JP7530231B2 - Prediction device, learning device and program - Google Patents

Description

Article 30, paragraph 2 of the Patent Act applies Publication name: 2019 Annual Conference Manzu, Abe! The Society of Mechanical Engineers! -Shittage Omoshie Technology Gathered in Akita- Sustainability, Mechanical Engineering in an AI Society Technology to Support the Low Birthrate, Aging Population, and Labor Shortage, Published: September 2, 2019, Publisher: The Japan Society of Mechanical Engineers (5F Shinanomachi Rengakan, 35 Shinanomachi, Shinjuku-ku, Tokyo), Disclosed by: Takeshi Fukazawa (Mitsubishi FBR Systems Co., Ltd.), Disclosed invention: Invention related to modeling of damping elements using artificial intelligence [Publications, etc.] Meeting name: 2019 Annual Conference, Date held: September 8-11, 2019, Organizer: The Japan Society of Mechanical Engineers (5F Shinanomachi Rengakan, 35 Shinanomachi, Shinjuku-ku, Tokyo), Venue: Akita University Tegata Campus (1-1 Tegata Gakuen-cho, Akita City, Akita Prefecture), Disclosed by: Takeshi Fukazawa (Mitsubishi FBR Systems Co., Ltd.), Publication date: September 9, 2019. Details of the disclosed invention: An invention related to modeling of damping elements using artificial intelligence.

The present invention relates to a prediction device, learning device, and program that utilize artificial intelligence, and in particular to a technology for predicting or learning output time series data from any two of input time series data of load, velocity, and displacement in a mechanical model.

Time history response analysis is used to verify the earthquake resistance of architectural structures and the like (see, for example, Non-Patent Document 1). When performing time history response analysis, it is necessary to understand the response of the structure from its elastic region to its ultimate region.

Fukazawa et al., “Study on analytical model of oil damper focusing on high vibration and low amplitude”, Proceedings of the Architectural Institute of Japan, Structural Design Journal, Architectural Institute of Japan, December 2018, Vol. 83, No. 754, pp. 1777-1787

However, since the behavior in the ultimate region is generally highly nonlinear, it is difficult to model the behavior in the ultimate region mathematically. Even if it were possible to model it, the constitutive equation would be complicated.

The present invention was made in consideration of the above circumstances, and aims to calculate the time history response even for mechanical models that cannot be modeled mathematically.

In order to solve the above problems, a prediction device that predicts output time series data from any two of input time series data of time series data of load, velocity, and displacement in a mechanical model that models a civil engineering and architectural structure includes an RNN construction unit that constructs a recurrent neural network having an input layer, a hidden layer, and an output layer in that order from input to output, and a prediction unit that inputs the two input time series data to the input layer, assigns a learned parameter group consisting of a plurality of weights obtained by prior learning to the hidden layer, and activates the hidden layer to calculate the output time series data in the output layer , and the RNN construction unit constructs a layer in which a plurality of long and short-term memory blocks are connected in time series as the hidden layer .

Further, a learning device for learning any two input time series data of time series data of load, velocity, and displacement in a mechanical model that models a civil engineering and architectural structure and another ground truth time series data includes an RNN construction unit that constructs a recurrent neural network having an input layer, a hidden layer, and an output layer in that order from input to output, and a learning unit that inputs the two input time series data to the input layer, assigns a learned parameter group consisting of a plurality of weights to the hidden layer and activates the hidden layer to calculate output time series data in the output layer, calculates losses of the output time series data and the ground truth time series data, calculates differential parameters by differentiating the losses with the plurality of weights by backpropagation processing, and updates the plurality of weights based on the differential parameters and the plurality of weights by a gradient descent method , and the RNN construction unit constructs a layer in which a plurality of long and short-term memory blocks are connected in time series as the hidden layer .

According to the present invention, even if the mechanical model cannot be modeled by a formula, it is possible to calculate another output time series data from any two of the input time series data of the load, velocity, and displacement. The output time series data is the response when any two of the load, velocity, and displacement that change over time are given to the mechanical model.

