapid development of intelligent control technology for buildings (office buildings, commercial facilities, etc.), air conditioning and ventilation intelligent systems are being installed in buildings. By monitoring temperature, humidity, _CO2 concentration, and room occupancy, the system can provide demand-based feedback control, such as adjusting the temperature setting of an air conditioner based on the number of people in the room.

To achieve this type of feedback control in a specific region, connect HVAC equipment and sensors that perform ventilation and air conditioning to a network, and set the network address of the HVAC equipment to the physical location of the HVAC equipment. need to be mapped. Mapping, also called address configuration, has traditionally been done manually. Work performed by workers on-site includes using a control device to operate HVAC equipment one at a time, and using the addresses of HVAC equipment, etc. displayed on the display of the control device on a layout map. This includes writing things down.

These operations are time consuming and may involve rework due to human error. In addition, the labor costs associated with the work performed by workers on-site are a large cost. For example, in the case of a large-scale building with a floor area of 50,000 square meters, it will take three months to set up an address even if two workers work on it.

(1-2) Basic Concept of Reducing Mapping Work in HVAC Management System In order to reduce the burden of this mapping (address setting) work, in this embodiment, an HVAC management system for performing automatic mapping will be described. In this HVAC management system, as shown in FIG. 1, an air conditioner A, which is an HVAC device, is equipped with a BLE (Bluetooth Low Energy) module M, and the RSSI (Received Signal Strength Indicator) of the BLE module M is used. Here, the installation position of the BLE module M equipped in the air conditioner A is estimated, and the radio wave propagation state between the BLE modules M is estimated. When the radio wave propagation state between each BLE module M is estimated, mapping can be easily performed from the estimated values.

Specifically, each air conditioner A is equipped with a BLE module M, and the BLE modules M communicate with each other in packets, and the actual received signal strength of the packet sent from the transmitting BLE module M to the receiving BLE module M is measured. Meanwhile, the physical arrangement of each air conditioner A, the distance between the air conditioners A, and the arrangement of obstacles that may affect the propagation of radio waves (signals) are automatically read from the layout drawing of each air conditioner A extracted from the design drawing (see FIG. 3) provided by the building designer, etc. The radio wave propagation state between the BLE module M and another BLE module M, for example, the amount of radio wave attenuation between both modules M, the ratio of the received signal strength in the receiving BLE module M to the transmitted signal strength of the transmitting BLE module M, etc., can be estimated by using the linear model 35 (see FIG. 3) of the learning unit 30. By using information on the location of each air conditioner A and obstacles read from the layout drawings of each air conditioner A, and a linear model 35 of the learning unit 30 trained using a huge learning data set sampled in the past in various buildings, it becomes possible to accurately estimate the radio wave propagation state between a BLE module M and another BLE module M.

As a result, it eventually becomes possible to automatically map the network address to the air conditioner A installed in the space above the ceiling of the building. Thereby, it is possible to shorten the time required for the initial setting work of an intelligent system for air conditioning and ventilation (HVAC management system) in a building, and to reduce labor costs.

(2) Configuration of the HVAC Management System The HVAC management system is a system that manages HVAC equipment that performs heating, ventilation, air conditioning, etc., but here, we will use air conditioner A, which is an HVAC equipment, as an example and explain it with reference to Figures 1 to 3.

(2-1) Installation location of air conditioner which is HVAC equipment Air conditioner A is an air conditioner indoor unit installed in the interior space of building 81, as shown in FIG. The plurality of air conditioners A are arranged in a space above the ceiling of each floor (room) of the building 81. FIG. 1 shows three air conditioners A1, A2, and A3 installed in a space S above the ceiling on the first floor of a building 81. These air conditioners A1, A2, A3 are shown in FIG. 2, which is a plan view including a space S above the ceiling on the first floor. A large number of beams B extend horizontally in the space S above the ceiling on the first floor. As shown in FIGS. 1 and 2, a beam B1 exists between the air conditioners A2 and A3.

(2-2) BLE module Each air conditioner A has a built-in BLE module M. The BLE module M has RSSI and can measure the strength of received radio waves (received signal strength) in addition to transmitting radio waves.

As shown in FIG. 1, air conditioner A1 has a built-in BLE module M1, air conditioner A2 has a built-in BLE module M2, and air conditioner A3 has a built-in BLE module M3.

