An adaptive sliding mode fault-tolerant control method for train wireless network delay compensation

Abstract

The present invention relates to the field of train control technology, and specifically to an adaptive sliding mode fault-tolerant control method for train wireless network delay compensation, including collecting wireless delay data, performing delay prediction, performing health detection of traction / brake actuators, establishing a dynamic model, and inputting the delay prediction results, health detection results, and basic resistance data as parameters into an adaptive sliding mode fault-tolerant controller to control the operation of the train. The present invention sets a CNN-GRU delay prediction model, which can improve the accuracy of delay prediction by performing experimental prediction after extracting feature information through a convolutional neural network; by establishing a multi-particle train dynamics model, considering basic resistance, additional resistance, force between carriages, mass slew coefficient, and the health status of the actuator, the train dynamics model is closer to reality, and an improved particle swarm algorithm is used to optimize the hyperparameters of the delay prediction model, thereby improving the accuracy of the particle swarm algorithm in optimizing the hyperparameters.

Description

Technical Field

[0001] The present invention relates to the technical field of train control, and in particular to an adaptive sliding mode fault-tolerant control method for time delay compensation of a train wireless network. Background Art

[0002] Wireless network communication is a major direction for the next generation of train network communication. Adaptive sliding mode fault-tolerant control is a common control method in train network control. The mixed approach rate in adaptive sliding mode fault-tolerant control is a convergence law used to improve the jitter phenomenon in sliding mode control. The addition of adaptation can make the controller more robust. The addition of fault-tolerant control can ensure the smooth operation of the train when the train actuator control fails. Delay compensation is used to eliminate the influence of the delay generated by the control information from the controller to the controlled object on the control system during the control process. Train wireless network control is a major direction for future train network control. Before practical application, it is necessary to fully consider the delay problem and actuator failure problem in the wireless transmission process. Wireless transmission is susceptible to external interference, which causes delay fluctuations. When the delay is large, it will lead to inaccurate control.

[0003] The delay problem in wireless network control in the prior art is usually dealt with in the form of state prediction, that is, predicting the next action, and then making corresponding compensation methods, mainly by processing the acquired delay sequence, and further using machine learning or deep learning algorithms to predict the delay at the next moment. A common method is to process the original delay sequence by decomposition, noise reduction, etc., so as to reduce the complexity of the sequence, and then use SVM, LSSVM, BP, Elman, LSTM and other algorithms to predict the delay sequence. However, the complexity of some data after a certain processing is often still high, which is not conducive to subsequent predictions.

[0004] Train network control in the existing technology includes operation control, air conditioning control, door control, etc. Operation control mainly controls the running speed of the train. After the train kinematic model is established, it is achieved through different control methods, such as PID control, sliding mode control, generalized control or various combination algorithms. However, the influence of various uncertain factors should be taken into account in the design process of the controller, which leads to a decrease in control effect. Summary of the invention

[0005] In order to overcome the time delay problem in wireless network control and the technical problem of poor controller control effect in the prior art, the present invention provides a train wireless network time delay compensation adaptive sliding mode fault-tolerant control method.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a train wireless network delay compensation adaptive sliding mode fault-tolerant control method, comprising the following steps:

[0007] S1: Collect the wireless delay data of the train wireless communication equipment by intercepting the timestamp of the data sending time;

[0008] S2: Build a CNN-GRU delay prediction model, input the wireless delay data into the CNN-GRU delay prediction model for prediction, and obtain the network delay;

[0009] S3: using a BP neural network to detect the health of the traction / brake actuator to obtain health data of the traction / brake actuator;

[0010] S4: Establish train dynamics model;

[0011] S5: Control the train according to the network delay obtained in step S2, the health data of the traction / brake actuator obtained in step S3, and the control rate of the adaptive sliding mode fault-tolerant controller improved by designing the train dynamics model established in step S4.

[0012] Preferably, the CNN-GRU delay prediction model in step S2 includes a CNN convolutional neural network and a gated recurrent unit, the CNN convolutional neural network and the gated recurrent unit are connected through a flattening layer, the flattening layer is used to convert the time series into an input format suitable for the subsequent neural network, the CNN convolutional neural network is used to extract features from the wireless delay data to obtain feature information of the wireless delay data, and the gated recurrent unit is used to predict the delay through the feature information.

[0013] Preferably, the CNN convolutional neural network includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer and an output layer connected in sequence, wherein the first convolutional layer and the second convolutional layer are used for feature information extraction, the first pooling layer and the second pooling layer are used for reducing the dimension of the feature information extracted by the convolutional layer, and the fully connected layer is used for classifying and outputting the feature information.

[0014] Preferably, the expression for the delay prediction process of the gated recurrent unit is:

[0015] z t = a ( U z [ h t − 1 , R t ] ) ;

[0016] r t = a ( U r [ h t − 1 , R t ] ) ;

[0017] h ˜ t = tanh( U h [ r t h t − 1 , R t ] ) ;

[0018] ;

[0019] in, To update the gate, To reset the gate, is the multiplication result of the input and the hidden layer at the previous moment, is the hidden layer output, is the activation function, is the hidden layer output at the previous moment, , , are weight matrices, is the input of the neural network.

