A piezoelectric ceramic actuator creep prediction method based on ABSO-LSTM

Abstract

The application discloses a piezoelectric ceramic actuator creep prediction method based on ABSO-LSTM. The method comprises the following steps: (1) collecting piezoelectric ceramic actuator creep time series data under different input voltage conditions to construct a data set, and dividing the data set into a training set and a test set; (2) setting the time step of the LSTM network as 10, taking the historical creep variable, the initial input voltage value and the input voltage transformation value as the input, and using the adaptive artificial bee optimization algorithm (ABSO) to optimize the long short-term memory neural network (LSTM) model hyperparameters on the randomly selected training set, so that the creep data characteristics and the network topology structure are matched; (3) based on the historical data, the optimal hyperparameter combination obtained by the ABSO algorithm is used to construct an LSTM creep prediction model to realize real-time prediction of the piezoelectric ceramic actuator creep variable. The method can effectively predict the creep trend of the piezoelectric ceramic actuator, and solves the problem that the LSTM creep prediction network hyperparameters cannot determine the optimal value, thereby reducing the prediction accuracy.

Description

Technical fields:

[0001] This invention relates to a piezoelectric ceramic actuator, and more specifically to a creep prediction method for piezoelectric ceramic actuators based on ABSO-LSTM. Background technology:

[0002] In fields such as satellite laser communication and space astronomical observation, large-scale photoelectric tracking systems widely employ compound axis control technology to achieve second-level or even millisecond-level accuracy. This is the most effective control structure for achieving large-scale, high-precision tracking. Compound axis control is a form of two-dimensional correlated control system. The control system using a fast steering mirror (FSM) as the actuator is a high-precision tracking system for compound axis systems, and the achievable accuracy is mainly determined by the fast steering mirror actuator. The fast steering mirror mechanism typically uses a voice coil motor (VCM) and a piezoelectric actuator (PZT) as drivers. Compared to the former, the latter has the advantages of high resonant frequency and large displacement resolution; however, its inherent creep nonlinearity can adversely affect positioning accuracy. Piezoelectric creep is characterized by the phenomenon that when a fixed voltage is applied to the piezoelectric actuator, the resulting displacement does not stabilize immediately but changes slowly over a subsequent period. The subsequent displacement is the piezoelectric actuator creep. Its microscopic mechanism is due to the influence of internal friction between the crystal lattices within the dielectric. When the crystal domains rotate, they influence each other, resulting in a certain hysteresis. The piezoelectric creep is greatly affected by the material of the piezoelectric actuator itself and the external environment. The values ​​of L0, γ, and t0 in the model will change due to the input voltage increment, the initial input voltage, the direction of voltage change, and the different materials of the piezoelectric actuator. Therefore, this model has a certain effect on modeling the piezoelectric creep characteristics under certain conditions, but it does not have universality. Therefore, there is still a lack of a high-precision, real-time method for predicting the creep of piezoelectric ceramic actuators. Summary of the Invention:

[0003] To address the aforementioned application background, this invention proposes a creep prediction method for piezoelectric ceramic actuators based on ABSO-LSTM, comprising the following steps:

[0004] 1) Collect creep time series data of piezoelectric ceramic actuators under different input voltage conditions to construct a dataset, and divide it into training set and test set;

[0005] 2) The time step of the LSTM network model is set to 10. The historical creep variables, the initial input voltage value, and the input voltage transformation value are used as inputs. On the randomly selected training set, the hyperparameters of the LSTM model are optimized using the ABSO optimization algorithm (Zhang Q, Gao Y, Li Q, et al. Adaptive compound control based on generalized Bouc-Wen inverse hysteresis modeling in piezoelectric actuators[J]. Review of Scientific Instruments, 2021, 92(11): 115004) to match the creep data characteristics with the network topology.

[0006] 3) Based on historical data, an LSTM creep prediction model is constructed using the optimal hyperparameter combination obtained by the ABSO optimization algorithm to predict the creep of piezoelectric ceramic actuators in real time.

[0007] Specifically, in step 1:

[0008] The specific process of constructing the dataset is as follows: set different starting drive voltages and voltage step changes for piezoelectric ceramic actuators, collect the corresponding creep variables, the duration of each set of input and output timing data is 1 second, the sampling rate is 100Hz, and the first 0.8 seconds of data in each set constitute the training set, and the last 0.2 seconds of data constitute the test set.