(2-3) Computer as a machine learning device The computer 10 that functions as a machine learning device is composed of one or more computers, and is connected to HVAC equipment such as air conditioners A in each building 81 via a communication network 80 such as the Internet. connected to. The computer 10 executes a cloud computing service built by an HVAC management system service provider to provide various services. The hardware configuration of the computer 10 does not need to be housed in one housing or provided as a unit of equipment.

As shown in FIG. 3, the computer 10 mainly includes an acquisition section 20, a learning section 30, an output section 40, an input section 50, and an update section 60. The computer 10 includes a control calculation device and a storage device. A processor such as a CPU or a GPU can be used as the control calculation device. The control arithmetic device reads a program stored in the storage device, and performs predetermined image processing and arithmetic processing according to this program. Furthermore, the control calculation device can write calculation results to the storage device and read information stored in the storage device according to the program. The acquisition section 20, the learning section 30, the output section 40, the input section 50, and the updating section 60 shown in FIG. 3 are various functional blocks realized by the control calculation device. These functional blocks appear when the control arithmetic unit executes the model creation program.

(2-3-1) Acquisition Unit The acquisition unit 20 acquires air conditioner arrangement information (first information) 21 and beam arrangement information (first information) 22 in the ceiling space from an external design drawing database 70 as information for obtaining state variables. The design drawing database 70 stores design drawings and the like for each floor of a building 81. The air conditioner arrangement information 21 is information relating to the location where the air conditioner A is arranged, as shown in Figs. 1 and 2. The air conditioner arrangement information 21 includes information such as data on the X and Y coordinates of each air conditioner A in a plan view and the distance between the air conditioners A. The beam arrangement information 22 includes data on the X and Y coordinates of both ends of each beam B and information indicating between which two air conditioners A each beam B is located.

In other words, these air conditioner arrangement information 21 and beam arrangement information 22 are information regarding between a certain BLE module M and another BLE module M. Information on the distance between the air conditioners A becomes information on the distance between the BLE modules M built in those two air conditioners A. Furthermore, from the data of the X coordinate and Y coordinate in the plan view of each air conditioner A and the data of the X coordinate and Y coordinate of both ends of each beam B, it is possible to determine the Information on the number of located beams B and air conditioners A can be calculated.

を取得部２０が状態変数として取得する。 Here, from the layout map of air conditioner A and beam B extracted from the design drawings,

The acquisition unit 20 acquires as a state variable.

For example, between the BLE module M1 and another BLE module M2 shown in Figures 1 and 2, there is one beam B1 and one air conditioner A3. Between the BLE module M2 and another BLE module M3, there is one beam B1 and no air conditioner A. Between the BLE module M1 and another BLE module M3, there is neither a beam B1 nor an air conditioner A.

(2-3-2) Learning Unit The learning unit 30 learns by associating the state variables with the radio wave propagation state between the BLE module M and another BLE module M. The learning unit 30 learns by associating the state variables with the radio wave propagation state based on a learning dataset. The learning dataset is composed of actual measurement results 55 of the radio wave propagation state, which are the results of actually measuring the radio wave propagation state, and the state variables at the time of the actual measurement of the radio wave propagation state.

The actual measurement result 55 of the radio wave propagation state is information collected from the BLE module M of each air conditioner A installed in the building 81 by the input unit 50, which will be described later, as shown in FIG. Specifically, the state variable used in the actual measurement of the radio wave propagation state is a value obtained from information on each air conditioner A and beam B installed in the building 81. Here, from the layout map of air conditioner A and beam B extracted from the design drawing, we calculate the distance between any two BLE modules M, and the beam B and air conditioner on the line segment connecting the two BLE modules M. The number A is used as a state variable when actually measuring the radio wave propagation state.

More specifically, the learning unit 30 adjusts the coefficients of a linear model 35 of the following equation 12, thereby learning the association between the state variables and the radio wave propagation state.
Equation 12:

(2-3-3) Output Unit The output unit 40 outputs the estimation result 45 of the radio wave propagation state obtained by the linear model 35 of the learning unit 30.

(2-3-4) Input Unit The input unit 50 collects the actual measurement results 55 of the radio wave propagation state from the BLE modules M of each air conditioner A installed in the building 81 via the communication network 80. The input unit 50 can also collect the actual measurement results 55 of the radio wave propagation state from a user terminal 90 via the communication network 80.

(2-3-5) Update Unit The update unit 60 obtains the difference between the estimation result 45 of the radio wave propagation state output by the output unit 40 and the actual measurement result 55 of the radio wave propagation state input to the input unit 50. Then, the update unit 60 updates the learning state of the learning unit 30 by evaluating the difference.

