A method for detecting air volume installed in smart doors and windows

Abstract

The present invention discloses a method for detecting wind volume installed on smart doors and windows, comprising the following steps: (1) installing at least two ultrasonic transducers on the window to collect ultrasonic signals within a time slot; (2) analyzing the ultrasonic signals received by the receiving ultrasonic transducer to obtain energy characteristics, time characteristics, and frequency characteristics; (3) static feature training to calibrate a data extraction deep neural network model; (4) pre-processing the received ultrasonic signal and sending it to the trained deep neural network model; (5) comparing the real-time energy characteristics, time characteristics, and frequency characteristics with a decision table to determine the wind volume. The present invention utilizes ultrasonic transducers to detect the wind volume blowing toward the window by extracting the static and dynamic characteristics of energy, time, and frequency of different ultrasonic transducers, thereby realizing intelligent wind volume sensing for the window without affecting the structural integrity and aesthetics of the window.

Description

Technical Field

[0001] The present invention relates to the field of intelligent doors and windows, and in particular to a method for detecting air volume installed in intelligent doors and windows. Background Art

[0002] The integration of sensing, interconnected, and intelligent technologies into architecture and homes is transforming the way people, buildings, nature, and living spaces interact. Windows are the gateway between living spaces and nature, and their intelligence significantly impacts the quality of living spaces. In the field of smart windows, for example, natural wind sensing measures wind volume and transmits this information to the smart home controller via a communication system. The controller then integrates this information and makes appropriate control decisions, which are then transmitted back via the communication system to the window's servo mechanism, ultimately controlling the window's behavior and maintaining comfort and safety within the building's interior.

[0003] Currently, the most commonly used wind speed measuring instruments include mechanical cup anemometers, laser Doppler anemometers, simple pendulum anemometers, differential pressure anemometers, thermal anemometers, and ultrasonic anemometers.

[0004] The cup anemometer is the most widely used method for measuring wind speed. This method is technically mature, simple to use, and low-cost. However, its measurement accuracy is limited. Due to the inertia and mechanical friction of the cup, the cup will only rotate when the wind speed exceeds the starting wind speed, making it only suitable for measuring relatively high wind speeds.

[0005] Thermal anemometers use a thermal probe to measure gas flow rate. A thermocouple wrapped around a hot wire measures the temperature change of the wire. Because there's a functional relationship between air volume and temperature, wind speed can be measured accordingly. This method of measuring wind speed is highly accurate at low flow rates, but it requires manual intervention, making it inconvenient. Furthermore, when gas flows in multiple directions within the flow field, measurement accuracy is poor.

[0006] An ultrasonic anemometer is a measuring device based on the principle that the propagation characteristics of ultrasonic waves are modulated by air flow. Methods for measuring wind speed using ultrasonic anemometers include time difference, phase difference, frequency difference, and Doppler methods. The phase difference method requires that the phase change must be within a single phase period, limiting the wind speed measurement range. The frequency difference method suffers from small frequency changes at low wind speeds, preventing high-precision measurements. The Doppler method is only suitable for applications where the measured medium contains certain impurities.

[0007] Ultrasonic waves, as high-frequency mechanical vibrations, propagate through elastic media at a speed determined by the medium's characteristic parameters. When propagating through air, the speed of sound is influenced by a combination of factors, including air density, temperature, humidity, air pressure, and air purity. Furthermore, the relationship between many air parameters and the speed of sound is nonlinear, making it difficult to express analytically. This typically requires at least four sets of ultrasonic transducers to be positioned as required, resulting in a bulky device.

[0008] Furthermore, existing wind speed measuring instruments, such as mechanical cup anemometers, laser Doppler anemometers, pendulum anemometers, differential pressure anemometers, thermal anemometers, and ultrasonic anemometers, must be installed protruding from the exterior of the mounting surface to ensure that the measuring instrument and mounting surface do not affect the wind field, thereby ensuring smooth, undistorted air flow through the measuring instrument. In the smart window industry, wind speed measurement is essential for ensuring the safety of windows, interior spaces, and buildings; at the same time, the installation of anemometers must not affect the structural integrity and aesthetics of the windows. These two contradictions make traditional anemometers unsuitable for use in the smart window and door industry. Summary of the Invention

[0009] The purpose of the present invention is to provide a method for detecting wind volume installed in smart doors and windows. The method uses ultrasonic transducers to detect the amount of wind blowing toward the windows by extracting the static and dynamic characteristics of different ultrasonic transducers in energy, time, and frequency. This method realizes intelligent wind volume perception of windows without affecting the integrity and aesthetics of the window structure.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] A method for detecting air volume installed in smart doors and windows comprises the following steps:

