Data-driven train fault-tolerant tracking control method

Abstract

The invention belongs to the technical field of train speed tracking control, and discloses a data-driven train fault-tolerant tracking control method. According to the method, data driving is combined under a fault-tolerant control framework, all-directional and high-precision monitoring of the running state of the train actuator is achieved, once an abnormal or potential fault is detected, a control system can immediately trigger an emergency response mechanism, accidents are effectively avoided, and the safety guarantee of the train is improved. Besides, fault-tolerant control is added on the basis of the data driving model, so that the expected speed curve can still be quickly adjusted and tracked in a fault scene, and accurate speed tracking is realized. In addition, by combining the data driving technology and fault-tolerant control, the optimal output driving force can be obtained by optimizing system output according to the fault condition, the energy consumption during train operation is remarkably reduced, the operation cost is reduced, the burden of the train system on the environment is relieved, and the train energy consumption is reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of train speed tracking control and relates to a data-driven fault-tolerant train tracking control method. Background Art

[0002] At present, rail transit trains detect faults by manually detecting vibration, abnormal sound, instrument panel indication, etc., or through the automated monitoring system equipped by the train. The monitoring system will monitor the speed, position, traction and other key parameters of the train according to the axle counter, transponder, speed sensor and other equipment, and compare them with the preset safety threshold. Once an abnormality is found, the system will immediately issue an alarm and display the type and location of the fault. When a system or component is detected to have a fault, the train will first try to automatically isolate the faulty component and dynamically adjust the control strategy of the train according to the type of fault and the scope of impact. In the fault-tolerant control mode, the train can continue to operate in a degraded mode, even if some functions are restricted, but still ensure the safety of passengers. For example, the system may limit the maximum speed of the train or adjust the acceleration / deceleration to adapt to the current fault condition. The communication between the train and the ground control center ensures the transmission of real-time information. The ground control center can dynamically adjust and optimize the driving strategy according to the latest status information and fault conditions, ensuring the safety and efficiency of the train operation.

[0003] In the current train fault-tolerant control mode, there is a delay in the isolation response after fault detection, and the system cannot quickly switch to the redundant module. It relies on an accurate physical model, which is difficult to implement in a complex rail transit system. The switching and coordination of redundant modules can easily lead to system instability, requiring additional control logic and resources, which has an unignorable impact on the safe operation of the train. Specifically, the existing train control methods mainly rely on the problems of classical control strategies such as traditional PID control and LQR control. These methods perform well under normal train operation, but when encountering faults, such as actuator failure or sensor failure, it is difficult to quickly adjust the control strategy, resulting in a decrease in the safety and stability of train operation. In addition, since these methods rely on accurate physical models to describe the dynamic behavior of the train, and establishing an accurate physical model requires a lot of prior knowledge and a complex parameter identification process. In practical applications, the operating environment of the train is complex and changeable, and the physical model is difficult to fully and accurately reflect the actual operating status of the train, which makes the traditional control method have great limitations in fault tolerance performance.

[0004] With the development of data-driven technology, data-driven models are becoming an important means in the field of control, which can enable trains to predict and optimize their operating status more accurately. Data-driven models collect train operating data, including position, speed, acceleration and other information, and use advanced algorithms to analyze and model these data. During the operation of the train, the data-driven model can update the train's status prediction in real time and generate the optimal control strategy based on the current operating conditions and expected future status. Train control based on these strategies can not only improve operating efficiency but also enhance safety.