[0020] Preferably, the hyperparameters of the gated recurrent unit are optimized using a particle swarm algorithm, and the optimization process is as follows:

[0021] Initial particles are randomly generated, and the positions of the randomly generated initial particles are mapped to new positions through Tent chaos mapping, where the Tent chaos mapping formula is:

[0022] ;

[0023] in, is the particle position after chaos, are randomly generated initial particles, is a constant, ;

[0024] Input the hyperparameters corresponding to the initial particles into the gated recurrent unit, output the prediction results, and calculate the current fitness;

[0025] Inertia weight is introduced, and the inertia weight is updated through the adaptive algorithm. Then the particle speed and its position are updated according to the individual extreme value and the global extreme value, and the hyperparameters are updated. The expression of inertia weight is:

[0026] ;

[0027] in, is the inertia weight, is the maximum inertia weight, is the minimum inertia weight, is the current fitness value, is the minimum value of fitness, is the average value of fitness;

[0028] The expression of particle update speed is:

[0029] ;

[0030] in, For the Particles in After iterations, The speed in dimension, is the inertia weight, For the Particles in After iterations, The speed in dimension, and are learning factors, and are randomly generated numbers. and For the Particles in The optimal solution after iteration is The position in dimension and the global optimal solution are The position in dimension, For the Particles in After iterations, Position in dimension;

[0031] The expression for updating the particle position is:

[0032] ;

[0033] in, For the Particles in After iterations, The position in dimension, For the Particles in After iterations, The position in dimension, For the Particles in After iterations, Speed ​​in dimension;

[0034] Use the current hyperparameters to predict the delay and calculate the fitness to determine whether the fitness is less than the previous fitness. If so, repeatedly update the inertia weight through the adaptive algorithm, and then update the particle's speed and position according to the individual extreme value and the global extreme value, and update the hyperparameters and use the current hyperparameters to predict the delay and calculate the fitness to determine whether the fitness is less than the previous fitness until the maximum number of iterations is reached. If not, output the current hyperparameters as the optimal hyperparameters.

[0035] Preferably, the fitness calculation expression is:

[0036] ;

[0037] in, is the current fitness value, is the difference between the ideal speed and the actual speed of the carriage, is the total number of iterations, is the current iteration number.

[0038] Preferably, the BP neural network is used in step S3 to detect the health of the traction / brake actuator. Specifically, the weight and offset of the BP neural network model are initialized, the actual speed, displacement and acceleration of the train are used as inputs of the BP neural network, and the health of the traction / brake actuator: 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100% is used as outputs of the BP neural network. The BP neural network is trained by the gradient descent method, and the trained BP neural network is used to detect the health of the traction / brake actuator to obtain the health data of the traction / brake actuator.

[0039] Preferably, the train dynamics model established in S4 is specifically: considering the force between carriages, the health data of the traction / brake actuator, the basic resistance data, the additional resistance data and the mass rotation coefficient, the train dynamics model is established, and its expression is:

[0040] ;

[0041] in, For the The displacement of the train, For the The speed of the train, For the The mass of a train carriage, For the The health status of the train traction / brake actuators, is the mass rotation coefficient, For the The force output by the train controller, For the The resistance of the train during operation, , is the force between adjacent carriages, is the elastic coupling coefficient of the spring of the spring-damper system, is the damping coefficient of the spring-damper system, and represent the displacement difference and speed difference between adjacent carriages, respectively. For the current moment, For time delay, is the force between adjacent carriages.

[0042] Preferably, step S5 specifically includes the following steps:

[0043] S501: According to the train dynamics model, the displacement error and speed error are obtained, and their expressions are:

[0044] ;

[0045] in, is the error between the ideal displacement and the actual displacement, is the error between the ideal speed and the actual speed, is the ideal displacement, is the actual displacement, For the ideal speed, is the actual speed;

[0046] S502: Obtain the sliding surface function, whose expression is:

[0047] ;

[0048] in, is the sliding surface function, Indicates the serial number, is the coefficient for adjusting the approach speed, and , is the error between the ideal displacement and the actual displacement, is the error between the ideal speed and the actual speed;

[0049] S503: Obtain the mixed approach rate, which is expressed as:

[0050] ;

[0051] in, Indicates the serial number, , , , , , are coefficients for adjusting the approach speed. is the sliding surface;

[0052] S504: Obtain the control rate of the adaptive sliding mode fault-tolerant controller according to the sliding surface function and the mixed approach rate, and its expression is:

[0053] ;

[0054] in, is the force output by the controller, For the current moment, is the predicted network delay, , It is the inverse of the product of the acceleration factor and the health of the traction / brake actuator. is the acceleration factor, The health of the traction / brake actuator. , , , , , To adjust the coefficient of approach speed, is the sliding surface function, To adjust the coefficient of approach speed, is the error between the ideal displacement and the actual displacement, is the intermediate quantity, For the ideal speed, is the basic resistance, is a symbolic function;

