Displacement control method and device of piezoelectric ceramic actuator, equipment and medium

Abstract

The invention relates to the technical field of computer simulation, in particular to a displacement control method and device of a piezoelectric ceramic actuator, equipment and a storage medium. The method comprises the following steps: acquiring relevant parameters of the piezoelectric ceramic actuator, wherein the relevant parameters at least comprise target displacement; the relevant parameters are input into a pre-training model, target electric field intensity corresponding to the target displacement is obtained, and the pre-training model comprises a Bi-LSTM module; and controlling the piezoelectric ceramic actuator to move based on the target electric field intensity.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer simulation, and more particularly, to a displacement control method, device, equipment and storage medium of a piezoelectric ceramic actuator. BACKGROUND

[0002] As a core component of precision driving and positioning, the precise characterization and control of the performance of a piezoelectric ceramic actuator (PCA) is the key to achieving a high-precision system. However, the accurate and efficient measurement of the displacement behavior of the PCA itself constitutes a major challenge. Due to the hysteresis characteristics of piezoelectric ceramics, when the applied voltage increases, the generated displacement will change along a path; but when the voltage starts to decrease from the maximum value, the displacement does not return along the original path, but along a different, lagging path. This forms a loop-shaped curve, called a "hysteresis loop". This phenomenon makes the displacement output not simply one-to-one corresponding to the input voltage or electric field strength.

[0003] The traditional measurement method of the displacement and electric field strength corresponding relationship usually relies on dense discrete point sampling. This not only requires expensive hardware support and high-frequency data acquisition systems, but also consumes a large amount of time and human resources in curve fitting. Especially when it is necessary to characterize the complete displacement and electric field strength corresponding response curve of the PCA in a complex, dynamically changing multi-physical field environment, such dense sampling requirements become particularly prominent and costly. Moreover, for piezoelectric ceramic actuators that are different in temperature, humidity, driving frequency, or even material composition, their displacement and electric field corresponding relationship are not the same, therefore, the displacement response curve reference model established by the traditional method will also decrease in accuracy during actual operation, and it is difficult to meet the needs of practical applications. SUMMARY

[0004] An object of the present disclosure is to provide a displacement control method for a piezoelectric ceramic actuator to solve the precision problem of the piezoelectric ceramic actuator.

[0005] According to a first aspect of the present disclosure, a displacement control method for a piezoelectric ceramic actuator is provided, comprising: obtaining related parameters of the piezoelectric ceramic actuator, the related parameters at least including a target displacement; inputting the related parameters into a pre-trained model to obtain a target electric field strength corresponding to the target displacement, wherein the pre-trained model comprises a Bi-LSTM module; controlling the piezoelectric ceramic actuator for displacement based on the target electric field strength.

[0006] According to a second aspect of the present disclosure, a displacement control device for a piezoelectric ceramic actuator is provided, comprising: An acquisition module is configured to acquire related parameters of the piezoelectric ceramic actuator, and the related parameters at least include a target displacement; An electric field determination module is configured to input the related parameters into a pre-trained model to obtain a target electric field intensity corresponding to the target displacement, wherein the pre-trained model includes a Bi-LSTM module. A control module is configured to control the piezoelectric ceramic actuator to perform displacement based on the target electric field intensity.

[0007] According to a third aspect of the present disclosure, an electronic device is provided, including a processor and a memory, the memory storing computer instructions, and the computer instructions are executed by the processor to implement the steps of the method of the first aspect.

[0008] According to a fourth aspect of the present disclosure, a storage medium is provided, and the storage medium stores computer instructions, and the computer instructions are executed by a processor to implement the steps of the method of the first aspect.

