Three-parameter control parameter tuning method for earthquake simulation shaking table based on deep learning

Abstract

A method for adjusting the three-parameter control parameters of an earthquake simulation vibration table based on deep learning belongs to the field of earthquake simulation vibration table control. The present invention includes: a system identification link, based on the response signal and input signal of the actual system, a numerical model that can correctly reflect the real input-output relationship is identified through a system identification method as a system closed-loop model; an offline tuning link, adding a three-parameter control as an incremental closed-loop control system outside the system closed-loop model, and selecting a deep learning algorithm to perform parameter tuning on the simulation model; a real machine testing link, obtaining the offline tuned control parameters, inputting them into the actual vibration table system for verification, and ending the tuning if the performance index requirements are met, otherwise executing the offline tuning link to re-tune the parameters. The present invention combines a deep learning algorithm to achieve the tuning of the three-parameter control parameters of the vibration table, which can improve the parameter tuning efficiency and accuracy, improve the safety of the parameter tuning, and ensure the smooth operation of the vibration table system.

Description

Technical Field

[0001] The invention relates to a three-parameter control parameter setting method for an earthquake simulation vibration table, and belongs to the field of earthquake simulation vibration table control. Background Art

[0002] An earthquake simulation shake table is an irreplaceable testing tool in the field of earthquake resistance. The shake table system itself possesses a certain bandwidth and stability. Currently, most earthquake simulation shake tables utilize a three-parameter control system, and the parameters of the three-parameter control are the decisive factors in determining the effectiveness of the shake table's control.

[0003] At present, the main methods for adjusting the parameters of three-parameter control are: manual parameter adjustment method, in which the parameter adjustment personnel continuously try different parameters until a satisfactory control effect is achieved. This method is simple and crude, but difficult to adjust; expert experience adjustment method, which gives parameter adjustment rules based on the experience of experts in system debugging, and performs fast and effective parameter adjustment based on the frequency response curve. This method clarifies the adjustment steps and is simple and effective; parameter adjustment method based on frequency domain identification, which identifies the true characteristics of the actual system and combines theoretical calculation methods to obtain control parameters with higher precision. This method requires fewer debugging times and has high accuracy. Although all of the above methods can complete parameter adjustment, the following problems still exist:

[0004] (1) The adjustment process is time-consuming and labor-intensive, requiring a large amount of manpower and time. Currently, the adjustment of vibration table parameters may take half a day at the fastest, or a week or even several weeks at the slowest.

[0005] (2) There are certain risks in the tuning process. For example, improper parameter input may cause the vibration table to become unstable and hit the cylinder, causing irreparable damage to the table or the test piece.

[0006] Currently, some applications of deep learning algorithms in control parameter adjustment have been achieved. For example, the invention patent with publication number CN109330 846A proposes a parameter optimization method for an air wave pressure massager based on a deep learning algorithm. This method collects the user's blood pressure and pulse data in real time and adjusts the control parameters of the massage chair in real time through a deep reinforcement learning algorithm. For example, the invention patent with publication number CN113193789B proposes a method for optimizing motor startup control parameters. This method randomly selects initial values based on the motor startup preset values to generate a training set and a test set, and then establishes a deep network model for training. Ultimately, the target deep network model is obtained, and the optimal parameters can be obtained based on the motor startup prediction values. Both examples can use deep learning or reinforcement learning to obtain the optimal control parameters, but both require real-time acquisition of object information and real-time adjustment of control parameters. This is not applicable to the earthquake simulation vibration table. The vibration table is a system with rapid response and extremely fast signal changes. The speed of real-time parameter adjustment cannot keep up with the response of the system. Moreover, due to the large and rapid amplitude of signal changes, it may cause system instability. Therefore, the vibration table requires a fixed control parameter. The present invention uses deep learning to perform offline parameter adjustment and conducts actual machine verification, which not only ensures safety but also saves time and effort to obtain the optimal control parameters. Summary of the Invention

[0007] The present invention proposes a three-parameter control parameter setting method for an earthquake simulation vibration table based on a deep learning algorithm.

[0008] refer to Figure 1 A method for tuning the three-parameter control parameters of an earthquake simulation vibration table includes system identification, offline parameter tuning, and real-machine testing. The specific steps are as follows:

[0009] Step 1: Use the preprocessed seismic wave signal as the input signal of the shaking table and determine the initial parameters of the three-parameter control through expert experience method, run the shaking table system, and obtain the actual response signal of the shaking table system.