[0009] Specifically, in step 2:

[0010] 1) The LSTM network model described above has a time step of 10 and input features are: historical creep variables, initial input voltage values, and input voltage transformation values;

[0011] 2) The specific process of optimizing the hyperparameters of the LSTM model using the ABSO optimization algorithm is as follows: Initialize the number of iterations of the adaptive beetle swarm optimization algorithm, the population size NP, the acceleration constants c1 and c2, and the maximum value of the inertia weight ω. max Minimum inertial weight ω min The constant λ, the initial step size step(1), the antenna distance attenuation factor c, the initial value of the attenuation factor eta(1), and the location range of each longhorn beetle are given. The BSO-LSTM model is trained on a randomly selected training set, with the root mean square error between the predicted output creep variable and the actual collected creep variable as the objective function for each individual. The hyperparameters of the LSTM network model are optimized through continuous iteration and updating.

[0012] 3) The hyperparameters of the LSTM network model to be optimized are: the number of neurons in the first LSTM layer, the number of neurons in the second LSTM layer, the learning rate of the LSTM, and the maximum number of iterations of the prediction model.

[0013] This invention establishes an ABSO-LSTM creep prediction model for piezoelectric ceramic actuators, which can predict the creep of the piezoelectric ceramic actuator in the next moment with high accuracy and in real time. This has great practical significance for suppressing creep nonlinearity and improving positioning accuracy in the control process of piezoelectric ceramic actuators. Attached image description:

[0014] Figure 1 This is a structural diagram for predicting creep in piezoelectric ceramic actuators.

[0015] Figure 2 This is a block diagram of a piezoelectric ceramic actuator data acquisition system.

[0016] Figure 3 It is an LSTM cell structure.

[0017] Figure 4 It is a creep curve for the 0-30V input voltage range.

[0018] Figure 5 This is the creep prediction curve for an LSTM network. Detailed implementation method:

[0019] The present invention will be further described below with reference to specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0020] A practical control system was built for data acquisition. The piezoelectric ceramic actuator control system consists of an SGS micro-displacement sensor, an SGS signal conditioning module, a main control module, a piezoelectric ceramic actuator drive module, and the piezoelectric ceramic actuator itself. The main control module utilizes a hardware-in-the-loop real-time simulation platform. The main control module generates drive signals for the piezoelectric ceramic actuator, which are then used by the drive module to control the actuator. The SGS signal conditioning module feeds back the detected actual displacement to the main control module. The data acquisition system block diagram is shown below. Figure 2 As shown.

[0021] The piezoelectric ceramic actuator used has an input voltage range of 0–100V, corresponding to an output displacement of 0–30μm. Creep time-series data of the piezoelectric ceramic actuator under different input voltage conditions were collected, with a creep time of 1s under each condition and a sampling rate of 100Hz. The specific drive voltage input methods were as follows: 0V was stepped to 1V, 2V, …, 100V; 1V was stepped to 2V, 3V, …, 100V; 2V was stepped to 3V, 4V, …, 100V, and so on; 100V was stepped to 99V, 98V, …, 0V; 99V was stepped to 98V, 97V, …, 0V; 98V was stepped to 97V, 96V, …, 0V, and so on. Corresponding piezoelectric ceramic actuator output displacement data were collected, totaling 10100 sets of time-creep data, with 100 data points per set. The first 0.8 seconds of data from each group constitute the training set, and the last 0.2 seconds of data from each group constitute the test set.

[0022] The model parameters α, β, γ, p0, p1, q1, q0 are identified using an adaptive beetle swarm optimization algorithm. Assume the adaptive beetle swarm optimization algorithm has 300 iterations, a population size NP = 120, speedup constants c1 = 2.8, c2 = 1.3, and a maximum weight ω. max =0.9, minimum weight ω min =0.4, λ=0.95, initial step size step(1)=2, c=2, initial value of decay factor eta(1)=0.95, the number of neurons in the first LSTM layer is set to [1,100], the number of neurons in the second LSTM layer is set to [1,100], the learning rate of the LSTM is set to [0.01,0.15], and the maximum number of iterations of the prediction model is set to [40,100]. The core problem of the optimization algorithm is to select the objective function:

[0023]

[0024] Where n is the number of samples, y i Represents the true value. This represents the predicted value.

[0025] LSTM is a variant of traditional recurrent neural networks. The LSTM unit structure is as follows: Figure 3 As shown.

[0026] The difference equations for forward propagation to update the cell state and compute the output are defined as shown in equations (2) to (7), which are the forget gate, candidate state, update gate, cell state, output gate and output result, respectively.

[0027] σ f [t]=σ(W f x[t]+R fy[t-1]+b f (2)

[0028]

[0029] σ u [t]=σ(W u x[t]+R u y[t-1]+b u (4)

[0030]

[0031] σ o [t]=σ(W o x[t]+R o y[t-1]+b o (6)

[0032] y[t]=σ o [t]⊙g2(h[t]) (7)

[0033] Where x[t] is the input vector at time t. W f W h W u and W o R is a rectangular weight matrix applied to the input of the LSTM unit. f R h R u and R o It is a square matrix that defines the weights of recursive connections, and b f b h b u and b o It is the deviation vector. σ(·) is the sigmoid function, ⊙ is the Hadamard product of the two vectors, and g1(·) and g2(·) are point-state hyperbolic tangent functions with a range of [-1,1].