FIG. 2 is a schematic diagram of a modeled damping device. FIG. 1 is a block diagram of an artificial intelligence device that functions as both a learning device and a prediction device. FIG. 1 is a schematic diagram of a recurrent neural network. FIG. 1 is a schematic diagram of long short-term memory blocks.

[1. Configuration of the artificial intelligence device]
1 is a schematic diagram of a dynamically modelled damping system 90. FIG. 2 is a block diagram of an artificial intelligence system 1.

The damping device 90 is, for example, an oil damper. When the damping device 90 is mechanically modeled, it is composed of a dashpot 91, an elastic spring 92, an elastic spring 93, and a friction resistance 94, and the dashpot 91, the elastic spring 92, and the elastic spring 93 are connected in series, and the series is connected in parallel with the friction resistance 94. The dashpot 91 and the elastic spring 92 are viscoelastic elements of the damping device 90, the elastic spring 93 is an elastic element of the damping device 90, and the friction resistance 94 is generated in a moving part (e.g., a rod) of the damping device 90. When an earthquake, vibration, collision, or the like occurs, a load and a speed that change over time are applied to the damping device 90, and the damping device 90 responds and displaces accordingly. The artificial intelligence device 1 shown in FIG. 2 predicts the response of the damping device 90, that is, the change in the displacement of the damping device 90 over time. In other words, the artificial intelligence device 1 uses artificial intelligence to calculate the time series data of the displacement from the time series data of the load and the time series data of the speed.

The artificial intelligence device 1 functions as both a prediction device and a learning device, and includes a processor 10, a display device 30, an input device 40, and a storage 50.
The arithmetic processing device 10 includes a CPU, a GPU, a RAM, a system bus, and various interfaces.

The display device 30 is, for example, a liquid crystal display device, an organic EL display device, or a projector. The arithmetic processing device 10 generates a video signal and outputs the video signal to the display device 12. Then, a screen according to the video signal is displayed by the display device 30.

The input device 40 is, for example, a switch, a keyboard, or a pointing device. When the input device 40 is operated, it outputs a signal corresponding to the operation to the arithmetic processing device 10.

The storage 50 stores a program 51, training data 52, a learned parameter set 56, and data to be analyzed 57.

The teacher data 52 is used for learning the recurrent neural network 200 (see FIG. 3) described below. The teacher data 52 is composed of load time series data 53, velocity time series data 54, and displacement time series data 55, which are associated with each other. These time series data 53, 54, and 55 are discrete data strings in which the load, velocity, and displacement values are arranged in time series, respectively. The displacement time series data 55 is used as the correct answer.

These time series data 53, 54, 55 were obtained by simulations or experiments carried out in advance. Assuming a range of events from extraordinary earthquakes or shocks to everyday micro-vibrations, the load, speed, and displacement conditions were changed for each simulation or experiment, and a large number of teacher data 52 obtained by performing multiple simulations or experiments are stored in storage 50.

The trained parameter group 56 is composed of weights and biases for a large number of neurons in the recurrent neural network 200. The trained parameter group 56 is updated each time the recurrent neural network 200 trains. When the recurrent neural network 200 is in an untrained state, the initial values of each weight and bias in the trained parameter group 56 are random numbers.

The analysis data 57 is used to predict the time series data of displacement using the recurrent neural network 200. The analysis data 57 is composed of time series data 58 of load and time series data 59 of velocity, which are associated with each other. These time series data 58 and 59 are discrete data strings in which the load and velocity values are respectively arranged in time series.

The program 51 is readable and executable by the arithmetic processing device 10. By executing the program 51, the arithmetic processing device 10 functions as an RNN model construction unit 11, a learning unit 12, and a prediction unit 13. The RNN model construction unit 11 and the learning unit 12 are functions of a learning device, and the RNN model construction unit 11 and the prediction unit 13 are functions of a prediction device. The functions of the RNN model construction unit 11, the learning unit 12, and the prediction unit 13 are described in detail below.

[2. RNN model construction section]
The RNN model construction unit 11 constructs a recurrent neural network 200 using a stacked long short-term memory (LSTM) architecture, as shown in Fig. 3. The recurrent neural network 200 is also called a recurrent neural network. The recurrent neural network 200 is composed of an input layer 210, four hidden layers 220, 230, 240, and 250, a fully connected layer 260, a dropout layer 270, and an output layer 280.