[0010] Step 2: Use the input signal of the vibration table as the input of the neural network, the response signal of the vibration table as the label signal of the neural network, and build a deep multi-layer neural network model for training.

[0011] Step 3: Introduce three-parameter control as a closed-loop control model on the deep multi-layer neural network model.

[0012] Step 4: Parameter tuning of the closed-loop model.

[0013] Step 5: Combine and verify the parameter tuning results with the parameters in the actual vibration table control system.

[0014] Step 6: Determine whether the parameter tuning results meet the vibration table control performance index δ [≥90%, ≤10%] in the closed-loop model. If the requirements are met, proceed to step 7. If the performance requirements are not met, repeat step 4.

[0015] Step 7: If the control parameters adjusted in step 6 meet the control index in the test model and the input actual vibration table system still meets the control index δ [≥90%, ≤10%], the control parameters are adjusted.

[0016] Furthermore, in step 1, the vibration table's input signal is a Gaussian white noise acceleration signal. Signal preprocessing includes verifying the displacement amplitude and filtering. The displacement amplitude must not exceed the vibration table's displacement limit. The signal frequency is 0.5 to 50 Hz, and both high and low frequencies must be filtered out.

[0017] Furthermore, in step 2, the input signal of the vibration table is set as the input layer of the neural network, the response signal of the vibration table is set as the output layer of the neural network, and two or more hidden layers are connected between the input layer and the output layer, where K neurons are set in each hidden layer.

[0018] Furthermore, in step 2, the input signal of the vibration table is set as the input layer of the neural network, the response signal of the vibration table is set as the output layer of the neural network, and two or more hidden layers are connected between the input layer and the output layer, where K neurons are set in each hidden layer.

[0019] Furthermore, in step 3, by adding three-parameter control to the deep network model and considering the second-order effect of the servo valve and the influence of the sensor, a closed-loop control model can be derived. The expression is as follows:

[0020]

[0021] Where x represents the displacement value measured by the displacement sensor of the vibration table, u is the input acceleration signal, and A p is the effective pressure bearing area of the vibration table actuator piston, G(s) is the deep network closed-loop model, and A d is the displacement feedforward gain, A′ a is the acceleration feedback gain, A′ d is the displacement feedback gain, A′ V is the speed feedback gain, K d is the displacement feedback normalized sensitivity, K A is the normalized sensitivity of acceleration feedback, K V is the normalized sensitivity of speed feedback, G q and G a They are the inherent second-order characteristics of the servo valve and the inherent second-order characteristics of the sensor, and the expressions are as follows:

[0022]

[0023] Among them, n q is the second-order system natural frequency of the servo valve, D q is the inherent damping ratio of the second-order system of the servo valve; n a is the second-order system natural frequency of the sensor, D a is the inherent damping ratio of the second-order system of the sensor. The above values can be found in the servo valve and sensor manuals. V , A d , A′ a , A′ d , A′ V It is the object of parameter adjustment, and its initial value is determined by expert experience.

[0024] Furthermore, the control index δ can be calculated from the input and output data of the model to obtain a value that can measure the control effect. It is expressed as follows:

[0025]

[0026] Where u i represents the acceleration input signal, represents the target acceleration signal, ε represents the correlation coefficient, and τ represents the peak error. Generally, the correlation coefficient ε is required to be greater than 90% and the peak error τ is required to be less than 10% to meet the control index requirements.

[0027] Furthermore, in step 4, the loss function μ is applied to the parameter K i =[A V , A d , A′ a , A′ d , A′ V ] Calculate the partial derivatives respectively and get the corresponding partial derivative function, which is the gradient value function. Update the control parameters according to the gradient of each parameter. The iterative termination condition of the gradient descent algorithm is that the number of training times reaches 1000. The parameter update formula is as follows:

[0028]

[0029] Where α is the learning rate, u i represents the acceleration input signal, Indicates the target acceleration signal.

[0030] Furthermore, in step 7, the control parameter K set in step 4 is iEnter the actual earthquake simulation shaking table system, feeding the preprocessed seismic waves as the input signal into the shaking table. Observe whether the shaking table's control index δ meets the control requirements to verify the effectiveness of the parameter tuning. If the control effect meets the control index requirements, the tuning is complete. Otherwise, proceed to step 4.

