CN113705084A - Seismic simulation vibration table closed-loop control method based on deep learning - Google Patents

Seismic simulation vibration table closed-loop control method based on deep learning Download PDF

Info

Publication number
CN113705084A
CN113705084A CN202110883227.5A CN202110883227A CN113705084A CN 113705084 A CN113705084 A CN 113705084A CN 202110883227 A CN202110883227 A CN 202110883227A CN 113705084 A CN113705084 A CN 113705084A
Authority
CN
China
Prior art keywords
controller
deep learning
training
closed
parameter
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110883227.5A
Other languages
Chinese (zh)
Inventor
纪金豹
胡宗祥
杨森
张伟祺
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing University of Technology
Original Assignee
Beijing University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing University of Technology filed Critical Beijing University of Technology
Priority to CN202110883227.5A priority Critical patent/CN113705084A/en
Publication of CN113705084A publication Critical patent/CN113705084A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V13/00—Manufacturing, calibrating, cleaning, or repairing instruments or devices covered by groups G01V1/00 – G01V11/00
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The invention provides a deep learning-based seismic simulation vibrating table closed-loop control method which comprises a deep learning-based vibrating table closed-loop controller and a deep learning-based vibrating table closed-loop controller training method. The method comprises the following steps: constructing a deep learning controller based on a cyclic neural network and a vibration table closed-loop control simulation model based on a three-parameter algorithm; acquiring input and output data of a three-parameter controller to train a deep learning controller; the deep learning controller after supervision training receives seismic wave acceleration signals and vibration table feedback acceleration signals, and outputs control signals to enter a vibration table open-loop model to form closed-loop control. The method can be used for carrying out closed-loop control on the earthquake simulation vibration table and improving the reproduction precision of the vibration table on the earthquake time course.

