Patent Application
20200401092
The present disclosure relates to the field of building automation, and more precisely the control of environmental conditions in an area of a building. More specifically, the present disclosure presents an environment controller and a method for predicting carbon dioxide (CO2) level variations based on sound level measurements.
Systems for controlling environmental conditions, for example in buildings, are becoming increasingly sophisticated. An environment control system may at once control heating and cooling, monitor air quality, detect hazardous conditions such as fire, carbon monoxide release, intrusion, and the like. Such environment control systems generally include at least one environment controller, which receives measured environmental values, generally from external sensors, and in turn determines set-points or command parameters to be sent to controlled appliances.
The environment controller and the devices under its control (sensors, controlled appliances, etc.) are generally referred to as Environment Control Devices (ECDs). An ECD comprises processing capabilities for processing data received via one or more communication interface and/or generating data transmitted via the one or more communication interface.
For example, the environment controller controls a heating, ventilating, and/or air-conditioning (HVAC) appliance, in order to regulate the temperature, humidity level and CO2 level in an area of a building. The temperature and CO2 level in the area depend on the number of persons present in the area. If the number of persons present in the area increases, the temperature and CO2 level in the area are likely to increase. Similarly, If the number of persons present in the area decreases, the temperature and CO2 level in the area are likely to decrease. By anticipating the variations of temperature and CO2 level in the area, the environment controller is capable of smoothly adjusting the operations of the HVAC appliance, to maintain a safe and comfortable environment for the persons present in the area.
The variations of temperature and CO2 level in the area are correlated to the variations in the number of persons present in the area. However, it is not always possible to directly track the number of persons present in an area. In this case, the evolution of the sound level in the area can be used as a proxy of the evolution of the number of persons present in the area.
Therefore, there is a need for an environment controller and a method for predicting CO2 level variations based on sound level measurements.
According to a first aspect, the present disclosure relates to an environment controller. The environment controller comprises at least one communication interface, memory for storing a predictive model, and a processing unit. The processing unit determines N consecutive sets of frequency domain sound level measurements. Each set of frequency domain sound level measurements comprises a given number M of sound level amplitudes at the corresponding given number M of frequencies. N and M are integers. The processing unit executes a neural network inference engine using the predictive model for inferring one or more output based on inputs. The inputs comprise the N consecutive sets of frequency domain sound level measurements. The one or more output comprises a predicted variation of a Carbon Dioxide (CO2) level.
According to a second aspect, the present disclosure relates to a method for predicting CO2 level variations based on sound level measurements. The method comprises storing a predictive model in a memory of a computing device. The method comprises determining, by a processing unit of the computing device, N consecutive sets of frequency domain sound level measurements. Each set of frequency domain sound level measurements comprises a given number M of sound level amplitudes at the corresponding given number M of frequencies. N and M are integers. The method comprises executing, by the processing unit of the computing device, a neural network inference engine using the predictive model for inferring one or more output based on inputs. The inputs comprise the N consecutive sets of frequency domain sound level measurements. The one or more output comprises a predicted variation of a CO2 level.
According to a third aspect, the present disclosure relates to a non-transitory computer program product comprising instructions executable by a processing unit of an environment controller. The execution of the instructions by the processing unit of the environment controller provides for predicting CO2 level variations based on sound level measurements, by implementing the aforementioned method.
In a particular aspect, determining the N consecutive sets of frequency domain sound level measurements comprises receiving a plurality of consecutive time domain sound level measurements and generating the N consecutive sets of frequency domain sound level measurements based on the plurality of consecutive time domain sound level measurements (for instance by using a Fast Fourier Transform algorithm).
Embodiments of the disclosure will be described by way of example only with reference to the accompanying drawings, in which:
The foregoing and other features will become more apparent upon reading of the following non-restrictive description of illustrative embodiments thereof, given by way of example only with reference to the accompanying drawings.