[0005] Compared with traditional train control methods, data-driven models can better adapt to complex operating environments. By mining and utilizing big data, they can achieve coordinated control between trains, thereby improving the overall operational efficiency of the rail transit system. Although the current data-driven model predictive control scheme for optimized train operation can use data-driven methods to build a linear approximate model of train dynamics, it still has deficiencies in fault detection and processing. Traditional data-driven train control models mainly rely on fixed training data sets. When a fault occurs, the control strategy is difficult to quickly adapt to the new operating state because the model fails to be updated in real time. In addition, when formulating the control strategy, this method fails to fully consider the changes in the real-time operating environment and the impact of different fault types, making it difficult to quickly respond to sudden faults that occur during train operation. Summary of the invention

[0006] In response to the above-mentioned problems existing in the existing data-driven control schemes, the present invention proposes a data-driven fault-tolerant train tracking control method, which enables the speed tracking controller to adapt to and compensate for faults in real time, so that when encountering sudden faults during train operation, the control model adjustment strategy can be changed in time, thereby ensuring the efficiency and safety of train operation.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme: A data-driven train fault-tolerant tracking control method comprises the following steps: Step 1. Obtain the initial historical operation data of the train, and use the data-driven method combined with the deep neural network to train the initial historical operation data to extract operators and build a linear dynamic system model in a high-dimensional space; Among them, the lifting function obtained by the DNN neural network maps the train nonlinear control model to a linear system in a high-dimensional space, and the dynamic mode decomposition method is used to iteratively solve the (A, B, C) matrix of the state space to construct the initial train data-driven linear state space model; A is the state transfer matrix, B is the control input matrix, and C is the state projection matrix; Step 2. Based on the train data driven linear state space model obtained in step 1, the train data driven linear state space model is updated according to the actuator failures and additional failures that may occur in the train to obtain a fault-tolerant train data driven linear state space model, and the model is integrated into the MPC framework to replace the fault detection unit of the train; Step 3. After the ATO system has been running for a period of time, the fault-tolerant control matrix is ​​initialized by the dynamic mode decomposition method using the train operation data obtained during this period. and , and establish a fixed-size sliding window of train operation status data, and recursively update the fault-tolerant control matrix in real time based on a set of system snapshots taken within a given time window; Step 4. For each train control cycle, the train sensors are used to obtain the state information required for training, and the state information is compared with the fault-tolerant control matrix at the previous moment. Recursively update the fault-tolerant control matrix at this moment and use it to update the train data driven linear state space model for model predictive control; Model predictive control is used to obtain the control law sequence by minimizing the cost function of MPC according to the train's motion objectives and constraints. The first element of the control law sequence is used as the traction / braking force to drive the train. After the train responds, the new state quantity is measured and the calculation of the next cycle is started for continuous rolling optimization.

[0008] In addition, the present invention also proposes a computer device, which includes a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement the data-driven train fault-tolerant tracking control method mentioned above.

[0009] In addition, the present invention also proposes a computer-readable storage medium on which a program is stored. When the program is executed by a processor, it is used to implement the data-driven fault-tolerant train tracking control method mentioned above.

[0010] The present invention has the following advantages: As described above, the present invention relates to a data-driven train fault-tolerant tracking control method, which combines data-driven technology under the fault-tolerant control framework to achieve all-round and high-precision monitoring of the operating status of the train actuator. Once an abnormality or potential fault is detected, the control system can immediately trigger an emergency response mechanism, thereby effectively avoiding the occurrence of an accident, providing passengers and staff with a higher level of safety protection, and improving train safety protection. In addition, compared with the traditional data-driven control method, when a train fails, it cannot judge the system fault, and will not adjust the control model that can use the fault, thereby resulting in the output driving force / braking force being unable to make the train unable to accurately track the speed curve. The method of the present invention adds fault-tolerant control on the basis of the data-driven model, and can still quickly adjust and track the desired speed curve in the fault scenario, providing passengers with a comfortable riding environment, and achieving accurate speed tracking. In addition, in view of the problem that the control system fails to identify the fault of the train actuator, which will lead to the output of excessive driving force, thereby leading to a decrease in energy utilization, the method of the present invention combines data-driven technology and fault-tolerant control, and can optimize the system output according to the fault situation to obtain the optimal output driving force, significantly reducing the energy consumption of the train during operation, reducing operating costs and reducing the burden of the train system on the environment, and reducing the energy consumption of the train. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flowchart of constructing a train linear model using a data-driven approach in an embodiment of the present invention.