[0012] (1) Install at least two ultrasonic transducers on the window, and s There is only one ultrasonic transducer that transmits ultrasonic signals, and the transmission time is T0, T0<T s ; Multiple ultrasonic transducers take turns to occupy time slots to generate ultrasonic signals, and do not occupy the time slots of other ultrasonic transducers;

[0013] (2) analyzing the ultrasonic signal received by the ultrasonic transducer to obtain energy characteristics, time characteristics, and frequency characteristics;

[0014] The peak value of the envelope of the ultrasonic signal is the energy characteristic;

[0015] The moment of the first zero crossing after the envelope of the ultrasonic signal exceeds the threshold is the time feature;

[0016] The amplitude ratio of the ultrasonic signal at the preset frequency points f1, f2 and f3 is the frequency characteristic. The ultrasonic signal is subjected to fast Fourier transform to extract the amplitudes A1, A2 and A3 of the preset frequency points f1, f2 and f3, and A2 / A1 and A2 / A3 are calculated to obtain the frequency characteristics.

[0017] (3) Static feature training, used to calibrate the data extraction deep neural network model;

[0018] (4) Dynamic feature learning: The received ultrasonic signal is pre-processed and sent to the deep neural network model extracted in step (3). The deep neural network model processes the signal in real time to obtain the energy feature, time feature, and frequency feature corresponding to the currently received ultrasonic signal.

[0019] (5) Air volume decision: Compare the real-time energy characteristics, time characteristics and frequency characteristics with the decision table, and calculate the air volume in the preset time window T w The energy characteristics, time characteristics, and frequency characteristics of each ultrasonic signal are compared with each row of the decision table. After the comparison is successful, the count is performed, and the row with the largest count value is found. The corresponding air volume value is output and the count value is reset to zero to make a new air volume decision in the next time window.

[0020] Furthermore, an envelope detection method is specifically used to calculate the envelope of the received ultrasonic signal and find the peak value of the envelope.

[0021] Furthermore, the threshold in step (2) is 60.

[0022] Furthermore, the step (3) for correcting the data extraction deep neural network model specifically includes the following steps:

[0023] Calibration data preparation: Calibration data includes wind volume data and label data. Wind volume data includes wind speed and wind direction. Label data is used to annotate the energy characteristics, time characteristics, and frequency characteristics of ultrasonic signals.

[0024] Deep learning network model: The deep learning network model consists of an input layer, four convolutional layers, a fully connected layer, and an output layer. The number of neurons in the four convolutional layers is 4, 8, 4, and 2, respectively. The input layer of the deep learning network models for energy and time features has two input nodes, while the input layer of the deep learning network model for frequency features has four input nodes. Normalized mean square error is used as the cost function for each network model.

[0025] Deep learning network training: The energy characteristics, time characteristics and frequency characteristics of the wind volume data in the correction data are respectively input into the deep neural network model to obtain the current output results, and the normalized mean square error is calculated between the current output results and the corresponding label data; the normalized mean square error is then used to adopt the back gradient propagation algorithm to update the network parameters until the performance of the deep neural network reaches the preset threshold, thereby obtaining the corresponding final deep neural network model.

[0026] Furthermore, the preset frequency points are 39.5kHz, 40kHz and 40.5kHz respectively.

[0027] The present invention uses two or more ultrasonic transducers installed on doors and windows, and analyzes the air volume through the signal between the two ultrasonic transducers that changes in air flow caused by the wind volume, so that the measured wind speed is not affected by the environment. In addition, it can adapt to a relatively cramped installation space and detect the correspondence between the wind direction, wind volume, field type and ultrasonic sound field parameters of the wind field, thereby realizing the detection of wind volume. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the process of the present invention.

[0029] Figure 2 Schematic diagram of ultrasonic signal of the present invention.

[0030] Figure 3 Schematic diagram of the deep learning network structure of the present invention. DETAILED DESCRIPTION

[0031] like Figure 1 As shown, the present embodiment provides a method for detecting air volume on smart doors and windows, comprising the following steps:

[0032] (1) Install at least two ultrasonic transducers on the window, and s There is only one ultrasonic transducer that transmits ultrasonic signals, and the transmission time is T0, T0<T s ; Multiple ultrasonic transducers take turns to occupy time slots to generate ultrasonic signals, and do not occupy the time slots of other ultrasonic transducers;

[0033] In this embodiment, two ultrasonic transducers are set up, and the two ultrasonic transducers transmit and receive signals in turn. When the number of ultrasonic transducers is greater than or equal to 3, one of them can be fixed as a receiving ultrasonic transducer, and the rest are transmitting ultrasonic transducers. The time slot T s = 100ms, T0 = 1ms, the ultrasonic signal received by the ultrasonic transducer is as follows Figure 1 shown.