[0009] One technical effect of the present disclosure is to provide a displacement control method of a piezoelectric ceramic actuator, the core of which is to directly predict the required target electric field intensity according to the target displacement and other parameters by using a pre-trained Bi-LSTM module. The Bi-LSTM module can accurately capture and represent the inherent hysteresis loop phenomenon of the piezoelectric ceramic that causes the displacement output and the input voltage to be non-corresponding by learning and modeling complex time-dependent relationships and nonlinear dynamic characteristics, thereby avoiding the errors caused by relying on simple static mapping or dense sampling in traditional methods. Secondly, the model can learn the complex correspondence between displacement and electric field intensity under various working conditions (such as different temperatures, humidity, driving frequency, and material composition) through one training, which significantly reduces the high hardware cost, high-frequency acquisition system, and long dense sampling time required to obtain complete and accurate response curves in complex multi-physical field environments. Finally, the target electric field intensity predicted based on the model can significantly improve the displacement control accuracy and adaptability of the piezoelectric ceramic actuator in actual dynamic operation, solve the problem that the traditional reference curve is invalid due to changes in working conditions, and meet the needs of high-precision applications.

[0010] Other features and advantages of the embodiments of the present disclosure will become clear from the following detailed description of exemplary embodiments of the present disclosure with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings, which form a part of the specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the principles of the embodiments of the present disclosure.

[0012] Figure 1 is a flowchart of a displacement control method of a piezoelectric ceramic actuator according to an embodiment; Figure 2 is a schematic diagram of state update and transmission of a long short-term memory network neuron of a Bi-LSTM module according to an embodiment; Figure 3 is a structural diagram of a pre-training model according to an embodiment Figure 4 is a comparison diagram of actual data and model predicted data according to an embodiment; Figure 5 is a schematic diagram of a displacement control device of a piezoelectric ceramic actuator according to an embodiment; Figure 6 is a structural diagram of an electronic device according to an embodiment. DETAILED DESCRIPTION

[0013] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. Note that the relative arrangement, numerical expressions, and numerical values of components and steps set forth in these embodiments are not limiting to the scope of the present disclosure unless otherwise specifically stated.

[0014] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way limiting to the scope of the application and its applications or uses.

[0015] Techniques and equipment known to those of ordinary skill in the relevant art can not be discussed in detail, but should be considered part of the specification where appropriate.

[0016] In all examples shown and discussed herein, any specific value should be interpreted as merely illustrative and not as a limitation. Thus, other examples of the exemplary embodiments can have different values.

[0017] Note that like reference numerals and letters indicate similar items in the accompanying drawings, and thus once an item is defined in one drawing, it need not be discussed further in subsequent drawings.

[0018] It should be noted that the collection, storage, use, processing, transmission, provision, disclosure, deletion, etc. of data in the present disclosure are carried out in compliance with the data protection-related regulations and policies of the place of residence and with the full authorization of the corresponding data owner.

[0019] The embodiment of the present application discloses a displacement control method of a piezoelectric ceramic actuator, as shown in Figure 1 comprises steps S11-S13.

[0020] Step S11, obtaining the related parameters of the piezoelectric ceramic actuator, the related parameters at least including the target displacement.

[0021] In the embodiment of the present application, the related parameters of the piezoelectric ceramic actuator can include a target displacement of the piezoelectric ceramic actuator, a material of the piezoelectric ceramic, for example, BZZ, PT, BS components, and a working condition corresponding to the piezoelectric ceramic actuator, for example, temperature, humidity, driving frequency, and the like, and a change trend of displacement or electric field.

[0022] In step S12, the related parameters are input into the pre-trained model to obtain a target electric field intensity corresponding to the target displacement, wherein the pre-trained model includes a Bi-LSTM module.

[0023] In the embodiment, the pre-trained model at least includes a Bi-LSTM (Bidirectional Long Short-Term Memory) module, which is a recurrent neural network structure combining forward / backward time series modeling, used to capture local features and bidirectional global dependencies of a sequence. Specifically, the asymmetric characteristics when the voltage rises or falls can be captured through the two different directions of forward and backward.

[0024] In one example of the embodiment, the pre-trained model includes a position encoding module for injecting position information into a feature element in an input feature sequence of the pre-trained model, and after injecting the position information, inputting the input feature sequence into the Bi-LSTM module. The position information includes The position information is determined based on the following formula:

[0025] wherein, is a dimension index of the encoding vector, is an embedding dimension, and the constant c is a decay rate, is a position of the feature element in the feature sequence.