を更新することにより、学習部３０の学習状態を更新する。
評価関数１２：

Specifically, the updating unit 60 updates the equation 12 above so that the following evaluation function 12 becomes small.

By updating , the learning state of the learning section 30 is updated.
Evaluation function 12:

(3) Characteristics of the HVAC management system computer (machine learning device) (3-1)
In general, when the distance between the transmitter and the receiver is the main factor of the propagation path loss, the radio wave propagation in free space can be described by the Friis Transmission Equation. However, the space above the ceiling of a building is usually 0.5 m or 1.5 m high and is a very limited space. In addition, beams and HVAC equipment exist in the space above the ceiling to maintain the strength of the building. In such a complex space, in addition to the distance, obstacles on the radio wave propagation path may also have a significant effect on the propagation path loss due to reflection and refraction. For this reason, the obstacles must also be considered as variables in the model. In particular, two types of obstacles made of metal and relatively large in volume, beams and HVAC equipment, should be considered as variables in the linear model.

In view of this, in the computer 10 as a machine learning device according to the present embodiment, the distance between the BLE module M and other BLE modules M, and the distance between the beam B and the air between the BLE module M and the other BLE modules M are determined. The focus is on the number of harmonizers A.

In the computer 10 of the HVAC management system of this embodiment, the acquisition unit 20 acquires air conditioner arrangement information (first information) 21 regarding between a BLE module M and another BLE module M, and beam arrangement information ( 1st information) 22 is acquired. Therefore, the computer 10 can obtain the state variables required by the learning section 30 from this information. Since the learning unit 30 learns the state variables and the radio wave propagation state in association with each other, the computer 10 can obtain the radio wave propagation state between the BLE module M and another BLE module M.

を採用しているので、精度の高い電波伝搬状態を得ることができている。 In addition, as state variables,

This enables highly accurate measurement of radio wave propagation conditions.

In this way, the computer 10 is able to obtain relatively accurate radio wave propagation conditions using the learning unit 30 in the narrow space S above the ceiling, where it is difficult to obtain accurate results using existing radio wave propagation simulation devices.

を採用している。このため、コンピュータ１０では、より精度の高い電波伝搬状態を得ることができている。 (3-2)
In the computer 10 as the machine learning device according to the present embodiment, the following state variables are further set:

This allows the computer 10 to obtain a more accurate picture of the radio wave propagation state.

が、その建物８１の天井の上の空間の梁Ｂや空気調和機Ａの配置に合った適切な数値になる。 (3-3)
In the computer 10 serving as the machine learning device according to this embodiment, actual measurements of radio wave propagation conditions are performed by each BLE module M in the building 81, and the learning unit 30 learns using a learning data set consisting of state variables and actual measurement results 55 of the radio wave propagation conditions at that time. Specifically, the update unit 60 updates the learning state of the learning unit 30 by evaluating the difference between the estimated results 45 of the radio wave propagation conditions and the actual measurement results 55 of the radio wave propagation conditions. As a result,

However, this value becomes an appropriate value that matches the arrangement of beams B and air conditioners A in the space above the ceiling of the building 81.

(4) An example of comparison between a model with one variable and a model with three variables

The computer (machine learning device) 10 described above focuses on two types of obstacles that are made of metal and have a relatively large volume: a beam and an air conditioner. First, from the layout map extracted from the design drawing, calculate the distance between any two BLE modules and the number of beams and air conditioners on the line connecting the two BLE modules (#beam and #machine). and have obtained. Next, a predictive variable set consisting of three variables (distance, #beam, #machine) is created.

Now, to evaluate the impact of #beam and #machine,
One variable (1-feature; distance) and
Three variables (3-feature; distance, #beam, #machine) and
The trainer constructed a linear model using both and compared the differences between them.

For the univariate model, OLS regression (Ordinary Linear Regression) with polynomials was chosen. Linear regression fits a linear model with coefficients that minimize the sum of squared residuals between the observed response in the data set and the response predicted by a linear fit. Mathematically, the model can be expressed as:

The residual sum of squares (sometimes called a loss function) is expressed by the following formula.

On the other hand, for models with three variables [distance, #beam, #machine], we chose Ridge Regression with polynomials. Ridge regression is also a type of generalized linear regression. Compared to OLS regression, Ridge regression has a loss function with an additional penalty term.

