A Nonlinear Compensation Method for Electromagnetic Actuators Based on Deep Learning

Abstract

The present invention relates to a non-linear compensation method for electromagnetic actuators based on deep learning, comprising: constructing a non-contact one-dimensional electromagnetic actuator data acquisition system based on Lorentz force, and collecting the output acceleration of the electromagnetic actuator at each position under a given voltage through an acceleration sensor and a displacement sensor; using the collected output acceleration data to obtain the compensation voltage at each position under the target acceleration, and forming an electromagnetic actuator position-compensation voltage data set; performing normalization and sliding time window method processing on the position-compensation voltage data; dividing the processed data into a training set and a test set; determining the initial parameters of a deep learning compensation model mainly composed of CNN and BiLSTM; training and testing the deep learning compensation model; and using the trained deep learning compensation model to perform non-linear compensation on the electromagnetic actuator. This method can not only improve the non-linear compensation effect of the electromagnetic actuator, but also simplify the modeling process, reduce the calculation amount and time cost.

Description

Technical Field

[0001] The present invention relates to the field of electromagnetic actuators, and particularly to a non-linear compensation method for electromagnetic actuators based on deep learning. Background Art

[0002] The non-linearity existing in electromagnetic actuators hinders the performance improvement of active vibration isolation systems. In the actual application environment of electromagnetic actuators, external factors such as umbilical cords will bring additional non-linear effects. Moreover, the accuracy and generalization of traditional models are also relatively low, resulting in limited accuracy of non-linear compensation for electromagnetic actuators. Currently, the main non-linear compensation methods for electromagnetic actuators are as follows:

[0003] (1) Model-based method: This method requires establishing an accurate mathematical model of the electromagnetic actuator, and then through analyzing and calculating the model, a non-linear compensation controller is obtained. The advantage of this method is that accurate non-linear compensation effects can be obtained, but it has poor effects on systems where it is difficult to establish an accurate physical model, and it requires consuming more computing resources, and it is difficult to handle practical problems such as system parameter changes and system dynamic responses.

[0004] (2) Data-based method: This method collects the operation data of the electromagnetic actuator, and uses technologies such as machine learning or deep learning to train a non-linear compensation model. This method does not require accurate physical modeling of the system, reducing the errors and uncertainties of the model. The actual operation data collected by it contains the actual non-linear characteristics of the system, which can more accurately reflect the actual operation of the system and has stronger adaptability to the system. Since the complex calculation and modeling processes are abandoned, the amount of calculation and time cost can be reduced, but there are certain requirements for the accuracy and quantity of data sources.

[0005] (3) Hybrid method: This method combines the model-based and data-based methods, and improves the robustness and performance of the non-linear compensation controller through the use of prior knowledge and data-driven methods. Compared with a single method, this method can obtain better compensation effects and performance, but its implementation process is more complex, and it also faces great challenges for systems where it is difficult to accurately establish a physical model, and more research and experimental verification are required. Summary of the Invention

[0006] The purpose of the present invention is to provide a non-linear compensation method for electromagnetic actuators based on deep learning, which can not only improve the non-linear compensation effect of electromagnetic actuators, but also simplify the modeling process and reduce the amount of calculation and time cost.

[0007] To achieve the above purpose, the technical solution adopted by the present invention is: a non-linear compensation method for electromagnetic actuators based on deep learning, including the following steps:

[0008] Step S1: Construct a non-contact one-dimensional electromagnetic actuator data acquisition system based on the Lorentz force, and collect the output accelerations at various positions of the electromagnetic actuator under a given fixed voltage through an acceleration sensor and a displacement sensor;

[0009] Step S2: Utilize the output acceleration data at various positions of the electromagnetic actuator under the fixed voltage collected in Step S1 to obtain the compensation voltages required at various positions under the target acceleration, and form a position-compensation voltage dataset of the electromagnetic actuator;

[0010] Step S3: Normalize the position-compensation voltage data in the position-compensation voltage dataset obtained in Step S2, and process it using the sliding time window method to form the input data of the deep learning compensation model;

[0011] Step S4: Divide the data processed in Step S3 into a training set and a test set according to a set ratio;

[0012] Step S5: Determine the initial parameters of the deep learning compensation model mainly composed of a convolutional neural network and a bidirectional long short-term memory neural network;

[0013] Step S6: Train the deep learning compensation model through the training set, test it through the test set, and then save the trained deep learning compensation model;

[0014] Step S7: Given the target output acceleration of the electromagnetic actuator, perform non-linear compensation on the electromagnetic actuator using the trained deep learning compensation model.