Various aspects of the present disclosure generally address one or more of the problems related to environment control systems for buildings. More particularly, the present disclosure aims at providing solutions for predicting the evolution of an environmental condition (e.g. CO2 level or temperature) in an area of a building in relation to the evolution of the number of persons present in the area. For this purpose, the evolution of the sound level in the area is used as a proxy of the evolution of the number of persons in the area. A neural network is used in this context.
Area of a building: the expression ‘area of a building’ is used throughout the present specification to refer to the interior of a whole building or a portion of the interior of the building such as, without limitation: a floor, a room, an aisle, etc.
Referring now concurrently to
The area under the control of the environment controller 100 is not represented in the Figures for simplification purposes. As mentioned previously, the area may consist of a room, a floor, an aisle, etc. However, any type of area located inside any type of building is considered within the scope of the present disclosure.
In the context of the present disclosure, a particular type of sensor is used: the sound sensor 200, which measures a sound level in the area and transmits the measured sound level to the environment controller 100.
Other examples of sensors include the CO2 sensor 210 (measuring a CO2 level in the area and transmitting the measured CO2 level to the environment controller 100). Still other examples of sensors not represented in
In a first implementation, a single sensor measures a given type of environmental characteristic value (e.g. CO2 level) for the whole area. In a second implementation, the area is divided into a plurality of zones, and a plurality of sensors measures the given type of environmental characteristic value (e.g. temperature) in the corresponding plurality of zones. In the second implementation, the environment controller 100 calculates an average environmental characteristic value in the area (e.g. an average temperature in the area) based on the environmental characteristic values transmitted by the plurality of sensors respectively located in the plurality of zones of the area.
Additional sensor(s) may be deployed outside of the area and report their measurement(s) to the environment controller 100. For example, the area is a room of a building. An external temperature sensor measures an external temperature outside the building and transmits the measured external temperature to the environment controller 100. Similarly, an external humidity sensor measures an external humidity level outside the building and transmits the measured external humidity level to the environment controller 100.
Each controlled appliance 300 comprises at least one actuation module, to control the operations of the controlled appliance 300 based on the commands received from the environment controller 100. The actuation module can be of one of the following type: mechanical, pneumatic, hydraulic, electrical, electronical, a combination thereof, etc. The commands control operations of the at least one actuation module. Although a single controlled appliance 300 is represented in
An example of a controlled appliance 300 consists of a VAV appliance. Examples of commands transmitted to the VAV appliance 300 include commands directed to one of the following: an actuation module controlling the speed of a fan, an actuation module controlling the pressure generated by a compressor, an actuation module controlling a valve defining the rate of an airflow, etc. This example is for illustration purposes only. Other types of controlled appliances 300 could be used in the context of an environment control system managed by the environment controller 100.
Details of the environment controller 100, sound sensor 200 and control appliance 300 will now be provided.
The environment controller 100 comprises a processing unit 110, memory 120, and a communication interface 130. The environment controller 100 may comprise additional components, such as another communication interface 130, a user interface 140, a display 150, etc.
The processing unit 110 comprises one or more processors (not represented in the Figures) capable of executing instructions of a computer program. Each processor may further comprise one or several cores. The processing unit 110 executes a neural network inference engine 112 and a control module 114, as will be detailed later in the description.
The memory 120 stores instructions of computer program(s) executed by the processing unit 110, data generated by the execution of the computer program(s), data received via the communication interface 130 (or another communication interface), etc. Only a single memory 120 is represented in
The communication interface 130 allows the environment controller 100 to exchange data with remote devices (e.g. the sound sensor 200, the controlled appliance 300, etc.) over a communication network (not represented in
A detailed representation of the components of the sound sensor 200 is not provided in
A detailed representation of the components of the controlled appliance 300 is not provided in
A detailed representation of the components of the training server 400 is not provided in
The execution of the neural network training engine 411 generates a predictive model, which is transmitted to the environment controller 100 via the communication interface of the training server 400. The predictive model is transmitted over a communication network and received via the communication interface 130 of the environment controller 100.
