JP7638802B2 - Control system, control method, and program for variable damping type active damper - Google Patents

Description

The present invention relates to a control system, control method, and program for a variable damping type active damper.

In mid- to high-rise, high-rise, and super-high-rise buildings equipped with active dampers, the so-called active control of the active dampers actively supplies energy from the outside to the building, thereby suppressing vibrations caused by earthquake loads and wind loads acting on the building. In order to improve the seismic isolation performance of buildings, variable damping active dampers that make it possible to exert multiple stages of damping performance are sometimes applied to these active dampers. Variable damping active dampers can effectively reduce deformation and acceleration occurring in buildings by changing the damping performance of the damper according to the response amount of the building, and seismic isolation technology and seismic isolation control that apply these dampers are called semi-active seismic isolation and semi-active seismic isolation control.

For example, a variable damping type active damper is formed by an oil damper in which a piston slides inside a cylinder filled with oil, and is configured so that the damping coefficient (damping force) can be freely switched between multiple stages by opening and closing an electromagnetic valve, etc. When the damping force of this damper is switched, a time delay occurs according to, for example, the sliding speed of the piston. In other words, since the sliding speed of the piston changes according to the damping force, the time delay changes according to the damping force. Therefore, even if a command signal corresponding to the damping force is sent to the active damper to exert a damping force suitable for reducing external forces such as earthquake loads and wind loads and building vibrations caused by these external forces, the above-mentioned time delay causes a problem in that the expected vibration reduction effect cannot be obtained. The above-mentioned time delay associated with switching the damping force of the damper is called the "time constant."

Here, Non-Patent Document 1 discloses a method for predicting and controlling the resonance response of a skyscraper against long-period earthquake motion, which assumes that the resonance component of ground acceleration can be extracted in real time, predicts the resonance component of ground acceleration several seconds ahead using a deep learning method for waves band-pass filtered in a narrow band around the resonance frequency, and feed-forward controls the active damper. Here, for deep learning, a recurrent neural network (RNN), a long short-term memory (LSTM), and a gated recurrent unit (GRU), which are suitable for time series learning, are applied as waveform prediction models, and band-pass filtering is performed in a band of 4.5 to 5.5 seconds including the natural frequency of the main system of the controlled model when learning the model. Then, the numerical data of the past 15 seconds is used as learning data, and the numerical data one step ahead from the current time is used as training data, and in order to predict several seconds ahead, a cycle of predicting the next one step ahead using the predicted information one step ahead as known is repeated to predict the numerical data 15 seconds ahead from the current time.

Abstracts of the Academic Lectures at the Architectural Institute of Japan Annual Meeting (Hokuriku), September 2019, p.909-p.910 Predictive Control of Resonant Responses of Skyscrapers to Long-Period Earthquake Motion

The resonance response prediction control method described in Non-Patent Document 1 is a technology that uses deep learning techniques to predict the resonance components of ground acceleration several seconds ahead and feedforward control the active damper, but it does not resolve the issue that variable damping type active dampers cannot achieve sufficient vibration reduction effects due to the time constant.

The present invention has been made in consideration of the above problems, and aims to provide a control system, control method, and program for a variable damping type active damper that can achieve the expected vibration reduction effect without being affected by the time constant.

In order to achieve the above object, one aspect of a control system for a variable damping type active damper according to the present invention comprises:
A control system for a variable damping active damper that controls a variable damping active damper installed in a building, comprising:
A vibration meter that measures earthquake waveforms or building vibrations;
a control device that receives measurement waveform data that changes from moment to moment and is measured by the vibration meter, creates predicted waveform data based on data that has been received up to a predetermined time of the measurement waveform data, and drives and controls the variable damping type active damper based on the predicted waveform data,
The control device includes:
a communication unit that receives the measurement waveform data from the vibration meter and transmits a damping force signal related to a damping force to the variable damping type active damper;
a prediction unit that predicts a waveform from a predetermined time onward based on data received up to a predetermined time of the measured waveform data, to generate the predicted waveform data;
a control signal generating unit that uses the predicted waveform data to specify the damping force at each time to generate the damping force signal and to specify a time constant associated with switching of the damping force,
the prediction unit creates the predicted waveform data from the measured waveform data based on a learning result obtained in advance by learning sine waves of a plurality of frequencies through machine learning;
The damping force signal is transmitted from the communication unit to the variable damping type active damper with a time corresponding to the time constant advanced.