[0026] In the embodiment, the position encoding module uses different frequency sine or cosine function values to generate absolute position encoding for each position in the input feature sequence, so that the model can perceive the input feature order. The module is used to inject position information into sequence data to solve the problem that the attention mechanism cannot perceive the order. The parameters can be set according to actual needs, wherein the parameter The decay rate for adjusting the frequency can be set to 10000 in one example.

[0027] In the embodiment of the present application, the output of the Bi-LSTM module is The output of the Bi-LSTM module is determined based on the following formula:

[0028] wherein, For the moment the forward state of the hidden layer, For the moment the reverse state of the hidden layer, the the moment the forward state of the hidden layer is determined based on the following formula:

[0029] wherein, For the moment the state of the forget gate of the hidden layer, For the moment the state of the input gate of the hidden layer, For the moment the candidate state of the hidden layer, For the moment the cell state of the hidden layer, For the moment the output gate state of the hidden layer, is a first activation function, is a second activation function, , , , is a preset weight matrix, , , a preset bias term, the forward and reverse states of the 0th hidden layer at the initial moment are the feature elements after the position information is injected, denotes a layer normalization operation; For the moment the reverse state of the hidden layer, the the moment the reverse state of the hidden layer is determined in the same manner as the the moment the forward state of the hidden layer . It can be determined through the reverse state of the forget gate, the input gate and the output gate of each hidden layer, and the reverse candidate state and the cell state.

[0030] In this embodiment, the state update and transmission diagram of the long short-term memory network neurons of the Bi-LSTM module is as shown in Figure 2As shown, in order to solve the problem of capturing long-range dependencies due to gradient disappearance, the state of the hidden layer neuron is updated by using the calculation mode of the forgetting gate, the input gate, the cell state update, and the output gate, and the state transmission is required between the neurons of the same hidden layer at different times and between the neurons of different layers at the same time.

[0031] In one example, the input feature can be regarded as the hidden layer state of the 0th layer, i.e. , wherein represents the input feature element. With the state transmission, the expansion of the sequence length and the deepening of the hidden layer, the final hidden layer will perform the concatenation operation in the direction and the sequence to obtain the final output when propagating to the final hidden layer state.

[0032] In this example, the Bi-LSTM module can accurately capture the asymmetry of the voltage rising / falling path in the hysteresis loop through bidirectional time series modeling, and generalize the complex hysteresis nonlinear dynamics under multiple physical fields (temperature, frequency, material composition) from sparse sampling data through long-term dependency learning, thereby replacing the traditional dense sampling to directly generate a high-precision electric field strength-displacement mapping relationship through model prediction, thereby reducing the hardware cost.

[0033] In one example of the embodiment, the pre-trained model further includes a multi-head attention mechanism module configured to determine a fusion attention output based on the output of the Bi-LSTM module, wherein the fusion attention output is determined based on the following formula:

[0034]

[0035] wherein, is the attention score of the jth attention head, is the number of attention heads, , , are the query matrix, the key matrix and the value matrix of the jth attention head, respectively, is a scaling factor used to stabilize the attention head score in a reasonable range and ensure the effective flow of the gradient during back propagation, is a normalization function.

[0036] In the embodiment of the application, the multi-head attention mechanism captures the dependency relationship of different positions in the sequence through parallel multiple attention heads. Each head learns different representation subspaces, and finally combines the results of each attention head to obtain a more rich semantic expression. In one example, the output of the Bi-LSTM can be linearly changed to obtain the query matrix , the key matrix query matrix, key matrix and value matrix i.e.

[0037]

[0038]

[0039] wherein, , , is a linear transformation matrix, which can be set based on actual conditions. , , are the query matrix, the key matrix and the value matrix of the jth attention head respectively, which can be obtained by dividing the query matrix, the key matrix and the value matrix of the multi-head attention. , , .