In the first attempt at training the model, the model with each of the three (distance, #beam, #machine) variables:

For both OLS regression and ridge regression, the polynomial order was varied from 1 to 4 and scaling was applied to each variable before regression. By using the functionality of the machine learning library, the variables are first transformed to a normal distribution and then the maximum absolute value of each variable is scaled to 1.0.

The accuracy of linear model prediction is mainly evaluated by RMSE (Root Mean Square Error) and R ² (coefficient of determination). Here, RMSE was adopted. RMSE is calculated using the absolute difference between the estimated and measured RSSI.

Here, RMSE calculation was performed using the following two methods.
1) Both model training and testing (RMSE calculation) were performed on the entire dataset (100% dataset).
2) The K-fold cross-validation (K=10) set was randomly divided into 10 sets, and the model was trained on 9 sets and tested on the remaining 1 set. The model was tested by repeating the test set and the RMSE was calculated 10 times. The average value of the 10 RMSEs was then used as an evaluation index.

In the experiment, data sampling was performed in a typical steel-framed building (office building) where all beams were made of metal.

The height of the space above the ceiling (from the bottom of the plasterboard to the top slab) is 0.85m.
The maximum height of the beam is 0.7m.
The average height of an air conditioner is 0.3m.
It is.

Each floor has a flat ceiling measuring 18m x 18m. 26 BLE modules are placed on the first and second floors. The distance between two BLE modules is 1.3m to 20m.

The data sampling process was conducted for one week, during which the BLE modules exchanged communication packets and packet records were collected. The packet record includes a timestamp, transmitter ID, receiver ID, transmit power (always set to 8 dBm for each BLE module), and RSSI. Wireless interference caused by Wi-Fi in buildings becomes more serious depending on the working hours. Taking this into account, we chose RSSI at night (from 10pm to 7pm) and on weekends for model training. The RSSI was further resampled at 5 minute time intervals for each pair of BLE modules, and the average value of RSSI at each time interval was used as the 100% data set described above.

The model evaluation results are shown in Table 1.

Here, two types of RMSE are calculated. K-fold cross-validation showed a larger RMSE than the 100% dataset. Both RMSEs showed the same trend.

For many models with the same function, the higher the order, the lower the RMSE. This model improves the accuracy of the radio wave propagation model in the space above the ceiling by adding the number of beams and the number of air conditioners as variables in addition to distance.

Next, Table 2 shows a comparison between the RSSI indicated by the model and the RSSI of the actual BLE module. Of the eight pairs of models, the three-variable fourth-order model showed the best estimation accuracy. Table 2 shows the differences for all pairs of BLE modules involving the BLE module "102A" (not shown).

In order to improve the process of network address configuration of HVAC equipment, it is useful to utilize the RSSI of the above BLE module. Through experiments, we found that the RMSE of RSSI prediction has the following tendency.
1) The RMSE is reduced by taking into account information about obstacles, such as the number of beams and the number of HVAC devices such as air conditioners.
2) The RMSE decreases with increasing degree of the polynomial.

(5) Modification example (5-1)
In the computer (machine learning device) 10 described above, the number of obstacles (beam B and air conditioner A) on the line segment connecting the two BLE modules M is used as a state variable. However, instead of or in addition to this, the size of the obstacle may be used as a state variable.

(5-2)
In the computer (machine learning device) 10 described above, the number of obstacles (beam B and air conditioner A) on the line segment connecting the two BLE modules M is used as a state variable. However, instead of or in addition to this, the position of the obstacle may be used as a state variable.

(5-3)
In the computer (machine learning device) 10 described above, the number of obstacles (beam B and air conditioner A) on the line segment connecting the two BLE modules M is used as a state variable. However, instead of or in addition to this, the orientation of the obstacle may be used as a state variable.

(5-4)
In the computer (machine learning device) 10 described above, the learning unit 30 learns in association with the state variable and the radio wave propagation state by adjusting the coefficients of the linear model 35 of the above equation 12. Alternatively, the learning unit 30 may learn the state variables and the radio wave propagation state in association with each other by adjusting the coefficients of the linear model of Equation 11 below.
Formula 11:

を更新することによって、学習部３０の学習状態を更新してもよい。
評価関数１１：

Then, the updating unit 60 updates the equation 11 so that the following evaluation function 11 becomes smaller instead of the evaluation function 12 described above.

The learning state of the learning section 30 may be updated by updating .
Evaluation function 11:

を更新することによって学習部３０の学習状態を更新する場合にも、精度の高い電波伝搬状態を得ることができるようになる。 In this way, the linear model of Equation 11

Even when updating the learning state of the learning section 30 by updating , a highly accurate radio wave propagation state can be obtained.