[0055] S505: Considering the influence of external environmental factors on basic resistance, the basic resistance considering external environmental factors is obtained by RBF neural network fitting, and its expression is:

[0056] ;

[0057] in, In order to consider the basic resistance of external environmental factors, is the estimated weight of the RBF neural network, is the activation function, Estimate the difference for the RBF neural network;

[0058] S506: Construct an improved adaptive sliding mode fault-tolerant controller control rate model taking into account external environmental factors, and its expression is:

[0059] ;

[0060] in, is the force output by the controller, For the current moment, is the predicted network delay, , It is the inverse of the product of the acceleration factor and the health of the traction / brake actuator. is the acceleration factor, The health of the traction / brake actuator. , , , , , To adjust the coefficient of approach speed, is the sliding surface function, To adjust the coefficient of approach speed, is the error between the ideal displacement and the actual displacement, For the ideal speed, is the basic resistance, is the symbolic function, is the estimated weight of the RBF neural network, is the activation function, Estimate the difference for the RBF neural network;

[0061] S507: Obtain the adaptive rate of the hyperparameter, the expression of which is:

[0062] ;

[0063] ;

[0064] ;

[0065] in, for Moment The adaptive rate, is the constant that controls the change of the adaptive rate, is the sliding surface function, is the symbolic function, The health of the traction / brake actuator. is the adaptive rate of the weight, is the constant that controls the change of the adaptive rate, is the acceleration factor, is the activation function, is the adaptive rate of the offset, A constant that controls the adaptation rate.

[0066] Preferably, after step S5, step S6 is further included: stability verification of the designed adaptive sliding mode fault-tolerant controller, and step S6 specifically includes the following steps:

[0067] S601: Design a Lyapunov function, whose expression is:

[0068] ;

[0069] in, is the Lyapunov function, is the sliding surface function, is the acceleration factor, is the constant that controls the change of the adaptive rate, For the The difference between the ideal value and the estimated value of the vehicle, is the constant that controls the change of the adaptive rate, is the transpose of the difference in RBFNN weights, is the difference of RBFNN weights, is the constant that controls the change of the adaptive rate, is the difference of RBFNN offset, is the ideal value, is an estimated value, is the difference between the ideal value and the estimated value, is the ideal weight of RBF neural network, Estimate weights for the RBF neural network, is the difference between the estimated weight and the ideal weight of the RBF neural network, is the ideal difference of RBF neural network, is the RBF neural network estimation difference, is the difference between the RBF neural network estimated difference and the ideal difference, for the current moment;

[0070] S602: Derivative of the Lyapunov function, the expression is:

[0071] ;

[0072] in, is the derivative of the Lyapunov function, is the sliding surface function, is the derivative of the sliding surface function, is the acceleration factor, is the constant that controls the change of the adaptive rate, For the The difference between the ideal value and the estimated value of the vehicle, for The adaptive rate, is the constant that controls the change of the adaptive rate, is the difference between the ideal weight and the actual weight of RBFNN, is the derivative of the expected weight of RBFNN, is the constant that controls the change of the adaptive rate, is the difference between the ideal offset and the actual offset, is the derivative of the ideal offset, for the current moment;

[0073] S603: Substitute the improved adaptive sliding mode fault-tolerant controller control rate model into the Lyapunov function derivative expression obtained in step S602 to obtain the following expression:

[0074] ;

[0075] in, is the derivative of the Lyapunov function, is the sliding surface function, To adjust the coefficient of approach speed, is the difference between the ideal speed and the actual speed, is the derivative of the ideal velocity, The health of the traction / brake actuator. is the force output by the controller, is a constant, is the resistance of the train running, is the constant that controls the change of the adaptive rate, is the constant that controls the change of the adaptive rate, is the constant that controls the change of the adaptive rate, For the The difference between the ideal value and the estimated value of the vehicle, for The adaptive rate, is the transpose of the difference in RBFNN weights, is the adaptive rate of the weight, is the difference between the ideal offset and the actual offset, is the adaptive rate of the offset, is the intermediate quantity, is the derivative of the ideal velocity, To adjust the coefficient of approach speed, is the expected train running resistance, is the symbolic function, is the acceleration factor, is the sliding surface, is the activation function, is the difference between the RBF neural network estimated difference and the ideal difference, For the current moment.

[0076] Compared with the prior art, the present invention sets a CNN-GRU delay prediction model, can perform experimental prediction after extracting feature information through a convolutional neural network, and can improve the accuracy of delay prediction; by establishing a multi-particle train dynamics model, the basic resistance, additional resistance, the force between carriages, the mass rotation coefficient and the health status of the traction / brake actuator are considered at the same time, so that the train dynamics model is closer to reality and the final control effect is more accurate; an improved particle swarm algorithm is used to optimize the memory hyperparameters of the delay prediction model, the Tent chaotic mapping is used to map the position of the initial particle to the new position, and the inertia weight is introduced to improve the accuracy of the particle swarm algorithm in optimizing the hyperparameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 The present invention is a flow chart of a train wireless network delay compensation adaptive sliding mode fault-tolerant control method.