[0040] In this example, the multi-head attention mechanism sets multiple independent attention heads working in parallel on the basis of high-order time sequence features output by the Bi-LSTM module, so that the system can extract and strengthen the complex correlation between features from different subspaces, effectively overcoming the limitations of Bi-LSTM in global dependence modeling and dynamic focusing of features. Firstly, the sub-attention matrix inside each attention head directly calculates the correlation weight of the feature vectors at any two time steps in the sequence, breaking through the distance limit of LSTM implicit state transmission, significantly enhancing the reliability of modeling long-range physical field coupling across time steps (such as the influence of historical temperature drift on the current hysteresis path). Secondly, different attention heads independently learn different feature combination interaction patterns (such as head 1 focusing on “voltage-temperature” coupling and head 2 analyzing “frequency-material component” correlation), and through parallel calculation and fusion, the collaborative interaction representation ability of multi-source heterogeneous features (electric field, environment, material) under dynamic working conditions is enhanced. Thirdly, the attention mechanism supports adaptive focusing and noise suppression on key time steps (such as voltage turning points and displacement saturation regions), amplifies features that significantly contribute to displacement prediction through soft selection weights, and improves the prediction accuracy of the model in the path switching and nonlinear sensitive region.

[0041] In one example of the embodiment, the pre-training model further comprises a feature concatenation module, an additive attention mechanism module, and a feature fusion module. The feature concatenation module is configured to concatenate the output of the Bi-LSTM module and the fused attention output of the multi-head attention mechanism module to obtain a dynamic query vector, and input the dynamic query vector into the additive attention mechanism module. The additive attention mechanism module is configured to determine a context vector based on the dynamic query vector and the first key matrix output by the multi-head attention mechanism module. The feature fusion module is configured to fuse the output of the Bi-LSTM module, the context vector of the additive attention mechanism module, and the fused attention output of the multi-head attention mechanism module to obtain an output result of the pre-training model.

[0042] In the embodiment, the architecture of the pre-training model can be as shown in Figure 3 The multi-head attention mechanism can be complementary to the LSTM, supplementing global dependencies by accepting high-order features of the LSTM output, and focusing on the most relevant spatiotemporal feature combinations based on the dynamic query vector, thereby significantly improving the prediction effect of the numerical value. Based on this, the feature concatenation module can combine the last time state of the bidirectional long short-term memory network and the average pooling features of the multi-head attention to obtain the corresponding dynamic query vector wherein the dynamic query vector is determined based on the following formula:

[0043] In one example of the embodiment, the context vector is determined based on the following formula:

[0044]

[0045] wherein, is a linear transformation matrix corresponding to the query matrix, is a linear transformation matrix corresponding to the key matrix, is the dynamic query vector, is the first key matrix, represents the transpose matrix of the linear layer, and is a time step index, is a decompression function, is a compression function, is an exponential function.

[0046] In one example of the embodiment, the first key matrix is determined based on the following formula:

[0047] wherein, is a random dropout function, is a linear projection layer matrix, used to integrate multi-head information, the projection layer matrix The fused features are mapped to a target space, and then connected with the original input in residual connection, and finally normalized.

[0048] In this embodiment, the additive attention mechanism can dynamically analyze the correlation between the current query state (such as the current predicted target displacement) and the key matrix state of each time step in the past (historical voltage, historical displacement, etc.). It can identify which features jointly affect the current state (such as actual displacement) the most under certain conditions (for example, the combination of "high voltage + high temperature" may cause a displacement deviation greater than the sum of each individual effect), more accurately model the multi-physical field coupling effect, and greatly improve the model's ability to capture complex hysteresis loop displacement.

[0049] In this embodiment, the model fusion module realizes collaborative modeling by integrating three heterogeneous features: Bi-LSTM features depict the long-term hysteresis path dependence of the voltage-displacement sequence, multi-head self-attention mechanism analyzes the global multi-physical field (such as temperature / frequency / material) coupling relationship, and additive attention mechanism focuses on the local nonlinear details of key time steps based on dynamic query vectors. Subsequently, a phased dimension reduction strategy is adopted: first, the high-dimensional features are projected to a low-dimensional space through a fully connected layer to learn the dynamic weight combination of the three features; second, the GELU activation function is applied for multi-level nonlinear transformation, and its smooth gradient characteristics enhance the expression ability for complex hysteresis loops; finally, Dropout and random position mask double regularization are introduced to effectively suppress the overfitting risk of multi-source features. This progressive dimension reduction structure gradually filters out noise redundancy, compared with direct dimension reduction, reduces the loss of key information, and significantly improves the robustness of displacement prediction.