The input layer 210 is composed of a plurality of time-series input units 211. Two types of time-series data, specifically, displacement time-series data and velocity time-series data, are input to the input layer 210. That is, two neurons are used for each input unit 211 of the input layer 210, and displacement data and velocity data are input to these neurons, respectively. Therefore, the input data _xt input to each input unit 211 is a vector, and the elements of the input data _xt consist of displacement data and velocity data of the same step (i.e., the same time). Here, the subscript "t" represents the order of the steps in the time series.

The first hidden layer 220 is composed of a time-series long short-term memory block 221. The second to fourth hidden layers 230, 240, and 250 are also composed of time-series long short-term memory blocks 231, 241, and 251, respectively. Here, if j is an arbitrary number from 1 to 4, the "j" in the subscript "(t,j)" represents the layer order of the hidden layers 210, 230, 240, and 250.

In each of the hidden layers 210, 230, 240, and 250, the output of a neuron in a certain step is connected to the input of a neuron in the next step and the input of a neuron in the next layer. Specifically, when each long short-term memory block 221 in the first layer inputs input data _xt from the input unit 211 of the same step and inputs hidden state data h _(t-1,1) from the long short-term memory block 221 of the previous step, it outputs the hidden state data h _(t,1) to the long short-term memory block 221 of the next step and outputs data y _(t,1) to the long short-term memory block 231 of the next layer. Furthermore, when the second-layer long short-term memory block 231 inputs data y _(t,1) from the first-layer long short-term memory block 221 and hidden state data h _(t-1,2) from the previous step long short-term memory block 231, it outputs the hidden state data h _(t,2) to the next step long short-term memory block 231 and outputs data y _(t,2) to the next layer long short-term memory block 241. Furthermore, when the third-layer long short-term memory block 241 inputs data y( _t,2) from the second-layer long short-term memory block 231 and hidden state data h _(t-1,3) from the previous step long short-term memory block 241, it outputs hidden state data h _(t,3) to the next step long short-term memory block 241 and outputs data y _(t,3) to the next layer long short-term memory block 251. Furthermore, when the long short-term memory block 251 in the fourth layer inputs data y _(t,3) from the long short-term memory block 241 in the third layer and inputs hidden state data h _(t-1,4) from the long short-term memory block 251 in the previous step, it outputs the hidden state data h _(t,4) to the long short-term memory block 251 in the next step and outputs data y _(t,4) to the fully connected block 261 of the same step in the fully connected layer 260. The numbers of long short-term memory blocks 221, 231, 241, 251 in one step are greatest in the order of long short-term memory blocks 221, 231, 241, 251. For example, the numbers of long short-term memory blocks 221, 231, 241, 251 in one step are 200, 125, 78, 49, respectively.

The fully connected layer 260 is composed of a time-series fully connected block 261. The fully connected block 261 has a plurality of neurons, for example 30 neurons, and the neurons of the fully connected block 261 are fully connected to the output of the long short-term memory block 251 in the fourth layer of the same step. When the fully connected block 261 receives the output data y _(t,4) of the long short-term memory block 251 in the same step, it multiplies them by weights and adds a bias to the product.

The dropout layer 270 is composed of a time-series dropout block 271. The dropout block 271 has multiple neurons, and the neurons of the dropout block 271 are fully connected to the neurons of the fully connected block 261 of the same step. During learning, some neurons are updated to be disabled, and during prediction, the input from the neurons of the fully connected block 261 of the same step to the disabled neurons of this dropout block 271 is set to zero. In this way, the dropout block 271 reduces the deterioration of accuracy due to overlearning.

The output layer 280 is composed of a plurality of time-series output units 281. One neuron is used for each output unit 281 in the input layer 280, and the neuron of the output unit 281 is connected to the neuron of the dropout block 271 of the same step, and when the output unit 281 receives the output of a valid neuron of the dropout block 271, it outputs output data _zt . Therefore, the output data _zt output to each output unit 281 is scalar weight data, and one type of time-series data, specifically, weight time-series data, is output from the output layer 280.

