A neural network-based anti-disturbance control method, device, equipment and medium for a maglev train suspension system

Abstract

The application discloses a neural network-based maglev train suspension system anti-disturbance control method and device, equipment and medium, and relates to the field of transportation, which comprises the following steps: determining a target space model, a target controlled object and a corresponding target nominal controller based on a target state variable and a suspension equilibrium point of a target maglev train, constructing a target loop through Youla parameterization and determining a corresponding target relationship matrix, and determining a corresponding target optimization function based on the target loop; training a parameter prediction model based on an original residual signal output by a residual generator in the target loop, the target relationship matrix and the target optimization function to obtain a target parameter prediction model, determining a prediction control parameter of a dynamic feedback gain based on a real-time residual signal by using the target parameter prediction model, and performing anti-disturbance control on the target maglev train based on the prediction control parameter. The application improves the wideband anti-disturbance capability of the maglev train control system.

Description

Technical Field

[0001] This invention relates to the field of transportation, and in particular to a method, apparatus, equipment and medium for anti-disturbance control of a maglev train suspension system based on neural networks. Background Technology

[0002] Maglev trains utilize electromagnetic force to overcome gravity and achieve contactless levitation. Their levitation control system is a typical open-loop unstable, nonlinear coupled system. As train operating speeds continue to increase, the external disturbance environment faced by the levitation system becomes increasingly severe. Among these disturbances, track irregularities are the most critical source affecting the stability of the levitation system. Track irregularities mainly originate from track laying errors, uneven settlement of bridge pier foundations, and track deformation under long-term dynamic loads. These disturbances are characterized by wide frequency band coverage (encompassing low-frequency long waves to high-frequency short waves), strong time-varying amplitude, and significant nonlinear characteristics. During high-speed operation, these complex track irregularities can easily induce severe fluctuations in the levitation gap, severely reducing passenger comfort and, under extreme conditions, even leading to levitation failure, posing a significant threat to train operation safety.

[0003] Therefore, designing a high-performance controller with wide-band interference immunity is a core requirement for the control system of maglev trains. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, equipment, and medium for anti-interference control of a maglev train levitation system based on neural networks, which can improve the broadband anti-interference capability of the maglev train control system. The specific solution is as follows: In a first aspect, this application discloses a disturbance rejection control method for a maglev train levitation system based on a neural network, comprising: The target state variables and levitation equilibrium point of the target maglev train are determined, and the target space model corresponding to the target maglev train is determined based on the target state variables and the levitation equilibrium point; the target space model is a linear state space model of the levitation system of the target maglev train. Based on the target space model, the target controlled object and the corresponding target nominal controller of the target maglev train are determined. Then, using Youla parameterization, a target loop is constructed based on the target controlled object and the target nominal controller, and the corresponding target relationship matrix is ​​determined. Based on the target loop, the corresponding target optimization function is determined. The target loop includes a target basic loop and a corresponding target disturbance rejection loop. The target basic loop includes the nominal controller and the controlled object. The target disturbance rejection loop includes an observer-based residual generator and a dynamic feedback gain. The target optimization function is used to adjust the control parameters of the dynamic feedback gain. Based on the original residual signal output by the residual generator in the target loop, a corresponding residual feature vector is constructed. The residual feature vector is then used to train a pre-constructed parameter prediction model based on the target relation matrix and the target optimization function to obtain a target parameter prediction model. The target parameter prediction model is then used to determine the predictive control parameters of the corresponding dynamic feedback gain based on the real-time residual signal of the suspension system. Based on the predictive control parameters, the target maglev train is subjected to disturbance rejection control. The parameter prediction model is a prediction model based on a preset neural network architecture.

[0005] Optionally, the step of constructing the target loop and determining the corresponding target relationship matrix based on the target controlled object and the target nominal controller through Youla parameterization includes: Determine the state feedback gain matrix and the observer gain matrix, use the state feedback gain matrix and the observer gain matrix to perform coprime decomposition on the target controlled object to determine the corresponding first coprime factor, and use the state feedback gain matrix and the observer gain matrix to perform coprime decomposition on the target nominal controller to determine the corresponding second coprime factor; The target loop is constructed based on the target controlled object and the target nominal controller through Youla parameterization, and the target equivalent Youla parameters corresponding to the target loop are determined based on the predetermined target transfer function matrix, the first coprime factor and the second coprime factor. The target relation matrix corresponding to the target loop is determined based on the target space model and the target equivalent Youla parameters.

[0006] Optionally, determining the corresponding objective optimization function based on the objective loop includes: The interference transfer function corresponding to the target loop is determined based on the target relationship matrix, and the corresponding target optimization function is determined based on the dynamic feedback gain of the target loop and the interference transfer function. The dynamic feedback gain is a first-order parameterized form that includes the control parameters.

[0007] Optionally, constructing the corresponding residual feature vector based on the original residual signal output by the residual generator in the target loop includes: Obtain the original residual signal output by the residual generator in the target loop, and use the Kalman filter algorithm to filter the original residual signal to obtain the corresponding target residual signal; The target residual signal is processed using a fast Fourier transform and a target spectral peak detection and extraction algorithm to determine the target main frequency component corresponding to the original residual signal, and to determine the target mean, target standard deviation, target amplitude, and target energy integral corresponding to the original residual signal within the target sliding window; The residual feature vector corresponding to the original residual signal is constructed based on the target dominant frequency component, the target mean, the target standard deviation, the target amplitude, and the target energy integral.

[0008] Optionally, the step of training a pre-built parameter prediction model using the residual feature vector based on the target relation matrix and the target optimization function to obtain a target parameter prediction model includes: Determine whether the target main frequency component corresponding to the residual feature vector is greater than a preset frequency threshold; If the target main frequency component is greater than the preset frequency threshold, then the residual feature vector is used to train the first parameter prediction model based on the target relation matrix and the target optimization function to obtain the corresponding first target parameter prediction model; If the target main frequency component is less than or equal to the preset frequency threshold, then the residual feature vector is used to train the second parameter prediction model based on the target relation matrix and the target optimization function to obtain the corresponding second target parameter prediction model. The parameter prediction model includes the first parameter prediction model and the second parameter prediction model.