[0012] Figure 2 Schematic diagram of the sliding window of train operation data in an embodiment of the present invention.

[0013] Figure 3 The figure is an overall flow chart of the data-driven train fault-tolerant tracking control method in an embodiment of the present invention.

[0014] Figure 4 Schematic diagram of train speed tracking comparison under normal conditions in an embodiment of the present invention.

[0015] Figure 5 Schematic diagram of train speed error comparison under normal conditions of the actuator in an embodiment of the present invention.

[0016] Figure 6 Schematic diagram of train displacement tracking comparison under normal conditions of the actuator in an embodiment of the present invention.

[0017] Figure 7 Schematic diagram of train speed tracking comparison in the case of actuator failure in an embodiment of the present invention.

[0018] Figure 8 Schematic diagram of train speed tracking error comparison in the case of actuator failure in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 In view of the above technical problems existing in the existing data-driven control schemes, the present invention discloses a data-driven train fault-tolerant tracking control method. The method of the present invention uses the historical operation data of the system to build a data-driven model to capture the nonlinear behavior and fault characteristics of the system. The fault impact is resolved through online estimation of the data-driven model. The method of the present invention allows the adjustment of the Model Predictive Control (MPC) control law to offset the fault impact. By adding an online update process, the data-driven model can characterize the dynamic characteristics of the faulty system. Therefore, the MPC controller can adapt to and compensate for the fault in real time, and integrate the data-driven linear state space model of the train with fault tolerance into the MPC controller without the need for a separate fault detection unit, highlighting the controller's ability to mitigate faults and maintain desired system behavior.

[0020] like Figure 1 and Figure 3 As shown, the data-driven train fault-tolerant tracking control method in this embodiment includes the following steps: Step 1. Obtain the initial historical operation data of the train, and use the data-driven method combined with the deep neural network to train the initial historical operation data to extract operators and build a linear dynamic system model in high-dimensional space.

[0021] The lifting function obtained by the DNN neural network is used to map the train nonlinear control model into a linear system in a high-dimensional space. The extended dynamic mode decomposition (EDMD) method is used to iteratively solve the (A, B, C) matrix of the state space to construct the initial train data-driven linear state space model.

[0022] like Figure 1 As shown, firstly, the historical operation data information of the ATO system when the train is running is obtained, including actual state information and control information, and the historical operation data information is used to form a training set.

[0023] The actual state information includes speed, position, etc., and the control information includes driving force / braking force and corresponding instruction percentage, etc.

[0024] The input of the neural network training is a combination of state information and control information, and the output is the state information at the next moment. The training set data is preprocessed, normalized, and denoised so that the nonlinear dynamic information of the train can be fully reflected.

[0025] Build a DNN neural network to learn the observation function to express system characteristics, set the training time, batch size, and learning rate parameters, and save the network parameters after determining whether the maximum training time has been reached or the training error is less than the set value.

[0026] After obtaining the observation function, define a finite observable vector for: (1) in, Indicates the extended state. Represents the first element of the input sequence, Indicates the train status; The state observation function can be expressed as follows: (2) in, It is a boosting function trained by the DNN neural network, used to capture the train operation characteristics. is the size of the dimension being promoted.

[0027] The extended dynamic mode decomposition method is used to calculate a finite-dimensional operator, which is approximated as , the dynamic mode decomposition method solves the least squares problem by minimizing the following cost function and obtains an approximate closed form estimate of the operator: (3) in represents the koopman operator, represents the operator dimension, represents the dimensional space of the operator, Indicates the size of the dataset, represents the train status at time i, represents the train status of i at the next moment, express The observation function of express The observation function of Represents the system output traction / braking force at time i.

[0028] In the optimization iteration process of the above formula, the state transfer matrix A and the control input matrix B of the train data-driven linear state space model are obtained. The state projection C of the model is optimized by minimizing the square difference between the predicted value and the actual output value. The expression is as follows: (4) The dataset obtained from the historical train operation data is described as: (5) (6) (7) in Represents the train status information sequence, Indicates the state at a certain moment, consisting of speed and position. It represents the train status information sequence at the next moment after U acts. Indicates the train output traction force / braking force sequence, Indicates the output traction / braking force at a certain moment.