The long short-term memory blocks 221, 231, 241, and 251 will now be described with reference to FIG.
The long short-term memory blocks 221, 231, 241, and 251 are composed of an input block 310, an input gate block 320, a forget gate block 330, an output gate block 340, a multiplication block 350, a multiplication block 360, an addition block 370, a memory cell 375, an activation function block 380, and a multiplication block 390. Here, in the case of the first-layer long short-term memory block 221, the input data v _(t,j-1) in Fig. 4 is the input data _xt in Fig. 3, and the output data v _(t,j) in Fig. 4 is the data y _(t,1) in Fig. 3. In the case of the second-layer long short-term memory block 231, the input data v _(t,j-1) in Fig. 4 is the data y _(t,1) in Fig. 3, and the output data v _(t,j) in Fig. 4 is the data y _(t,2) in Fig. 3. In the case of the third-layer long short-term memory block 241, the input data v _(t,j-1) in Fig. 4 is the data y _(t,2) in Fig. 3, and the output data v _(t,j) in Fig. 4 is the data y _(t,3) in Fig. 3. In the case of the fourth-layer long short-term memory block 251, the input data v _(t,j-1) in Fig. 4 is the data y _(t,3) in Fig. 3, and the output data v _(t,j) in Fig. 4 is the data y _(t,4) in Fig. 3.

When the input data v _(t,j-1) and the hidden state data h _(t-1,j) are input to the input block 310, the input block 310 calculates the state data s _(t,j) of the input block 310 by the following formula (1). Here, the input block 310 is composed of a weighted sum block 311 and an activation function block 312. The weighted sum block 311 calculates the sum of the product of the input data v _(t,j-1) and the weight Us _(t, j), the product of the hidden state data h _(t-1,j) and the weight Ws _{(t, j), and the bias bs(t,j)} as a net value by adding them. The activation function block 312 calculates the state data s _(t,j) of the input block 310 by fitting the net value calculated by the weighted sum block 311 to an activation function (specifically, a hyperbolic tangent function).

where U _{s (t,j)} and W _{s (t,j)} are the weights at the input block 310 and b _{s (t,j)} is the bias at the input block 310 .

When the input data v _(t,j-1) and the hidden state data h _(t-1,j) are input, the input gate block 320 calculates the state data i _(t,j) of the input gate block 320 by the following formula (2). Here, the input gate block 320 is composed of a weighted sum block 321 and an activation function block 322. The weighted sum block 321 calculates the sum as a net value by adding the product of the input data v _{(t,j-1) and} the weight U i(t,j), the product of the hidden state data h _{(t-1,j )} and the weight W _{i(t,j), and the bias b i(} t,j). The activation function block 322 calculates the state data i _(t,j) of the input gate block 320 by fitting the net value calculated by the weighted sum block 321 to an activation function (specifically, a sigmoid function).

where U _{i (t,j)} and W _{i (t,j)} are the weights at the input gate block 320 and b _{i (t,j)} is the bias at the input gate block 320 .

When the forgetting gate block 330 receives the input data v _(t,j-1) and the hidden state data h _(t-1,j) , it calculates the state data f _(t,j) of the forgetting gate block 330 by the following formula (3). Here, the forgetting gate block 330 is composed of a weighted sum block 331 and an activation function block 332. The weighted sum block 331 adds the product of the input data v _(t,j-1) and the weight Uf _(t,j) , the product of the hidden state data h _(t-1,j) and the weight Wf _(t,j) , and a bias bf _(t,j) , and calculates the sum as a net value. The activation function block 332 calculates the state data i _(t,j) of the input gate block 320 by fitting the net value calculated by the weighted sum block 331 to an activation function (specifically, a sigmoid function).

where U _f(t,j) and W _f(t,j) are the weights in the forget gate block 330 and b _f(t,j) is the bias in the forget gate block 330.