According to this aspect, the prediction section of the control device creates predicted waveform data based on data received up to a specified time of the measured waveform data, the control signal generation section of the control device uses the predicted waveform data to identify the damping force at each time and the time constant associated with switching the damping force, and the communication section of the control device advances the time equivalent to the time constant and transmits the damping force signal to the variable damping type active damper, thereby achieving the vibration reduction effect expected of the active damper without being affected by the time constant. In addition, the prediction section creates predicted waveform data from the measured waveform data based on the learning results obtained in advance by learning sine waves of multiple frequencies through machine learning, thereby creating highly accurate predicted waveform data. Furthermore, by using sine waves of multiple frequencies during machine learning, it becomes possible to create highly versatile predicted waveform data.

Here, "learning results" refers to weights or parameters obtained, for example, when machine learning various earthquake waveforms and building vibration waveforms (vibration waveforms of buildings determined by the height and stiffness of a building) measured in the past. In this example, the weight of the entire waveform of the received measured waveform data up to a specified time (for example, weight A) and the weight of the waveform of the step just before the time to be predicted (for example, weight B) are obtained as learning results, and predicted waveform data is created by taking into account the weights A and B, etc., which are learning results, for the entire waveform and the waveform of the step just before.

The buildings that can be controlled using this method include mid-rise, high-rise, and ultra-high-rise buildings as well as low-rise buildings, and the building vibrations that can be controlled using the variable damping active damper include vibrations caused by earthquake loads and wind loads resulting from strong winds such as typhoons, as well as various types of mechanical vibrations and traffic vibrations.

Another aspect of the control system for an active damper according to the present invention is
The control device further has a machine learning unit, and is characterized in that the learning result is obtained by learning the sine waves of multiple frequencies through machine learning in the machine learning unit.

According to this aspect, by performing machine learning in the machine learning unit of the control device, the creation of predicted waveform data based on the received measured waveform data and the learning results obtained by machine learning can be executed by a common control device, and a control system with as simple a configuration as possible can be constructed.

In another aspect of the control system for the active damper according to the present invention,
The control system for the variable damping active damper has a machine learning device separate from the control device,
The learning result is obtained by learning the sine waves of a plurality of frequencies through machine learning in the machine learning device.

According to this aspect, by performing machine learning in a machine learning device separate from the control device, the control device that creates predicted waveform data and transmits a damping force signal to the damper can be made into a device with as simple a configuration as possible.

In another aspect of the control system for the active damper according to the present invention,
The length of the sine wave is at least one wavelength.

According to this aspect, the sine waves of each frequency used in machine learning have a length of one or more wavelengths, making it easier to create highly accurate predicted waveform data.

In another aspect of the control system for the active damper according to the present invention,
The machine learning unit or the machine learning device refers to natural frequency data regarding the natural frequency of the building during machine learning,
During machine learning, a sine wave having a frequency close to the natural frequency is selected from the plurality of sine waves and machine learning is performed.

According to this aspect, by selecting a sine wave with a frequency close to the natural frequency of the building to be controlled and performing machine learning, predicted waveform data can be created efficiently and effectively, and vibration suppression of the building to be controlled can be performed. Here, the natural frequency of the building is specified in advance based on the height and rigidity of the building, an analytical model of the building, etc., and, for example, the natural frequency data is input and stored in the machine learning device, and a sine wave with a frequency that matches or is close to the stored natural frequency data can be selected.

In another aspect of the control system for the active damper according to the present invention,
A method for controlling a variable damping active damper mounted in a building, comprising the steps of:
A receiving step of receiving measurement waveform data relating to an earthquake waveform or building vibration that changes from moment to moment, the measurement data being measured by a vibration meter;
a machine learning process for performing machine learning to obtain a learning result by superimposing sine waves of multiple frequencies;
a prediction step of predicting a waveform from a predetermined time onward based on data received up to a predetermined time of the measured waveform data and the learning result obtained in the machine learning step, thereby generating predicted waveform data;
a control signal generating step of generating a damping force signal by specifying a damping force at each time using the predicted waveform data, and specifying a time constant associated with switching of the damping force;
and a damper driving step of transmitting the damping force signal to the variable damping type active damper by advancing the time equivalent to the time constant, and driving the variable damping type active damper while switching the damping force of the variable damping type active damper.