[0009] Optionally, the total loss function of the parameter prediction model includes a data-driven loss function and a physical constraint loss function; the data-driven loss function is a loss function determined based on the target dimension weighted vector parameters, the target loss threshold, the prediction control parameters corresponding to the original residual signal, and the theoretical control parameters; the physical constraint loss function is a loss function determined based on the target relation matrix and the target loss threshold.

[0010] Optionally, training the pre-built parameter prediction model to obtain the target parameter prediction model includes: The pre-built parameter prediction model is trained based on the target adaptive weight mechanism and the target preheating training mechanism to obtain the target parameter prediction model.

[0011] Secondly, this application discloses a disturbance rejection control device for a maglev train levitation system based on a neural network, comprising: The spatial model determination module is used to determine the target state variables and suspension equilibrium point of the target maglev train, and to determine the target spatial model corresponding to the target maglev train based on the target state variables and the suspension equilibrium point; the target spatial model is a linear state-space model of the suspension system of the target maglev train. The optimization function determination module is used to determine the target controlled object and the corresponding target nominal controller of the target maglev train based on the target space model, and to construct a target loop and determine the corresponding target relationship matrix based on the target controlled object and the target nominal controller through Youla parameterization, so as to determine the corresponding target optimization function based on the target loop; the target loop includes a target basic loop and a corresponding target disturbance rejection loop, the target basic loop includes a nominal controller and a controlled object, the target disturbance rejection loop includes an observer-based residual generator and a dynamic feedback gain, and the target optimization function is used to adjust the control parameters of the dynamic feedback gain; An anti-disturbance control module is used to construct a corresponding residual feature vector based on the original residual signal output by the residual generator in the target loop, and to train a pre-constructed parameter prediction model using the residual feature vector based on the target relation matrix and the target optimization function to obtain a target parameter prediction model. The target parameter prediction model is then used to determine the predictive control parameters of the corresponding dynamic feedback gain based on the real-time residual signal of the suspension system, and the target maglev train is subjected to anti-disturbance control based on the predictive control parameters. The parameter prediction model is a prediction model based on a preset neural network architecture.

[0012] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned disturbance rejection control method for a neural network-based maglev train suspension system.

[0013] Fourthly, this application discloses a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned disturbance rejection control method for a neural network-based maglev train suspension system.

[0014] In this application, when performing disturbance rejection control on the maglev train levitation system based on a neural network, the target state variables and levitation equilibrium point of the target maglev train are determined, and the target space model corresponding to the target maglev train is determined based on the target state variables and the levitation equilibrium point; the target space model is a linear state space model of the levitation system of the target maglev train; based on the target space model, the target controlled object and the corresponding target nominal controller of the target maglev train are determined, and a target loop is constructed based on the target controlled object and the target nominal controller using Youla parameterization, and the corresponding target relation matrix is ​​determined, so as to determine the corresponding target optimization function based on the target loop; the target loop includes a target basic loop and a corresponding target disturbance rejection loop, the target basic loop including... The system includes a nominal controller and a controlled object. The target disturbance rejection loop includes an observer-based residual generator and a dynamic feedback gain. The target optimization function is used to adjust the control parameters of the dynamic feedback gain. Based on the original residual signal output by the residual generator in the target loop, a corresponding residual feature vector is constructed. The residual feature vector is used to train a pre-constructed parameter prediction model based on the target relation matrix and the target optimization function to obtain a target parameter prediction model. The target parameter prediction model is then used to determine the corresponding predictive control parameters of the dynamic feedback gain based on the real-time residual signal of the levitation system. The target maglev train is then subjected to disturbance rejection control based on the predictive control parameters. The parameter prediction model is a prediction model based on a preset neural network architecture. As can be seen, in this application, the target space model corresponding to the target maglev train is first determined based on the target state variables and the levitation equilibrium point. Then, the target controlled object is determined based on the target space model, and the corresponding target nominal controller is designed. The target controlled object and the target nominal controller are then used through Youla parameterization. On the basis that the target nominal controller can make the system stable, a dynamic feedback gain is embedded to form a plug-and-play target disturbance rejection loop corresponding to the target basic loop. The construction of the target loop is completed, the corresponding target relationship matrix is ​​determined, and the corresponding target optimization function is obtained. Next, based on the original residual signal output by the residual generator in the target loop, a corresponding residual feature vector is constructed. Then, the residual feature vector is used to train the pre-constructed parameter prediction model based on the target relation matrix and the target optimization function to obtain the target parameter prediction model. Finally, the target parameter prediction model is used to determine the predictive control parameters of the corresponding dynamic feedback gain based on the real-time residual signal of the suspension system. Based on the predictive control parameters, the target maglev train is subjected to anti-disturbance control, which forces the target parameter prediction model to strictly follow the physical laws of control while fitting the data. This ensures that the predictive control parameters have both high data accuracy and physical stability, enabling the suspension system of the target maglev train to have wideband anti-disturbance capability and achieve adaptive suppression of track irregularities. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a flowchart of a disturbance rejection control method for a maglev train suspension system based on a neural network, as disclosed in this application. Figure 2 This is a schematic diagram of a specific target circuit disclosed in this application; Figure 3 This is a schematic diagram of a specific residual signal feature extraction process disclosed in this application; Figure 4 This is a schematic diagram of a specific parameter prediction model training process disclosed in this application; Figure 5 This is a schematic diagram of the anti-disturbance control device for a maglev train suspension system based on a neural network disclosed in this application; Figure 6 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Maglev trains utilize electromagnetic force to overcome gravity and achieve contactless levitation. Their levitation control system is a typical open-loop unstable, nonlinear coupled system. With the continuous increase in train operating speed, the external disturbance environment faced by the levitation system becomes increasingly severe. Track irregularities are the most critical disturbance source affecting the stability of the levitation system. Track irregularities mainly originate from track laying errors, uneven settlement of bridge pier foundations, and track deformation under long-term dynamic loads. These disturbances are characterized by wide frequency band coverage (encompassing low-frequency long waves to high-frequency short waves), strong amplitude time-varying characteristics, and significant nonlinear features. During high-speed operation, these complex track irregularities can easily induce severe fluctuations in the levitation gap, severely reducing passenger comfort and potentially leading to levitation failure under extreme conditions, posing a significant threat to train operation safety. To address these technical problems, this application discloses a neural network-based disturbance rejection control method for maglev train levitation systems, which can improve the broadband disturbance rejection capability of the maglev train control system.