Description

Seismic simulation vibration table closed-loop control method based on deep learning
Technical Field
The invention relates to the field of test technology and control, in particular to a vibrating table closed-loop controller based on deep learning and a vibrating table closed-loop controller training method based on deep learning.
Background
The earthquake simulation shaking table test is the most convenient and most direct test method in the field of engineering earthquake resistance at present, but due to the nonlinear influence of a servo valve and the like and the combined action of a table body and a test piece, the shaking table is difficult to obtain high-precision earthquake waveform reproduction capability, and although the three-parameter control algorithm and feedforward compensation improve the reproduction precision of an earthquake time course to a certain extent, the complexity of system transfer function influence factors, particularly the influence of test piece characteristics, causes that the correlation of input and output waveforms is still not high.
The deep learning is a deep neural network in nature, has more excellent nonlinear processing capability, and is expected to have better effect when processing the time sequence problem of seismic wave reproduction according to the superiority of time sequence deep learning algorithms such as a Recurrent Neural Network (RNN) in the fields of speech recognition, language modeling, machine translation and the like. Therefore, the research of theories and application technologies in the field of the control of the vibration table based on the artificial intelligent algorithm such as deep learning and the like has higher feasibility and necessity.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a deep learning-based seismic simulation vibration table closed-loop control method and a deep learning-based vibration table closed-loop controller training method.
The technical scheme adopted by the invention is as follows:
the deep learning controller is a deep network model built based on a Recurrent Neural Network (RNN), takes seismic wave acceleration signals and vibrating table feedback acceleration signals as input, outputs control signals to enter a vibrating table open-loop model, and accordingly improves the reproduction accuracy of the vibrating table on the seismic time course.
The Recurrent Neural Network (RNN) can only store memory depending on a complete and continuous track, and because an acceleration feedback signal output by a vibration table system is generated in real time, and the RNN is difficult to establish time sequence correlation between front and rear signals when processing a non-completely observed signal, the invention provides a discontinuous track input signal processing method, so that the RNN can process the feedback signal generated in real time.
The input signal processing method of the discontinuous track specifically comprises the following steps: each set of input and feedback signals (a) is processed as the RNN receives the discontinuous feedback signalt,a’t) Then, the long-term memory unit c and hidden state h inside RNN are saved, and the next set of input and feedback signals (a) are processedt+1,a’t+1) C and h are also used as input, so that the RNN is prevented from not observing past memory due to multiple loop iterations. After the feedback signal is introduced, the input-output relationship of the deep learning controller is as follows:
Ot，C1，h1＝LSTM((at,a’t))
Ot+1，C2，h2＝LSTM((at+1,a’t+1)，C1，h1)
Ot+2，C3，h3＝LSTM((at+2,a’t+2)，C2，h2)
...........
the deep learning-based vibration table closed-loop controller training method comprises the following steps:
s1, constructing a deep network model based on a Recurrent Neural Network (RNN) to serve as a deep learning controller;
s2, building a closed-loop control simulation model of the seismic simulation vibrating table based on a three-parameter control algorithm, and collecting seismic wave acceleration signals, vibrating table feedback acceleration signals and output signals of a three-parameter controller as training data of a depth network model;
s3, dividing the training data into a training set, a verification set and a test set, and training the deep network model through supervised learning until the input-output relationship can approach the performance of the three-parameter controller;
in step S1, the deep network model adopts a long-short term memory network (LSTM) that is a variant of a recurrent neural network, and is constructed by using a pytorch deep learning framework based on python language, and the model structure and the super-parameter setting are determined and adjusted according to the complexity and the training effect of the simulation object.
The network structure of the LSTM model comprises an input layer, a plurality of hidden layers, a full-connection layer and an output layer, wherein the hidden layers adopt 40-80 nodes, Adam is adopted as a parameter optimization algorithm, and MSE is adopted as a loss function.
According to the real-time control requirement of the vibrating table, the input layer and the output layer of the LSTM model are set to be of a single-input single-output structure, so that the model can output control signals in real time along with sequential input of discretized seismic wave signals.
In step S2, the three-parameter control algorithm based vibration table closed-loop control simulation model is built by a Simulink simulation tool in Matlab, the three-parameter control algorithm includes two parts, namely a three-parameter generator and three-parameter feedback, and the three-parameter control algorithm changes the pole configuration of the control system on the whole by using the combination of displacement, speed and acceleration signals, compensates the resonance frequency of the system, and improves the frequency band characteristics.
The open-loop model of the vibration table is modeled based on a servo valve three-continuous equation, and only the characteristics of the empty table and the vibration condition of single degree of freedom are considered in the open-loop model of the vibration table.
The method comprises the steps of performing parameter setting on a vibrating table closed-loop control simulation model based on three-parameter control, inputting a plurality of seismic waves for real-time simulation after the output of a vibrating table can reproduce seismic wave signals accurately, and collecting input-output data generated by a three-parameter controller to serve as a training sample of a deep learning controller.
In order to improve the learning efficiency of the deep learning controller, a smaller sampling step length is selected when training samples are collected so as to ensure the time sequence correlation degree between data, and the convergence rate of the network can be effectively improved.
In step S3, the training data is divided into a training set, a validation set and a test set according to a ratio of 6:1:3, where the training set is used for supervised training of the model, the validation set is used for preliminary validation after network training is completed, and the validation set is used as a basis for adjusting the hyper-parameters, and the test set is used for final testing of the network to determine the generalization of the model.
The supervision type training comprises the following specific implementation steps: the deep learning controller takes a seismic wave acceleration signal and a vibration table feedback acceleration signal as input, takes an output signal of the three-parameter controller as a label, takes a difference value output by the label and the deep learning controller as a loss function, and updates network parameters through a gradient descent method, so that the deep learning controller has control performance approaching to that of the three-parameter controller.