[0015] Furthermore, the deep learning compensation model mainly uses two channels to mine the features of displacement data with time series characteristics. The inputs of both channels are the data reconstructed from the original data according to a pre-determined time step and the sliding time window method. Among them, the BiLSTM channel uses the bidirectional long short-term memory neural network BiLSTM to fully extract the bidirectional global time features, and the CNN channel uses the one-dimensional convolutional neural network 1DCNN to extract the local non-correlated features of the time series signal along the positive and negative directions of the time axis to obtain more time features. Both channels assign weights to each time step through the attention mechanism; then fuse the data features of the CNN channel and the BiLSTM channel to achieve feature merging; finally, perform regression layer prediction through the fully connected layer to obtain the predicted compensation voltages at different positions.

[0016] Furthermore, in the CNN channel, the 1DCNN network extracts features from the input one-dimensional signal. The process is to perform local convolution operations on the input signal using a convolution kernel at a set step size, sequentially traverse and extract the local features of the input data, and then use an activation function to perform category mapping on the features to generate a one-dimensional feature map. The expression of the one-dimensional convolution operation is as follows:

[0017] c i = σ(∑(L i * w i + b i ))

[0018] Among them, c i is the output feature map after convolution, L i is the input local data, w i is the weight of the convolution kernel, b i is the weight bias value, and σ is the activation function;

[0019] Then, the one-dimensional pooling operation is used to reduce the feature map size and the model computation amount, reduce the model complexity, increase the receptive field, and extract the main features after the convolution operation; the calculation formula of the one-dimensional pooling operation is as follows:

[0020]

[0021] Among them, is the m-th eigenvalue in the i-th group divided in the l-th layer; f is the maximum value function or the mean value function. If f is the maximum value function, then y l is the feature map obtained by the max pooling operation. If f is the mean value function, then y l is the feature map obtained by the average pooling operation.

[0022] Furthermore, the BiLSTM network includes a forward LSTM layer and a backward LSTM layer, which horizontally process the forward and backward LSTM hidden vectors at each time step in parallel, and vertically process them unidirectionally from the input layer to the hidden layer and then to the output layer to obtain more data features through the BiLSTM network; BiLSTM concatenates the forward and backward outputs of the LSTM unit at time t, and the update process is as follows:

[0023]

[0024] Among them, are respectively the forward and backward outputs of the LSTM unit at time t, h t is the output of the BiLSTM network, and R n is the n-dimensional vector set.

[0025] Furthermore, the deep learning compensation model introduces a soft attention mechanism into two channels to identify the importance of historical data segments at different times, so as to reasonably allocate weights to each time step; the implementation process of the soft attention mechanism includes the following steps:

[0026] (1) Calculate the attention scoring function s(X i, q), where X is the input vector and q is the query vector, and the relevance between each input vector and the query vector is calculated using a scoring function;

[0027] (2) Calculate the attention distribution α i , and use the attention vector z ∈ [1, N] to represent the index position of the selected information, where z = i represents the i-th input information;

[0028]

[0029] (3) Calculate the weighted average of the input data according to the attention distribution to obtain the attention value;

[0030]

[0031] Furthermore, the calculation methods of the scoring function include: additive model, dot product model, scaled dot product model, and bilinear model;

[0032] Additive model: s(X i , q) = v T tanh(WX i + Uq)

[0033] Dot product model:

[0034] Scaled dot product model:

[0035] Bilinear model:

[0036] Among them, W, U, and v are learnable network parameters, and d is the dimension of the input data.

[0037] Furthermore, the activation function of the deep learning compensation model selects the Relu activation function, the model compilation optimizer selects Adam, and the number of iterations is 200 times.

[0038] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method for compensating the nonlinearity of an electromagnetic actuator that does not require complex calculations and modeling processes, can reduce the amount of calculation and time cost. This method does not require accurate analytical modeling of the physical model of the electromagnetic actuator, greatly simplifies the modeling process, and reduces the amount of calculation and time cost; compared with common deep learning compensation models, this method has better prediction accuracy and can achieve better nonlinear compensation effects for electromagnetic actuators. The method of the present invention is simple to operate, has high accuracy, is easy to implement, and has strong practicability and broad application prospects. Description of the Drawings

[0039] Figure 1 It is a flowchart of the method of the embodiment of the present invention.