(5-5)
In the above-described HVAC management system, the air conditioner A, which is an HVAC device, is equipped with a BLE module M, but instead of the BLE module, other wireless devices may be equipped. For example, a ZigBee module may be adopted.

(6) Comparison of estimated value of radio field strength using machine learning device and estimated value of radio field strength using conventional simulation (6-1)
As described above, in order to reduce the work (address setting work) of mapping the network address of HVAC equipment, etc. to the physical location of the HVAC equipment, etc., the computer 10 as a machine learning device can perform a Estimating radio wave propagation conditions is useful. If it is possible to detect the installation location of each air conditioner and automate address settings, we can expect the development of feedback control, energy saving, and intelligent control technology for buildings.

Therefore, each air conditioner is equipped with a BLE module, and the above-mentioned machine learning device uses the reception strength of the radio waves output from the BLE module. The reception strength of radio waves (radio wave intensity) is one of the indicators representing the state of radio wave propagation.

The radio wave strength between two points tends to become weaker the farther the distance is, due to radio wave attenuation over distance. However, if there is something between the two points that blocks the propagation of radio waves, this will have a large effect on the attenuation of the radio wave strength. Traditionally, simulations based on physical models have been used to estimate radio wave propagation strength, but these require many and complicated input conditions (input contents). When setting addresses using conventional physical models, there was an issue that accuracy would deteriorate significantly if the input conditions were narrowed down to a level that could be used for routine operations.

In view of this, the machine learning device described above uses a radio wave propagation model that predicts the measured radio field strength based on the distance between two BLE modules and obstacles that obstruct radio wave propagation between the two BLE modules. is constructed using machine learning.

Below, in order to verify the effectiveness of the model using machine learning, we show the results of a comparison with a conventional physical model through simulation.

(6-2) Comparative evaluation First, in the machine learning prediction, the radio wave strength of the BLE module is used as the explained variable. In addition, the distance between the BLE modules and the number of air conditioners and beams are used as explanatory variables (state variables). After learning the radio wave propagation model using the actual measured values of the radio wave strength of a BLE module installed in the ceiling of a building, the radio wave strength was estimated using the learning model. In the evaluation of the accuracy of the machine learning prediction, the accuracy was evaluated using the radio wave strength between 15 BLE modules installed in the ceiling of a building.

For simulation predictions, we created a 3D model of the actual attic environment of a certain building (a model that includes air conditioners and beams), entered reference values for material parameters, performed a radio wave propagation simulation, and calculated the radio wave intensity. estimated. The simulation software used was a combination of a commercially available discrete event simulator and an expansion module for high-definition radio wave propagation. Simulations using this software can be performed in a detailed radio wave propagation environment, taking into account the effects of radio wave reflection, shielding, and diffraction from buildings and other objects. In the prediction accuracy evaluation by simulation, the radio wave intensity between five BLE modules included in the building part where the 3D model was constructed was evaluated.

In addition, when making simulation predictions, in addition to inputting material parameters, if a BIM model (a 3D digital model of a building) does not exist, it is also necessary to create the model, which requires more tedious work than using a learning model.

The results of accuracy evaluation performed according to the above procedure are shown in Table 3 below.

In the simulation, we can see that although the standard error of the actual measured signal strength is small, the estimated error is large.

These results show that machine learning methods can predict radio wave strength with better accuracy than simulation methods. Here, learning was performed using fewer state variables than were used as input conditions for the simulation, but higher accuracy was achieved using the learning model.

(7) Matching of the installation position of each air conditioner and the BLE module The radio wave strength as the radio wave propagation state obtained by the above machine learning device is used in the next step, a matching algorithm between the equipment installation position and the BLE module to identify the position of the BLE module. In matching between the equipment installation position and the BLE module, first, an undirected graph GE is obtained in which the estimated received radio wave strength obtained by the above machine learning method is set as the edge value and the position ID of the air conditioner is set as the vertex. Next, the BLE modules built into the air conditioners of the building (site), which is the target property, transmit to each other. The actual received radio wave strength of the BLE module is collected, and an undirected graph GM is created in which the actual received strength is set as the edge value and the vertex is the ID of the transmitting BLE module and the receiving BLE module. However, there is always an error between the estimated value and the actual measured value of the radio wave strength. Therefore, by setting an error tolerance (slack value), it is determined that there is a possibility that the edge in the undirected graph GM in which the error is equal to or less than the tolerance value and the edge in the undirected graph GE are the same. Then, by performing a matching algorithm of the undirected graph GM to the undirected graph GE, BLE modules (multiple) that are matching candidates are determined for each installation position of the air conditioner.