Reference is now made concurrently to
A dedicated computer program has instructions for implementing at least some of the steps of the method 500. The instructions are comprised in a non-transitory computer program product (e.g. the memory 120) of the environment controller 100. The instructions provide for predicting CO2 level variations based on sound level measurements, when executed by the processing unit 110 of the environment controller 100. The instructions are deliverable to the environment controller 100 via an electronically-readable media such as a storage media (e.g. CD-ROM, USB key, etc.), or via communication links (e.g. via a communication network through the communication interface 130).
The instructions of the dedicated computer program executed by the processing unit 110 implement the neural network inference engine 112 and the control module 114. The neural network inference engine 112 provides functionalities of a neural network, allowing to infer output(s) based on inputs using the predictive model, as is well known in the art. The control module 114 provides functionalities allowing the environment controller 100 to interact with and control other devices (e.g. the sound sensor 200 and the controlled appliance 300).
The method 500 comprises the step 505 of executing the neural network training engine 411 to generate the predictive model. Step 505 is performed by the processing unit of the training server 400. This step will be further detailed later in the description.
The method 500 comprises the step 510 of transmitting the predictive model generated at step 505 to the environment controller 100, via the communication interface of the training server 400. Step 510 is performed by the processing unit of the training server 400.
The method 500 comprises the step 515 of receiving the predictive model from the training server 400, via the communication interface 130 of the environment controller 100. Step 515 is performed by the processing unit 110 of the environment controller 100.
The method 500 comprises the step 520 of storing the predictive model in the memory 120 of the environment controller 100. Step 520 is performed by the processing unit 110 of the environment controller 100.
The method 500 comprises the step 525 of determining N consecutive sets of frequency domain sound level measurements Step 525 is performed by the control module 114 executed by the processing unit 110. Each set of frequency domain sound level measurements comprises a given number M of sound level amplitudes at the corresponding given number M of frequencies. N and M are integers.
In a first exemplary implementation, the determination performed at step 525 comprises receiving a plurality of consecutive time domain sound level measurements for an area, via the communication interface 130. The plurality of consecutive time domain sound level measurements are measured by the sound sensor 200 and transmitted to the environment controller 100.
The sound sensor 200 is located in the area for which the time domain sound level measurements are performed. As mentioned previously, examples of areas include a room of a building, an aisle of the building, a floor of the building etc. Alternatively, the plurality of consecutive time domain sound level measurements are not transmitted directly from the sound sensor 200 to the environment controller 100, but transit through one or more intermediate device.
The determination performed at step 525 further comprises generating the N consecutive sets of frequency domain sound level measurements based on the plurality of consecutive time domain sound level measurements received from the sound sensor 200.
As mentioned previously, each sound level measurement of the time domain sound level measurements consists of one of the following: a sound pressure, a sound pressure level, a sound power, etc. Each sound level amplitude of the frequency domain sound level measurements consists of one of the corresponding following: a sound pressure at a given frequency, a sound pressure level at a given frequency, a sound power at a given frequency, etc.
For example, the determination of the N consecutive sets of frequency domain sound level measurements based on the plurality of consecutive time domain sound level measurements uses a Fast Fourier Transform (FFT) algorithm. The FFT algorithm is well known in the art for transforming a time domain signal into a frequency domain signal. It is an adaptation of the Discrete Fourier Transform (DFT) algorithm, which is also well known in the art.
The FFT algorithm operates with two parameters. The first parameter is the sampling rate or sampling frequency Fs, which defines the average number of time domain sound level measurements per second. For example, Fs is equal to 48000 time domain measurements per second (48 kilohertz or kHz). For illustration purposes only,
The second parameter is the selected number of samples or blocklength BL, which is always an integer power to the base 2 in the FFT. For example, BL is equal to 1024 samples.
The following parameters of the FFT are deducted from Fs and BL. The bandwidth Fn=Fs/2 indicates the theoretical maximum frequency that can be determined by the FFT. In the current example, Fn is equal to 24 kHz. The measurement duration D=BL/Fs. In the current example, D is equal to 1024/48000=21.3 milliseconds. The frequency resolution df=Fs/BL indicates the frequency spacing between two measurement results. In the current example, df is equal to 48000/1024=46.88 Hz.