When the output gate block 340 receives input data v _(t,j-1) and hidden state data h _(t-1,j) , it calculates state data o _(t,j) of the output gate block 340 by the following formula. Here, the output gate block 340 is composed of a weighted sum block 341 and an activation function block 342. The weighted sum block 341 adds the product of the input data v _(t,j-1) and the weight _Uo(t,j) , the product of the hidden state data h _(t-1,j) and the weight Wo _(t,j) , and a bias b _o(t,j) to calculate the sum as a net value. The activation function block 342 calculates state data o( _{t,j) of the input gate block 320 by fitting the net value calculated by the weighted sum block 341 to an activation function (specifically, a sigmoid function)} .

where U _{o (t, j)} and W _{o (t, j)} are the weights at the output gate block 340 and bo _{(t, j)} is the bias at the output gate block 340 .

The multiplication block 350 receives the state data i _(t,j) calculated by the input block 310 and the state data i _(t,j) calculated by the input gate block 320, and calculates the Hadamard product by Hadamard multiplication of these data as shown in the following equation (5). The Hadamard product is the product of each element of matrices of the same size.

Multiplication block 360 receives state data f _(t,j) calculated by forget gate block 330 and the calculation result of addition block 370 in the previous step (stored data C _(t-1,j) ) held in memory cell 375, and calculates the Hadamard product by Hadamard multiplication of these data as shown in the following equation (6).

Addition block 370 receives the Hadamard product calculated by multiplication block 350 and the Hadamard product calculated by multiplication block 360, adds them together as shown in the following equation (7), and stores the calculation result (storage data C _{(t, j)} ) in memory cell 375.

Equation (7) is an update equation for memory cell 375. That is, memory cell 375 updates the calculation result (stored data C _(t-1,j) ) of addition block 370 in the previous step that it has held to the calculation result (stored data C _(t,j) ) of addition block 370 in the current step, and holds it.

The activation function block 380 calculates activation state data by fitting the calculation result (stored data C _(t,j) ) of the addition block 370 updated and stored in the memory cell 375 to an activation function (specifically, a hyperbolic tangent function).

The multiplication block 330 inputs the state data f _{(t, j)} calculated by the forget gate block 330 and the activation state data calculated by the activation function block 380, and calculates hidden state data h _{(t, j)} by Hadamard multiplication of these as shown in the following equation (8).

3 and 4, the hidden state data h _(t,j) calculated by the multiplication block 330 is output to the long short-term memory block in the next step of the same layer. Data y _{(t,j) equal to the hidden state data h(t,j )} calculated by the multiplication block 330 is output to the long short-term block or the full connection block 261 in the next layer.

The weights Us _(t,j) , Ws _{(t,j), bs(t,j)} , Ui(t,j), Wi _{(t,j), bi(t,j)} , _Uf(t,j) , Wf _(t,j) , bf _(t,j) , Uo _{(t, j), Wo(t,j) ,} and bo _{(t,j) in each long and short- term memory block 221, 231, 241, and} 251 of the recurrent neural network 200 configured as described above are stored in the storage 50 as a learned parameter set 56. In the unlearned state, the values of weights _Us(t,j) , Ws _{(t,j), bs(t,j)} , Ui _{(t,j), Wi(t,j )} , _bi(t,j) , Uf(t,j), _Wf(t,j) , _{bf( t,j)} , _{Uo( t,j)} , Wo _(t,j) , and bo _(t,j) pre-stored in storage 50 as learned parameter set 56 are random numbers.

[3. Learning Section]
The learning unit 12 causes the recurrent neural network 200 constructed by the RNN model construction unit 11 to learn a large amount of teacher data 52, and updates the learned parameter group 56. A specific description will be given below.

First, the learning unit 12 retrieves the current learned parameter set 56 from the storage 50 and assigns weights Us _{(t,j), Ws(t,j)} , bs _{(t,j) ,} Ui _{(t,j), Wi(t,j)} , bi _(t,j) , Uf(t,j ₎ , _Wf(t,j) , _bf(t,j) , Uo _(t,j) , Wo _{(t ,j)} , _bo(t,j) of the learned parameter set 56 to each long and short-term memory block 221, 231, 241, 251 of the recurrent neural network 200.