According to this aspect, in the prediction process, predicted waveform data is created based on the data received up to a specified time of the measured waveform data and the learning results obtained in the machine learning process, and in the control signal generation process, the predicted waveform data is used to identify the damping force at each time and the time constant associated with switching the damping force. In the damper drive process, the damping force signal is sent to the variable damping type active damper ahead of time equivalent to the time constant, thereby achieving the vibration reduction effect expected of the active damper without being affected by the time constant. Note that since the predicted waveform data changes from moment to moment, when the prediction process, control signal generation process, and damper drive process are treated as a series of sequential process units, these process units are repeatedly executed until, for example, the building vibration is reduced to a certain level.

Moreover, one aspect of the program according to the present invention is
A program for causing a computer that controls a variable damping type active damper installed in a building to execute the following processes:
a machine learning process for performing machine learning to obtain a learning result by superimposing sine waves of multiple frequencies;
a prediction step of predicting a waveform after a predetermined time based on received data up to a predetermined time of the earthquake waveform or measured waveform data related to building vibration, which are measured and received by a vibration meter and change from moment to moment, and the learning result, to generate predicted waveform data;
a control signal generating step of generating a damping force signal by specifying a damping force at each time using the predicted waveform data, and specifying a time constant associated with switching of the damping force;
and a damper driving step of transmitting the damping force signal to the variable damping type active damper by advancing the time equivalent to the time constant and driving the variable damping type active damper while switching the damping force of the variable damping type active damper.

According to this aspect, by having a computer execute the steps of: obtaining learning results through machine learning; creating predicted waveform data based on the learning results and data received up to a specified time of the measured waveform data; generating a damping force signal using the predicted waveform data and identifying the time constant associated with switching the damping force; and transmitting the damping force signal to a variable damping type active damper ahead of the time equivalent to the time constant, and driving the damping force while switching the damping force, via a program, it is possible to obtain the vibration reduction effect expected of an active damper without being affected by the time constant.

As can be understood from the above explanation, the control system, control method, and program for a variable damping active damper of the present invention make it possible to obtain the expected vibration reduction effect in a variable damping active damper without being affected by the time constant.

1 is a schematic diagram showing a building equipped with an example of a control system for a variable damping type active damper according to an embodiment. FIG. 2 is a hydraulic circuit diagram of a variable damping type active damper. FIG. 2 illustrates an example of a hardware configuration of a computer. FIG. 2A is a diagram showing an example of the functional configuration of a control device that constitutes the control system, and FIG. 2B is a diagram showing an example of the functional configuration of a machine learning device that constitutes the control system. FIG. 1 is a diagram illustrating an example of a machine learning method in a machine learning device. FIG. 11 is a diagram illustrating an example of a method for creating predicted waveform data. FIG. 4 is a diagram showing an example of the relationship between the piston speed and the time constant of an oil damper. This figure shows both past earthquake waveforms (Elcentro) and predicted waveforms created using machine learning. This figure shows both the previous earthquake waveform (Hachinohe) and the predicted waveform created by machine learning. This figure shows both the previous earthquake waveform (Taft) and the predicted waveform created by machine learning. This figure shows both past earthquake waveforms (JMA Kobe) and predicted waveforms created using machine learning.

Below, an example of a control system, control method, and program for a variable damping type active damper according to an embodiment will be described with reference to the attached drawings. Note that in this specification and the drawings, substantially identical components may be designated by the same reference numerals to avoid redundant description.

[Control system, control method, and program for variable damping type active damper according to the embodiment]
An example of a control system, a control method, and a program for a variable damping active damper according to an embodiment will be described with reference to Fig. 1 to Fig. 7. Here, Fig. 1 is a schematic diagram showing a building equipped with an example of a control system for a variable damping active damper according to an embodiment, and Fig. 2 is a hydraulic circuit diagram of the variable damping active damper.