[0019] See Figure 1 As shown in the figure, this invention discloses a disturbance rejection control method for a maglev train levitation system based on a neural network, comprising: Step S11: Determine the target state variables and suspension equilibrium point of the target maglev train, and determine the target space model corresponding to the target maglev train based on the target state variables and the suspension equilibrium point; the target space model is a linear state space model of the suspension system of the target maglev train.

[0020] In this embodiment, the target state variables and levitation equilibrium point of the target maglev train are first determined. Taking a single-iron levitation electromagnet system as an example, the levitation gap is selected. Vertical velocity and coil current As state variables, i.e., state vectors Based on Newton's second law and Kirchhoff's voltage law, a nonlinear state-space model of the levitation system of the target maglev train is established, and a model is established at the rated levitation equilibrium point. Linearization is performed at the point to obtain a linear state-space model of the suspension system, which serves as the target space model for the target maglev train. : ; in, To control the voltage input, The output is the suspension gap measured by the sensor. Under this definition, the system matrix... Input matrix Output matrix and feedforward matrix All are constant matrices of the corresponding dimension.

[0021] Step S12: Based on the target space model, determine the target controlled object and the corresponding target nominal controller of the target maglev train. Construct a target loop and determine the corresponding target relationship matrix based on the target controlled object and the target nominal controller using Youla parameterization, and determine the corresponding target optimization function based on the target loop. The target loop includes a target basic loop and a corresponding target disturbance rejection loop. The target basic loop includes a nominal controller and a controlled object. The target disturbance rejection loop includes an observer-based residual generator and a dynamic feedback gain. The target optimization function is used to adjust the control parameters of the dynamic feedback gain.

[0022] In this embodiment, the target controlled object and the corresponding target nominal controller of the target maglev train are determined based on the target space model. Simultaneously, to introduce dynamic feedback gain adjustment while maintaining system stability, the controlled object... and nominal controller Perform coprime decomposition separately.

[0023] In one specific implementation, the state feedback gain matrix is ​​designed. and observer gain matrix , make the matrix and The eigenvalues ​​of the controlled object are all located in the left half of the complex plane. The controlled object is decomposed into right-coprime components. The left coprime decomposes into ,in (i.e., a stable, regular rational function space), specifically expressed as: ; In one specific implementation, a nominal controller is designed. To stabilize the system, it is decomposed into right coprime components. Coprime decomposition with left ,in To specifically implement the coprime decomposition described above and provide a computational basis for constructing the Youla parameterized controller, this embodiment utilizes the aforementioned state feedback gain matrix. and observer gain matrix The following formula is defined as the specific coprime factor implementation of the nominal controller: ; in, The above decomposition factors satisfy the generalized Bezout equation, ensuring the internal stability of the system. The specific identity relationships are as follows: ; This equation shows that a stable coprime relationship exists between the controlled object and the controller, thereby ensuring the stability of the bounded input bounded output (BIBO) of the closed-loop system.

[0024] In this embodiment, after determining the target controlled object and the corresponding target nominal controller, the target loop can be constructed and the corresponding target relationship matrix can be determined based on the target controlled object and the target nominal controller through Youla parameterization, so as to determine the corresponding target optimization function based on the target loop. The target loop includes the target basic loop and the corresponding target disturbance rejection loop. The target basic loop includes the nominal controller and the controlled object. The target disturbance rejection loop includes the observer-based residual generator and the dynamic feedback gain. The target optimization function is used to adjust the control parameters of the dynamic feedback gain.

[0025] In one specific implementation, the target loop is as follows: Figure 2 As shown, constructing a target loop and determining the corresponding target relation matrix based on the target controlled object and the target nominal controller through Youla parameterization can specifically include: determining the state feedback gain matrix and the observer gain matrix; using the state feedback gain matrix and the observer gain matrix to perform coprime decomposition on the target controlled object to determine the corresponding first coprime factor; and using the state feedback gain matrix and the observer gain matrix to perform coprime decomposition on the target nominal controller to determine the corresponding second coprime factor; constructing a target loop based on the target controlled object and the target nominal controller through Youla parameterization; and determining the target equivalent Youla parameters corresponding to the target loop based on a pre-determined target transfer function matrix, the first coprime factor, and the second coprime factor; and determining the target relation matrix corresponding to the target loop based on the target space model and the target equivalent Youla parameters. Specifically, determining the corresponding target optimization function based on the target loop can include: determining the disturbance transfer function corresponding to the target loop based on the target relation matrix; and determining the corresponding target optimization function based on the dynamic feedback gain of the target loop and the disturbance transfer function; wherein the dynamic feedback gain is a first-order parameterized form including the control parameters.