Next, the learning unit 56 acquires any one of the teacher data 52 from the storage 50, and inputs the load and speed time series data 53, 54 of the teacher data 52 to the input layer 210 of the recurrent neural network 200 to execute forward processing of the recurrent neural network 200. As a result, the learning unit 56 activates the hidden layers 220, 230, 240, 250, the fully connected layer 260, and the dropout layer 270 of the recurrent neural network 200 to calculate output data _zt in each output unit 281 of the output layer 280, and acquires output time series data in which the output data _zt is arranged in time series.

Next, the learning unit 56 fits the displacement data in the displacement time-series data 53 of the teacher data 52 and the output data _zt to a loss function for each step, and calculates the loss between the displacement data in the displacement time-series data 53 of the teacher data 52 and the output data _zt . The loss function is preferably a half mean squared error function, but may also be a mean squared error function, a mean absolute error function, a mean logarithmic error function, or a cross entropy error function.

Next, the learning unit 56 calculates differential parameters (gradients) by differentiating the losses with the current weights Us _(t,j) , Ws _(t,j) , bs _(t, j), Ui _(t,j) , _Wi(t,j) , bi _(t,j) , Uf( _t, j), _Wf(t,j) , bf _(t,j) , Uo _(t,j) , _Wo(t,j) , and _bo(t,j) in each long short-term memory block 221, 231, 241, and 251 through backpropagation processing. Here, a gradient clipping method is used when calculating the differential parameters. In other words, if the differential parameters exceed a predetermined threshold, the differential parameters are normalized with a predetermined threshold (for example, 1.0).

Next, the learning unit 56 calculates new weights Us(t,j) based on the current weights Us _(t,j) and their differential parameters in each of the long short-term memory blocks 221, 231, 241, and 251 by using a gradient descent method, particularly a gradient descent method using Adam's optimization algorithm. Similarly, the learning unit 56 calculates new weights Ws _{(t,j) , bs(t,j)} , Ui _(t,j) , _{Wi(t,j), bi(t,j)} , Uf _(t,j) , Wf(t _,j) , _bf(t,j) , Uo _{(t,j) ,} Wo _(t,j) , _{and bo(t,j) in each of the long short-term memory blocks 221, 231, 241, and 251} . In the case of Adam's optimization algorithm, the value of the coefficient related to the first moment (average value) of the differential parameter is, for example, 0.900, the value of the coefficient related to the second moment (variance) of the differential parameter is, for example, 0.999, and the value of the initial learning rate is 0.005, but these values are not limited to these.

Next, the learning unit 56 updates the learned parameter set 56 stored in the storage 50 to the calculated new weights Us _(t,j) , Ws _{(t,j), bs(t,j) ,} Ui _(t,j) , _Wi(t,j) , bi _(t,j ), Uf _(t,j) , _Wf(t,j) , bf _(t,j) , _Uo(t,j) , Wo _(t,j) , bo _(t,j) .

The above is one learning process, and when the learning unit 56 repeatedly executes the same process using other teacher data 52, the learned parameter set 56 is optimized. Here, because the gradient clipping method is adopted as described above, gradient vanishing and gradient explosion are suppressed even when learning is repeated.

[4. Prediction Section]
The prediction unit 13 predicts the time series data of the displacement from the analysis data 57 using the recurrent neural network 200 constructed by the RNN model construction unit 11. This will be specifically described below.

First, the prediction unit 13 retrieves the learned parameter group 56 from the storage 50 and assigns weights Us _{(t,j), Ws(t,j)} , bs _(t,j) , Ui _{(t ,j)} , Wi(t,j), bi _(t,j) , Uf _(t,j) , Wf _(t, j), _bf(t,j) , Uo _(t,j) , _{Wo(t,j) , bo(t,j)} of the learned parameter group 56 to each long and short-term memory block 221, 231, 241, 251 of the recurrent neural network 200.

Next, the prediction unit 13 acquires the analysis data 57 from the storage 50, and inputs the time series data 58, 59 of the load and speed of the analysis data 57 to the input layer 210 of the recurrent neural network 200. As a result, the prediction unit 56 activates the hidden layers 220, 230, 240, 250, the fully connected layer 260, and the dropout layer 270 of the recurrent neural network 200 to calculate the output data _zt in each output unit 281 of the output layer 280, and acquires time series data of predicted displacement in which the output data _zt is arranged in time series. This time series data is a discrete data string in which the predicted displacement values are arranged in time series.