[0050] In the embodiment, the pre-trained model can be trained in the following way: first, through feature engineering, on the basis of initial data features, in-depth mining of data is carried out to create more features. It can include: dynamic features, statistical features, mutation features, trend features, lag features. Based on the corresponding relationship between the strain of the material and the displacement, each kind of feature is calculated and analyzed based on the original strain size value and the relationship between the strain size and the electric field intensity. The obtained features are tested by ablation experiment, and the best features are finally selected to participate in the calculation of the subsequent neural network. In this example, dynamic features can include: numerical gradient of strain, energy accumulation of strain, global standardized strain, energy gradient of strain, etc., trend features can include absolute value of trend change, duration of current trend, statistical features can include 25% and 75% quantile of strain in window, skewness of strain in window, moving average of strain in window, standard deviation of strain in window, etc., and then features include 1, 2 and 3 order historical observation values, electric field change rate with strain, electric field acceleration with strain, etc., and mutation features can include Z-score exceeding 3 markers, strain mean in mutation point window, etc.

[0051] In the embodiment, before training the pre-trained model, the method further comprises model hyper-tuning the pre-trained model, selecting a group of hyper-parameter sets into the model framework to instantiate the model on the basis of the existing model framework, then training and verifying the prediction effect of the model, and obtaining the best hyper-parameter set of the model by continuously changing the hyper-parameter selection method. The optimization target of this step is to find the minimum loss function value of the verification data set, which is mathematically expressed as follows, wherein is the best hyper-parameter combination, is the tuning target, represents the loss function of the verification set:

[0052] In one example, a group of hyper-parameter data sets can be first selected to participate in the instantiation of the model, and the instantiated model is trained through the training data to obtain the best verification set loss. The verification set loss will participate in the construction of the tree-like Parson estimator model, divide the performance and calculate the improvement amount EI. The maximum EI value is used to guide the hyper-parameter sampling to determine the specific value of the next group of hyper-parameter tuning, until all the scheduled hyper-tuning rounds are completed. The group of hyper-parameter combinations with the lowest verification set loss in the hyper-tuning process is output to instantiate the final model. The specific formula is as follows:

[0053]

[0054]

[0055] wherein, denotes any set of hyperparameter combinations, denotes the best loss function, denotes the high loss set, denotes the low loss set, denotes the density estimation of high performance group parameters, denotes the density estimation of low performance group parameters.

[0056] After the completion of the model hyper-tuning, based on the final determined hyperparameter combination, the model is parameter optimized and weight updated using the training set data. According to the specific selected hyperparameter combination, the instance displacement prediction model, learning machine and learning rate scheduler are instantiated. During the training process, gradient clipping and dynamic early stopping mechanism are combined to prevent overfitting, and finally the model parameters and standardized configuration at the time of the lowest validation set loss are saved. Then, based on the test set data, the model performance is quantitatively analyzed and visually verified, the black box test set is predicted, the error index, correlation index and robustness index are combined for evaluation, the comparison curve of true value and predicted value is drawn, and the accuracy performance of the model in the single-loop hysteresis trajectory is intuitively displayed, and finally the pre-trained model is obtained.

[0057] In this embodiment, various parameters in each module can be set based on actual conditions.