[0026] In this embodiment, as Figure 2 As shown, the target loop consists of two parts: a basic stabilization loop (i.e., the target basic loop) and a plug-and-play anti-interference module (i.e., the target anti-interference loop). The basic stabilization loop includes a nominal controller. and the controlled object The plug-and-play disturbance rejection module consists of an observer-based residual generator and a dynamic feedback gain. Composition. In particular, the plug-and-play disturbance rejection module is embedded in the basic stabilization loop in parallel as an independent functional unit in a non-intrusive "external" form. This unique topology decouples the dynamic disturbance rejection function from the basic stabilization function in terms of physical logic, enabling the disturbance rejection module to be flexibly connected without changing the original structural parameters of the basic loop. The reference signal is the specified suspension gap value, with a dimension of 1. The error signal is the difference between the reference signal and the feedback signal of the closed loop, with a dimension of 1. The output signal of the nominal controller has a dimension of 1; This is an interference signal with a dimension of 1. This is the actual system output, i.e., the actual suspension gap, with a dimension of 1; The observer output signal is the system state estimate output value calculated by the observer, with a dimension of 1. This is the residual signal, generated from the difference between the observer's state estimate and the actual output. , with a dimension of 1.

[0027] In this embodiment, the control output It consists of a nominal control section and a dynamic compensation section, and the control law expression is: ; Based on the coprime decomposition results in the preceding steps, an intermediate variable is introduced. The controlled object can be represented as: ; Substituting the control law into the expression for the controlled object above, and simplifying, we get: ; Define the key transfer function matrix : ; Since all decomposition factors belong to Space, according to Bezout's equation, It must be an invertible matrix. Therefore, the intermediate variable can be solved. This leads to the system input / output expression: ; Based on the standard form of Youla parameterization, assume that there exists an equivalent Youla parameter. This makes the plug-and-play architecture equivalent to the standard Youla controller. Based on the standard form of the closed-loop transfer function in Youla's parameterization theory, and combined with the input-output relationship of the plug-and-play architecture, the following equation is established through coefficient comparison: ; By applying the left-hand multiplication and combining the equations using Bezout's equations, we can obtain the solution. The parsing expression: ; The obtained Substituting into the system equations and rearranging, we obtain the final closed-loop input-output relationship matrix (i.e., the target relationship matrix): ; Among them, the interference transfer function matrix Represented as: ; Understandably, to eliminate the impact of track irregularities on the suspension gap, the system needs to achieve zero steady-state gain, meaning that at a specific frequency, the system output's frequency response amplitude to the interference input is zero. Reflected in the above algebraic equations, this is represented as the residual signal. To system output signal Transfer function of dynamic mapping relationship between It is always equal to zero, that is: .

[0028] It is understandable that increasing the order of the dynamic feedback gain has a very limited impact on the accuracy of parameter prediction. Therefore, in order to reduce computational complexity and meet the real-time requirements of the system, this embodiment uses the dynamic feedback gain... The design is in the following first-order parameterized form: ; in, and These are the design parameters to be optimized. Under this structural constraint, in order to approximate the theoretical goal of "zero steady-state gain" as closely as possible, the original algebraic equation solving problem is transformed into a mathematical optimization problem, namely, finding the optimal combination of parameters. This minimizes the infinite norm of the disturbance transfer function of the closed-loop system under these parameters. The objective function is as follows: ; By solving this optimization problem, the optimal control parameters corresponding to the current interference frequency can be obtained. and .

[0029] Step S13: Construct a corresponding residual feature vector based on the original residual signal output by the residual generator in the target loop, and use the residual feature vector to train a pre-constructed parameter prediction model based on the target relation matrix and the target optimization function to obtain a target parameter prediction model. Use the target parameter prediction model to determine the predictive control parameters of the corresponding dynamic feedback gain based on the real-time residual signal of the suspension system, and perform anti-disturbance control on the target maglev train based on the predictive control parameters; wherein, the parameter prediction model is a prediction model based on a preset neural network architecture.

[0030] In this embodiment, as Figure 3 As shown, the dominant angular frequency That is, the target dominant frequency component, the residual signal dominant frequency component extracted through Fast Fourier Transform (FFT) and spectral peak detection, is used to characterize the frequency attributes of the interference and is the basis for switching the frequency band model of the neural network; target mean It is the arithmetic mean of the residual signal, used to characterize the DC offset of the signal; target standard deviation. It is the standard deviation of the residual signal, used to characterize the degree of fluctuation of the interference signal around the mean; maximum amplitude. That is, the target amplitude, which is the maximum absolute value of the residual signal within the sliding window, used to characterize the limiting intensity of the interference at the current moment; target energy integral. It is the integral of the square of the residual signal, used to characterize the total energy of the interference signal within the time window.

[0031] In one specific implementation, a corresponding residual feature vector is constructed based on the original residual signal output by the residual generator in the target loop. Specifically, this may include: acquiring the original residual signal output by the residual generator in the target loop; filtering the original residual signal using a Kalman filter algorithm to obtain the corresponding target residual signal; processing the target residual signal using a fast Fourier transform and target peak detection extraction algorithm to determine the target dominant frequency component corresponding to the original residual signal, and determining the target mean, target standard deviation, target amplitude, and target energy integral corresponding to the original residual signal within the target sliding window; and constructing the residual feature vector corresponding to the original residual signal based on the target dominant frequency component, target mean, target standard deviation, target amplitude, and target energy integral. In other words, firstly, the original residual signal output by the observer is acquired in real time. Given that the original signal may contain sensor measurement noise, a Kalman filter algorithm is used to denoise the original residual signal to improve the signal-to-noise ratio and accuracy of subsequent feature extraction. The denoised residual signal is then split into two paths, which are processed in parallel through time-domain analysis and frequency-domain analysis, respectively. The time-domain analysis process may include: using a sliding window technique to truncate the continuous residual signal to obtain the signal segment within the current time window. Subsequently, statistical feature calculations are performed on the signal within this window, specifically extracting the mean (…). ), standard deviation ), maximum amplitude ( ) and energy integral ( Four key physical characteristics. The frequency domain analysis process may include: performing a Fast Fourier Transform (FFT) on the signal to convert the time-domain signal into a frequency-domain signal. Subsequently, spectral analysis is performed, and peak detection algorithms are used to identify energy extrema in the spectrum, thereby extracting the dominant angular frequency of the current residual signal. ), used to characterize the main frequency components of the interference. Finally, the dominant angular frequency ( ) extracted by frequency domain analysis is used. ) and statistical features extracted by time-domain analysis The features are then fused to construct a five-dimensional feature vector. This is the residual feature vector corresponding to the original residual signal. This data will be used as input to the neural network model in subsequent steps for training the target parameter prediction model.