[0078] Figure 2It is a comparison chart of wireless delay data prediction results in a specific implementation manner of the present invention.

[0079] Figure 3 It is a speed tracking curve diagram of vehicles 1 and 2 in a specific implementation manner of the present invention.

[0080] Figure 4 It is a speed tracking curve diagram of vehicles 3 and 4 in a specific implementation manner of the present invention.

[0081] Figure 5 It is a speed tracking curve diagram of vehicles 5 and 6 in a specific embodiment of the present invention.

[0082] Figure 6 It is a speed tracking curve diagram of vehicles 7 and 8 in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0083] The specific implementation method of the adaptive sliding mode fault-tolerant control method for train wireless network delay compensation of the present invention is as follows Figure 1 As shown, the following steps are included:

[0084] S1: Collect wireless delay data of wireless communication devices of 8 vehicles by intercepting the timestamp of sending data;

[0085] S2: Construct a CNN-GRU delay prediction model, input the wireless delay data into the CNN-GRU delay prediction model for prediction, and obtain the network delay. The comparison between the prediction results and the PSO-LSTM prediction results is shown in the figure below. Figure 2 As shown;

[0086] The CNN-GRU delay prediction model includes a CNN convolutional neural network and a gated recurrent unit, wherein the CNN convolutional neural network and the gated recurrent unit are connected through a flattening layer, wherein the flattening layer is used to convert the time series into an input format suitable for a subsequent neural network, the CNN convolutional neural network is used to extract features from wireless delay data to obtain feature information of the wireless delay data, and the gated recurrent unit is used to predict delays based on the feature information;

[0087] The CNN convolutional neural network includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer and an output layer connected in sequence, wherein the first convolutional layer and the second convolutional layer are used for feature information extraction, the first pooling layer and the second pooling layer are used for dimensionality reduction of the feature information extracted by the convolutional layer, and the fully connected layer is used for classifying and outputting the feature information;

[0088] The expression of the delay prediction process of the gated recurrent unit is:

[0089] z t = a ( U z [ h t − 1 , R t ] ) ;

[0090] r t = a ( U r [ h t − 1 , R t ] ) ;

[0091] h ˜ t = tanh( U h [ r t h t − 1 , R t ] ) ;

[0092] ;

[0093] in, To update the gate, To reset the gate, is the multiplication result of the input and the hidden layer at the previous moment, is the hidden layer output, is the activation function, is the hidden layer output at the previous moment, , , are weight matrices, is the input of the neural network;

[0094] The hyperparameters of the gated recurrent unit are optimized using the particle swarm algorithm, and the optimization process is as follows:

[0095] Initial particles are randomly generated, and the positions of the randomly generated initial particles are mapped to new positions through Tent chaos mapping, where the Tent chaos mapping formula is:

[0096] ;

[0097] in, is the particle position after chaos, are randomly generated initial particles, is a constant, ;

[0098] Input the hyperparameters corresponding to the initial particles into the gated recurrent unit, output the prediction results, and calculate the current fitness;

[0099] Inertia weight is introduced, and the inertia weight is updated through the adaptive algorithm. Then the particle speed and its position are updated according to the individual extreme value and the global extreme value, and the hyperparameters are updated. The expression of inertia weight is:

[0100] ;

[0101] in, is the inertia weight, is the maximum inertia weight, is the minimum inertia weight, is the current fitness value, is the minimum value of fitness, is the average value of fitness;

[0102] The expression of particle update speed is:

[0103] ;

[0104] in, For the Particles in After iterations, The speed in dimension, is the inertia weight, For the Particles in After iterations, The speed in dimension, and are learning factors, and are randomly generated numbers. and For the Particles in The optimal solution after iteration is The position in dimension and the global optimal solution are The position in dimension, For the Particles in After iterations, Position in dimension;

[0105] The expression for updating the particle position is:

[0106] ;

[0107] in, For the Particles in After iterations, The position in dimension, For the Particles in After iterations, The position in dimension, For the Particles in After iterations, Speed ​​in dimension;

[0108] Use the current hyperparameters to predict the delay and calculate the fitness to determine whether the fitness is less than the previous fitness. If so, repeatedly update the inertia weight through the adaptive algorithm, and then update the particle speed and its own position according to the individual extreme value and the global extreme value, and update the hyperparameters and use the current hyperparameters to predict the delay and calculate the fitness to determine whether the fitness is less than the previous fitness until the maximum number of iterations is reached. If not, output the current hyperparameters as the optimal hyperparameters;

[0109] The fitness calculation expression is:

[0110] ;

[0111] in, is the current fitness value, is the difference between the ideal speed and the actual speed of the carriage, is the total number of iterations, is the current iteration number;