For illustration purposes only,
The sampling frequency Fs is determined by the capabilities of the sound sensor 200, since the sound sensor 200 is the equipment which performs the consecutive sound level measurements in the time domain. The blocklength BL can be adjusted accordingly, for example to reach a target value for the measurement duration D and/or frequency resolution df. For instance, increasing the blocklength BL for a given sampling frequency Fs increases the measurement duration D and decreases the frequency resolution df.
A first plurality of samples of consecutive time domain sound level measurements is used for generating a first set of frequency domain sound level measurements by applying the FFT.
A second plurality of samples of consecutive time domain sound level measurements is used for generating a second set of frequency domain sound level measurements by applying the FFT.
A third plurality of samples of consecutive time domain sound level measurements is used for generating a third set of frequency domain sound level measurements by applying the FFT.
The fist, second and third sets of frequency domain sound level measurements are used for the inputs of the neural network inference engine 112 to infer a predicted variation of a CO2 level in the area where the sound sensor 200 is deployed.
In the previous example, the measurement duration D is equal to 21.3 milliseconds. Thus, a new set of frequency domain sound level measurements is generated approximately every 21.3 milliseconds. Consequently, step 530 is repeated approximately every D*N=21.3*3=64 milliseconds.
Alternatively, a new set of frequency domain sound level measurements is generated periodically at a configured period P. For example, the period P is set to 1 second. At a time t, a first set of frequency domain sound level measurements is generated. At a time t+P (e.g. t+1 second), a second set of frequency domain sound level measurements is generated. At a time t+2*P (e.g. t+2 second), a third set of frequency domain sound level measurements is generated. The three generated sets are then used at step 530. In this case, step 530 is repeated every P*N=1*3=3 seconds. Thus, a predicted variation of a CO2 level is calculated every 3 seconds.
In the previous example, Fn is equal to 24 kHz and df is equal to 48000/1024=46.88 Hz. Thus, each set of frequency domain sound level measurements comprises approximately Fn/df=24000/46.88=512 measurements (each measurement corresponds to an amplitude of the sound level at a given frequency). If the number of measurements is too high to be used as inputs of the neural network inference engine 112 at step 530, a subset of the measurement is selected for being used at step 530. For example, only 20 measurements among the 512 measurements are selected for each set of frequency domain sound level measurements. Thus, with N=3, the number of values used for the inputs of the neural network inference engine 112 at step 530 is 20*3=60 values. The selection of the 20 measurements may be random. Alternatively, 20 frequencies are selected via an experimental process consisting in testing the accuracy of the output of the neural network inference engine 112 for various candidate frequencies. Alternatively, a basic sampling mechanism is used, which consists in selecting one measurements every 26 (approximately 512/20) measurements of the 512 measurements of a given set.
In another exemplary implementation, the determination performed at step 525 comprises directly receiving the N consecutive sets of frequency domain sound level measurements via the communication interface 130.
For example, the sound sensor 200 has the capability of generating the sets of frequency domain sound level measurements (e.g. by applying the previously described FFT to time domain sound level measurements collected by the sound sensor 200). The frequency domain sound level measurements are transmitted by the sensor 200 to the environment controller 100.
In another example, an intermediate computing device (not represented in the Figures) has the capability of generating the sets of frequency domain sound level measurements based on data received from the sound sensor 200 (e.g. based on a plurality of time domain sound level measurement). The intermediate computing device transmits the generated sets of frequency domain sound level measurements to the environment controller 100.