[0032] In this embodiment, a nonlinear mapping model is established as the parameter prediction model to be trained. This parameter prediction model is then used to predict the dynamic feedback gain in real time based on the residual feature vector extracted in the aforementioned steps. Optimal parameters ( To address the safety hazard of purely data-driven neural networks, which lack physical constraints and are prone to outputting parameters that violate system stability under extreme conditions, this embodiment introduces a physical information constraint mechanism. By embedding the closed-loop input-output equation (i.e., the target relation matrix) obtained in the preceding steps into the loss function as a physical mechanism regularization term, the network is forced to strictly adhere to the laws of control physics while fitting the data. This fundamentally eliminates prediction parameters that violate system stability, thereby ensuring that the prediction parameters possess both high data accuracy and physical stability, providing a highly reliable theoretical guarantee for the safe operation of the maglev system.

[0033] In this embodiment, the parameter prediction model is a prediction model based on a preset neural network architecture, including convolutional neural networks, multilayer perceptron models, etc. In one specific implementation, a multilayer perceptron model is selected, which includes: an input layer, and a time-frequency domain feature vector of the residual signal. The hidden layer has 5 layers, each containing 128 neurons, and uses ReLU activation to enhance non-linear expression. The output layer outputs dynamic feedback gain through linear transformation. Two-dimensional parameter prediction values The total loss function of the parametric prediction model is a composite physical constraint loss function, which includes a data-driven loss function and a physical constraint loss function. The data-driven loss function is determined based on the target dimension weighted vector parameters, the target loss threshold, the prediction control parameters corresponding to the original residual signal, and the theoretical control parameters. The physical constraint loss function is determined based on the target relation matrix and the target loss threshold.

[0034] In one specific implementation, to balance convergence speed and robustness, a weighted combination of mean squared error (MSE), mean absolute error (MAE), and Huber loss is used to construct the data-driven loss. Considering The impact on system performance is greater, so dimension-weighted vector parameters are introduced. .in Corresponding to parameters loss weights, Corresponding parameters The loss weight. In this embodiment, considering... The sensitivity to system disturbance rejection performance is much higher than To force the network to prioritize the fitting accuracy of key parameters, it is preferable to set... (For example, take) This overcomes the convergence challenges caused by orders-of-magnitude differences. Data-driven loss. It can be represented as: ; in, These correspond to the first dimension of the neural network output ( ) and the second dimension ( ); Let be the weighting coefficients for each component loss, and they typically satisfy . To simplify the description, let any parameter to be predicted be denoted as . (Right now and The corresponding network prediction value is (Right now and The theoretical truth value is (Right now and The calculation formulas for each component loss function are standardized as follows: Mean Squared Error (MSE): ; Mean Absolute Error (MAE): ; Huber loss: ; in, This is the threshold parameter for Huber loss (i.e., the target loss threshold).

[0035] Understandably, traditional neural networks only fit the data and are prone to generating parameters that do not conform to the physical characteristics of the system. Based on the Youla parameterized closed-loop equation (i.e., the target relation matrix) derived in the aforementioned steps, this embodiment introduces closed-loop physical residuals. Parameters used to quantify predictions made by the current network. The residual value represents the degree to which the closed-loop system deviates from the theoretically optimal solution of "zero steady-state gain". The closer this residual value is to zero, the more physically the predictive control parameters output by the parameter prediction model can suppress the current frequency disturbance with "zero gain", that is, the more it meets the disturbance rejection design requirements of the control system. Closed-loop physical residual The expression is as follows: ; The dynamic feedback gain form is as follows: ; To prevent tiny constants with a denominator of zero; , All of these are known transfer functions obtained from the coprime decomposition in the preceding steps. Based on the above physical residuals, a physical constraint loss function is defined. Let this be the average Huber Loss of the physical residual on the training sample set. By minimizing this loss, the neural network is forced to not only learn the data distribution but also follow the intrinsic physical mechanism of the control system. Specifically, it is expressed as follows: ; in, This represents the total number of samples in the current training batch. For sample index, Indicates the first The physical residual values ​​corresponding to each sample; This is used for Huber loss function calculation, designed to reduce sensitivity to outliers when residuals are large. Total loss function. Data-driven loss and physical constraint loss Weighted composition: ; in, This is the physical information weighting coefficient, used to adjust the degree to which physical constraints dominate during training.

[0036] In one specific implementation, training a pre-built parameter prediction model to obtain a target parameter prediction model can specifically include: training the pre-built parameter prediction model based on a target adaptive weighting mechanism and a target preheating training mechanism to obtain the target parameter prediction model. That is, for the data-driven loss in the total loss function... With physical constraint loss The invention addresses the problems of large differences in numerical magnitudes and difficulty in convergence due to physical constraints by designing the following dual optimization strategy: ① Adaptive weight mechanism: During training, weights are calculated in real time. and The numerical ratio, dynamically adjusting the physical weight. To address the magnitude imbalance problem, ensure that data fitting accuracy and physical consistency are given equal importance at every step of network optimization.

[0037] ② Target warm-up training mechanism: Because physical constraints involve complex differential equations, parameter prediction networks have difficulty learning directly under initial random conditions, which can easily lead to training oscillations. Therefore, in the early stages of training, the physical weights are first set... Setting the weights to 0 allows the network to focus on learning simple data distributions; as training progresses, the physical weights are then linearly increased to address the instability of the cold start and ensure smooth convergence during training.