[0058] In one example, a certain type of piezoelectric ceramic actuator is taken as the research object, 4000 complete hysteresis loop strain-electric field intensity data are collected under the condition of fixed environmental temperature, fixed driving frequency and 10 different material components. The data set is established by random division or fixed interval division, the training set accounts for 0.16, the validation set accounts for 0.04, and the test set accounts for 0.80. Through wavelet denoising and binning standardization processing, noise interference is eliminated, dynamic features (strain rate, energy gradient), statistical features (quantile, skewness) and hysteresis features (historical strain value) are extracted to construct a composite feature space. After hyperparameter tuning, the unique prediction model adopts 4-layer Bi-LSTM (hidden layer dimension 256), 4-head attention mechanism and phased dimension reduction fusion layer, and the optimal hyperparameter combination (learning rate 3e-4, Batch_Size=32, gradient clipping threshold 3.3) is determined by Bayesian optimization. The AdamW learning machine is used in the training process, and the SGDR (Stochastic Gradient Descent with Restarts, Stochastic Gradient Descent with Restarts) scheduling strategy is adopted. After 8000 iterations, the validation set Huber loss converges to 0.0000565, the best model parameters are saved and used for prediction of the test set, and the results of the piezoelectric ceramic actuator with 10 different components are as follows Figure 4The evaluation of the test set shows that the model R² (goodness of fit) reaches 0.9999, NRMSE (normalized root mean square error) = 0.0025, SMAPE (symmetric mean absolute percentage error) = 1.18%, and the prediction ability of R²> 0.995 is still maintained in the cross-condition test of unknown material components. In engineering applications, the prediction results are used as feedforward signals combined with PID (Proportional-Integral-Derivative) controllers, and the closed-loop tracking error is reduced by 92% compared with using only PID. It still maintains high precision operation under extreme conditions of temperature fluctuation ± 5℃ and humidity change ± 10%, meeting the real-time needs of micro-nano positioning systems.

[0059] Step S13, controlling the piezoelectric ceramic actuator to displace based on the target electric field intensity.

[0060] After obtaining the target electric field intensity through the pre-trained model, the target electric field intensity is used as a control reference to drive the piezoelectric ceramic actuator to apply the corresponding voltage waveform, and the actuator is controlled to move the target displacement distance.

[0061] In this example, a displacement control method for a piezoelectric ceramic actuator is provided, which is based on predicting the required target electric field intensity directly from parameters such as target displacement using a pre-trained Bi-LSTM module. The Bi-LSTM module can accurately capture and represent the inherent hysteresis loop phenomenon of piezoelectric ceramics, which causes the displacement output to be non-corresponding to the input voltage (including the difference between the voltage rising path and the falling path), by learning and modeling complex time-dependent relationships and nonlinear dynamic characteristics, thereby avoiding the errors caused by relying on simple static mapping or dense sampling in traditional methods. Secondly, this model can learn the complex correspondence between displacement and electric field intensity under various conditions (such as different temperatures, humidity, driving frequency, material components) through one training, significantly reducing the high hardware cost and lengthy dense sampling time required to obtain complete and accurate response curves in complex multi-physical field environments. Finally, based on the target electric field intensity predicted by the model for control, the displacement control accuracy and adaptability of the piezoelectric ceramic actuator in actual dynamic operation can be significantly improved, solving the problem of invalidation of traditional reference curves due to changes in working conditions, and meeting the needs of high-precision applications.

[0062] As Figure 5As shown, an embodiment of the present application also provides a displacement control device 100 for a piezoelectric ceramic actuator, including: an acquisition module 101, used to obtain relevant parameters of the piezoelectric ceramic actuator, the relevant parameters including at least a target displacement; an electric field determination module 102, used to input the parameters into a pre-trained model to obtain a target electric field strength corresponding to the target displacement, wherein the pre-trained model includes a Bi-LSTM module; and a control module 103, used to control the displacement of the piezoelectric ceramic actuator based on the target electric field strength.

[0063] Optionally, the pre-trained model includes: a position encoding module, configured to inject position information into feature elements in an input feature sequence of the pre-trained model, and input the input feature sequence into the Bi-LSTM module after injecting the position information; Among them, location information Determined based on the following formula:

[0064] in, is the dimension index of the encoding vector, is the embedding dimension, the constant c is the decay rate, is the position of the feature element in the feature sequence.

[0065] Optionally, the output of the Bi-LSTM module is determined based on the following formula:

[0066] in, for Moment The forward state of the hidden layer, for Moment The reverse state of the hidden layer, Moment Forward state of the hidden layer Determined based on the following formula:

[0067] in, for Moment The state of the hidden layer forget gate, for Moment The state of the hidden layer input gate, for Moment hidden layer candidate states, for Moment Hidden layer cell states, is the hidden layer output gate state, is a first activation function, is a second activation function, , , , is a preset weight matrix, , , a preset bias term, the forward and reverse states of the 0th hidden layer at the initial moment are the feature elements after the position information is injected, denotes a layer normalization operation; is the reverse state of the hidden layer at the moment, the reverse state of the hidden layer at the moment is determined in the same way as the forward state of the hidden layer at the moment .