The method 500 comprises the step 530 of executing the neural network inference engine 112 using the predictive model (stored at step 520) for inferring a predicted variation of a CO2 level in the area, based on the N consecutive sets of frequency domain sound level measurements (determined at step 525). Each one of the N (e.g. 3) consecutive sets of frequency domain sound level measurements respectively includes the aforementioned given number M (e.g. 20) of sound level amplitudes at the corresponding given number M of frequencies. Step 530 is performed by the processing unit 110 of the environment controller 100. Details of how the N consecutive sets of frequency domain sound level measurements are used to generate the inputs of the neural network inference engine 112 will be provided later in the description, in relation to
The inputs used by the neural network inference engine 112 may include other parameter(s), in addition to the N consecutive sets of frequency domain sound level measurements. Similarly, the outputs generated by the neural network inference engine 112 may include other predicted data, in addition to the predicted variation of a CO2 level in the area.
For example, the inputs further include N consecutive CO2 level measurements corresponding to the N consecutive sets of frequency domain sound level measurements. The method 500 comprises the optional step (not represented in
The determination of each consecutive CO2 level measurement can be implemented in different ways. We consider a reference interval of time, during which a given set of frequency domain sound level measurements is determined. During the reference interval of time, the CO2 sensor 210 is configured to make a single CO2 level measurement, which is transmitted to the environment controller 100 and used as input (at step 530) for the reference interval of time. Alternatively, during the reference interval of time, the CO2 sensor 210 is configured to make several CO2 level measurements, the average of the several CO2 level measurements being calculated and transmitted by the CO2 sensor 210 to the environment controller 100, and the calculated average is used as input (at step 530) for the reference interval of time. In still another alternative implementation, the CO2 sensor 210 has no knowledge of the reference interval of time and simply transmits CO2 level measurements to the environment controller 100. In this case, for each reference interval of time, the environment controller 100 sends a request to the CO2 sensor 210 to transmit a CO2 level measurement. The CO2 sensor 210 sends the requested CO2 level measurement to the environment controller 100, which is used as input (at step 530) for the reference interval of time. Instead of a single CO2 level measurement for each reference interval of time, the environment controller 100 may request and receive a plurality of CO2 level measurements from the CO2 sensor 210. The environment controller 100 uses as input (at step 530) the average of the plurality of received CO2 level measurements for the reference interval of time. In yet another alternative implementation, during each reference interval of time, the CO2 sensor 210 spontaneously transmits one or more CO2 level measurement to the environment controller 100. The environment controller 100 uses for the reference interval of time as input (at step 530) a single one of the CO2 level measurement(s) received from the CO2 sensor 210 or an average of several CO2 level measurements received from the CO2 sensor 210. The predictive model has been trained at step 505 to use N consecutive CO2 level measurements as inputs (corresponding to the N consecutive sets of frequency domain sound level measurements).
Alternatively, a single CO2 level measurement is used as input at step 530. The single CO2 level measurement is determined during a reference interval of time, during which the N (e.g. 3) sets of frequency domain sound level measurements are determined. The determination of the single CO2 level measurement (based on measurements made by the CO2 sensor 210) is made in a manner similar to one of the previously mentioned alternatives.
In another example, the outputs further include a predicted variation of temperature in the area. The predictive model has been trained at step 505 to predict both a variation of CO2 level and a variation of temperature in the area.
In this case, the inputs may further include N consecutive temperature measurements corresponding to the N consecutive sets of frequency domain sound level measurements. The determination of the N consecutive temperature measurements is similar to the aforementioned determination of the N consecutive CO2 level measurements, and uses a temperature sensor (not represented in the Figures) located in the area. The predictive model has been trained at step 505 to further use N consecutive temperature measurements as inputs. As mentioned previously in relation to the CO2 level measurement, a single temperature measurement corresponding to the N consecutive sets of frequency domain sound level measurements may also be used as input.
The method 500 comprises the step 535 of generating a command for controlling the controlled appliance 300 based on the predicted variation of the CO2 level in the area. Step 535 is performed by the control module 114 executed by the processing unit 110. Optionally, the generation of the command uses one or more additional parameter, such as a current CO2 level in the area transmitted by the CO2 sensor 210.
As mentioned previously, an example of controlled appliance 300 is a VAV appliance. Examples of commands for controlling the VAV appliance 300 include commands directed to one of the following actuation modules of the VAV appliance 300: an actuation module controlling the speed of a fan, an actuation module controlling the pressure generated by a compressor, an actuation module controlling a valve defining the rate of an airflow, etc.