[0029] Using the above least squares problem formula, we get: (8) in , Represents the use of observation function Centralize the data , The data after dimensionality upgrade. express and Construct the pseudo-inverse of the matrix, express The pseudoinverse is the inverse matrix of a non-square matrix.

[0030] After obtaining the matrices of the train data-driven linear state space model, we define , construct the train data driven linear state space model as follows: (9) Among them, A is the state transfer matrix, B is the control input matrix, C is the state projection matrix, is the state variable of the system after dimensionality increase, , where N is The dimension of is an estimate of the original nonlinear system state x(k).

[0031] , , is a linear time invariant matrix.

[0032] is the system input, which indicates the traction or braking force received by the train during operation. . Where m represents the dimension of the control input, n is the dimension of the train state, and k is the discrete time step.

[0033] The train data-driven linear state space model can be used to design an MPC controller to obtain a stable control effect.

[0034] Step 2. Based on the train data-driven linear state space model obtained by using the train's historical operation data, the train data-driven linear state space model is updated according to the actuator failures and additional failures that may occur in the train to obtain a fault-tolerant train data-driven linear state space model, which is integrated into the MPC framework to replace the train's fault detection unit.

[0035] In order to add fault-tolerant control effects to the control system, we first obtain the information of the train actuator, consider the actuator failure, and model the system input to obtain the following fault input model: (10) in Represents the equivalent input when the actuator fails, the matrix Used to represent the actuator state, which is defined as: (11) matrix is a diagonal matrix, Indicates an additional fault of the corresponding actuator.

[0036] matrix The value of each diagonal element in The value between 0 and 1 indicates the degree of the corresponding actuator failure, where 0 indicates complete failure and 1 indicates that the actuator is working normally.

[0037] Substituting the fault input model in formula (10) into the train data driven linear state space model in step 1, we obtain a fault-tolerant train data driven linear state space model, which is expressed as follows: (12) (13) in, represents the state variable of the system at time k+1 after dimensionality increase, represents the state variable of the system at time k after dimensionality increase, Represents the fault-tolerant control matrix; At this point, the fault-tolerant train data-driven control model of the data-driven MPC control framework is obtained, and the matrix As the train's fault status is continuously updated to respond to sudden failures.

[0038] Step 3. After the ATO system has been running for a period of time, the fault-tolerant control matrix is ​​initialized by the dynamic mode decomposition method using the train operation data obtained during this period. and , and establish a fixed-size sliding window of train operation status data, such as Figure 2As shown, the fault-tolerant control matrix is ​​recursively updated in real time according to a set of system snapshots taken within a given time window.

[0039] Run the train system for a while The running data of time constructs an observable sequence of length a+1 , arrange the observable sequence into the following matrix: (14) (15) in, and Represent the state operation data matrix and the input sequence data matrix respectively, which are used to calculate the fault-tolerant control matrix ; , , Indicates the weight of the data in the calculation.

[0040] The dynamic mode decomposition (EDMD) method is used. The EDMD method uses the lifting function trained by the neural network to map the low-dimensional state space to the high-dimensional space, which can better describe the dynamics of the system in the high-dimensional space.

[0041] The EDMD method first uses the following formula at each time step k: (16) Solving the least squares problem by the above formula gives the matrix The calculation formula is as follows: (17) Initialize the fault-tolerant control matrix using the above formula After that, a data sliding window is established for Matrix update.

[0042] When accessing new data generated during train operation, the system is updated based on a set of system snapshots taken within a given time window. Matrix, most of the data in the window remains unchanged, only the first and last elements of the window are changed, and the remaining data elements are only shifted left by one unit.

[0043] definition: (18) (19) The input matrix is ​​represented as: (20) renew , which can be updated recursively and , new data in the observable sequence The matrix is ​​accessed at the next moment Can be recursively computed with new data.