[0034] (2) Analyze the ultrasonic signal received by the receiving ultrasonic transducer to obtain energy characteristics, time characteristics and frequency characteristics, as follows:

[0035] The peak value of the envelope of the ultrasonic signal is the energy characteristic; specifically, the envelope of the received ultrasonic signal is calculated using the envelope detection method, and the peak value of the envelope is found, which is the energy characteristic e;

[0036] The time feature is the moment when the ultrasonic signal envelope first crosses zero after exceeding the threshold. In this embodiment, the threshold is 60. The envelope of the ultrasonic signal exceeding the threshold is compared with the threshold 60. When the instantaneous envelope is greater than the threshold, the moment when the next most recent ultrasonic echo signal changes from positive to negative or from negative to positive is recorded as the time feature t.

[0037] The amplitude ratio of the ultrasonic signal at the preset frequencies f1, f2 and f3 is the frequency characteristic. The preset frequencies are 39.5kHz, 40kHz and 40.5kHz. The ultrasonic signal is subjected to fast Fourier transform to extract the amplitudes A1, A2 and A3 of the preset frequencies f1, f2 and f3, and A2 / A1 and A2 / A3 are calculated, thus obtaining the frequency characteristics fu and fd.

[0038] (3) Static feature training: used to calibrate the feature extraction deep neural network. The specific process is as follows:

[0039] Calibration data preparation: Multiple ultrasonic transducers are installed on the windows, and calibration data is input. The calibration data includes wind volume data and label data. The wind volume data are generated by existing equipment (such as hair dryers, fans, etc.), and multiple sets of wind volumes with different wind speeds and directions can be generated by controlling the equipment. The label data is used to mark the energy characteristics, time characteristics and frequency characteristics in the ultrasonic signal.

[0040] Deep learning network model: The energy features and time features adopt a deep learning network model with the same structure. The network model includes an input layer, four convolutional layers, a fully connected layer and an output layer. The number of neurons in the four convolutional layers are 4, 8, 4 and 2 respectively, and the input layer has 2 input nodes; the 2 input nodes correspond to the energy features e or time features t of the ultrasonic signals received in two consecutive time slots. This embodiment is specifically defined as the first input node corresponding to the energy features e or time features t of the odd time slots, and the second input node corresponding to the energy features e or time features t of the even time slots; the output of the deep neural network model of energy features and time features is the fused energy features e or time features t.

[0041] The frequency-feature deep neural network model also includes an input layer, four convolutional layers, a fully connected layer, and an output layer. The number of neurons in the four convolutional layers is 4, 8, 4, and 2, respectively. The input layer has four input nodes, corresponding to the frequency features fu and fd of receiving ultrasound in two consecutive time slots. The specific correspondence is: the first input node corresponds to the frequency feature fu of odd time slots, the second input node corresponds to the frequency feature fd of odd time slots, the third input node corresponds to the frequency feature fu of even time slots, and the fourth input node corresponds to the frequency feature fd of even time slots. The output of the frequency-dimensional deep neural network model is the fused frequency feature f.

[0042] The normalized mean square error is used as the cost function of each network model, and the standard error RMSE = (A-Label) 2 / Label, where A is the energy feature e, time feature t, and frequency feature f of the received ultrasonic signal output by the deep learning network model, and Label is the energy feature, time feature, and frequency feature of the label data.

[0043] Deep learning network training: The energy characteristics, time characteristics and frequency characteristics of the wind volume data in the correction data are respectively input into the deep neural network model to obtain the current output results, and the normalized mean square error is calculated between the current output results and the corresponding label data; the normalized mean square error is then used to adopt the back gradient propagation algorithm to update the network parameters. The training is continuously carried out until the performance of the deep neural network reaches the preset threshold, thereby obtaining the corresponding final deep neural network model.

[0044] (4) Dynamic feature learning: The received ultrasonic signal is pre-processed and then sent to the trained deep neural network model. The deep neural network model processes the signal in real time to obtain the energy, time, and frequency characteristics corresponding to the currently received ultrasonic signal.

[0045] (5) Air volume decision: Compare the real-time energy characteristics, time characteristics and frequency characteristics with the decision table, and calculate the air volume in the preset time window T w The energy characteristics, time characteristics, and frequency characteristics of each ultrasonic signal are compared with each row of the decision table. After the comparison is successful, the count is performed, and the row with the largest count value is found. The corresponding air volume value is output and the count value is reset to zero to make a new air volume decision in the next time window.