[0068] Optionally, the pre-training model further comprises a multi-head attention mechanism module, the multi-head attention mechanism module determines a fusion attention output based on the output of the Bi-LSTM module, and the fusion attention output is determined based on the following formula:

[0069]

[0070] wherein, is an attention score of the jth attention head, is the number of attention heads, , , are respectively a query matrix, a key matrix and a value matrix of the jth attention head, is a scaling factor, is a normalization function.

[0071] Optionally, the pre-training model further comprises a feature splicing module, an additive attention mechanism module and a feature fusion module, the feature splicing module is used for splicing the output of the Bi-LSTM module and the fusion attention output of the multi-head attention mechanism module to obtain a dynamic query vector, and inputting the dynamic query vector into the additive attention mechanism module; ​The additive attention mechanism module determines a context vector based on the dynamic query vector and a first key matrix output by the multi-head attention mechanism module; The feature fusion module is configured to fuse the output of the Bi-LSTM module, the context vector of the additive attention mechanism module, and the fused attention output of the multi-head attention mechanism module to obtain an output result of the pre-training model.

[0072] Optionally, the context vector is determined based on the following formula:

[0073]

[0074] wherein, is a linear transformation matrix corresponding to the query matrix, is a linear transformation matrix corresponding to the key matrix, is the dynamic query vector, is the first key matrix, denotes a transpose matrix of the linear layer, and is a time step index, is a decompression function, is a compression function, is an exponential function.

[0075] Optionally, the first key matrix is determined based on the following formula:

[0076] wherein, is a random dropout function, is a linear projection layer matrix.

[0077] As Figure 6 shown, the embodiments of the present application also provide an electronic device 200, comprising a processor 201 and a memory 202, the memory 202 stores computer instructions, and the computer instructions are executed by the processor 201 to implement the steps of any one of the piezoelectric ceramic actuator displacement control method embodiments.

[0078] The embodiments of the present application also provide a storage medium having computer instructions stored thereon, the computer instructions are executed by a processor to implement any one of the piezoelectric ceramic actuator displacement control embodiments described above and achieve the same technical effects. To avoid repetition, details are not repeated here.

[0079] The various embodiments in this disclosure are described in a progressive manner, and the same or similar parts among the various embodiments can be referred to each other, and each embodiment focuses on the difference from other embodiments. Especially, for the device and equipment embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiments.

[0080] The above describes specific embodiments of the present disclosure. In some cases, the actions or steps described can be performed in an order different than the order described in the embodiments and still achieve the desired result. Also, the processes depicted in the figures do not necessarily require the particular order shown, or sequential order, to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous.

[0081] Embodiments of the present disclosure can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of embodiments of the present disclosure.

[0082] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or punched tape, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0083] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network can comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0084] Computer readable program instructions for carrying out embodiments of the disclosure can be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computing / processing device, partly on the user's computing / processing device, as a stand-alone software package, partly on the user's computing / processing device and partly on a remote computing / processing device or entirely on the remote computing / processing device or server. In the latter scenario, the remote computing / processing device can be connected to the user's computing / processing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing / processing device, for example, through the Internet using an Internet Service Provider. In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate array (FPGA), or programmable logic array (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of embodiments of the present disclosure.

[0085] Aspects of the embodiments of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.

[0086] These computer readable program instructions can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. These computer readable program instructions can also be stored in a computer readable storage medium that can include a non-transitory computer readable storage medium that can be a computer- readable storage medium having no data, programs, program modules, e.g., instructions for operation, or digital content stored thereon or therein for a short time or not at all. The computer readable storage medium can also have instructions stored thereon or therein which may

[0087] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0088] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0089] Having described above several embodiments of the disclosure, any modifications and variations that fall within the scope of the described embodiments are also intended to be within the scope of the disclosure. As will be apparent to those skilled in the art, some modifications and variations to the embodiments described herein can be practiced while staying within the scope of the claimed subject matter. The present disclosure is to be considered as illustrative and not restrictive, and the scope of these embodiments are therefore intended to be, and will be, set to the breadth and scope of any claims that can be filed claiming priority to this disclosure. The claims as filed and permitted to be amended, which can be presented with or after the filing of the application as a continuation, continuation-in-part, divisional, or terminal claim depending on the provisions of 35 U.S.C. § 120, are intended to define and protect the scope of these embodiments.