[0044] Step 4. For each train control cycle, the train sensors are used to obtain the state information required for training, and the state information is compared with the fault-tolerant control matrix at the previous moment. Recursively update the fault-tolerant control matrix at this moment and use it to update the train data-driven linear state-space model for model predictive control.

[0045] Model predictive control is adopted. According to the motion objectives and constraints of the train, the control law sequence is obtained by minimizing the cost function of MPC. The first element of the sequence is used as the traction / braking force to drive the train. After the train responds, the new state quantity is measured and the calculation of the next cycle is entered for continuous rolling optimization.

[0046] After the calculation formula is simplified, the new train operation data of the window is updated through the following formula: (twenty one) Among them, a represents the window size, b represents the state subscript, and u represents the input subscript.

[0047] (twenty two) in Represents the identity matrix.

[0048] (twenty three) (twenty four) (25) (26) matrix Update using the following formula: (27) Introducing the Fault Tolerant Control Matrix Later, the controller was made more tolerant to any unnecessary deviations. This fault-tolerant control matrix It consists of two parts, namely and , represents the actual input matrix, represents the additive term for estimating the actuator state, can be considered as a known disturbance and is input into the MPC controller.

[0049] This therefore enhances the robustness of the controller to unexpected deviations, including additive faults and modeling errors.

[0050] In finding the fault tolerance matrix Finally, a fault-tolerant train data-driven linear state space model is obtained as the MPC control model, and then the objective function and constraints of the MPC optimization problem are defined according to the train operation control objectives.

[0051] (28) (29) Among them, J is the objective function, N represents the prediction window size of MPC, , and Used to penalize tracking errors and control input changes, represents the reference trajectory, represents the MPC prediction trajectory, Indicates the train traction / braking force change constraint; Indicates the actual running trajectory. It represents the dimension increase of the actual running trajectory. , , and Represents control input and its changes minimum and maximum boundaries.

[0052] By minimizing the objective function J, the optimal control law sequence for a period of time is obtained. , Indicates input quantities, select the first input quantity in the sequence As the driving force input of the train.

[0053] After the train control system responds, the status information is obtained again to enter the next cycle of calculation.

[0054] In addition, in order to enable those skilled in the art to more clearly understand the technical solution of the present invention, a simulation case is designed below to further illustrate the technical solution of the present invention, which mainly includes the following two parts: 1. Verify the performance of the established high-dimensional train model, compare the speed tracking effect with the general linear model of the train under normal operating conditions of the actuator, and evaluate indicators such as the accuracy and error of the model.

[0055] 2. Verify the performance of the data-driven fault-tolerant predictive controller and compare it with the ordinary data-driven MPC controller under the condition of 50% actuator failure to evaluate the speed and stability of the controller's active fault tolerance.

[0056] In terms of data-driven model training and comparison, the root mean square error is used to evaluate the performance of the proposed model.

[0057] (30) In the formula, RMSE represents the root mean square error, and are the simulation state and measurement state of the train speed, is the sampling time, and k is the number of measurement points. In order to obtain operating data, the sampling time is = 0.01 s, a discrete-time train model is used to represent the actual train dynamics, where the train mass m = 194.295 t, and the model is simplified to =-(2.031 +0.0622 * + 0.001807* ) / 194.295 + u / 194295, and then simulate 500 trajectories in 1000 sampling periods.

[0058] To obtain more comprehensive data, the control inputs are randomly generated uniformly in the interval [−1, 2], and the obtained control inputs are multiplied by 50 to be applied to the train. The initial state of each trajectory is randomly generated and uniformly distributed on the box [0, 70]. This data collection process generates and stores 500,000 data points. Construct matrices X and X of size 2×2×51 + and a matrix U of size 1 × 2 × 51. The general linear MPC method is used for comparison during the simulation. For the general linear MPC method, the prediction model is constructed as a local linearized model of the dynamics at the origin, and the linearized model is updated every 100 steps.