The building 10 in which the variable damping active damper 20 is mounted is a 30-story high-rise building on which the variable damping active damper 20 is mounted on the top floor. The buildings to be vibration-controlled are not limited to the number of floors and proportions (slenderness ratio) shown in the illustrated example, and include a variety of buildings, from low-rise buildings to super-high-rise buildings, such as houses and condominiums, office buildings, logistics warehouses, and gymnasiums. In addition, the building may be provided with a seismic isolation device such as a laminated rubber device or a sliding seismic isolation device on the foundation or intermediate floors, or the building may have a basement or may have an underground seismic isolation structure in which a seismic isolation device is mounted in the basement. In addition, the building vibration controlled by the variable damping active damper in the illustrated example is described as vibration due to earthquake load, but the building vibration may also be vibration due to wind load caused by strong winds such as typhoons, vibration due to various mechanical vibrations, vibration due to traffic vibrations, etc.

The control system 60 for the variable damping active damper has multiple vibration meters 50 (two in the illustrated example), a control device 30, and a machine learning device 40. Note that the control device may also be equipped with a machine learning unit, which may eliminate the need for a machine learning device.

On the top floor of the building 10, a variable damping active damper 20 consisting of an oil damper whose main components are a cylinder 21 and a piston 22, and one of the vibration meters 50A are placed. On the first floor of the building 10, the control device 30, the machine learning device 40, and the other of the vibration meters 50B are placed. Here, the vibration meter 50 may be placed only on the top floor, or may be placed on each floor, or may be placed on any one or more floors.

For example, when earthquake motion E acts on building 10 through ground G, building 10 vibrates in the X1 direction at its natural frequency. An active damper suppresses or reduces the vibration of a building by actively supplying energy to the building through active control, while a variable damping active damper is a damper that can generate multiple damping forces and suppresses building vibration with a damping force suitable for the building vibration that changes from moment to moment.

On the top floor of the building 10, the cylinder 21 and piston 22 that make up the active damper 20 are fixed to a pair of fixing fixtures 28, and when the building 10 vibrates in the X1 direction, the piston 22 slides in the X2 direction relative to the cylinder 21, allowing energy to be freely supplied to the building 10.

The building vibration on the top floor is measured by the vibration meter 50A, and the measurement waveform data for the top floor is transmitted to the control device 30 in the Z1 direction.

Meanwhile, the building vibration on the first floor is measured by a vibration meter 50B, and the measured waveform data is similarly transmitted in the Z2 direction to the control device 30. Here, the building vibration on the first floor may be identified as the earthquake waveform input to the building, but a vibration meter (not shown) may be installed as a seismometer on the ground G or the like outside the building 10, and the control device 30 may receive the measured earthquake waveform data from this seismometer.

The machine learning device 40 performs intensive machine learning to obtain learning results in advance, and transmits the obtained learning results to the control device 30 in the Z3 direction. Here, the vibration meters 50A, 50B and the machine learning device 40 transmit and receive measurement waveform data, learning result data, and the like, to the control device 30 via wireless communication or wired communication, but in the illustrated example, the system is described as transmitting and receiving data via wireless communication.

The control device 30 receives the measurement waveform data that changes from moment to moment measured by the vibration meters 50A and 50B, and creates predicted waveform data by predicting the waveform from a predetermined time onward based on the data already received up to a predetermined time of the received measurement waveform data and the learning results obtained from the machine learning device 40. The control device 30 then uses the created predicted waveform data to identify the damping force at each time of the predicted waveform data to generate a damping force signal, and transmits the damping force signal to the variable damping active damper 20 in the Z4 direction to perform drive control of the variable damping active damper 20.

Here, when the damping force at each time is identified and a damping force signal is generated, the time constant associated with switching the damping force from one damping force to another is simultaneously identified, and the damping force signal is transmitted to the variable damping type active damper 20 ahead of the time equivalent to the time constant.

As shown in FIG. 2, the variable damping active damper 20 is a variable damping oil damper. The oil damper 20 in the illustrated example has a hydraulic circuit 23, which is a closed circuit that leads to a cylinder 21, two oil passages consisting of two solenoid valves 24 and pressure regulating valves 26 in parallel with each other, and further has a relief valve 25.

The state shown in Figure 2 shows that each solenoid valve 24 is closed, oil flows through the main circuit in the Y1 direction, and this is the circuit state when the damping force (damping coefficient) is large. In contrast, by opening each solenoid valve 24, oil flows through the solenoid valve 24 in the Y2 direction and through the corresponding pressure regulating valve 26 in the Y3 direction, thereby making it possible to reduce the damping force (damping coefficient). By setting the number of solenoid valves (number of oil passage systems) to the desired number, it becomes possible to switch between two or several stages of damping force, or even a dozen or so stages.