[0038] In one specific implementation, such as Figure 4 As shown, the method involves training a pre-constructed parameter prediction model using residual feature vectors based on the target relation matrix and the target optimization function to obtain a target parameter prediction model. This includes: determining whether the target dominant frequency component corresponding to the residual feature vector is greater than a preset frequency threshold; if the target dominant frequency component is greater than the preset frequency threshold, then training a first parameter prediction model using the residual feature vector based on the target relation matrix and the target optimization function to obtain a corresponding first target parameter prediction model; if the target dominant frequency component is less than or equal to the preset frequency threshold, then training a second parameter prediction model using the residual feature vector based on the target relation matrix and the target optimization function to obtain a corresponding second target parameter prediction model. The parameter prediction model includes both the first and second parameter prediction models. In other words, in this embodiment, based on prior experience, a frequency division threshold (i.e., a preset frequency threshold) is first set in the residual signal frequency band. Using this threshold as a boundary, the interference signal is divided into low-frequency and high-frequency bands. The high-frequency model (i.e., the first parameter prediction model) covers... Frequency band, low-frequency model (i.e., second parameter prediction model) coverage Frequency band. This segmentation strategy effectively solves the problem of insufficient fitting accuracy of a single network for different dynamic characteristics in a wide frequency domain. Specifically, by combining time-frequency feature extraction and frequency band prediction techniques during model training and prediction, it can accurately perceive and distinguish the low-frequency long-wave and high-frequency short-wave components in track irregularities. At the same time, combined with the generalization ability brought by physical constraints, the target parameter prediction model can generate control parameters that conform to the physical optimal solution in real time for composite interference with different frequency characteristics. Thus, without compromising the basic stability of the system, it solves the safety hazards of pure data-driven models and achieves accurate, stable, and efficient suppression of broadband, time-varying track irregularities of maglev trains.

[0039] In this embodiment, after training the target parameter prediction model, the predictive control parameters for the corresponding dynamic feedback gain can be determined based on the real-time residual signal of the suspension system using the target parameter prediction model. Based on these predictive control parameters, the target maglev train can then be subjected to disturbance rejection control. In one specific implementation, the trained target parameter prediction model is embedded into the maglev train system loop. During train operation, the system residual signal data is first collected in real-time and preprocessed. The corresponding residual feature vector is then obtained by feature extraction from this real-time residual signal. The parameters are input into the target parameter prediction model for online inference to obtain the predictive control parameters corresponding to the current disturbance. And update the dynamic feedback gain in real time. The parameters are used to generate compensating control force, thereby achieving adaptive suppression of track irregularities.

[0040] As can be seen, in this application, the target space model corresponding to the target maglev train is first determined based on the target state variables and the levitation equilibrium point. Then, the target controlled object is determined based on the target space model, and the corresponding target nominal controller is designed. The target controlled object and the target nominal controller are then used through Youla parameterization. On the basis that the target nominal controller can make the system stable, a dynamic feedback gain is embedded to form a plug-and-play target disturbance rejection loop corresponding to the target basic loop. The construction of the target loop is completed, the corresponding target relationship matrix is ​​determined, and the corresponding target optimization function is obtained. Next, based on the original residual signal output by the residual generator in the target loop, a corresponding residual feature vector is constructed. Then, the residual feature vector is used to train the pre-constructed parameter prediction model based on the target relation matrix and the target optimization function to obtain the target parameter prediction model. Finally, the target parameter prediction model is used to determine the predictive control parameters of the corresponding dynamic feedback gain based on the real-time residual signal of the suspension system. Based on the predictive control parameters, the target maglev train is subjected to anti-disturbance control, which forces the target parameter prediction model to strictly follow the physical laws of control while fitting the data. This ensures that the predictive control parameters have both high data accuracy and physical stability, enabling the suspension system of the target maglev train to have wideband anti-disturbance capability and achieve adaptive suppression of track irregularities.

[0041] See Figure 5 As shown, this application discloses an anti-disturbance control device for a maglev train levitation system based on a neural network, comprising: The spatial model determination module 11 is used to determine the target state variables and suspension equilibrium point of the target maglev train, and to determine the target spatial model corresponding to the target maglev train based on the target state variables and the suspension equilibrium point; the target spatial model is a linear state-space model of the suspension system of the target maglev train. The optimization function determination module 12 is used to determine the target controlled object and the corresponding target nominal controller of the target maglev train based on the target space model, and to construct a target loop and determine the corresponding target relationship matrix based on the target controlled object and the target nominal controller through Youla parameterization, so as to determine the corresponding target optimization function based on the target loop; the target loop includes a target basic loop and a corresponding target disturbance rejection loop, the target basic loop includes a nominal controller and a controlled object, the target disturbance rejection loop includes an observer-based residual generator and a dynamic feedback gain, and the target optimization function is used to adjust the control parameters of the dynamic feedback gain; The disturbance rejection control module 13 is used to construct a corresponding residual feature vector based on the original residual signal output by the residual generator in the target loop, and to train a pre-constructed parameter prediction model using the residual feature vector based on the target relation matrix and the target optimization function to obtain a target parameter prediction model. The target parameter prediction model is then used to determine the predictive control parameters of the corresponding dynamic feedback gain based on the real-time residual signal of the suspension system, and the target maglev train is subjected to disturbance rejection control based on the predictive control parameters. The parameter prediction model is a prediction model based on a preset neural network architecture.