In an alternative implementation, steps 525-530 are repeated several times before performing step 535. The command generated at step 535 is based on several consecutive predicted variations of the CO2 level in the area (generated by the repetition of step 525).
The method 500 comprises the step 540 of transmitting the command (generated at step 535) to the controlled appliance 300 via the communication interface 130. Step 540 is performed by the control module 114 executed by the processing unit 110.
The method 500 comprises the step 545 of receiving the command at the controlled appliance 300, via the communication interface of the controlled appliance 300. Step 545 is performed by the processing unit of the controlled appliance 300.
The method 500 comprises the step 550 of applying the command at the controlled appliance 300. Step 550 is performed by the processing unit of the controlled appliance 300. Applying the command consists in controlling one or more actuation module of the controlled appliance 300 based on the received command.
In an alternative implementation, instead of performing steps 535 and 540, the environment controller 100 transmits (via the communication interface 130) the predicted variation of the CO2 level in the area to another device (not represented in the Figures). The other device performs steps 535 and 540 instead of the environment controller 100.
A plurality of commands may be generated at step 535 and transmitted at step 540 to the same controlled appliance 300. Alternatively, the same command may be generated at step 535 and transmitted at step 540 to a plurality of controlled appliances 300. In yet another alternative, a plurality of commands may be generated at step 535 and transmitted at step 540 to a plurality of controlled appliances 300. In any case, the command(s) are always based on the predicted variation of the CO2 level (inferred at step 530).
Various algorithms may be used for generating the command based on one or more parameter comprising the predicted variation of the CO2 level in the area. These algorithms are out of the scope of the present disclosure. However, examples of such algorithms are well known in the art of environment control.
If the one or more output generated at step 530 further comprises a predicted variation of temperature in the area, the generation of a particular command may use the predicted variation of CO2 level in the area only, the predicted variation of temperature in the area only, or a combination of the predicted variation of CO2 level in the area and the predicted variation of temperature in the area. In this case, the generation of a particular command optionally uses a current temperature in the area transmitted by a temperature sensor located in the area, in addition to the predicted variation of temperature in the area.
The steps of the method 500 involving the reception or the transmission of data by the environment controller 100 may use the same communication interface 130 or different communication interfaces 130. For example, step 515 uses a first communication interface 130 of the Ethernet type, while steps 525 and 540 use a second communication interface 130 of the Wi-Fi type. In another example, steps 515, 525 and 540 use the same communication interface 130 of the Wi-Fi type.
As mentioned previously, each one of the N consecutive sets of frequency domain sound level measurements determined at step 525 of the method 500 comprises the number M of sound level amplitudes at the corresponding number M of frequencies.
The first set of frequency domain sound level measurements consists of a first set of M sound level amplitude values at respective frequencies F1, F2 . . . FM. The second set of frequency domain sound level measurements consists of a second set of M sound level amplitude values at respective frequencies F1, F2 . . . FM. The last set (N) of frequency domain sound level measurements consists of a last set of M sound level amplitude values at respective frequencies F1, F2 . . . FM.
As illustrated in
If the value of N is low (e.g. 3, 4 or 5), the neural network inference engine 112 implements a neural network comprising an input layer, one or more intermediate layer, and an output layer; where all the layers are fully connected. The input layer comprises at least N*M neurons, where each one among the N*M neurons of the input layer receives an amplitude Ai,j (i varies from 1 to N and j varies from 1 to M). For example, if N=5 and M=20, the input layer comprises at least 100 neurons respectively receiving the amplitudes Aij. The output layer comprises at least one neuron outputting the predicted variation of CO2 level. The generation of the outputs based on the inputs using weights allocated to the neurons of the neural network is well known in the art for a neural network with only fully connected layers. The architecture of the neural network, where each neuron of a layer (except for the first layer) is connected to all the neurons of the previous layer is also well known in the art.