[0040] Figure 2 This is the structural diagram of the deep learning compensation model in the embodiments of the present invention.

[0041] Figure 3 This is the displacement-acceleration curve of the electromagnetic actuator within ±20 mm in the embodiments of the present invention.

[0042] Figure 4 This is the compensation voltage curve of the electromagnetic actuator within ±20 mm with the target acceleration a = 0.6 m / s in the embodiments of the present invention 2 in the following.

[0043] Figure 5 This is the compensation voltage prediction curve of each model under the same validation set in the embodiments of the present invention.

[0044] Figure 6 This is the compensation error of each model with the target acceleration a = 0.6 m / s in the compensation experiment in the embodiments of the present invention 2 in the following.

[0045] Figure 7 This is the comparison of the average error before and after compensation using the proposed method with the target acceleration a = 0.1 - 0.6 m / s in the compensation experiment in the embodiments of the present invention 2 in the following. Detailed implementation manners

[0046] The following further describes the present invention in conjunction with the accompanying drawings and embodiments.

[0047] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present application belongs.

[0048] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0049] As Figure 1 shown, this embodiment provides a non-linear compensation method for an electromagnetic actuator based on deep learning, including the following steps:

[0050] Step S1: Construct a non-contact one-dimensional electromagnetic actuator data acquisition system based on the Lorentz force, and collect the output acceleration of the electromagnetic actuator at each position under a given fixed voltage through an acceleration sensor and a displacement sensor.

[0051] Step S2: The acceleration output by the electromagnetic actuator is: a = BIL / m, where B is the magnetic field strength, I is the current of the actuator excitation coil, L is the length of the excitation coil, and m is the mass of the excitation coil and its attached devices. When the excitation coil is at the same position under the magnetic field, the magnetic field strength can be regarded as a constant value at this time, that is, the acceleration output by the electromagnetic actuator can be regarded as only related to the input current. Therefore, by giving the actuator corresponding compensation current when the excitation coil is at different positions, the non-linear compensation of the acceleration output by the electromagnetic actuator can be realized. The compensation current is converted from the compensation voltage by the drive board, and the conversion relationship is ±10V to ±3A. Therefore, it can be regarded that the acceleration output by the electromagnetic actuator is related to the compensation voltage. Using the output acceleration data of the electromagnetic actuator at each position under the fixed input voltage collected in Step S1, the compensation voltage required at each position under the target acceleration is obtained, and a position-compensation voltage data set of the electromagnetic actuator is formed.

[0052] Step S3: Normalize the position-compensation voltage data in the position-compensation voltage data set obtained in Step S2, and process it using the sliding time window method to form the input data of the deep learning compensation model.

[0053] Step S4: Divide the data processed in Step S3 into a training set and a test set according to a set ratio.

[0054] Step S5: Determine the initial parameters of the deep learning compensation model mainly composed of a convolutional neural network and a bidirectional long short-term memory neural network.

[0055] Step S6: Train the deep learning compensation model with the training set, test it with the test set, and then save the trained deep learning compensation model.

[0056] Step S7: Given the target output acceleration of the electromagnetic actuator, use the trained deep learning compensation model to perform non-linear compensation on the electromagnetic actuator.

[0057] In this embodiment, the structure of the deep learning compensation model is as Figure 2As shown. The deep learning compensation model mainly uses two channels to mine features of displacement data with time series features. The input of both channels is the data reconstructed from the original data according to a pre-determined time step and the sliding time window method. Among them, the BiLSTM channel uses the bidirectional long short-term memory neural network BiLSTM to fully extract bidirectional global time features, and the CNN channel uses the one-dimensional convolutional neural network 1DCNN to extract local uncorrelated features of the time series signal along the positive and negative directions of the time axis to obtain more time features. Both channels assign weights to each time step through the attention mechanism. Then, the data features of the CNN channel and the BiLSTM channel are fused to achieve feature merging. Finally, regression layer prediction is performed through the fully connected layer to obtain the predicted compensation voltage at different positions.

[0058] The CNN channel and the BiLSTM channel are the core parts of the entire model, consisting of the 1DCNN and BiLSTM networks, and the network input is the time series matrix.