[0031] The beneficial effects of the present invention are:

[0032] The present invention includes a system identification link, an offline tuning link, and a real machine verification link. The identification link is based on the real vibration table system for system identification. The identification results obtained can also reflect the input-output relationship of the real system. On this basis, the three-parameter control parameters are offline tuned. The tuning algorithm can select a deep learning algorithm, a BP neural network algorithm, a reinforcement learning algorithm, etc. Compared with the traditional three-parameter control parameter tuning algorithm, the tuning algorithm adopted by the present invention saves a lot of manual parameter adjustment process. It only needs to set the control index and let the deep learning algorithm perform autonomous training. Finally, the control parameters that meet the control index requirements can be obtained. After the offline tuning is completed, the tuned parameters are re-input into the real vibration table. If the control effect can meet the requirements of the control index, the tuning is completed. If the control effect is not satisfied, the number of training times or learning rate of the deep learning is modified, and retraining is performed until the ideal control parameters are obtained. Since the process is offline tuning, it will not cause damage to the actual table body or test piece, and the autonomous training of deep learning saves the time cost and labor cost of manual parameter adjustment.

[0033] Combining deep learning with the three-parameter control parameter tuning problem of a vibration table.

[0034] The present invention can also adopt the method of multi-network parallel training. When multiple tuning algorithms perform parameter tuning at the same time, the tuning parameters with the best effect are finally selected.

[0035] The present invention establishes a complete set of practical parameter setting methods based on the characteristics of the vibration table system itself, which can not only save a lot of parameter setting time for parameter setting personnel and improve the accuracy of control parameters, but also ensure the safety of the vibration table system and test pieces. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Flowchart of a three-parameter control parameter tuning method for earthquake simulation shaking table based on deep learning

[0037] Figure 2 Flowchart of parameter tuning of three-parameter control system based on deep learning algorithm mentioned in the embodiment

[0038] Figure 3 Schematic diagram of the three-parameter control system based on the deep learning algorithm mentioned in the embodiment DETAILED DESCRIPTION

[0039] The accompanying drawings are only for illustrative purposes and should not be construed as limiting this patent.

[0040] The present invention proposes a parameter setting method to solve the problem that the setting of three-parameter control parameters of an existing earthquake simulation vibration table is time-consuming and labor-intensive.

[0041] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0042] See also Figure 1 and Figure 3 , the present invention proposes a parameter setting method, comprising the following steps:

[0043] Step 1: Use the preprocessed seismic wave signal as the input signal of the shaking table and determine the initial parameters of the three-parameter control through expert experience method, run the shaking table system, and obtain the actual response signal of the shaking table system.

[0044] Step 2: Use the input signal of the vibration table as the input of the neural network, the response signal of the vibration table as the label signal of the neural network, and build a deep multi-layer neural network model for training.

[0045] Step 3: Introduce three-parameter control as a closed-loop control model on the deep multi-layer neural network model.

[0046] Step 4: Parameter tuning of the closed-loop model.

[0047] Step 5: Verify the parameter tuning results in the closed-loop model system.

[0048] Step 6: Determine whether the parameter tuning results meet the vibration table control performance index δ [≥90%, ≤10%] in the closed-loop model. If the requirements are met, proceed to step 7. If the performance requirements are not met, repeat step 4.

[0049] Step 7: If the control parameters adjusted in step 6 meet the control index in the test model and the input actual vibration table system still meets the control index δ [≥90%, ≤10%], the control parameters are adjusted.

[0050] In the selection of the input signal described in step 1 of this embodiment, since the actual vibration table system is a system with a certain bandwidth and rapid response, a white noise acceleration signal is selected as the input signal with a bandwidth of 0.1-50 Hz and an amplitude of 0.3g.

[0051] The initial parameters of the vibration table are selected according to the expert experience method, and the acceleration feedforward A a Take 0.03, speed feedforward A VTake 0.05, displacement feedforward A d Take 0.295, acceleration feedback A′ a Take -0.028, displacement feedback A′ d Take 0, speed feedback A' V Take 0.005.