Claims

1. A displacement control method for a piezoelectric ceramic actuator, characterized in that: include: Acquiring relevant parameters of the piezoelectric ceramic actuator, wherein the relevant parameters at least include a target displacement; Inputting the relevant parameters into a pre-training model to obtain a target electric field intensity corresponding to the target displacement, wherein the pre-training model includes a Bi-LSTM module; The piezoelectric ceramic actuator is controlled to move based on the target electric field strength.

2. The method according to claim 1, characterized in that The pre-training model includes: a position encoding module, which is used to inject position information into the feature elements in the input feature sequence of the pre-training model, and after injecting the position information, input the input feature sequence into the Bi-LSTM module; Among them, location information Determined based on the following formula: in, is the dimension index of the encoding vector, is the embedding dimension, the constant c is the decay rate, is the position of the feature element in the feature sequence.

3. The method according to claim 2, characterized in that The output of the Bi-LSTM module Determined based on the following formula: in, for Moment The forward state of the hidden layer, for Moment The reverse state of the hidden layer, Moment Forward state of the hidden layer Determined based on the following formula: in, for Moment The state of the hidden layer forget gate, for Moment The state of the hidden layer input gate, for Moment hidden layer candidate states, for Moment Hidden layer cell states, for Moment The hidden layer output gate state, is the first activation function, is the second activation function, 、 、 、 is the preset weight matrix, 、 、 The bias term is preset, and the forward and reverse states of the 0th hidden layer at the initial moment are the characteristic elements after the position information is injected. Representation layer normalization operation; for Moment The reverse state of the hidden layer, Moment The reverse state of the hidden layer With the Moment Forward state of the hidden layer is determined in the same way.

4. The method according to claim 3, characterized in that The pre-training model also includes a multi-head attention mechanism module, which determines a fused attention output based on the output of the Bi-LSTM module. Determined based on the following formula: in, is the attention score of the j-th attention head, is the number of attention heads, 、 、 are the query matrix, key matrix, and value matrix of the j-th attention head, respectively. is the scaling factor, is the normalization function.

5. The method according to claim 4, characterized in that The pre-training model also includes a feature splicing module, an additive attention mechanism module and a feature fusion module. The feature splicing module is used to splice the output of the Bi-LSTM module and the fused attention output of the multi-head attention mechanism module to obtain a dynamic query vector, and input the dynamic query vector into the additive attention mechanism module; The additive attention mechanism module determines a context vector based on the dynamic query vector and a first key matrix output by the multi-head attention mechanism module; The feature fusion module is used to fuse the output of the Bi-LSTM module, the context vector of the additive attention mechanism module, and the fused attention output of the multi-head attention mechanism module to obtain the output result of the pre-trained model.

6. The method according to claim 5, characterized in that The context vector Determined based on the following formula: in, is the linear transformation matrix corresponding to the query matrix, is the linear transformation matrix corresponding to the key matrix, is the dynamic query vector, is the first bond matrix, represents the transposed matrix of the linear layer, and is the time step index, is the inverse compression function, is the compression function, is an exponential function.

7. The method according to claim 6, characterized in that The first bond matrix is ​​determined based on the following formula: in, is the random dropout function, Linear projection layer matrix.

8. A displacement control device for a piezoelectric ceramic actuator, characterized in that: include: an acquisition module, configured to acquire relevant parameters of the piezoelectric ceramic actuator, wherein the relevant parameters include at least a target displacement; an electric field determination module, configured to input the relevant parameters into a pre-trained model to obtain a target electric field strength corresponding to the target displacement, wherein the pre-trained model includes a Bi-LSTM module; A control module is configured to control the piezoelectric ceramic actuator to move based on the target electric field strength.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer instructions, and when the computer instructions are executed by the processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A storage medium, characterized in that: Computer instructions are stored thereon, and when the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images