[0112] S3: Use the BP neural network to detect the health of the traction / brake actuator and obtain the health data of the traction / brake actuator. Specifically, initialize the weight and offset of the BP neural network model, use the actual speed, displacement and acceleration of the train as the input of the BP neural network, use the health of the traction / brake actuator: 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100% as the output of the BP neural network, use the gradient descent method to train the BP neural network, use the trained BP neural network to detect the health of the traction / brake actuator, and obtain the health data of the traction / brake actuator;

[0113] S4: Establish a train dynamics model. Specifically, consider the force between carriages, the health data of traction / brake actuators, basic resistance data, additional resistance data and mass rotation coefficient to establish a train dynamics model, which is expressed as follows:

[0114] ;

[0115] in, For the The displacement of the train, For the The speed of the train, For the The mass of a train carriage, For the The health status of the train traction / brake actuators, is the mass rotation coefficient, For the The force output by the train controller, For the The resistance of the train during operation, , is the force between adjacent carriages, is the elastic coupling coefficient of the spring of the spring-damper system, is the damping coefficient of the spring-damper system, and represent the displacement difference and speed difference between adjacent carriages, respectively. For the current moment, For time delay, is the force between adjacent carriages;

[0116] S5: Control the train according to the network delay obtained in step S2, the health data of the traction / brake actuator obtained in step S3, and the control rate of the adaptive sliding mode fault-tolerant controller designed and improved by the train dynamics model established in step S4; specifically, the following steps are included:

[0117] S501: According to the train dynamics model, the displacement error and speed error are obtained, and their expressions are:

[0118] ;

[0119] in, is the error between the ideal displacement and the actual displacement, is the error between the ideal speed and the actual speed, is the ideal displacement, is the actual displacement, For the ideal speed, is the actual speed;

[0120] S502: Obtain the sliding surface function, whose expression is:

[0121] ;

[0122] in, is the sliding surface function, Indicates the serial number, is the coefficient for adjusting the approach speed, and , is the error between the ideal displacement and the actual displacement, is the error between the ideal speed and the actual speed;

[0123] S503: Obtain the mixed approach rate, which is expressed as:

[0124] ;

[0125] in, Indicates the serial number, , , , , , are coefficients for adjusting the approach speed. is the sliding surface;

[0126] S504: Obtain the control rate of the adaptive sliding mode fault-tolerant controller according to the sliding surface function and the mixed approach rate, and its expression is:

[0127] ;

[0128] in, is the force output by the controller, For the current moment, is the predicted network delay, , It is the inverse of the product of the acceleration factor and the health of the traction / brake actuator. is the acceleration factor, The health of the traction / brake actuator. , , , , , To adjust the coefficient of approach speed, is the sliding surface function, To adjust the coefficient of approach speed, is the error between the ideal displacement and the actual displacement, is the intermediate quantity, For the ideal speed, is the basic resistance, is a symbolic function;

[0129] S505: Considering the influence of external environmental factors on basic resistance, the basic resistance considering external environmental factors is obtained by RBF neural network fitting, and its expression is:

[0130] ;

[0131] in, In order to consider the basic resistance of external environmental factors, is the estimated weight of the RBF neural network, is the activation function, Estimate the difference for the RBF neural network;

[0132] S506: Construct an improved adaptive sliding mode fault-tolerant controller control rate model taking into account external environmental factors, and its expression is:

[0133] ;

[0134] in, is the force output by the controller, For the current moment, is the predicted network delay, , It is the inverse of the product of the acceleration factor and the health of the traction / brake actuator. is the acceleration factor, The health of the traction / brake actuator. , , , , , To adjust the coefficient of approach speed, is the sliding surface function, To adjust the coefficient of approach speed, is the error between the ideal displacement and the actual displacement, For the ideal speed, is the basic resistance, is the symbolic function, is the estimated weight of the RBF neural network, is the activation function, Estimate the difference for the RBF neural network;

[0135] S507: Obtain the adaptive rate of the hyperparameter, the expression of which is:

[0136] ;

[0137] ;

[0138] ;

[0139] in, for Moment The adaptive rate, is the constant that controls the change of the adaptive rate, is the sliding surface function, is the symbolic function, The health of the traction / brake actuator. is the adaptive rate of the weight, is the constant that controls the change of the adaptive rate, is the acceleration factor, is the activation function, is the adaptive rate of the offset, To control the change of the adaptation rate;

[0140] The step S6 is also included: performing stability verification on the designed adaptive sliding mode fault-tolerant controller. The step S6 specifically includes the following steps:

[0141] S601: Design a Lyapunov function, whose expression is:

[0142] ;

[0143] in, is the Lyapunov function, is the sliding surface function, is the acceleration factor, is the constant that controls the change of the adaptive rate, For the The difference between the ideal value and the estimated value of the vehicle, is the constant that controls the change of the adaptive rate, is the transpose of the difference in RBFNN weights, is the difference of RBFNN weights, is the constant that controls the change of the adaptive rate, is the difference of RBFNN offset, is the ideal value, is an estimated value, is the difference between the ideal value and the estimated value, is the ideal weight of RBF neural network, Estimate weights for the RBF neural network, is the difference between the estimated weight and the ideal weight of the RBF neural network, is the ideal difference of RBF neural network, is the RBF neural network estimation difference, is the difference between the RBF neural network estimated difference and the ideal difference, for the current moment;