[0052] The closed-loop identification model G(s) of the actual system described in step 2 of this embodiment is based on the acceleration response signal of the actual system and the system acceleration input signal. It uses 50 seismic wave inputs to obtain 50 response signals, with each input and output serving as a set of network input data. The LSTM network uses a training data to test data ratio of 7:3, i.e., 35 sets of data are used for training and 15 sets of data are used for testing. The LSTM network model uses a two-layer network with 40 hidden nodes per layer. The learning rate ranges from 0.01 to 0.0001, varying with the number of training cycles. After 1000 training cycles, the resulting LSTM network model achieves an average correlation coefficient ε of 97% and an average peak error τ of 9% when running the test data, meeting the training requirements. Therefore, this network is considered an LSTM network closed-loop model.

[0053] In step 3 of this embodiment, three-parameter control is added to the LSTM network closed-loop model, and the influence of the servo valve second-order effect and the sensor is considered. The expression is as follows:

[0054]

[0055] Where x is the displacement of the vibration table, u is the acceleration control command signal, and A p is the effective pressure bearing area of the piston, K d is the displacement feedback normalized sensitivity, K A is the normalized sensitivity of acceleration feedback, K V is the normalized sensitivity of velocity feedback. The above parameters are the inherent parameters of the vibration table. Here we take A p =1.1×10 -3 m 2 , the three sensitivities of the feedback sensor are: K d =100 / ,K V =16.67 / / , K A =0.5V / m / s 2 , G q is the second-order characteristic of the servo valve, G a is the second-order characteristic of the sensor.

[0056]

[0057] Among them, n q is the second-order system natural frequency of the servo valve, D q is the inherent damping ratio of the second-order system of the servo valve; n ais the second-order system natural frequency of the sensor, D a is the inherent damping ratio of the second-order system of the sensor. The above values can be found in the servo valve and sensor manuals, where n q =628.31,D q =0.7,n a =942.48, D a =0.7.

[0058] In step 4 of this embodiment, the gradient descent algorithm is selected as the tuning algorithm. Figure 2 shown

[0059] The method for implementing the gradient descent algorithm training in step 4 of this embodiment is as follows:

[0060] Step 4.1. Set the initialization system control parameter K i =[A V , A d , A′ a , A′ d , A′ v ]

[0061] Step 4.2: Input seismic wave signal u i , run the LSTM network model and get the model output

[0062] Step 4.3. Input u according to the model i and output Calculate the control index δ and loss function μ. The specific calculation is as follows:

[0063]

[0064] Where ε represents the correlation coefficient and ε represents the peak error.

[0065]

[0066] Where u i Represents the output signal, represents the target signal, and n represents the length of the signal.

[0067] Step 4. Compare the loss function μ to the parameter K i =[A V , A d , A′ a , A′ d , A′ V ] Calculate the partial derivatives respectively and get the corresponding partial derivative function, which is the gradient value function. Update the control parameters according to the gradient of each parameter. The iterative termination condition of the gradient descent algorithm is that the number of training times reaches 1000 times. The formula is as follows:

[0068]

[0069] Where α is the learning rate, which defaults to 0.001.

[0070] Step 45: Input the updated parameters into the system model, repeat steps 42 and 43, and check whether ε>90% and τ<10%. If both meet the requirements, proceed to step 46, otherwise repeat steps 42 to 45.

[0071] Step 46: After the operation is completed, output the adjusted control parameters

[0072] In step five of this embodiment, the result of the adjustment in step four is combined with the initial parameters determined in step one as new control parameters and input into the actual vibration table. The vibration table system is run to check whether ε is greater than 90% and whether τ is less than 10%. If so, the adjustment is completed; otherwise, step four is executed.

[0073] Although the present invention has been described herein with reference to specific embodiments, it should be understood that the embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than as described in the original claims.