[0059] In the CNN channel, the 1DCNN network extracts features from the input one-dimensional signal. The process is the same as that of the two-dimensional convolutional layer, which is to perform local convolution operations on the input signal using the convolution kernel at the set step size, sequentially traverse and extract the local features of the input data, and then use the activation function to perform class mapping on the features to generate a one-dimensional feature map. The features extracted by the lower-level convolution operations are more basic textures and other features, and the features extracted by the deep-layer feature extraction are more abstract and representative. However, unlike two-dimensional convolution, one-dimensional convolution can only traverse the input features in one direction. The expression of the one-dimensional convolution operation is as follows:

[0060] c i =σ(∑(L i *w i +b i ))

[0061] Among them, c i is the output feature map after convolution, L i is the input local data, w i is the weight value of the convolution kernel, b i is the weight bias value, and σ is the activation function.

[0062] Then, through one-dimensional pooling operation, the size of the feature map and the computational amount of the model are reduced, the model complexity is decreased, the receptive field is increased, and the main features after convolution operation are extracted. One-dimensional pooling operation can only perform window sliding in one direction of the input data. The commonly used one-dimensional pooling methods can be divided into two types: average pooling and max pooling. Average pooling calculates the mean value of the divided feature map data blocks and uses the obtained mean value to represent the feature block. Max pooling, on the other hand, obtains the maximum value of the divided feature map data blocks to represent the feature block. Pooling operation does not increase the training parameters of the network, can effectively reduce the overfitting phenomenon of the model, and can accelerate the training process of the network. The calculation formula of one-dimensional pooling operation is as follows:

[0063]

[0064] Among them, is the m-th eigenvalue in the i-th group divided in the l-th layer; f is the maximum value function or the mean value function. If f is the maximum value function, then y l is the feature map obtained by max pooling operation. If f is the mean value function, then y l is the feature map obtained by average pooling operation.

[0065] The LSTM network uses three control gates to control the flow of historical information, thus having the function of utilizing long-term historical data and can well solve the long-term dependence problem of RNN. The BiLSTM network includes a forward LSTM layer and a backward LSTM layer, parallelly processes the forward and backward LSTM hidden vectors at each time step in the horizontal direction, and unidirectionally goes from the input layer to the hidden layer and then to the output layer in the vertical direction to obtain more data features through the BiLSTM network. BiLSTM concatenates the forward and backward outputs of the LSTM unit at time t, and the update process is as follows:

[0066]

[0067] Among them, are the forward and backward outputs of the LSTM unit at time t respectively, h t is the output of the BiLSTM network, and R n is the n-dimensional vector set.

[0068] This method improves the CNN and BiLSTM channels, introduces the soft attention mechanism into the two channels in the deep learning compensation model, discriminates the importance degree of historical data segments at different times, and reasonably assigns weights to each time step. The implementation process of the soft attention mechanism includes the following steps:

[0069] (1) Calculate the attention scoring function s(X i, q), where X is the input vector and q is the query vector, and the correlation between each input vector and the query vector is calculated using a scoring function. The calculation methods of the scoring function include: additive model, dot product model, scaled dot product model, and bilinear model;

[0070] Additive model: s(X i , q) = v T tanh(WX i + Uq)

[0071] Dot product model:

[0072] Scaled dot product model:

[0073] Bilinear model:

[0074] Among them, W, U, and v are learnable network parameters, and d is the dimension of the input data.

[0075] (2) Calculate the attention distribution α i , and use the attention vector z ∈ [1, N] to represent the index position of the selected information, where z = i represents the i-th input information;

[0076]

[0077] (3) Calculate the weighted average of the input data according to the attention distribution to obtain the attention value;

[0078]

[0079] To verify the effectiveness of this method, a non-contact one-dimensional electromagnetic actuator data acquisition system was built for testing. A constant current of 1.2 A was applied to the excitation coil, and 577 groups of position-acceleration data were collected within the range of ±20 mm along the direction parallel to the working air gap. The displacement-acceleration curve is as Figure 3 shown.

[0080] Since the output acceleration of the electromagnetic actuator and the input voltage are linearly related at the same position, the compensation voltage required for each position under the target acceleration can be obtained from the displacement-acceleration data, where the target acceleration is 0.6 m / s 2 The compensation voltage values for each position under are as Figure 4 shown. The original data set of position-compensation voltage at an acceleration of 0.6 m / s 2 was processed into a data set in the input form required by the model using the sliding time window method after being normalized by the maximum and minimum values.