Next, the prediction unit 13 records the time series data of the calculated predicted displacement in the storage 50 as time series data 60 of the prediction result.

The prediction unit 13 also generates a chart from the time series data of the predicted displacement, and outputs a video signal including the chart on the screen to the display device 30. The chart is then displayed on the display device 30. The chart, for example, represents time on the horizontal axis and predicted displacement on the vertical axis, with the values of the predicted displacement plotted in time series.

The prediction unit 13 also associates each value of the time series data of the predicted displacement with each value of the time series data of the load 58 at the same time, generates a chart showing the correspondence between the time series data of the predicted displacement and the time series data of the load 58, and outputs a video signal including the chart on the screen to the display device 30. The chart is then displayed on the display device 30. The chart, for example, shows the predicted displacement on the horizontal axis and the load on the vertical axis, with the associated load values and predicted displacement values plotted.

The prediction unit 13 also associates each value of the time series data of the predicted displacement with each value of the time series data of the speed 58 at the same time, generates a chart showing the correspondence between the time series data of the predicted displacement and the time series data of the speed 58, and outputs a video signal including the chart on the screen to the display device 30. The chart is then displayed on the display device 30. For example, the chart shows the predicted displacement on the horizontal axis and the speed on the vertical axis, with the associated speed values and predicted displacement values plotted.

5. Advantageous Effects
Even without modeling the damping device 90 with a mathematical formula, it is possible to calculate the time series data of the displacement from the time series data 58, 59 of the load and the velocity by using the trained recurrent neural network 200. In other words, it is possible to calculate the response when the load and the velocity that change with time are applied to the damping device 90 as the time series data of the displacement. In particular, it is possible to calculate the response displacement even for small loads that occur in daily life and large loads that reach an ultimate state.

The use of recurrent neural network 200 enables regression modeling of time series. In other words, in each of hidden layers 210, 230, 240, and 250, the input of a neuron at a certain step is connected to the output of a neuron at the previous step, so that the calculation result of the previous step regresses and affects the calculation of the next step, making recurrent neural network 200 suitable for calculating time series data.

Because the recurrent neural network 200 uses a long short-term memory (LSTM) architecture, it realizes long-term memory in time series and is less susceptible to gradient vanishing. Therefore, the recurrent neural network 200 using the long short-term memory architecture is suitable for response analysis in cases where various loads of different sizes are excited over a long period of time, such as an earthquake.

In addition, because the recurrent neural network 200 has multiple hidden layers 210, 230, 240, and 250, prediction accuracy is improved, but the multiple hidden layers 210, 230, 240, and 250 can cause gradient vanishing and other problems. Such problems are solved by the long short-term memory architecture.

6. Modifications
In the above embodiment, the analysis data 57 is composed of the time series data 58 of the load and the time series data 59 of the speed, and at the time of prediction, the time series data 58 of the load and the time series data 59 of the speed are input to the input layer 210 of the recurrent neural network 200, and the time series data of the displacement is output to the output layer 280. On the other hand, the analysis data 57 may be composed of the time series data 58 of the load and the time series data of the displacement, and the time series data 58 of the load and the time series data of the displacement are input to the input layer 210 of the recurrent neural network 200, and the time series data of the speed is output to the output layer 280. In this case, the correct answer in the teacher data 52 is the time series data 54 of the speed, and at the time of learning, the time series data 53 and 55 of the load and the displacement are input to the input layer 210 of the recurrent neural network 200, and the loss between the speed data in the speed time series data 54 of the teacher data 52 and the output data _zt is calculated.

Alternatively, the analysis data 57 may be composed of time series data 59 of velocity and time series data of displacement, and the time series data 59 of velocity and the time series data of displacement may be input to the input layer 210 of the recurrent neural network 200, and the time series data of load may be output to the output layer 280. In this case, the correct answer in the teacher data 52 is the time series data 53 of load, and during learning, the time series data 54 and 55 of velocity and displacement are input to the input layer 210 of the recurrent neural network 200, and the loss between the load data in the load time series data 53 of the teacher data 52 and the output data _zt is calculated.