Next, referring to Figures 3 to 7, the hardware configuration and functional configuration of the control device 30 and the machine learning device 40 included in the computer will be described, along with an example of a machine learning method in the machine learning device 40 and an example of a method of creating predicted waveform data in the control device 30. Here, Figure 3 is a diagram showing an example of the hardware configuration of a computer, Figure 4 (a) is a diagram showing an example of the functional configuration of a control device constituting a control system, and Figure 4 (b) is a diagram showing an example of the functional configuration of a machine learning device constituting a control system. Also, Figure 5 is a diagram explaining an example of a machine learning method in the machine learning device, and Figure 6 is a diagram explaining an example of a method of creating predicted waveform data. Furthermore, Figure 7 is a diagram showing an example of the relationship between the piston speed and the time constant of an oil damper.

The control device 30 and the machine learning device 40 are composed of information processing devices such as personal computers and workstations (WS: Work Stations), and both are composed of the computers shown in Figure 2.

A computer such as the control device 30 includes a CPU (Central Processing Unit) 31, a main memory device 32, an auxiliary memory device 33, a communication IF (interface) 34, and an input/output IF 35, which are interconnected by a connection bus 36. The main memory device 32 and the auxiliary memory device 33 are computer-readable recording media. Note that each of the above components may be provided separately, or some of the components may not be provided.

The CPU 31 is a central processing unit that performs overall control of the control device 30, which is a computer. The CPU 31, for example, deploys a program stored in the auxiliary storage device 33 in an executable manner in the working area of the main storage device 32, and provides functions that meet a specific purpose by controlling peripheral devices through the execution of the program.

The main storage device 32 stores the computer programs executed by the CPU 31 and the data processed by the CPU 31. The main storage device 32 includes, for example, a flash memory, a RAM (Random Access Memory), and a ROM (Read Only Memory). The auxiliary storage device 33 stores various programs and various data in a readable and writable recording medium, and is also called an external storage device. The auxiliary storage device 33 stores, for example, an OS (Operating System), various programs, various tables, and the like. The OS includes, for example, a communication interface program that transfers data to and from an external device connected via the communication IF 34. The external device includes, for example, a personal computer (PC), a workstation (WS), a gateway, and the like that connect to a network. Here, the network includes a public network such as the Internet, a wireless network such as a mobile phone network, a dedicated network such as a VPN (Virtual Private Network), and a LAN (Local Area Network).

The auxiliary storage device 33 is used, for example, as a storage area that supports the main storage device 32, and stores the computer programs executed by the CPU 31 and the data processed by the CPU 31. The auxiliary storage device 33 is a silicon disk including a non-volatile semiconductor memory (flash memory, EPROM (Erasable Programmable ROM)), a hard disk drive (HDD: Hard Disk Drive) device, a solid state drive device, etc. Examples of the auxiliary storage device 33 include drives for removable recording media such as CD drives, DVD drives, and BD drives, and examples of the removable recording media include CDs, DVDs, BDs, USB (Universal Serial Bus) memories, and SD (Secure Digital) memory cards.

The communication IF 34 is an interface with the network to which the control device 30 and the like are connected. The communication IF 34 receives measurement waveform data from the vibration meter 50 mounted on the building 10 via the network, and also receives learning result data from the machine learning device 40, while transmitting a control signal generated by the control device 30 to the variable damping type active damper 20.

The input/output IF 35 is an interface that inputs and outputs data between devices connected to the control device 30, etc. Input devices such as a keyboard, a pointing device such as a touch panel or a mouse, and a microphone are connected to the input/output IF 35. The control device 30, etc. accepts operation instructions, etc. from an operator who operates the input device via the input/output IF 35.

The input/output IF 35 is also connected to display devices such as liquid crystal displays (LCDs) and organic electroluminescence (EL) panels, as well as output devices such as printers and speakers. For example, in the machine learning device 40, the natural frequency data of the building 10 is input via the input/output IF 35, and the machine learning device 40 selects one or more sine waves with a frequency close to the natural frequency from among sine waves of multiple frequencies, and performs machine learning. In the machine learning device 40 and the control device 30, the learning result data from the machine learning and the created predicted waveform data are displayed on a display device constituting the input/output IF 35.