[0059] In the simulation test, constraints are imposed on the range of control variables and the corresponding increments. The prediction level and control level are = 20 and = 10, and the weights of model predictive control are , .

[0060] The fault-tolerant controller window size =200, weight factor =0.5.

[0061] The effectiveness and superiority of the controller are evaluated based on displacement and velocity tracking as well as velocity tracking error.

[0062] The speed tracking and error results of the fault-tolerant data-driven (KMPC) controller and the general linear MPC (LMPC) controller, as well as the displacement tracking such as speed tracking Figure 4 ,error Figure 5 and displacement tracking Figure 6 As shown in the figure, the global linear modeling strategy of high-speed trains based on the data-driven model has a higher modeling accuracy and is superior to the local linear modeling method in the state space. During the entire observation period, the data-driven fault-tolerant model is highly consistent with the actual data. The following Table 1 shows a comparison of the RMSE tracking of the actuator under normal conditions.

[0063] Table 1. Comparison of RMSE tracking of actuators under normal conditions

[0064] It can be seen from Table 1 that the displacement and velocity RMSE of the KMPC model are 0.0098 and 0.0036, respectively. The prediction accuracy of displacement and velocity is improved by 59.5% and 70%, respectively, compared with the linear state space model LMPC.

[0065] In terms of the comparison of the effect of data-driven fault-tolerant controller, this part evaluates the control effect of data-driven fault-tolerant predictive controller under the condition of actuator failure. Based on the above experiments, the actuator is simulated with a 50% effectiveness loss when the step length is 200 steps, and additional ( -100) additional faults.

[0066] Compared with the ordinary data-driven controller, that is, the controller without a fault-tolerant model, the speed tracking effect of the data-driven fault-tolerant predictive controller (referred to as fault-tolerant KMPC) and the ordinary data-driven controller (referred to as ordinary KMPC) under the same fault scenario is as follows: Figure 7 As shown, the speed tracking error is Figure 8 As shown in the figure, when the fault occurs in the simulation, the fault-tolerant control model can reduce the error caused by the fault through model update relatively quickly, and because the fault information is detected through data, it is more accurate than the ordinary data-driven model in the subsequent speed tracking. Table 2 shows the RMSE comparison of actuator fault tracking.

[0067] Table 2 Comparison of RMSE of actuator fault tracking

[0068] It can be seen from Table 2 that the displacement and velocity RMSE of the fault-tolerant KMPC model are 0.0063 and 0.0064, respectively, which are 67.9% and 73.8% higher than those of the ordinary KMPC. Therefore, in the case of actuator failure, the data-driven fault-tolerant train tracking control method described in the present invention exhibits excellent accuracy and stability.

[0069] Example 2 This embodiment 2 describes a computer device, which includes a memory and one or more processors. An executable code is stored in the memory. When the processor executes the executable code, it is used to implement the steps of the data-driven train fault-tolerant tracking control method in the above embodiment 1.

[0070] In this embodiment, the computer device is any device or apparatus with data processing capability, which will not be described in detail here.

[0071] Example 3 This embodiment 3 describes a computer-readable storage medium on which a program is stored. When the program is executed by a processor, it is used to implement the steps of the data-driven train fault-tolerant tracking control method in the above-mentioned embodiment 1.

[0072] The computer-readable storage medium may be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc., equipped on the device.

[0073] Of course, the above description is only a preferred embodiment of the present invention, and the present invention is not limited to the above embodiments. It should be noted that all equivalent substitutions and obvious deformation forms made by any technician familiar with the field under the guidance of this specification fall within the essential scope of this specification and should be protected by the present invention.