If the value of N is higher (e.g. 10, 20 or 30), the neural network inference engine 112 implements a convolutional neural network comprising an input layer, several intermediate layers, and an output layer. The layers immediately after the input layer comprise one or more convolutional layer. Optionally, each convolutional layer is followed by a pooling layer. The rest of the layers up to the output layer consists of fully connected layers.
In a first implementation, one dimensional (1D) convolution is used. The input layer comprises at least M neurons, where each one among the M neurons of the input layer receives a one-dimension matrix with amplitude values (i varies from 1 to N for each matrix and j varies from 1 to M across the matrices). For example, if N=25 and M=20, the input layer comprises at least 20 neurons respectively receiving the jth matrixes [A1,j, A2,j, . . . A25,j]. Each matrix [A1,j, A2,j, . . . A25,j] represents the evolution of the amplitude at a given frequency Fj over the 25 consecutives sets determined at step 525 of the method 500.
The first layer following the input layer is a 1D convolutional layer applying a 1D convolution to the M matrixes [A1,j, A2,j, . . . AN,j]. The 1D convolution uses a one dimension filter of size S lower than N. The output of the 1D convolutional layer consists in M resulting matrixes [B1,j, B2,j, . . . BN,j]. As mentioned previously, the 1D convolutional layer may be followed by a pooling layer for reducing the size of the M resulting matrixes [B1,j, B2,j, . . . BN,j] into respective reduced matrixes [C1,j, C2,j, . . . CO,j] where O is lower than N. Various algorithms (e.g. maximum value, minimum value, average value, etc.) can be used for implementing the pooling layer, as is well known in the art (a one dimension filter of given size is also used by the pooling layer).
The neural network may include several consecutive 1D convolutional layers, optionally respectively followed by pooling layers. The M matrixes [A1,j, A2,j, . . . AN,j] are processed independently of one another along the chain of 1D convolutional layer(s) and optional pooling layer(s).
The chain of 1D convolutional layer(s) and optional pooling layer(s) is followed by one or more fully connected layer, which operates with weights associated to neurons, as is well known in the art.
In a second implementation, two dimensional (2D) convolution is used. The input layer comprises at least one neuron receiving a two-dimensions matrix with amplitude values Ai,j (i varies from 1 to N and j varies from 1 to M). For example, if N=25 and M=20, the input layer comprises a neuron receiving the 25×20 input matrix. Following is a representation of the input matrix:
Each line [A1,j, A2,j, . . . AN,j] of the matrix represents the evolution of the amplitude at a given frequency Fj over the N (e.g. 25) consecutives sets determined at step 525 of the method 500.
The first layer following the input layer is a 2D convolutional layer applying a 2D convolution to the N×M (e.g. 25×20) matrix. The 2D convolution uses a two-dimensions filter of size S×T, where S is lower than N and T is lower than M. The output of the 2D convolutional layer consists in a resulting matrix:
As mentioned previously, the 2D convolutional layer may be followed by a pooling layer for reducing the size of the resulting matrix into a reduced matrix:
where O is lower than N and P is lower than M. Various algorithms can be used for implementing the pooling layer, as is well known in the art (a two-dimensions filter of given size is also used by the pooling layer).
The neural network may include several consecutive 2D convolutional layers, optionally respectively followed by pooling layers.
The chain of 2D convolutional layer(s) and optional pooling layer(s) is followed by one or more fully connected layer, which operates with weights associated to neurons, as is well known in the art.
The usage of one or more 1D convolutional layer allows to detect patterns between values of the amplitude at a given frequency, but not across frequencies. The usage of one or more 2D convolutional layer allows to detect patterns between values of the amplitude at different frequencies.
As mentioned previously and optionally, a plurality of N consecutive CO2 level measurements in the area is also used for the inputs of the neural network inference engine 112 when performing step 530.
In the case where the neural network inference engine 112 implements a neural network only comprising fully connected layers, the input layer comprises N additional neurons for receiving the N consecutive CO2 level measurements.