[0034] Each door in the unit has a specific and unique function. Forget Door σ f Determine the information to be filtered out from the previous unit state h[t-1]. Input gate σ u Acting on h[t-1], after being modified by the forget gate, it determines the new candidate state. The degree of update of the new cell state h[t]. To obtain the output y[t], g2(·) is used to filter the current cell state, and the output gate σ is... oA portion of the state is selected as the output and returned. Each gate depends on the current external input x[t] and the previous unit output y[t-1]. The gate mechanism in the LSTM structure avoids repeated multiplication operations on a single matrix to some extent, alleviating the vanishing gradient problem, but it also preserves the error that appears over time.

[0035] The creep variable of a piezoelectric ceramic actuator at any given time depends not only on the current time but also on its historical creep variable. The initial input voltage, voltage transition value, and historical creep variable of the piezoelectric ceramic actuator are selected as input features. The time step of the LSTM network is set to 10, and the historical creep variable and the initial input voltage and voltage transition value from the previous 10 time steps are used to predict the creep variable of the piezoelectric ceramic actuator at the next time step. The initial input voltage and voltage step transition value are normalized to [-1, 1].

[0036] The hardware platform for training the model is an Intel i7-12700k CPU and an NVIDIA 3080ti graphics card. The software uses Python 3.6 and PyTorch 1.1 deep learning frameworks to build the network.

[0037] The activation function of the fully connected layer is set to a linear function for the neural network regression output. The performance of the trained network is evaluated by comparing the trained network output with the actual values ​​using statistical evaluation metrics. The Adam adaptive learning rate method is used to update the model parameters. Essentially, the Adam method is a parameter optimization method that incorporates a momentum term; its learning rate is determined by the moment estimate of the gradient, resulting in smooth parameter updates and minimizing the risk of getting trapped in local optima.

[0038] The input voltage signal is set to 0-30V, the creep time is 1s, and the sampling rate is 100Hz. The creep-time curve is shown below. Figure 4 As shown in the figure. The measured creep data of the piezoelectric ceramic actuator from 0.79s to 0.99s were sequentially input into the model to obtain the creep data from 0.8s to 1s. The above neural network creep prediction curve is shown in the figure. Figure 5 As shown in Table 1, the root mean square error of the prediction is as follows.

[0039] Table 1. Prediction Accuracy of LSTM Networks

[0040]

Claims

1.A piezoelectric ceramic actuator creep prediction method based on ABSO-LSTM, characterized by The method comprises the following steps: 1) Collecting time series data of piezoelectric ceramic actuator creep under different input voltage conditions to construct a data set, and dividing it into a training set and a test set; 2) The time step of the LSTM network model is set to 10, and the historical creep, the initial input voltage value and the input voltage change value are used as the input. On the randomly selected training set, the ABSO optimization algorithm is used to optimize the hyperparameters of the LSTM network model, so that the creep data characteristics and the network topology structure are matched; 3) Based on the historical data, the optimal hyperparameter combination obtained by the ABSO optimization algorithm is used to construct an LSTM creep prediction model to predict the piezoelectric ceramic actuator creep in real time; The specific process of constructing the data set in step 1) is as follows: set different initial driving voltages and voltage step changes of the piezoelectric ceramic actuator, and collect the corresponding creep. The time length of each input-output time series data is 1s, and the sampling rate is 100Hz. The first 0.8s data of each group of data constitutes the training set, and the last 0.2s data constitutes the test set; In step 2), the time step of the LSTM network model is 10, and the input features include: historical creep, initial input voltage value, and input voltage change value; The specific process of optimizing the hyperparameters of the LSTM network model by using the ABSO optimization algorithm in step 2) is as follows: initializing the iteration number of the self-adaptive artificial bee optimization algorithm, the population size NP, the acceleration constant c 1 and c 2, the maximum inertia weight , the minimum inertia weight , the constant lambda , the initial step size step is 1, the antenna distance attenuation factor c , the initial value of the attenuation factor eta is 1, and the position range of each bee; training the ABSO-LSTM model on a randomly selected training set to predict the root mean square error of the predicted output creep variable and the actual collected creep variable as the target function of each individual, and optimizing the hyperparameters of the LSTM network by continuously iterating and updating. In step 2), the hyperparameters of the LSTM network model are: the number of neuron nodes of the first layer of LSTM, the number of neuron nodes of the second layer of LSTM, the learning rate of LSTM, and the maximum iteration number of the prediction model.

Images