[0042] As can be seen, in this application, the target space model corresponding to the target maglev train is first determined based on the target state variables and the levitation equilibrium point. Then, the target controlled object is determined based on the target space model, and the corresponding target nominal controller is designed. The target controlled object and the target nominal controller are then used through Youla parameterization. On the basis that the target nominal controller can make the system stable, a dynamic feedback gain is embedded to form a plug-and-play target disturbance rejection loop corresponding to the target basic loop. The construction of the target loop is completed, the corresponding target relationship matrix is ​​determined, and the corresponding target optimization function is obtained. Next, based on the original residual signal output by the residual generator in the target loop, a corresponding residual feature vector is constructed. Then, the residual feature vector is used to train the pre-constructed parameter prediction model based on the target relation matrix and the target optimization function to obtain the target parameter prediction model. Finally, the target parameter prediction model is used to determine the predictive control parameters of the corresponding dynamic feedback gain based on the real-time residual signal of the suspension system. Based on the predictive control parameters, the target maglev train is subjected to anti-disturbance control, which forces the target parameter prediction model to strictly follow the physical laws of control while fitting the data. This ensures that the predictive control parameters have both high data accuracy and physical stability, enabling the suspension system of the target maglev train to have wideband anti-disturbance capability and achieve adaptive suppression of track irregularities.

[0043] In one specific implementation, the optimization function determination module 12 includes: The coprime decomposition unit is used to determine the state feedback gain matrix and the observer gain matrix, and to perform coprime decomposition on the target controlled object using the state feedback gain matrix and the observer gain matrix to determine the corresponding first coprime factor, and to perform coprime decomposition on the target nominal controller using the state feedback gain matrix and the observer gain matrix to determine the corresponding second coprime factor. The Youla equivalent unit is used to construct a target loop based on the target controlled object and the target nominal controller through Youla parameterization, and to determine the target equivalent Youla parameters corresponding to the target loop based on the predetermined target transfer function matrix, the first coprime factor and the second coprime factor; The relation matrix determination unit is used to determine the target relation matrix corresponding to the target loop based on the target space model and the target equivalent Youla parameters.

[0044] In one specific implementation, the optimization function determination module 12 includes: The optimization function determination unit is used to determine the interference transfer function corresponding to the target loop based on the target relationship matrix, and to determine the corresponding target optimization function based on the dynamic feedback gain of the target loop and the interference transfer function; The dynamic feedback gain is a first-order parameterized form that includes the control parameters.

[0045] In one specific implementation, the disturbance rejection control module 13 includes: The signal filtering unit is used to obtain the original residual signal output by the residual generator in the target loop, and to use the Kalman filtering algorithm to filter the original residual signal to obtain the corresponding target residual signal. The signal analysis unit is used to process the target residual signal using fast Fourier transform and target peak detection and extraction algorithms to determine the target main frequency component corresponding to the original residual signal, and to determine the target mean, target standard deviation, target amplitude and target energy integral corresponding to the original residual signal within the target sliding window; The vector construction unit is used to construct the residual feature vector corresponding to the original residual signal based on the target main frequency component, the target mean, the target standard deviation, the target amplitude, and the target energy integral.

[0046] In one specific implementation, the disturbance rejection control module 13 includes: A threshold determination unit is used to determine whether the target main frequency component corresponding to the residual feature vector is greater than a preset frequency threshold. The first model training unit is used to train the first parameter prediction model based on the target relation matrix and the target optimization function using the residual feature vector if the target main frequency component is greater than the preset frequency threshold, so as to obtain the corresponding first target parameter prediction model. The second model training unit is used to train the second parameter prediction model based on the target relation matrix and the target optimization function using the residual feature vector if the target main frequency component is less than or equal to the preset frequency threshold, so as to obtain the corresponding second target parameter prediction model. The parameter prediction model includes the first parameter prediction model and the second parameter prediction model.

[0047] In one specific implementation, the disturbance rejection control module 13 includes: The third model training unit is used to train the pre-built parameter prediction model based on the target adaptive weight mechanism and the target preheating training mechanism to obtain the target parameter prediction model.

[0048] Furthermore, embodiments of this application also disclose an electronic device, Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0049] Figure 6 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the disturbance rejection control method for the maglev train suspension system based on neural networks disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0050] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0051] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk, or optical disk, etc. The resources stored thereon can include an operating system 221, computer programs 222, etc., and the storage method can be temporary storage or permanent storage.

[0052] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the neural network-based anti-disturbance control method for the maglev train suspension system executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0053] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disturbance rejection control method for a maglev train levitation system based on a neural network. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0054] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0055] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0056] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0057] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0058] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A disturbance rejection control method for a maglev train levitation system based on neural networks, characterized in that, include: Determine the target state variables and levitation equilibrium point of the target maglev train, and determine the target space model corresponding to the target maglev train based on the target state variables and the levitation equilibrium point; The target space model is a linear state-space model of the suspension system of the target maglev train; Based on the target space model, the target controlled object and the corresponding target nominal controller of the target maglev train are determined. Then, using Youla parameterization, a target loop is constructed based on the target controlled object and the target nominal controller, and the corresponding target relationship matrix is ​​determined. Based on the target loop, the corresponding target optimization function is determined. The target loop includes a target basic loop and a corresponding target disturbance rejection loop. The target basic loop includes the nominal controller and the controlled object. The target disturbance rejection loop includes an observer-based residual generator and a dynamic feedback gain. The target optimization function is used to adjust the control parameters of the dynamic feedback gain. Based on the original residual signal output by the residual generator in the target loop, a corresponding residual feature vector is constructed. The residual feature vector is then used to train a pre-constructed parameter prediction model based on the target relation matrix and the target optimization function to obtain a target parameter prediction model. The target parameter prediction model is then used to determine the predictive control parameters of the corresponding dynamic feedback gain based on the real-time residual signal of the suspension system. Based on the predictive control parameters, the target maglev train is subjected to disturbance rejection control. The parameter prediction model is a prediction model based on a preset neural network architecture. The construction of the corresponding residual feature vector based on the original residual signal output by the residual generator in the target loop includes: Obtain the original residual signal output by the residual generator in the target loop, and use the Kalman filter algorithm to filter the original residual signal to obtain the corresponding target residual signal; The target residual signal is processed using a fast Fourier transform and a target spectral peak detection and extraction algorithm to determine the target main frequency component corresponding to the original residual signal, and to determine the target mean, target standard deviation, target amplitude, and target energy integral corresponding to the original residual signal within the target sliding window; Based on the target main frequency component, the target mean, the target standard deviation, the target amplitude, and the target energy integral, construct the residual feature vector corresponding to the original residual signal; The total loss function of the parameter prediction model includes a data-driven loss function and a physical constraint loss function. The data-driven loss function is a loss function determined based on the target dimension weighted vector parameters, the target loss threshold, the prediction control parameters corresponding to the original residual signal, and the theoretical control parameters. The physical constraint loss function is a loss function determined based on the target relation matrix and the target loss threshold.