[0144] S602: Derivative of the Lyapunov function, the expression is:

[0145] ;

[0146] in, is the derivative of the Lyapunov function, is the sliding surface function, is the derivative of the sliding surface function, is the acceleration factor, is the constant that controls the change of the adaptive rate, For the The difference between the ideal value and the estimated value of the vehicle, for The adaptive rate, is the constant that controls the change of the adaptive rate, is the difference between the ideal weight and the actual weight of RBFNN, is the derivative of the expected weight of RBFNN, is the constant that controls the change of the adaptive rate, is the difference between the ideal offset and the actual offset, is the derivative of the ideal offset, for the current moment;

[0147] S603: Substitute the improved adaptive sliding mode fault-tolerant controller control rate model into the Lyapunov function derivative expression obtained in step S602 to obtain the following expression:

[0148] ;

[0149] in, is the derivative of the Lyapunov function, is the sliding surface function, To adjust the coefficient of approach speed, is the difference between the ideal speed and the actual speed, is the derivative of the ideal velocity, The health of the traction / brake actuator. is the force output by the controller, is a constant, is the resistance of the train running, is the constant that controls the change of the adaptive rate, is the constant that controls the change of the adaptive rate, is the constant that controls the change of the adaptive rate, For the The difference between the ideal value and the estimated value of the vehicle, for The adaptive rate, is the transpose of the difference in RBFNN weights, is the adaptive rate of the weight, is the difference between the ideal offset and the actual offset, is the adaptive rate of the offset, is the intermediate quantity, is the derivative of the ideal velocity, To adjust the coefficient of approach speed, is the expected train running resistance, is the symbolic function, is the acceleration factor, is the sliding surface, is the activation function, is the difference between the RBF neural network estimated difference and the ideal difference, For the current moment.

[0150] After the 8 vehicles are controlled by this method, their speed tracking curves are as follows: Figure 3-Figure 6 Shown by Figure 3-Figure 6 It can be seen that the actual speed is closer to the expected speed, which proves that the speed tracking of the train is more accurate when the basic resistance, additional resistance and actuator health are combined.

[0151] The present invention is described by way of embodiments, and those skilled in the art will appreciate that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the protection scope of the present invention.