As shown in FIG. 4(a), the control device 30 provides various functions of at least a first communication unit 302, a prediction unit 304, a control signal generation unit 306, and a first storage unit 308 (an example of a storage unit) by executing a program by a CPU 31. At least some of the above processing functions may be provided by a DSP (Digital Signal Processor), a GPU (Graphics Processing Unit), etc. Similarly, at least some of the above processing functions may be provided by a dedicated LSI (large scale integration) such as an FPGA (Field-Programmable Gate Array), a numerical calculation processor, an image processing processor, or other digital circuits, etc.

The first communication unit 302 receives measurement waveform data caused by building vibrations, transmitted from vibration meters 50A and 50B installed on the top floor and first floor of the building 10, via the network, and stores the data in the first storage unit 308 as needed. Furthermore, the first communication unit 302 receives learning result data that has been machine-learned by the machine learning device 40, and similarly stores the data in the first storage unit 308.

The prediction unit 304 reads out the measured waveform data stored in the first storage unit 308, and creates predicted waveform data by predicting the waveform from a specified time onwards based on the measured waveform data that has been received up to the specified time. More specifically, the predicted waveform data is created based on the received data and the learning results in the machine learning device 40. Below, the machine learning method in the machine learning device 40 will be explained, and then the method of creating predicted waveform data in the prediction unit 304 will be explained in detail. The created predicted waveform data is stored in the first storage unit 308.

The control signal generating unit 306 reads out the predicted waveform data from the first storage unit 308, identifies the damping force at each time in the predicted waveform data to generate a damping force signal, and further identifies the time constant associated with the switching of the damping force.

The sliding speed (piston speed) of the piston 22 that constitutes the oil damper differs depending on the damping force of the oil damper, and the time constant, which is the time delay according to the sliding speed of the piston (i.e., the time constant according to the damping force), differs.

In the example shown in Figure 7, the time constant is 0.16 s when the piston speed is less than 20 cm/s (17 cm/s or less), decreases linearly to 0.03 s when the piston speed is in the range of 17 cm/s to 40 cm/s, and becomes constant at 0.03 s when the piston speed is 40 cm/s or more.

For example, the relationship graph data of FIG. 7 showing the relationship between the damping force of the oil damper 20 and the time constant is stored in the first storage unit 308. Then, when the control signal generating unit 306 identifies the damping force at each time in the predicted waveform data to generate a damping force signal, it identifies the time constant corresponding to this identified damping force from the relationship graph.

By transmitting the damping force signal to the variable damping active damper 20 from the first communication unit 302 ahead of the time corresponding to the time constant identified by the control signal generation unit 306, it becomes possible to obtain the expected vibration reduction effect in the variable damping active damper 20 without being affected by the time constant.

On the other hand, as shown in FIG. 4(b), the machine learning device 40 provides at least various functions of the second communication unit 402, the machine learning unit 404, and the second storage unit 406 by executing a program by the CPU.

The second communication unit 402 transmits the learning result data that has been machine-learned by the machine learning unit 404 and stored in the second storage unit 406 to the first communication unit 302 of the control device 30.

The machine learning unit 404 performs intensive machine learning using a specified machine learning algorithm to obtain learning results.

The machine learning method used here is the Long Short-Term Memory (LSTM) model, which is one of the algorithms suitable for predicting time-series data, since earthquake waveforms are time-series data.

In machine learning, the length of one time series data set input to the LSTM model can be set to, for example, 100, the batch size during machine learning can be set to, for example, 32, the number of hidden layers can be set to 500, optimization can be performed using the Adam optimizer, and the MSE loss function can be set as the loss function for the LSTM model.

In addition, machine learning uses multiple sine waves with different frequencies, each of which has a length of at least one wavelength. For example, 39 types of sine waves with frequencies ranging from 0.1 Hz to 30 Hz can be used. In machine learning, the order of the sine waves to be learned can be random.

As shown in FIG. 5, in machine learning, the measured waveform is input into the LSTM model to predict the value of the next step. Here, various types of measured waveforms can be used, including representative earthquake waveforms measured in the past, such as the Elcentro earthquake waveform and the Hachinohhe earthquake waveform, as well as earthquake waveforms measured in the past near the building 10 to be controlled.

By conducting intensive machine learning, a "learning result" is obtained. Here, the learning result means, for example, weights and parameters, and in predicting earthquake waveforms, the degree of influence of the entire earthquake waveform of the learning subject and the degree of influence of the step just before the time to be predicted are used as weights, and these weights are calculated as the learning result.