Under the test of a plurality of seismic waves in the test set, if the correlation degree between the output of the deep network model after training and the corresponding label can reach more than 99%, the deep network controller can replace a three-parameter controller to carry out closed-loop control on the vibrating table.
The invention provides a deep learning-based seismic simulation vibration table closed-loop control method and a deep learning-based vibration table closed-loop controller training method, which utilize time sequence deep learning algorithms such as a Recurrent Neural Network (RNN) to simulate the control effect of a traditional three-parameter algorithm, are beneficial to promoting the construction of a high-performance, high-precision and high-reliability intelligent anti-seismic test platform, and can also provide bottom-layer algorithm support for researches such as a substructure mixed test and the like, thereby improving the international comprehensive competitiveness of China in the field of structural tests.
Drawings
FIG. 1 is a flow chart of a deep learning-based method for training a closed-loop controller of a vibrating table according to the present invention.
FIG. 2 is a schematic diagram of a vibration table closed-loop controller training method based on a three-parameter algorithm and deep learning.
FIG. 3 is a schematic diagram of the application of the seismic modeling vibration table closed-loop control based on deep learning.
Detailed Description
Embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
As shown in fig. 1, an embodiment of the deep learning-based shaking table closed-loop controller training method includes the steps of:
s1, constructing a deep network model based on a Recurrent Neural Network (RNN) to serve as a deep learning controller;
s2, building a closed-loop control simulation model of the seismic simulation vibrating table based on a three-parameter control algorithm, and collecting seismic wave acceleration signals, vibrating table feedback acceleration signals and output signals of a three-parameter controller as training data of a depth network model;
s3, dividing the training data into a training set, a verification set and a test set, and training the deep network model through supervised learning until the input-output relationship can approach the performance of the three-parameter controller;
in step S1, the depth network model adopts a long-short time memory network (LSTM) that is a variant of a recurrent neural network, and is constructed by using a pytorch depth learning framework based on python language, the network structure of the LSTM model is designed to be an input layer, two hidden layers, a fully-connected layer and an output layer, the hidden layers respectively adopt 40 and 60 nodes, the parameter optimization algorithm adopts Adam, and the loss function adopts MSE.
According to the real-time control requirement of the vibrating table, the input layer and the output layer of the LSTM model are set to be of a single-input single-output structure, so that the model can output control signals in real time along with sequential input of discretized seismic wave signals.
In step S2, the three-parameter algorithm-based vibration table closed-loop control simulation model passes through Matlab
The Simulink simulation tool is built, the three-parameter control algorithm changes the pole configuration of the control system on the whole by utilizing the combination of displacement, speed and acceleration signals, and compensates the resonance frequency of the system, so that the frequency band characteristic is improved.
The vibration table open-loop model is modeled based on a servo valve three-continuous equation, and the vibration table open-loop model related to the embodiment only considers the characteristics of an empty table and the vibration condition of single degree of freedom.
Through parameter setting of a vibrating table closed-loop control simulation model based on three-parameter control, after the output of a vibrating table can reproduce seismic wave signals accurately, more than 30 seismic waves are input for real-time simulation, and input-output data generated by a three-parameter controller are collected to serve as training samples of a deep learning controller.
In order to improve the learning efficiency of the deep learning controller, when training samples are collected, the sampling step length of 0.001s is selected to ensure the time sequence correlation degree between data, and the convergence rate of the network is effectively improved.
As shown in fig. 2, in step S2, the three-parameter controller includes two parts, namely a three-parameter generator and a three-parameter feedback link: the three-parameter generator takes seismic wave acceleration signals as input, the three-parameter feedback link takes vibration table feedback signals as input, and the output of the three-parameter generator and the output of the vibration table feedback signals are converged and then enter an open-loop model of the vibration table as control signals.
In step S3, the training data is divided into a training set, a validation set and a test set according to a ratio of 6:1:3, where the training set is used for supervised training of the model, the validation set is used for preliminary validation after network training is completed, and the validation set is used as a basis for adjusting the hyper-parameters, and the test set is used for final testing of the network to determine the generalization of the model.
The supervision type training comprises the following specific implementation steps: the deep learning controller takes a seismic wave acceleration signal and a vibration table feedback acceleration signal as input, takes an output signal of the three-parameter controller as a label, takes a difference value output by the label and the deep learning controller as a loss function, and updates network parameters through a gradient descent method, so that the deep learning controller has control performance approaching to that of the three-parameter controller.
Under the test of more than 10 seismic waves in the test set, the correlation degree between the output of the deep learning controller after training and the corresponding label can reach more than 99%, and the deep network controller can replace a three-parameter controller to carry out closed-loop control on the vibrating table.
The trained deep learning controller is connected with a vibration table system in an abutting mode, when the deep learning controller receives discontinuous feedback signals, after one group of input and feedback signals are processed each time, a long-term memory unit c and a hidden state h in the RNN are stored, and when the next group of input and feedback signals are processed, the c and the h are also used as input, so that the RNN is prevented from observing the past memory due to multiple times of cycle iteration.
As shown in fig. 3, in the present embodiment, the seismic simulation vibrating table closed-loop control method based on deep learning includes: the deep learning controller replaces an original three-parameter controller and a vibrating table open-loop model to form closed-loop control, takes seismic wave acceleration signals and vibrating table feedback acceleration signals as input, and outputs control signals to enter the vibrating table open-loop model, so that the reproduction precision of the vibrating table to the seismic time course is effectively improved.