Claims

1. A train wireless network delay compensation adaptive sliding mode fault-tolerant control method, characterized in that: The following steps are involved: S1: Collect the wireless delay data of the train wireless communication equipment by intercepting the timestamp of the data sending time; S2: Build a CNN-GRU delay prediction model, input the wireless delay data into the CNN-GRU delay prediction model for prediction, and obtain the network delay; S3: using a BP neural network to detect the health of the traction / brake actuator to obtain health data of the traction / brake actuator; S4: Establishing a train dynamics model; Establishing a train dynamics model specifically includes: considering the force between carriages, the health data of the traction / brake actuator, the basic resistance data, the additional resistance data and the mass rotation coefficient, establishing a train dynamics model, and its expression is: in, is the displacement of the jth train, v j is the speed of the jth train, m i is the mass of the i-th train carriage, λ i is the health status of the traction / brake actuator of the i-th train, r is the mass rotation coefficient, F i is the force output by the controller of the i-th train, f i is the resistance of the ith train during its operation, g i(i+1) =αΔx+βΔv,g i(i+1) is the force between adjacent carriages, α is the elastic coupling coefficient of the spring of the spring-damper system, β is the damping coefficient of the spring-damper system, Δx and Δv represent the displacement difference and velocity difference between adjacent carriages, t is the current moment, τ is the time delay, and g (i-1)i is the force between adjacent carriages; S5: Control the train according to the network delay obtained in step S2, the health data of the traction / brake actuator obtained in step S3, and the control rate of the adaptive sliding mode fault-tolerant controller designed and improved by the train dynamics model established in step S4, specifically including the following steps: S501: According to the train dynamics model, the displacement error and speed error are obtained, and their expressions are: Among them, e i (t) is the error between the ideal displacement and the actual displacement, is the error between the ideal speed and the actual speed, x di (t) is the ideal displacement, x i (t) is the actual displacement, v di (t) is the ideal speed, v i (t) is the actual speed; S502: Obtain the sliding surface function, whose expression is: Among them, s i (t) is the sliding surface function, i represents the serial number, c is the coefficient for adjusting the approach speed, and c>0, e i (t) is the error between the ideal displacement and the actual displacement, is the error between the ideal speed and the actual speed; S503: Obtain the mixed approach rate, which is expressed as: Among them, i represents the sequence number, xite, k, eq 1i ,eq 2i 、h 1i 、h 2i are coefficients for adjusting the approach speed, and s is the sliding surface; S504: Obtain the control rate of the adaptive sliding mode fault-tolerant controller according to the sliding surface function and the mixed approach rate, and its expression is: Among them, u i (t+τ) is the force output by the controller, t is the current time, τ is the predicted network delay, is the inverse of the product of the acceleration coefficient and the health of the traction / brake actuator, ψ is the acceleration coefficient, δ i (t) is the health status of the traction / brake actuator, xite, k, eq 1i ,eq 2i 、h 1i 、h 2i is the coefficient for adjusting the approach speed, s i (t) is the sliding surface function, c is the coefficient for adjusting the approach speed, e i (t) is the error between the ideal displacement and the actual displacement, θ is the intermediate value, v di (t) is the ideal speed, f b (t) is the basic resistance, sgn() is the sign function; S505: Considering the influence of external environmental factors on basic resistance, the basic resistance considering external environmental factors is obtained by RBF neural network fitting, and its expression is: Among them, f bi (t) is the basic resistance to external environmental factors. is the estimated weight of the RBF neural network, φ i (t) is the activation function, Estimate the difference for the RBF neural network; S506: Design an improved adaptive sliding mode fault-tolerant controller control rate taking into account external environmental factors, and its expression is: Among them, u i (t+τ) is the force output by the controller, t is the current time, τ is the predicted network delay, is the inverse of the product of the acceleration coefficient and the health of the traction / brake actuator, ψ is the acceleration coefficient, δ i (t) is the health status of the traction / brake actuator, xite, k, eq 1i ,eq 2i 、h 1i 、h 2i is the coefficient for adjusting the approach speed, s i (t) is the sliding surface function, c is the coefficient for adjusting the approach speed, e i (t) is the error between the ideal displacement and the actual displacement, v di (t) is the ideal speed, f b (t) is the basic resistance, sgn() is the sign function, is the estimated weight of the RBF neural network, φ i (t) is the activation function, Estimate the difference for the RBF neural network; S507: Obtain the adaptive rate of the hyperparameter, the expression of which is: in, is the time t The adaptation rate, r ζ is the constant that controls the change of the adaptive rate, s i (t) is the sliding surface function, sgn() is the sign function, δ i (t) is the health of the traction / brake actuator, is the adaptive rate of the weight, r w is the constant that controls the change of the adaptive rate, ψ is the acceleration coefficient, φ i (t) is the activation function, is the adaptive rate of the offset, r ε A constant that controls the adaptation rate.

2. The method for adaptive sliding mode fault-tolerant control of train wireless network delay compensation according to claim 1, characterized in that: The CNN-GRU delay prediction model in step S2 includes a CNN convolutional neural network and a gated recurrent unit, which are connected through a flattening layer. The flattening layer is used to convert the time series into an input format suitable for the subsequent neural network. The CNN convolutional neural network is used to extract features from the wireless delay data to obtain feature information of the wireless delay data, and the gated recurrent unit is used to predict the delay through the feature information.

3. The method for adaptive sliding mode fault-tolerant control of train wireless network delay compensation according to claim 2, characterized in that: The CNN convolutional neural network includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer and an output layer connected in sequence, wherein the first convolutional layer and the second convolutional layer are used to extract feature information, the first pooling layer and the second pooling layer are used to reduce the dimension of the feature information extracted by the convolutional layer, and the fully connected layer is used to classify and output the feature information.

4. The method for adaptive sliding mode fault-tolerant control of train wireless network delay compensation according to claim 2, characterized in that: The expression of the delay prediction process of the gated recurrent unit is: z t =a(U z [h t-1 ,R t ]); r t =a(U r [h t-1 ,R t ]); Among them, z t is the update gate, r t To reset the gate, is the multiplication result of the input and the hidden layer at the previous moment, h t is the hidden layer output, a is the activation function, h t-1 is the hidden layer output at the previous moment, U z , U r , U h are weight matrices, R t is the input of the neural network.

5. The method for adaptive sliding mode fault-tolerant control of train wireless network delay compensation according to claim 2, characterized in that: The hyperparameters of the gated recurrent unit are optimized using the particle swarm algorithm, and the optimization process is as follows: Initial particles are randomly generated, and the positions of the randomly generated initial particles are mapped to new positions through Tent chaos mapping, where the Tent chaos mapping formula is: in, is the particle position after chaos, is the initial particle generated randomly, Θ is a constant, Θ = 0.5; Input the hyperparameters corresponding to the initial particles into the gated recurrent unit, output the prediction results, and calculate the current fitness; Inertia weight is introduced, and the inertia weight is updated through the adaptive algorithm. Then the particle speed and its position are updated according to the individual extreme value and the global extreme value, and the hyperparameters are updated. The expression of inertia weight is: Among them, ω is the inertia weight, ω max is the maximum inertia weight, ω min is the minimum inertia weight, fit is the current fitness value, fit min is the minimum value of fitness, fit avg is the average value of fitness; The expression of particle update speed is: in, is the velocity of the βth particle in the η dimension after B+1 iterations, ω is the inertia weight, is the velocity of the βth particle in the η dimension after B iterations, C1 and C2 are both learning factors, rand1 and rand2 are randomly generated numbers, and is the position of the optimal solution of the βth particle after B iterations in the η dimension and the position of the global optimal solution in the η dimension, is the position of the βth particle in the η dimension after B iterations; The expression for updating the particle position is: in, is the position of the βth particle in the η dimension after B+1 iterations, is the position of the βth particle in the η dimension after B iterations, is the velocity of the βth particle in the η dimension after B+1 iterations; Use the current hyperparameters to predict the delay and calculate the fitness to determine whether the fitness is less than the previous fitness. If so, repeatedly update the inertia weight through the adaptive algorithm, and then update the particle's speed and position according to the individual extreme value and the global extreme value, and update the hyperparameters and use the current hyperparameters to predict the delay and calculate the fitness to determine whether the fitness is less than the previous fitness until the maximum number of iterations is reached. If not, output the current hyperparameters as the optimal hyperparameters.