2. The disturbance rejection control method for a maglev train levitation system based on neural networks according to claim 1, characterized in that, The step of constructing a target loop and determining the corresponding target relationship matrix based on the target controlled object and the target nominal controller using Youla parameterization includes: Determine the state feedback gain matrix and the observer gain matrix, use the state feedback gain matrix and the observer gain matrix to perform coprime decomposition on the target controlled object to determine the corresponding first coprime factor, and use the state feedback gain matrix and the observer gain matrix to perform coprime decomposition on the target nominal controller to determine the corresponding second coprime factor; The target loop is constructed based on the target controlled object and the target nominal controller through Youla parameterization, and the target equivalent Youla parameters corresponding to the target loop are determined based on the predetermined target transfer function matrix, the first coprime factor and the second coprime factor. The target relation matrix corresponding to the target loop is determined based on the target space model and the target equivalent Youla parameters.

3. The disturbance rejection control method for a maglev train levitation system based on neural networks according to claim 1, characterized in that, The step of determining the corresponding target optimization function based on the target loop includes: The interference transfer function corresponding to the target loop is determined based on the target relationship matrix, and the corresponding target optimization function is determined based on the dynamic feedback gain of the target loop and the interference transfer function. The dynamic feedback gain is a first-order parameterized form that includes the control parameters.

4. The disturbance rejection control method for a maglev train levitation system based on neural networks according to claim 1, characterized in that, The step of training a pre-built parameter prediction model using the residual feature vector based on the target relation matrix and the target optimization function to obtain a target parameter prediction model includes: Determine whether the target main frequency component corresponding to the residual feature vector is greater than a preset frequency threshold; If the target main frequency component is greater than the preset frequency threshold, then the residual feature vector is used to train the first parameter prediction model based on the target relation matrix and the target optimization function to obtain the corresponding first target parameter prediction model; If the target main frequency component is less than or equal to the preset frequency threshold, then the residual feature vector is used to train the second parameter prediction model based on the target relation matrix and the target optimization function to obtain the corresponding second target parameter prediction model. The parameter prediction model includes the first parameter prediction model and the second parameter prediction model.

5. The disturbance rejection control method for a maglev train levitation system based on neural networks according to any one of claims 1 to 4, characterized in that, The step of training a pre-built parameter prediction model to obtain a target parameter prediction model includes: The pre-built parameter prediction model is trained based on the target adaptive weight mechanism and the target preheating training mechanism to obtain the target parameter prediction model.

6. A disturbance rejection control device for a maglev train levitation system based on a neural network, characterized in that, include: The spatial model determination module is used to determine the target state variables and levitation equilibrium point of the target maglev train, and to determine the target spatial model corresponding to the target maglev train based on the target state variables and the levitation equilibrium point. The target space model is a linear state-space model of the suspension system of the target maglev train; The optimization function determination module is used to determine the target controlled object and the corresponding target nominal controller of the target maglev train based on the target space model, and to construct a target loop and determine the corresponding target relationship matrix based on the target controlled object and the target nominal controller through Youla parameterization, so as to determine the corresponding target optimization function based on the target loop; the target loop includes a target basic loop and a corresponding target disturbance rejection loop, the target basic loop includes a nominal controller and a controlled object, the target disturbance rejection loop includes an observer-based residual generator and a dynamic feedback gain, and the target optimization function is used to adjust the control parameters of the dynamic feedback gain; An anti-disturbance control module is used to construct a corresponding residual feature vector based on the original residual signal output by the residual generator in the target loop, and to train a pre-constructed parameter prediction model using the residual feature vector based on the target relation matrix and the target optimization function to obtain a target parameter prediction model. The target parameter prediction model is then used to determine the predictive control parameters of the corresponding dynamic feedback gain based on the real-time residual signal of the levitation system, and the target maglev train is subjected to anti-disturbance control based on the predictive control parameters. The parameter prediction model is a prediction model based on a preset neural network architecture. The disturbance rejection control module specifically includes: The signal filtering unit is used to obtain the original residual signal output by the residual generator in the target loop, and to filter the original residual signal using the Kalman filtering algorithm to obtain the corresponding target residual signal. The signal analysis unit is used to process the target residual signal using fast Fourier transform and target peak detection and extraction algorithms to determine the target main frequency component corresponding to the original residual signal, and to determine the target mean, target standard deviation, target amplitude and target energy integral corresponding to the original residual signal within the target sliding window; A vector construction unit is used to construct a residual feature vector corresponding to the original residual signal based on the target main frequency component, the target mean, the target standard deviation, the target amplitude, and the target energy integral. The total loss function of the parameter prediction model includes a data-driven loss function and a physical constraint loss function. The data-driven loss function is a loss function determined based on the target dimension weighted vector parameters, the target loss threshold, the prediction control parameters corresponding to the original residual signal, and the theoretical control parameters. The physical constraint loss function is a loss function determined based on the target relation matrix and the target loss threshold.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the disturbance rejection control method for a maglev train suspension system based on a neural network as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the disturbance rejection control method for a maglev train suspension system based on a neural network as described in any one of claims 1 to 5.

Images