Claims

1. A data-driven fault-tolerant train tracking control method, characterized in that: The steps include: Step 1. Obtain the initial historical operation data of the train, and use the data-driven method combined with the deep neural network to train the initial historical operation data to extract operators and build a linear dynamic system model in a high-dimensional space; Among them, the lifting function obtained by the DNN neural network maps the train nonlinear control model to a linear system in a high-dimensional space, and the dynamic mode decomposition method is used to iteratively solve the (A, B, C) matrix of the state space to construct the initial train data-driven linear state space model; A is the state transfer matrix, B is the control input matrix, and C is the state projection matrix; Step 2. Based on the train data driven linear state space model obtained in step 1, the train data driven linear state space model is updated according to the actuator failures and additional failures that may occur in the train to obtain a fault-tolerant train data driven linear state space model, and the model is integrated into the MPC framework to replace the fault detection unit of the train; Step 3. After the ATO system has been running for a period of time, the fault-tolerant control matrix is ​​initialized by the dynamic mode decomposition method using the train operation data obtained during this period. and , and establish a fixed-size sliding window of train operation status data, and recursively update the fault-tolerant control matrix in real time based on a set of system snapshots taken within a given time window; Step 4. For each train control cycle, the train sensors are used to obtain the state information required for training, and the state information is compared with the fault-tolerant control matrix at the previous moment. Recursively update the fault-tolerant control matrix at this moment and use it to update the train data driven linear state space model for model predictive control; Model predictive control is used to obtain the control law sequence by minimizing the cost function of MPC according to the train's motion objectives and constraints. The first element of the control law sequence is used as the traction / braking force to drive the train. After the train responds, the new state quantity is measured and the calculation of the next cycle is started for continuous rolling optimization.

2. The data-driven fault-tolerant train tracking control method according to claim 1, characterized in that: The step 1 is specifically as follows: First, the historical operation data information of the ATO system when the train is running is obtained, including the actual state information, i.e., speed, position, and control information, i.e., driving force / braking force and corresponding instruction percentage, and the historical operation data information is used to form a training set; The input of neural network training is a combination of state information and control information, and the output is the state information at the next moment; the training set data is preprocessed, normalized, and denoised so that the nonlinear dynamic information of the train can be fully reflected; Build a DNN neural network to learn the characteristics of the observation function, set the training time, batch size, and learning rate parameters, and save the network parameters after judging whether the maximum training time has been reached or the training error is less than the set value; After obtaining the observation function, define a finite observable vector for: (1) in, Indicates the extended state. Represents the first element of the input sequence, Indicates the train status; is the state observation function, expressed as follows: (2) in It is the boost function trained by the DNN neural network. is the size of the dimension to be promoted; The extended dynamic mode decomposition method is used to calculate a finite-dimensional operator, which is approximated as , dynamic mode decomposition solves the least squares problem by minimizing the following cost function to obtain an approximate closed form estimate of the operator: (3) in represents the koopman operator, represents the operator dimension, represents the dimensional space of the operator, Indicates the size of the dataset, represents the train status at time i, represents the train status of i at the next moment, express The observation function of express The observation function of represents the system output traction / braking force at time i; In the optimization iteration process of the above formula (3), the state transfer matrix A and the control input matrix B of the train data driven linear state space model are obtained; the state projection C of the train data driven linear state space model is optimized by minimizing the square difference between the predicted value and the actual output value, which is expressed as follows: (4) The dataset obtained from the historical train operation data is described as: (5) (6) (7) in, Represents the train status information sequence, Indicates the state at a certain moment. The state information at each moment consists of speed and position. Indicates the train status information sequence at the next moment after U acts; Indicates the train output traction force / braking force sequence, Indicates the output traction / braking force at a certain moment; Using the least squares problem formula in the above formulas (3) and (4), we can obtain: (8) in , Represents the use of observation function Centralize the data , The data after dimensionality upgrade; express and Construct the pseudo-inverse of the matrix, express The pseudo-inverse is the inverse matrix of a non-square matrix; After obtaining the A, B, and C matrices of the train data-driven linear state space model, we define , construct the train data driven linear state space model as follows: (9) Among them, A is the state transfer matrix, B is the control input matrix, and C is the state projection matrix; is the state variable of the system after dimensionality increase, , N is Dimensions; is the original nonlinear system state An estimated value of , , is a linear time-invariant matrix; is the system input, which indicates the traction or braking force received by the train during operation. ; Where m represents the dimension of the control input, n is the dimension of the train state, and k is the discrete time step; The train data-driven linear state-space model can be used to design an MPC controller to obtain stable control results.