By using sine waves of multiple frequencies during machine learning, it becomes possible to create highly versatile predicted waveform data. Furthermore, by selecting a sine wave of a frequency that matches the natural frequency of the building 10 or a frequency close to that natural frequency from among the sine waves of multiple frequencies stored in the second storage unit 406 and performing machine learning, predicted waveform data can be created efficiently and effectively.

As shown in FIG. 5, predicted waveform data for a certain period of time can be created by sequentially predicting the value of the next step. In making this prediction, the prediction unit 304 may refer to the learning result data transmitted from the machine learning device 40 to the control device 30 and stored in the first storage unit 308, or the prediction unit 304 may directly refer to the learning result data stored in the second storage unit 406 of the machine learning device 40.

As shown in FIG. 6, the prediction unit 304 of the control device 30 predicts the predicted value of the black circle (●) shown in FIG. 6 based on the received data (measured waveform data shown by the solid line up to 1.7 seconds) stored in the first storage unit 308 and the learning results by the machine learning unit 404 of the machine learning device 40.

The prediction unit 304 then predicts the next predicted value based on the measured waveform data shown by the solid line and the predicted value (black circle (●)), and repeats this process to sequentially predict the predicted value for the next step, thereby creating predicted waveform data as shown by the dotted line in Figure 6.

The predicted waveform data is created continuously, for example, until the vibration of the building 10 is reduced to a certain level, and the predicted waveform data stored in the first storage unit 308 is updated as needed.

The control signal generating unit 306 generates a damping force signal by identifying the predicted waveform data stored in the first storage unit 308 and the damping force at each time of each predicted waveform data, and further identifies the time constant associated with switching of the identified damping force based on the relationship graph data illustrated in FIG. 7.

Although not shown in the figures, the control method for the variable damping active damper 20 according to the embodiment includes a receiving step of receiving measurement waveform data related to earthquake waveforms or building vibrations that change from moment to moment, measured by a vibration meter 50, and a machine learning step of performing machine learning to obtain a learning result by superimposing sine waves of multiple frequencies. It also includes a prediction step of performing a prediction of the waveform after a predetermined time based on the received measurement waveform data up to a predetermined time and the learning result obtained in the machine learning step to create predicted waveform data. It also includes a control signal generation step of using the predicted waveform data to identify the damping force at each time to generate a damping force signal and to identify a time constant associated with the switching of the damping force. It also includes a damper drive step of transmitting a damping force signal to the variable damping active damper 20 ahead of the time equivalent to the time constant, and driving the variable damping active damper 20 while switching the damping force.

The above series of controls can also be executed by a program installed in a control device, such as a computer. This program causes the control device to execute a machine learning process of performing machine learning to obtain learning results by superimposing sine waves of multiple frequencies, a prediction process of predicting waveforms from a predetermined time onward based on received data up to a predetermined time of earthquake waveforms or measurement waveform data related to building vibrations that are measured and received by the vibration meter 50 and change from moment to moment, to create predicted waveform data, a control signal generation process of using the predicted waveform data to identify the damping force at each time and generate a damping force signal, and to identify a time constant associated with switching the damping force, and a damper drive process of transmitting a damping force signal to the variable damping active damper 20 ahead of the time equivalent to the time constant, to drive the variable damping active damper 20 while switching the damping force.

[Verification analysis of reproducibility of predicted waveforms using machine learning]
In order to verify the accuracy of machine learning using the LSTM model, the inventors conducted an analysis to create a predicted waveform by machine learning using four representative earthquake waveforms from the past. Figures 8A to 8D show the predicted waveforms (solid lines) and actual earthquake waveforms (dotted lines) for the four earthquake waveforms, and the earthquake waveforms are, in the order of Figures 8A to 8D, the Elcentro earthquake waveform, the Hachinoh earthquake waveform, the Taft earthquake waveform, and the JMA Kobe earthquake waveform.

As can be seen from each figure, the reproducibility of the machine learning predicted waveforms is good for all four types of earthquake waveforms, and it has been verified that the reproducibility of the Hachinohhe earthquake waveform, the Taft earthquake waveform, and the JMA Kobe earthquake waveform is particularly high.