6. The method for adaptive sliding mode fault-tolerant control of train wireless network delay compensation according to claim 5, characterized in that: The fitness calculation expression is: Among them, fit is the current fitness value, ξ i is the difference between the ideal speed and the actual speed of the carriage, n is the total number of iterations, and i is the current number of iterations.

7. The method for adaptive sliding mode fault-tolerant control of train wireless network delay compensation according to claim 1, characterized in that: In step S3, the BP neural network is used to detect the health of the traction / brake actuator. Specifically, the weight and offset of the BP neural network model are initialized, the actual speed, displacement and acceleration of the train are used as the input of the BP neural network, and the health of the traction / brake actuator: 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100% is used as the output of the BP neural network. The BP neural network is trained by the gradient descent method, and the trained BP neural network is used to detect the health of the traction / brake actuator to obtain the health data of the traction / brake actuator.

8. The method for adaptive sliding mode fault-tolerant control of train wireless network delay compensation according to claim 1, characterized in that: After step S5, the method further includes step S6: performing stability verification on the designed adaptive sliding mode fault-tolerant controller. Step S6 specifically includes the following steps: S601: Design a Lyapunov function, whose expression is: Where V is the Lyapunov function, s i (t) is the sliding surface function, ψ is the acceleration coefficient, r ζ is the constant that controls the change of the adaptive rate, is the difference between the ideal value and the estimated value of the i-th vehicle, r w is the constant that controls the change of the adaptive rate, is the transpose of the difference in RBF neural network weights, is the difference of RBF neural network weights, r ε is the constant that controls the change of the adaptive rate, is the difference of RBF neural network offset, ζ is the ideal value, is an estimated value, is the difference between the ideal value and the estimated value, W is the ideal weight of the RBF neural network, Estimate weights for the RBF neural network, is the difference between the estimated weight and the ideal weight of the RBF neural network, ε is the ideal difference of the RBF neural network, is the RBF neural network estimation difference, is the difference between the RBF neural network estimated difference and the ideal difference, and t is the current moment; S602: Derivative of the Lyapunov function, the expression is: in, is the derivative of the Lyapunov function, s i (t) is the sliding surface function, is the derivative of the sliding surface function, ψ is the acceleration coefficient, r ζ is the constant that controls the change of the adaptive rate, is the difference between the ideal value and the estimated value of the i-th vehicle, for The adaptation rate, r w is the constant that controls the change of the adaptive rate, is the difference between the ideal weight and the actual weight of the RBF neural network, is the derivative of the expected weight of the RBF neural network, r ε is the constant that controls the change of the adaptive rate, is the difference between the ideal offset and the actual offset, is the derivative of the ideal offset, t is the current moment; S603: Substitute the improved adaptive sliding mode fault-tolerant controller control rate model into the Lyapunov function derivative expression obtained in step S602 to obtain the following expression: in, is the derivative of the Lyapunov function, s i (t) is the sliding surface function, c is the coefficient for adjusting the approach speed, is the difference between the ideal speed and the actual speed, is the derivative of the ideal velocity, δ i (t) is the health of the traction / brake actuator, u i (t) is the force output by the controller, ψ i is a constant, f i (t) is the resistance of train operation, r ζ is the constant that controls the adaptation rate change, r w is the constant that controls the change of the adaptive rate, r ε is the constant that controls the change of the adaptive rate, is the difference between the ideal value and the estimated value of the i-th vehicle, for The adaptive rate, is the transpose of the difference in RBF neural network weights, is the adaptive rate of the weight, is the difference between the ideal offset and the actual offset, is the adaptive rate of the offset, θ i (t) is the intermediate quantity, is the derivative of the ideal speed, k is the coefficient for adjusting the approach speed, is the expected train running resistance, sgn() is the sign function, ψ is the acceleration coefficient, s i is the sliding surface, φ i (t) is the activation function, is the difference between the RBF neural network estimated difference and the ideal difference, and t is the current moment.

Images