3. The data-driven fault-tolerant train tracking control method according to claim 2 is characterized in that: The step 2 is specifically as follows: First, the information of the train actuator is obtained. Considering the failure of the actuator, the system input is modeled to obtain the fault input model, as shown in formula (10): (10) in Represents the equivalent input when the actuator fails, the matrix Used to represent the actuator state, which is defined as: (11) matrix is a diagonal matrix, Indicates additional faults of the corresponding actuator; matrix The value of each diagonal element in , between 0 and 1, indicating the degree of failure of the corresponding actuator, where 0 indicates complete failure and 1 indicates normal operation of the actuator; Substituting the fault input model in formula (10) into the train data driven linear state space model in step 1, we obtain a fault-tolerant train data driven linear state space model, which is expressed as follows: (12) (13) in , represents the state variables of the system at time k+1 and k after dimensionality increase, represents the fault-tolerant control matrix.

4. The data-driven fault-tolerant train tracking control method according to claim 3 is characterized in that: The step 3 is specifically as follows: Run the train system for a while The running data of time constructs an observable sequence of length a+1 , arrange the observable sequence into the following matrix: (14) (15) in, and Represent the state operation data matrix and input sequence data matrix respectively; , , Indicates the weight of data in calculation; Use the dynamic mode decomposition (EDMD) method, which uses the lifting function trained by the neural network to map the low-dimensional state space to the high-dimensional space, so that the system can be better described dynamically in the high-dimensional space; The EDMD method first uses the following formula at each time step k: (16) By solving the least squares problem above, we can get the initial fault-tolerant control matrix. The calculation formula is as follows: (17) Initialize the fault-tolerant control matrix using the above formula After that, a data sliding window is established for Matrix update; When accessing new data generated during train operation, the system is updated based on a set of system snapshots taken within a given time window. Matrix, only the first and last elements of the window are changed, and the remaining data elements of the window are only shifted left by one unit; definition: (18) (19) The input matrix is ​​represented as: (20) renew , which can be updated recursively and , new data in the observable sequence The matrix is ​​accessed at the next moment. Can be recursively computed with new data.

5. The data-driven fault-tolerant train tracking control method according to claim 4 is characterized in that: The step 4 is specifically as follows: After the calculation formula is simplified, the new train operation data of the window is updated through the following formula: (21) Among them, a represents the window size, b represents the state subscript, and u represents the input subscript; (22) in represents the identity matrix; (23) (24) (25) (26) matrix Update using the following formula: (27) Fault-Tolerant Control Matrix It consists of two parts: and , represents the actual input matrix, represents the additive term for estimating the actuator state, is regarded as a known disturbance and is input into the MPC controller; In finding the fault tolerance matrix After that, the fault-tolerant train data-driven linear state space model is obtained as the MPC control model. Then, according to the train operation control target, the objective function and constraint conditions of the MPC optimization problem are defined, as shown in formula (28) and formula (29), respectively. (28) (29) Among them, J is the objective function, N represents the prediction window size of MPC, , and Used to penalize tracking errors and control input changes, represents the reference trajectory, represents the MPC prediction trajectory, Indicates the train traction / braking force change constraint; Indicates the actual running trajectory. It represents the dimension increase of the actual running trajectory. , , and Represents control input and its changes minimum and maximum boundaries of ; By minimizing the objective function J, the optimal control law sequence for a period of time is obtained. , Indicates input quantity, select the first input quantity in the control law sequence As driving force input for the train; After the train control system responds, the status information is obtained again to enter the next cycle of calculation.

6. A computer device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, it is used to implement the data-driven fault-tolerant train tracking control method as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, it is used to implement the data-driven fault-tolerant train tracking control method as described in any one of claims 1 to 5.

Images