[0046] The decision table is shown in Table 1. There is a one-to-one correspondence between the value ranges of energy characteristics, time characteristics, and frequency characteristics and the wind volume levels. Each row specifies the value range of energy characteristics, time characteristics, and frequency characteristics corresponding to a wind volume level. When the values ​​of real-time energy characteristics, time characteristics, and frequency characteristics fall within the value range corresponding to each row of the decision table, the row is successfully compared.

[0047] Table 1 Air volume decision table

[0048] Serial number E T F Air volume 1 >800 [50,70] [2.1，2.6] 1 2 [600,800] [35,50] or [70,85] [2.6，3.1] 2 3 [500,600] [27,35] or [85,103] [3.1，3.7] 3 4 [430,500] [21,27] or [103,109] [3.7，4.7] 4 5 [390,430] [16,21] or [109,114] [4.7，6.0] 5 6 [360,390] [12,16] or [114,119] [6.0，8.0] 6 7 <360 <12 or >119 >8.0 7

[0049] The preset time window Tw in this embodiment is Ts = N × Ts, where N is a constant and takes the value of 600, that is, the ultrasonic signal within 1 minute is calculated, and the maximum wind volume within 1 minute is obtained as the current wind volume. When the wind volume exceeds the set value, the controller controls the smart doors and windows to close.

[0050] The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any modification and replacement based on the technical solution and inventive concept provided by the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for detecting air volume installed in smart doors and windows, characterized in that The steps include: (1) Install at least two ultrasonic transducers on the window, and s There is only one ultrasonic transducer that transmits ultrasonic signals, and the transmission time is T0, T0<T s ; Multiple ultrasonic transducers take turns to occupy time slots to generate ultrasonic signals, and do not occupy the time slots of other ultrasonic transducers; (2) analyzing the ultrasonic signal received by the ultrasonic transducer to obtain energy characteristics, time characteristics, and frequency characteristics; The peak value of the envelope of the ultrasonic signal is the energy characteristic; The moment of the first zero crossing after the envelope of the ultrasonic signal exceeds the threshold is the time feature; The amplitude ratio of the ultrasonic signal at the preset frequency points f1, f2 and f3 is the frequency characteristic. The ultrasonic signal is subjected to fast Fourier transform to extract the amplitudes A1, A2 and A3 of the preset frequency points f1, f2 and f3, and A2 / A1 and A2 / A3 are calculated to obtain the frequency characteristics. (3) Static feature training, used to calibrate the data extraction deep neural network model; (4) Dynamic feature learning: The received ultrasonic signal is pre-processed and sent to the deep neural network model extracted in step (3). The deep neural network model processes the signal in real time to obtain the energy feature, time feature, and frequency feature corresponding to the currently received ultrasonic signal. (5) Air volume decision: Compare the real-time energy characteristics, time characteristics and frequency characteristics with the decision table, and calculate the air volume in the preset time window T w The energy characteristics, time characteristics, and frequency characteristics of each ultrasonic signal are compared with each row of the decision table. After the comparison is successful, the count is performed, and the row with the largest count value is found. The corresponding air volume value is output and the count value is reset to zero to make a new air volume decision in the next time window.

2. The method for detecting air volume installed in smart doors and windows according to claim 1, characterized in that: Specifically, the envelope detection method is used to calculate the envelope of the received ultrasonic signal and find the peak value of the envelope.

3. The method for detecting air volume installed in smart doors and windows according to claim 1, characterized in that: The threshold in step (2) is 60.

4. The method for detecting air volume installed in smart doors and windows according to claim 1, characterized in that: The step (3) for correcting the data extraction deep neural network model specifically includes the following steps: Calibration data preparation: Calibration data includes wind volume data and label data. Wind volume data includes wind speed and wind direction. Label data is used to annotate the energy characteristics, time characteristics, and frequency characteristics of ultrasonic signals. Deep learning network model: The deep learning network model consists of an input layer, four convolutional layers, a fully connected layer, and an output layer. The number of neurons in the four convolutional layers is 4, 8, 4, and 2, respectively. The input layer of the deep learning network models for energy and time features has two input nodes, while the input layer of the deep learning network model for frequency features has four input nodes. Normalized mean square error is used as the cost function for each network model. Deep learning network training: The energy characteristics, time characteristics and frequency characteristics of the wind volume data in the correction data are respectively input into the deep neural network model to obtain the current output results, and the normalized mean square error is calculated between the current output results and the corresponding label data; the normalized mean square error is then used to adopt the back gradient propagation algorithm to update the network parameters until the performance of the deep neural network reaches the preset threshold, thereby obtaining the corresponding final deep neural network model.

5. The method for detecting air volume installed in smart doors and windows according to claim 4, characterized in that: The preset frequency points are 39.5kHz, 40kHz and 40.5kHz respectively.

Images