[0081] Before the model is trained, parameters need to be set in advance. Select some parameter combinations for training, and use the parameter combination with the optimal result as the initial parameters of the model. In this embodiment, the activation function selects the Relu activation function, the model compilation optimizer selects Adam, and the number of iterations is 200 times.

[0082] The following takes a specific embodiment to further illustrate the present invention.

[0083] In this embodiment, the non-linear compensation method for electromagnetic actuators based on deep learning includes the following steps:

[0084] Step S1: Build a non-contact one-dimensional electromagnetic actuator data acquisition system based on Lorentz force, and collect position-acceleration data within the range of ±20 mm along the direction parallel to the working air gap of the excitation coil.

[0085] Step S2: Use the output acceleration data of the electromagnetic actuator at each position under the collected fixed voltage to process the position-acceleration data into position-compensation voltage data, obtain the compensation voltage required at each position under the target acceleration, and form an electromagnetic actuator position-compensation voltage data set.

[0086] Step S3: Perform maximum-minimum normalization processing on the position-compensation voltage data. The normalization formula can be expressed as:

[0087]

[0088] Among them, x' is the normalized data, and x is the data before normalization.

[0089] Use the sliding time window method to further process the normalized data into model inputs, where the window size is 10, that is, use the position data of the previous 10 moments to predict the compensation voltage of the next moment.

[0090] Step S4: Divide the data processed in step S3 into a training set and a test set according to a ratio of 7:3.

[0091] Step S5: Determine the initial parameters of the one-dimensional convolutional neural network. Before the network is trained, parameters need to be set in advance. The main parameter information of the model is shown in Table 1. In addition, the activation function selects the Relu activation function, the model compilation optimizer selects Adam, and the number of iterations is 200 times.

[0092] Table 1 Main parameter information of the model

[0093]

[0094]

[0095] Step S6: Train the deep learning compensation model using the training set, test it using the test set, and then save the trained deep learning compensation model.

[0096] Step S7: Conduct an electromagnetic actuator nonlinear compensation experiment using the trained deep learning compensation model.

[0097] To verify the superiority of the established model, the model established by this method is compared with common nonlinear compensation deep learning models for compensation voltage prediction, including: The CNN model is a convolutional neural network model with only convolutional channels; the BiLSTM model is a recurrent neural network with only bidirectional LSTM units; the CNN-BiLSTM model is a two-channel neural network without an attention mechanism; the above three models are all trained using the same learning rate and the Adam random optimization algorithm, and the training dataset settings and various parameter settings are the same as those of the method model proposed in this paper, and all are made optimal during the training process. The compensation prediction curves of each model under the same test set are as Figure 5 shown. It can be observed from the figure that the model established by this method has the highest accuracy. And the model established by this method is used in the electromagnetic actuator nonlinear compensation experiment, and the compensation errors of each model under the target acceleration a = 0.6m / s 2 are as Figure 6 shown, and the average compensation errors before and after compensation of this method under the target acceleration of a = 0.1 - 0.6m / s 2 are as Figure 7 shown. From Figure 6 and Figure 7 it can be seen that the overall error of this method in the ±20mm section is the lowest among all models, and the average error is reduced from nearly 15% before compensation to about 3%, and the compensation effect is significant. Thus, it is proved that the method of the present invention is effective for the nonlinear compensation of electromagnetic actuators.

[0098] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0099] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce means for implementing the functions specified in one or more flows and / or blocks in the flowchart. Figure 1 in one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0100] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means for implementing the functions specified in one or more flows and / or blocks Figure 1 in one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0101] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operational steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more flows and / or blocks Figure 1 in one or more flows and / or blocks Figure 1 or means for implementing the functions specified in one or more blocks.