Note that the configurations described in the above embodiments may be combined with other components, and the present invention is not limited to the configurations shown here. In this regard, changes can be made without departing from the spirit of the present invention, and can be determined appropriately according to the application form.

10: Building 20: Variable damping type active damper (active damper, oil damper)
21: Cylinder 22: Piston 23: Hydraulic circuit 24: Solenoid valve 25: Relief valve 26: Pressure regulating valve 30: Control device 40: Machine learning device 50, 50A, 50B: Vibration meter 60: Control system (control system for variable damping type active damper)
302: First communication unit (communication unit)
304: Prediction unit 306: Control signal generation unit 308: First storage unit 402: Second communication unit 404: Machine learning unit 406: Second storage unit G: Ground E: Earthquake motion

Claims (7)

A control system for a variable damping active damper that controls a variable damping active damper installed in a building, comprising:
A vibration meter that measures earthquake waveforms or building vibrations;
a control device that receives measurement waveform data that changes from moment to moment and is measured by the vibration meter, creates predicted waveform data based on data that has been received up to a predetermined time of the measurement waveform data, and drives and controls the variable damping type active damper based on the predicted waveform data,
The control device includes:
a communication unit that receives the measurement waveform data from the vibration meter and transmits a damping force signal related to a damping force to the variable damping type active damper;
a prediction unit that predicts a waveform from a predetermined time onward based on data received up to a predetermined time of the measured waveform data, to generate the predicted waveform data;
a control signal generating unit that uses the predicted waveform data to specify the damping force at each time to generate the damping force signal and to specify a time constant associated with switching of the damping force,
the prediction unit creates the predicted waveform data from the measured waveform data based on a learning result obtained in advance by learning sine waves of a plurality of frequencies through machine learning;
A control system for a variable damping type active damper, characterized in that the damping force signal is transmitted to the variable damping type active damper from the communication unit with a time equivalent to the time constant advanced. The control system for a variable damping active damper according to claim 1, characterized in that the control device further has a machine learning unit, and the learning result is obtained by learning the sine waves of multiple frequencies through machine learning in the machine learning unit. The control system for the variable damping active damper has a machine learning device separate from the control device,
2. The control system for a variable damping type active damper according to claim 1, wherein the learning result is obtained by learning the sine waves of a plurality of frequencies through machine learning in the machine learning device. The control system for a variable damping type active damper according to any one of claims 1 to 3, characterized in that the length of the sine wave is one wavelength or longer. The machine learning unit or the machine learning device refers to natural frequency data regarding the natural frequency of the building during machine learning,
A control system for a variable damping active damper as described in any one of claims 2, 3, and claim 4 dependent on claim 2 or 3, characterized in that, during machine learning, a sine wave having a frequency close to the natural frequency is selected from the multiple sine waves and machine learning is performed. A method for controlling a variable damping active damper mounted in a building, comprising the steps of:
A receiving step of receiving measurement waveform data relating to an earthquake waveform or building vibration that changes from moment to moment, the measurement data being measured by a vibration meter;
A machine learning process for performing machine learning to obtain a learning result by superimposing sine waves of multiple frequencies;
a prediction step of predicting a waveform from a predetermined time onward based on data received up to a predetermined time of the measured waveform data and the learning result obtained in the machine learning step, thereby generating predicted waveform data;
a control signal generating step of generating a damping force signal by specifying a damping force at each time using the predicted waveform data, and specifying a time constant associated with switching of the damping force;
and a damper driving step of transmitting the damping force signal to the variable damping type active damper by advancing a time equivalent to the time constant and driving the variable damping type active damper while switching the damping force of the variable damping type active damper. A program for causing a computer that controls a variable damping type active damper installed in a building to execute the following processes:
a machine learning process for performing machine learning to obtain a learning result by superimposing sine waves of multiple frequencies;
a prediction step of predicting a waveform after a predetermined time based on received data up to a predetermined time of the earthquake waveform or measured waveform data related to building vibration, which are measured and received by a vibration meter and change from moment to moment, and the learning result, to generate predicted waveform data;
a control signal generating step of generating a damping force signal by specifying a damping force at each time using the predicted waveform data, and specifying a time constant associated with switching of the damping force;
and a damper driving step of transmitting the damping force signal to the variable damping type active damper by advancing a time equivalent to the time constant and driving the variable damping type active damper while switching the damping force of the variable damping type active damper.

Images