In the above embodiment, a damping device 90 is given as an example of a mechanical model, and time series data 53, 54, and 55 obtained by simulation or experiment for the damping device 90 are used as teacher data 52. The mechanical model is not limited to the damping device 90, but may be a civil engineering or architectural structure (e.g., a building, a bridge, a beam, a column, etc.) or a mechanical structure (e.g., an automobile body, an elastic device, a spring element, etc.), and time series data of the speed, load, and displacement obtained by simulation or experiment for the civil engineering or architectural structure or the mechanical structure may be used as teacher data 52.

10: Processing device 11: RNN model construction unit 12: Learning unit 13: Prediction unit 51: Program 52: Teacher data 57: Analysis data 200: Recurrent neural network 210: Input layer 220, 230, 240, 250: Hidden layer 260: Fully connected layer 270: Dropout layer 280: Output layer 221, 231, 241, 251: Long short-term memory block

Claims (6)

A prediction device for predicting output time series data from any two of input time series data of load, velocity, and displacement in a mechanical model of a civil engineering structure, comprising:
An RNN construction unit that constructs a recurrent neural network having an input layer, a hidden layer, and an output layer in this order from the input to the output;
a prediction unit that inputs the two input time series data to the input layer, assigns a learned parameter group consisting of a plurality of weights obtained by prior learning to the hidden layer, and activates the hidden layer to calculate the output time series data in the output layer ,
The RNN construction unit is a prediction device that constructs a layer in which a plurality of long-term short-term memory blocks are connected in a time series as the hidden layer . The prediction device according to claim 1 , wherein the RNN construction unit constructs the hidden layer into a plurality of layers. A learning device for learning two input time series data of time series data of load, velocity, and displacement in a mechanical model of a civil engineering and architectural structure and another correct time series data,
An RNN construction unit that constructs a recurrent neural network having an input layer, a hidden layer, and an output layer in this order from the input to the output;
a learning unit that inputs the two input time series data to the input layer, assigns a learned parameter group consisting of a plurality of weights to the hidden layer and activates the hidden layer to calculate output time series data in the output layer, calculates losses of the output time series data and the ground truth time series data, calculates differential parameters by differentiating the losses with the plurality of weights by a backpropagation process, and updates the plurality of weights based on the differential parameters and the plurality of weights by a gradient descent method ,
The RNN construction unit is a learning device that constructs a layer in which a plurality of long-term short-term memory blocks are connected in a time series as the hidden layer . The learning device according to claim 3 , wherein the RNN construction unit constructs the hidden layer in a plurality of layers.
A calculation processing device of a prediction device for predicting output time series data from any two of input time series data of load, velocity, and displacement in a mechanical model of a civil engineering and architectural structure, comprising :
an RNN constructing unit that constructs a recurrent neural network having an input layer, a hidden layer, and an output layer in this order from input to output , and constructs a layer in which a plurality of long short-term memory blocks are connected in time series as the hidden layer ;
a prediction unit that inputs the two input time series data to the input layer, assigns a learned parameter group consisting of a plurality of weights obtained by prior learning to the hidden layer, and activates the hidden layer to calculate the output time series data in the output layer. A processing unit of a learning device that learns two input time series data of time series data of load, velocity, and displacement in a mechanical model that models a civil engineering and architectural structure and another correct time series data,
an RNN constructing unit that constructs a recurrent neural network having an input layer, a hidden layer, and an output layer in this order from input to output , and constructs a layer in which a plurality of long short-term memory blocks are connected in time series as the hidden layer ;
a learning unit that inputs the two input time series data to the input layer, calculates output time series data in the output layer by assigning a learned parameter group consisting of a plurality of weights to the hidden layer and activating the hidden layer, calculates losses of the output time series data and the ground truth time series data, calculates differential parameters by differentiating the losses with the plurality of weights by backpropagation processing, and updates the plurality of weights based on the differential parameters and the plurality of weights by gradient descent processing.

Images