[0102] As described above, the above are only preferred embodiments of the present invention, and are not intended to limit the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution content of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A non - linear compensation method for electromagnetic actuators based on deep learning, characterized in that, it includes the following steps: Step S1: Construct a non - contact one - dimensional electromagnetic actuator data acquisition system based on Lorentz force, and collect the output acceleration of the electromagnetic actuator at each position under a given fixed voltage through an acceleration sensor and a displacement sensor; Step S2: Use the output acceleration data of the electromagnetic actuator at each position under the fixed voltage collected in Step S1 to obtain the compensation voltage required at each position under the target acceleration, and form an electromagnetic actuator position - compensation voltage data set; Step S3: Normalize the position - compensation voltage data in the position - compensation voltage data set obtained in Step S2, and process it using the sliding time window method to form the input data of the deep learning compensation model; Step S4: Divide the data processed in Step S3 into a training set and a test set according to a set ratio; Step S5: Determine the initial parameters of the deep learning compensation model mainly composed of a convolutional neural network and a bidirectional long - short - term memory neural network; Step S6: Train the deep learning compensation model through the training set, test it through the test set, and then save the trained deep learning compensation model; Step S7: Given the target output acceleration of the electromagnetic actuator, use the trained deep learning compensation model to perform non - linear compensation on the electromagnetic actuator; The deep learning compensation model mainly uses two channels to mine the features of displacement data with time - series characteristics. The inputs of the two channels are both the data reconstructed from the original data according to a pre - determined time step and the sliding time window method. Among them, the BiLSTM channel uses the bidirectional long - short - term memory neural network BiLSTM to fully extract the bidirectional global time features, and the CNN channel uses the one - dimensional convolutional neural network 1DCNN to extract the local non - correlated features of the time - series signal along the positive and negative directions of the time axis to obtain more time features; both channels assign weights to each time step through the attention mechanism; then the data features of the CNN channel and the BiLSTM channel are fused to achieve feature merging; finally, regression layer prediction is performed through the fully - connected layer to obtain the predicted compensation voltage at different positions; In the CNN channel, the 1DCNN network extracts features from the input one - dimensional signal. The process is to perform local convolution operations on the input signal using the convolution kernel at a set step size, sequentially traverse and extract the local features of the input data, and then use the activation function to perform category mapping on the features to generate a one - dimensional feature map; the expression of the one - dimensional convolution operation is as follows: c i = σ(Σ(L i * w i + b i )) Among them, c i is the output feature map after convolution, L i is the input local data, w i is the weight of the convolution kernel, b i is the weight bias value, and σ is the activation function; Then, the one - dimensional pooling operation is used to reduce the size of the feature map and the model operation amount, reduce the model complexity, increase the receptive field, and extract the main features after the convolution operation; the calculation formula of the one - dimensional pooling operation is as follows: Among them, is the m-th eigenvalue in the i-th group divided in the l-th layer; f is the maximum function or the mean function. If f is the maximum function, then y l is the feature map obtained by the maximum pooling operation. If f is the mean function, then y l is the feature map obtained by the mean pooling operation; The BiLSTM network includes a forward LSTM layer and a backward LSTM layer, which horizontally process the forward and backward LSTM hidden vectors at each time step in parallel, and vertically go unidirectionally from the input layer to the hidden layer and then to the output layer to obtain more data features through the BiLSTM network; BiLSTM concatenates the forward and backward outputs of the LSTM cell at time t, and the update process is as follows: Among them, are the forward and backward outputs of the LSTM cell at time t, h t is the output of the BiLSTM network, R n is a set of n-dimensional vectors; The deep learning compensation model introduces a soft attention mechanism into two channels to identify the importance of historical data segments at different times, so as to reasonably allocate weights to each time step; the implementation process of the soft attention mechanism includes the following steps: (1) Calculate the attention scoring function s(X i , q), where X is the input vector and q is the query vector, and use the scoring function to calculate the correlation between each input vector and the query vector; (2) Calculate the attention distribution α i , use the attention vector z ∈ [1, N] to represent the index position of the selected information, where z = i represents the i-th input information; (3) Calculate the weighted average of the input data according to the attention distribution to obtain the attention value; The calculation methods of the scoring function include: additive model, dot product model, scaled dot product model and bilinear model; Additive model: s(X i ,q) = v T tanh(WX i +Uq) Dot product model: Scaled dot product model: Bilinear model: Among them, W, U and v are learnable network parameters, and d is the dimension of the input data.

2. According to the method for non-linear compensation of an electromagnetic actuator based on deep learning described in claim 1, characterized in that the activation function of the deep learning compensation model selects the Relu activation function, the model compilation optimizer selects Adam, and the number of iterations is 200 times.

Images