In the case where the neural network inference engine 112 implements a neural network comprising 1D convolutional layer(s) and optionally pooling layer(s), the input layer comprises one additional neuron for receiving a matrix comprising the N consecutive CO2 level measurements. The processing of the matrix comprising the N consecutive CO2 level measurements is similar to the processing of the M matrixes comprising the amplitude values [Ai,1, Ai,2, . . . Ai,N]. This use case is not represented in
In the case where the neural network inference engine 112 implements a neural network comprising 2D convolutional layer(s) and optionally pooling layer(s), the input layer comprises one additional neuron for receiving a matrix comprising the N consecutive CO2 level measurements. The matrix comprising the N consecutive CO2 level measurements is treated by one or more 1D convolutional layer operating in parallel with the one or more 2D convolutional layer treating the amplitudes. Alternatively, one line comprising the N consecutive CO2 level measurements is added to the matrix comprising the amplitudes, and the CO2 level measurements are also treated by the one or more 2D convolutional layer. This use case is not represented in
The usage of optional inputs (e.g. the plurality of consecutive CO2 level measurements in the area) may improve the accuracy and resiliency of the inferences performed by the neural network inference engine 112 (at the cost of complexifying the predictive models used by the neural network inference engine 112). The relevance of using some optional inputs is generally evaluated during the training phase, when the predictive models are generated (and tested) with a set of training (and testing) inputs and outputs dedicated to the training (and testing) phase.
The optional output consisting of the predicted variation of temperature and the optional input(s) consisting of the temperature measurement(s) are not represented in
Referring back to
The inputs and output(s) for the training phase of the neural network can be collected through an experimental process. For example, a plurality of combinations of N sets of frequency domain sound level measurements are collected for different combination of persons present in the area, entering the area, leaving the area, etc. The collection is made by deploying the sound level sensor 200 in the area; and further deploying a computing device (e.g. the training server 400) capable of performing the determination of the N sets of frequency domain sound level measurements based on data provided by the sound level sensor 200. For each combination of N sets of frequency domain sound level measurements, a corresponding variation of the CO2 level in the area is determined using data provided by the CO2 sensor 210. The combinations of N sets of frequency domain sound level measurements and the corresponding variations of the CO2 level in the area are respectively used as inputs and output(s) by the neural network training engine 411 for generating the predictive model by adjusting the weights associated to the neurons of the fully connected layers.
Various techniques well known in the art of neural networks are used for performing (and improving) the generation of the predictive model, such as forward and backward propagation, usage of bias in addition to the weights (bias and weights are generally collectively referred to as weights in the neural network terminology), reinforcement training, etc.
In the case where a convolutional layer is used, parameters of the convolutional layer are also adapted during the training phase. For example, the size of the filter used for the convolution is determined during the training period. The parameters of the convolutional layer are included in the predictive model.
Similarly, in the case where a pooling layer is used, parameters of the pooling layer are also adapted during the training phase. For example, the algorithm and the size of the filter used for the pooling operation are determined during the training period. The parameters of the polling layer are included in the predictive model.
Reference is now made concurrently to
A first plurality of environment controllers 100 implementing the method 500 are deployed at a first location. Only two environment controllers 100 are represented for illustration purposes, but any number of environment controllers 100 may be deployed.
A second plurality of environment controllers 100 implementing the method 500 are deployed at a second location. Only one environment controller 100 is represented for illustration purposes, but any number of environment controllers 100 may be deployed.
The first and second locations may consist of different buildings, different floors of the same building, etc. Only two locations are represented for illustration purposes, but any number of locations may be considered.
Each environment controller 100 represented in
In an alternative configuration, at least some of the environment controllers 100 also execute the neural network training engine 411. In this case, the training phase (performed by the neural network training engine 411 for generating the predictive model) and the operational phase (performed by the neural network inference engine 112) are both executed by the environment controller 100.
Although the present disclosure has been described hereinabove by way of non-restrictive, illustrative embodiments thereof, these embodiments may be modified at will within the scope of the appended claims without departing from the spirit and nature of the present disclosure.