Claims

1. A method for setting three-parameter control parameters of an earthquake simulation vibration table based on deep learning, characterized by: The process includes system identification, offline parameter tuning, and actual machine testing. First, the system identification model of the actual vibration table is obtained through testing. Then, offline parameter tuning of the deep learning tuning algorithm is performed on the system identification model. Finally, the tuned parameters are tested in the actual vibration table system. The tuning method includes the following steps: Step 1: Use the pre-processed seismic wave signal as the input signal of the shaking table and determine the initial parameters of the three-parameter control, operate the shaking table system, and obtain the actual response signal of the shaking table system; Step 2: Use the vibration table's input signal as the input of the neural network, the vibration table's response signal as the label signal of the neural network, and build a deep multi-layer neural network model for training; Step 3: Introduce three-parameter control as a closed-loop control model on the deep multi-layer neural network model; Step 4: Parameter tuning of the closed-loop control model; Step 5: Verify the parameter tuning results in the closed-loop model system; Step 6: Determine whether the parameter tuning results meet the vibration table control performance index δ [≥90%, ≤10%] in the closed-loop model. If the requirements are met, proceed to step 7. If the performance requirements are not met, repeat step 4. Step 7: If the control parameters adjusted in step 6 meet the control index in the test model and the actual vibration table system input still meets the control index δ [≥90%, ≤10%], the control parameters are considered to be adjusted; In the second step, the input signal of the vibration table is set as the input layer of the neural network, the response signal of the vibration table is set as the output layer of the neural network, and two or more hidden layers are connected between the input layer and the output layer, wherein each hidden layer is provided with K neurons; the neural network is specifically: a two-layer LSTM network, 40 hidden nodes per layer, a learning rate of 0.01-0.0001, and training 1000 times. When the obtained LSTM network model runs the test data, the correlation coefficient ε is greater than 97% and the peak error τ is less than 9%, which meets the training requirements; In step 3, three-parameter control is added to the deep network model, and the second-order effect of the servo valve and the influence of the sensor are considered to introduce a seventh-order closed-loop control model, which is expressed as follows: Where x represents the displacement value measured by the displacement sensor of the vibration table, u is the acceleration input signal, and A p is the effective pressure bearing area of the vibration table actuator piston, G(s) is the deep network closed-loop control model, and A d is the displacement feedforward gain, A ′ a is the acceleration feedback gain, A ′ d is the displacement feedback gain, A ′ V is the speed feedback gain, K d is the displacement feedback normalized sensitivity, K A is the normalized sensitivity of acceleration feedback, K V is the normalized sensitivity of speed feedback, G q and G a are the inherent second-order characteristics of the servo valve and the sensor respectively; A in the derived closed-loop control model V , A d , A ′ a , A ′ d , A ′ V It is the object of parameter tuning.

2. A method for setting three-parameter control parameters of an earthquake simulation vibration table based on deep learning according to claim 1, characterized in that: In step 1, the input signal of the vibration table is a Gaussian white noise acceleration signal, and the signal preprocessing includes verifying the displacement amplitude and filtering; the displacement amplitude cannot exceed the displacement limit of the vibration table, the signal frequency is 0.5 to 50 Hz, and both high-frequency and low-frequency parts need to be filtered out.

3. The method for setting three-parameter control parameters of an earthquake simulation vibration table based on deep learning according to claim 1, characterized in that: The control index δ is a value that measures the control effect calculated by the input and output data of the model, which is expressed as follows: Where u i represents the white noise acceleration input signal, represents the system acceleration output signal, ε represents the correlation coefficient, τ represents the peak error; and n represents the length of the signal.

4. A method for setting three-parameter control parameters of an earthquake simulation vibration table based on deep learning according to claim 1, characterized in that: In step 4, the gradient descent algorithm is used to optimize the deep network closed-loop control model. The loss function μ is constructed based on the input and output signals of the closed-loop control model. The loss function is calculated point by point during operation. When the deep network updates the parameters, the total loss function of the entire training data is used. After each round of training data, the control parameters are updated once. The loss function expression is as follows: Where u i represents the acceleration input signal, Represents the system acceleration output signal, and n represents the length of the signal.

5. A method for setting three-parameter control parameters of an earthquake simulation vibration table based on deep learning as claimed in claim 4, characterized in that: In the step 4, the loss function μ is applied to the parameter K i =[A V , A d , A ′ a , A ′ d , A ′ V ] Calculate the partial derivatives respectively and get the corresponding partial derivative function, which is the gradient value function. Update the control parameters according to the gradient of each parameter. The iterative termination condition of the gradient descent algorithm is that the number of training times reaches more than 1000 times. The parameter update formula is as follows: Where α is the learning rate, u i represents the acceleration output signal, Represents the target acceleration signal; the control parameter K will be updated i Input the actual earthquake simulation vibration table system, input the preprocessed seismic wave as the input signal into the vibration table, and observe whether the control index δ of the vibration table meets the control requirements [≥90%, ≤10%]. If the control effect meets the control index requirements, the tuning is completed; otherwise, execute step 4.

Images