Claims (8)

1. A seismic simulation vibration table closed-loop control method based on deep learning is characterized by comprising the following steps:
s1, constructing a deep network model based on a Recurrent Neural Network (RNN) to serve as a deep learning controller;
s2, building a closed-loop control simulation model of the seismic simulation vibrating table based on a three-parameter control algorithm, and collecting seismic wave acceleration signals, vibrating table feedback acceleration signals and output signals of a three-parameter controller as training data of a depth network model;
s3, dividing the training data into a training set, a verification set and a test set, and training the deep network model through supervised learning until the input-output relationship can approach the performance of the three-parameter controller;
2. the method of claim 1, further comprising: because the acceleration feedback signal output by the vibration table system is generated in real time, the input signal processing method of the discontinuous track is provided aiming at the characteristic that the RNN can only store memory depending on the complete and continuous track, so that the RNN can process the feedback signal generated in real time.
3. The method of claim 2, wherein the input signal processing method comprises: each set of input and feedback signals (a) is processed as the RNN receives the discontinuous feedback signalt,a’t) Then, the long-term memory unit c and hidden state h inside RNN are saved, and the next set of input and feedback signals (a) are processedt+1,a’t+1) C and h are also used as input, so that the RNN is prevented from not observing past memory due to multiple loop iterations.After the feedback signal is introduced, the input-output relationship of the deep learning controller is as follows:
Ot，C1，h1＝LSTM((at,a’t))
Ot+1，C2，h2＝LSTM((at+1,a’t+1)，C1，h1)
Ot+2，C3，h3＝LSTM((at+2,a’t+2)，C2，h2)
...........
4. the deep learning based vibration table closed-loop controller training method according to claim 1, wherein: the deep network model in the step S1 adopts a long-short time memory network (LSTM) which is a variant of a recurrent neural network, the deep network model built based on the long-short time memory network (LSTM) is a single-input single-output model, the network is provided with a plurality of hidden layers, each layer adopts 40-80 nodes, a parameter optimization algorithm adopts Adam, and a loss function adopts MSE.
5. The deep learning-based vibrating table closed-loop controller training method as claimed in claim 1, wherein in step S2 the vibrating table simulation model performs closed-loop control using a conventional three-parameter algorithm, performs real-time simulation by inputting a plurality of seismic waves after parameter setting, and collects input-output data generated by the three-parameter controller and uses the input-output data as a training sample of the deep learning controller.
6. The deep learning-based vibrating table closed-loop controller training method as claimed in claim 1, wherein in step S3, the training data is divided into a training set, a validation set and a test set according to a ratio of 6:1:3, wherein the training set is used for supervised training of the model, the validation set is used for preliminary validation after network training is completed, and the preliminary validation is used as a basis for adjusting the hyper-parameters, and the test set is used for final testing of the network to determine the generalization of the model.
7. The deep learning based vibration table closed-loop controller training method according to claim 1, wherein: the supervision type training comprises the following specific implementation steps: the deep learning controller takes a seismic wave acceleration signal and a vibration table feedback acceleration signal as input, takes an output signal of the three-parameter controller as a label, takes a difference value output by the label and the deep learning controller as a loss function, and updates network parameters through a gradient descent method, so that the deep learning controller has control performance approaching to that of the three-parameter controller.
8. The deep network-based vibrating table closed-loop controller training method as recited in claim 1, wherein the deep learning controller after completion of supervised learning training performs closed-loop control on the vibrating table instead of a three-parameter controller if the correlation degree between the output of the deep learning controller and a corresponding label can reach more than 99% under the test of a plurality of seismic waves in a test set.
CN202110883227.5A 2021-08-02 2021-08-02 Seismic simulation vibration table closed-loop control method based on deep learning Pending CN113705084A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110883227.5A CN113705084A (en) 2021-08-02 2021-08-02 Seismic simulation vibration table closed-loop control method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110883227.5A CN113705084A (en) 2021-08-02 2021-08-02 Seismic simulation vibration table closed-loop control method based on deep learning