If, when performing the above matching, an estimated signal strength value obtained by simulation is used, and the same algorithm and the same error tolerance are used to determine BLE module candidates for each installation location, the correct answer will often not be included among the determined BLE module candidates. Conversely, if a larger error tolerance is used so that the correct answer is included, the number of determined BLE module candidates will increase, making it difficult to narrow down the correct answer.

(8)
The above describes an embodiment of an HVAC management system having a computer 10 as a machine learning device, but it will be understood that various changes in form and details are possible without departing from the spirit and scope of the present disclosure described in the claims.

10 Computer (machine learning device)
20 Acquisition unit 21 Air conditioner placement information (first information; article information)
22 Beam arrangement information in the attic space (first information; article information)
30 Learning section 35 Linear model 40 Output section 45 Estimation result of radio wave propagation state 55 Actual measurement result of radio wave propagation state 60 Update section A (A1, A2, A3) Air conditioner (predetermined item)
B (B1) Beam (specified item)
M (M1, M2, M3) BLE module (wireless device)

JP 2016-208265 A

Claims (15)

A machine learning device (10) that learns a radio wave propagation state between a wireless device (M1) and another wireless device (M2) to support mapping work of the wireless device and the other wireless device ,
an acquisition unit (20) that acquires first information (21, 22) regarding between the wireless device and the other wireless device as information for obtaining a state variable;
a learning unit (30) that learns in association with the state variable and the radio wave propagation state;
Equipped with
Machine learning device. The first information includes distance information (21) between the wireless device and the other wireless device,
a distance between the wireless device and the other wireless device is at least one of the state variables;
The machine learning device according to claim 1. The first information includes article information (21, 22) regarding a predetermined article located between the wireless device and the other wireless device.
The machine learning device according to claim 1 or 2. The article information includes information on the number of the predetermined articles,
The machine learning device according to claim 3. The predetermined article is an air conditioner (A) and/or a beam (B) arranged in a space above the ceiling.
The machine learning device according to claim 3 or 4. The article information includes information regarding the size of the predetermined article.
The machine learning device according to any one of claims 3 to 5. The item information includes information regarding the location of the specified item.
The machine learning device according to claim 3 . The article information includes information regarding the orientation of the predetermined article.
The machine learning device according to any one of claims 3 to 7. The learning unit is
A learning data set including the result of performing actual measurement of the radio wave propagation state (55) and the state variables at the time of the actual measurement of the radio wave propagation state;
and learning the state variables and the radio wave propagation state in association with each other based on the above.
The machine learning device according to claim 1 or 2. The learning department is
a learning data set consisting of the result of actually measuring the radio wave propagation state (55) and the state variable obtained from the article information at the time of the actual measurement of the radio wave propagation state;
learning by associating the state variable with the radio wave propagation state based on
The machine learning device according to any one of claims 3 to 8. The learning unit adjusts coefficients of a linear model (35) of the following equation 1 to associate and learn the state variables with the radio wave propagation state.
The machine learning device according to claim 1 , 2 , or 9 .
Formula 1:

The learning unit adjusts coefficients of a linear model (35) of the following equation 2 to associate and learn the state variables with the radio wave propagation state.
The machine learning device according to claim 3 ,
Formula 2:

an output unit (40) that outputs the estimation result of the radio wave propagation state;
an updating unit (60) that updates the learning state of the learning unit;
Furthermore,
The updating unit updates the learning state of the learning unit by evaluating the difference between the estimation result (45) of the radio wave propagation state and the result (55) of actually measuring the radio wave propagation state.
A machine learning device according to any one of claims 1 to 12. an output unit (40) that outputs the estimation result of the radio wave propagation state;
an updating unit (60) that updates the learning state of the learning unit;
Furthermore,
The update section is
In the above formula 1, so that the following evaluation function 1 becomes small,

By updating the
updating the learning state of the learning section;
The machine learning device according to claim 11.
Evaluation function 1:

an output unit (40) that outputs the estimation result of the radio wave propagation state;
an updating unit (60) that updates the learning state of the learning unit;
Furthermore,
The update section is
In the above formula 2, so that the following evaluation function 2 is small,

By updating the
updating the learning state of the learning section;
The machine learning device according to claim 12.
Evaluation function 2:

Images