Publications (1)

Publication Number Publication Date
CN113705084A true CN113705084A (en) 2021-11-26

Family

ID=78651336

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110883227.5A Pending CN113705084A (en) 2021-08-02 2021-08-02 Seismic simulation vibration table closed-loop control method based on deep learning

Country Status (1)

Country Link
CN (1) CN113705084A (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108596327A (en) * 2018-03-27 2018-09-28 中国地质大学（武汉） A kind of seismic velocity spectrum artificial intelligence pick-up method based on deep learning
CN112285776A (en) * 2020-10-23 2021-01-29 中国矿业大学(北京) Seismic velocity automatic picking method based on deep learning
US20210199828A1 (en) * 2018-09-12 2021-07-01 Korea Institute Of Geoscience And Mineral Resources Seismic vulnerability analysis system of user's living space and seismic vulnerability analysis method of user's living space using the same

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108596327A (en) * 2018-03-27 2018-09-28 中国地质大学（武汉） A kind of seismic velocity spectrum artificial intelligence pick-up method based on deep learning
US20210199828A1 (en) * 2018-09-12 2021-07-01 Korea Institute Of Geoscience And Mineral Resources Seismic vulnerability analysis system of user's living space and seismic vulnerability analysis method of user's living space using the same
CN112285776A (en) * 2020-10-23 2021-01-29 中国矿业大学(北京) Seismic velocity automatic picking method based on deep learning

Similar Documents

Publication Publication Date Title
CN106843172B (en) Complex industrial process On-line quality prediction method based on JY-KPLS
CN104035337B (en) A kind of flexible mechanical arm Sliding Mode Control Design Method based on singular perturbation theory
CN113359212B (en) Reservoir characteristic prediction method and model based on deep learning
CN110286592A (en) A kind of multi-modal movement technique of machine fish based on BP neural network and system
CN109783910A (en) It is a kind of to utilize the optimum structure design method for generating confrontation network acceleration
CN108562811A (en) Complicated electrical energy power quality disturbance analysis method based on the memory of two-way shot and long term
CN103559537B (en) Based on the template matching method of error back propagation in a kind of out of order data stream
CN111191785A (en) Structure searching method based on expanded search space
CN107894618A (en) A kind of full waveform inversion gradient preprocess method based on model smoothing algorithm
CN112529184A (en) Industrial process optimization decision method fusing domain knowledge and multi-source data
CN104616072A (en) Method for improving concentration of glutamic acid fermented product based on interval optimization
Ge et al. Computation of synchronized periodic solution in a BAM network with two delays
CN115600669A (en) High-efficiency deep pulse neural network learning method based on local classifier
CN113468466B (en) One-dimensional wave equation solving method based on neural network
Cheng et al. An on-line self-learning power system stabilizer using a neural network method
CN104238367A (en) Method for controlling consistency of vibration of surfaces of shell structures on basis of neural networks
CN113705084A (en) Seismic simulation vibration table closed-loop control method based on deep learning
CN112149255A (en) Coal face hydraulic support working resistance prediction method based on LSTM neural network
CN104668493A (en) Combined vibration control system and control method for crystallizer of continuous casting machine
CN114398819A (en) Method and system for predicting effective wave height of unstructured grid based on deep learning
CN112907698A (en) Logging curve generation method dynamically fusing time sequence and non-time sequence characteristics
Li et al. Parameter identification based on pso algorithm for piezoelectric actuating system with rate-dependent prandtl-ishlinskii hysteresis modeling method
AU2021102281A4 (en) Inverse identification method of aquifer structure based on octave convolution residual network
CN105353614A (en) Control method for speed-variable frequency-constant double-feed bidirectional tidal power generator system
Lei et al. Research on Intelligent PID Control Algorithm Based on Neural Network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination