High-speed magnetic levitation suspension system control method and system based on edge calculation and Transform prediction

Abstract

The invention provides a high-speed magnetic levitation suspension system control method and system based on edge calculation and Transform prediction, and the method comprises the steps: constructing a Transform prediction model with a space-time attention mechanism and an autoregression mechanism based on obtained local low-delay calculation resources and train real-time sensing data, and deploying the Transform prediction model in a vehicle-mounted edge calculation unit; performing short-term high-precision prediction on the gap, the acceleration and the disturbance trend at a plurality of sampling moments in the future through a Transform prediction model to obtain a prediction result; processing actuator current saturation and gap safety threshold hard constraints in a limited prediction domain by using a model prediction controller, and solving an optimization control sequence in real time in combination with a prediction result; and overlapping a control barrier function as a safety filter of the model prediction controller, correcting the optimized control sequence to obtain an optimal control sequence, and controlling the high-speed magnetic suspension system. According to the method, the cloud communication delay and jitter are reduced, and the robustness and security of the system under uncertain disturbance are improved.

Description

Technical Field

[0001] This invention relates to the field of high-speed maglev suspension technology, and in particular to a control method and system for a high-speed maglev suspension system based on edge computing and Transformer prediction. Background Technology

[0002] High-speed maglev transportation, as a future-oriented ground transportation mode, has the core advantage of eliminating wheel-rail adhesion and achieving contactless operation, avoiding the constraints of mechanical friction and adhesion limits, thus possessing the potential to operate at speeds exceeding 500 km / h. The suspension system, as the core subsystem of the high-speed maglev train, directly determines whether the train can safely and stably levitate above the track. However, the suspension system is a strongly nonlinear, open-loop unstable electromechanical coupling system, extremely sensitive to minute disturbances, which can be amplified by even slight deviations. As the speed increases to 600 km / h, multi-source disturbances such as aerodynamic lift pulsations, minor track irregularities, vehicle-rail coupling, and sensor noise significantly intensify. Simultaneously, actuator saturation and communication delays continuously compress the safety margin, and traditional control methods have significant bottlenecks in response speed, predictive ability, and constraint handling, lacking forward-looking judgment of disturbances and state evolution, easily leading to instability of the suspension system at higher speeds. Summary of the Invention

[0003] In view of this, the present invention provides a control method and system for a high-speed maglev suspension system based on edge computing and Transformer prediction to solve the above problems.

[0004] This invention provides a control method for a high-speed maglev suspension system based on edge computing and Transformer prediction, comprising: constructing a Transformer prediction model with spatiotemporal attention and autoregressive mechanisms based on acquired local low-latency computing resources and real-time train sensing data; deploying the Transformer prediction model on an onboard edge computing unit; performing short-term high-precision predictions of gaps, accelerations, and disturbance trends at several future sampling times using the Transformer prediction model to obtain prediction results; using a model predictive controller to handle actuator current saturation and gap safety threshold hard constraints within a finite prediction domain, and solving for and optimizing the control sequence in real time based on the prediction results; superimposing a control barrier function as a safety filter for the model predictive controller to correct the optimized control sequence to obtain the optimal control sequence; and using the optimal control sequence to control the high-speed maglev suspension system.

[0005] In another implementation of the present invention, the spliced ​​outputs of the spatiotemporal attention mechanism, after being linearly mapped, yield the following multi-head attention result:

[0006] in, These represent the query, key, and value matrices, respectively. For the first The projection matrix of each attention head. This is for outputting the projection matrix.

[0007] In another implementation of the present invention, the prediction result is expressed as:

[0008] in, For the future The system state prediction sequence of the step, It is a nonlinear mapping. Indicates from arrive The input signal sequence at time t, This represents the corresponding control input sequence.

[0009] In another implementation of the present invention, the cost function of the model prediction controller is defined as follows:

[0010] in, For the future k +1 step system prediction state, For the first k +1 step reference status, For the first k Step control input, To predict the terminal reference state at the Nth step in the time domain, Indicates a weighted quadratic form. Let be a given positive definite weight matrix.

[0011] In another implementation of the present invention, the constrained optimal control problem solved by the model predictive controller is:

[0012] in, For the first k The system status of the step, The state matrix, For the control matrix, For the disturbance term, For the influence matrix, and Separately define the constraint boundaries for control variables and state variables.

[0013] In another implementation of the present invention, the control barrier function is:

[0014] in, For gap deviation, h e This is the system's rated balance clearance. This is the minimum safe clearance.

[0015] In another implementation of the present invention, the optimal control sequence is:

[0016] in, These are the decision variables obtained through QP optimization. For candidate control inputs, These are system drift term and control term pair. Lie derivative, for Function-like.

[0017] Another aspect of the present invention provides a control system for a high-speed maglev suspension system based on edge computing and Transformer prediction, comprising: a model building module: constructing a Transformer prediction model with spatiotemporal attention and autoregressive mechanisms based on acquired local low-latency computing resources and real-time train sensing data, and deploying the Transformer prediction model on an on-board edge computing unit; a parameter prediction module: performing short-term high-precision prediction of gaps, accelerations, and disturbance trends at several future sampling times using the Transformer prediction model to obtain prediction results; a sequence optimization module: using a model prediction controller to handle actuator current saturation and gap safety threshold hard constraints within a finite prediction domain, and solving for and optimizing the control sequence in real time based on the prediction results; a safety filtering module: superimposing a control barrier function as a safety filter for the model prediction controller to correct the optimized control sequence and obtain the optimal control sequence; and a control module: using the optimal control sequence to control the high-speed maglev suspension system.

[0018] In another aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a high-speed maglev suspension system control method based on edge computing and Transformer prediction as described in any of the preceding claims. In another aspect, the present invention provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of a high-speed maglev suspension system control method based on edge computing and Transformer prediction as described in any of the preceding claims.

[0019] The high-speed maglev suspension system control method based on edge computing and Transformer prediction of the present invention forms a closed-loop chain of "predictive guidance - optimized decision-making - safety protection". The fast response at the edge end significantly reduces control loop delay and jitter, and reduces gap fluctuation and overshoot caused by phase lag. The prediction compensation based on deep learning can adaptively compensate for the impact of model mismatch and environmental changes on control performance, and improve the robustness and stability of the system under strong disturbance and high uncertainty conditions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. By reading the detailed description of the embodiments below, the advantages and benefits of the solutions will become clear to those skilled in the art. The accompanying drawings are only for illustrating preferred embodiments and are not intended to limit the present invention. In the accompanying drawings: Figure 1 This is a schematic diagram of a control method for a high-speed maglev suspension system based on edge computing and Transformer prediction, according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the overall framework of a suspension control system based on IoT edge computing and transformer-based prediction, according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of a high-speed maglev train suspension system according to an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of a single floating point frame according to an embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram of the Transformer-Encoder structure according to an embodiment of the present invention.

[0025] Figure 6 This is a schematic diagram of a predictor architecture according to an embodiment of the present invention.

[0026] Figure 7 This is a schematic diagram of the control structure for security filtering according to an embodiment of the present invention.

[0027] Figure 8 This is a schematic diagram of an experimental platform device according to an embodiment of the present invention.

[0028] Figure 9 This is a schematic diagram of the pulse perturbation response according to an embodiment of the present invention.

[0029] Figure 10 This is a schematic diagram of the periodic vibration response according to an embodiment of the present invention.

[0030] Figure 11 This is a schematic diagram of the model mismatch response according to an embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art should fall within the protection scope of the present invention.

[0032] Figure 1 A schematic flowchart of a high-speed maglev suspension system control method based on edge computing and Transformer prediction is provided for an embodiment of the present invention, as shown below. Figure 1 As shown, this embodiment mainly includes: S101. Based on the acquired local low-latency computing resources and real-time train sensing data, construct a Transformer prediction model with spatiotemporal attention mechanism and autoregressive mechanism, and deploy the Transformer prediction model on the vehicle edge computing unit. S102. The Transformer prediction model is used to make short-term high-precision predictions of the gaps, accelerations and disturbance trends at several future sampling times, and the prediction results are obtained. S103. The model predictive controller is used to handle actuator current saturation and hard constraints of gap safety threshold within a finite prediction domain, and the prediction results are combined to solve the optimized control sequence in real time.

[0033] S104. The superimposed control barrier function serves as a safety filter for the model predictive controller, correcting the optimized control sequence to obtain the optimal control sequence.

[0034] S105. The high-speed maglev suspension system is controlled using the optimal control sequence.

[0035] For example, such as Figure 2As shown, to achieve real-time performance of high-speed maglev trains operating at 600 km / h, a suspension control framework was constructed based on the Internet of Things (IoT) and onboard edge computing technology. This framework utilizes edge computing nodes deployed locally on the train to achieve real-time acquisition, synchronization, and preprocessing of data from multiple sensor sources. Leveraging their low-latency and highly reliable computing capabilities, a lightweight Transformer predictive model and Model Predictive Controller (MPC) are run to perform real-time state prediction, optimization, and safety verification, forming a closed-loop control chain of "perception-prediction-decision-execution".

[0036] like Figure 3 As shown, the suspension points of high-speed maglev trains are mechanically decoupled, and there is almost no interference between the suspension points. Generally, a single suspension point is used as the modeling object.

[0037] like Figure 4 As shown, a levitation point typically employs a closed loop consisting of an electromagnet, the levied object, and a position signal sensor. The system parameters are shown in Table 1.

[0038] Table 1 defines the system parameters and states as follows:

[0039] The system dynamic equations are:

[0040] At the equilibrium suspension point Place, satisfy The equilibrium current can be obtained. The control input is defined as the increment of the current relative to the equilibrium value. By linearizing the system with small perturbations, a locally linear model can be approximated:

[0041] The above equation neglects higher-order small quantities. As a predictive model for the MPC controller, this linear model is only approximately effective within a small range due to the strong nonlinearity and parameter uncertainties of the suspension system; under large deviations, nonlinear predictive and control strategies are needed to ensure performance. Therefore, a time-series predictor is incorporated into traditional model predictive control to estimate and compensate for the future evolution trend of the system.

[0042] The high-speed maglev suspension system control method based on edge computing and Transformer prediction of the present invention forms a closed-loop chain of "predictive guidance - optimized decision-making - safety protection". The fast response at the edge end significantly reduces control loop delay and jitter, and reduces gap fluctuation and overshoot caused by phase lag. The prediction compensation based on deep learning can adaptively compensate for the impact of model mismatch and environmental changes on control performance, and improve the robustness and stability of the system under strong disturbance and high uncertainty conditions.

[0043] In another implementation of the present invention, the spliced ​​outputs of the spatiotemporal attention mechanism, after being linearly mapped, yield the following multi-head attention result:

[0044] in, These represent the query, key, and value matrices, respectively. For the first The projection matrix of each attention head. This is for outputting the projection matrix.

[0045] For example, such as Figure 5 As shown, to improve prediction accuracy, a lightweight temporal prediction model with a Transformer-Encoder structure was designed and deployed on an edge computing unit. The prediction model utilizes historical sensor data and real-time streaming data stored locally on the edge side, constructing an input sequence through a sliding window mechanism to achieve high-precision prediction of the system state in multiple future steps. Edge deployment not only avoids the communication overhead of uploading data to the cloud but also enables efficient model inference through hardware acceleration, meeting the real-time requirements of the control system. Furthermore, compared to conventional RNNs or LSTMs, the Transformer, based on a self-attention mechanism, can extract complex temporal dependencies over longer time windows without being easily forgotten.

[0046] The prediction model employs an encoder-based multi-layer Transformer block structure, performing spatial and temporal dual-path attention modeling on the input temporal features. First, the input multidimensional feature sequence undergoes linear projection to obtain a high-dimensional embedding representation, and positional encoding is added to preserve temporal order information. Subsequently, the embedding sequence is fed in parallel into the spatial and temporal attention branches: the spatial branch obtains the query, key, and value matrix through linear mapping. The scaling dot product attention method captures the correlation between features from different sensors; its calculation formula is as follows:

[0047] in, These represent the query, key, and value matrices, respectively. The scaling factor (key vector dimension) is used. Parallel multi-head attention allows the model to focus on feature relationships at different positions and in different dimensions within a sequence. The concatenated outputs are then linearly mapped to obtain the multi-head attention result.

[0048] in, For the first The projection matrix of each attention head. The output projection matrix is ​​generated. The multi-head attention output of the spatial branch passes through residual connections and layer normalization before entering the feedforward network (FFN) to extract higher-level nonlinear features. Similarly, the temporal branch reuses the query from the spatial branch. This is compared with another set of keys and values ​​that characterize time series patterns. Cross-attention computation is performed to extract the dynamic dependencies of sensor data over time. The computational form of temporal attention is similar to that of spatial attention, except that the key values ​​are derived from time-series features. After the dual-path attention mechanism, the outputs of both the spatial and temporal branches are residually connected and normalized before being fed into the FFN layer for further feature extraction. The output features of both branches are then concatenated and fused as the input to the next Transformer coding block. Through this cascaded spatial-temporal dual-path attention structure, the correlation between multiple sensor channels and the time-varying patterns of single-channel signals can be modeled simultaneously.

[0049] Considering the limited edge computing resources in vehicles, the Transformer prediction model was miniaturized by reducing the number of network layers and hidden unit dimensions, and by employing methods such as knowledge distillation to compress model parameters. The resulting model can complete inference at extremely high speeds, meeting real-time requirements. Furthermore, with the acceleration of onboard GPUs / FPGAs, the prediction model can calculate the future... The system output of each step provides forward-looking information for MPC. The model is trained using offline historical data, and the loss function is the mean squared error (MSE) between the predicted and actual outputs, plus... Regularization terms prevent overfitting. After training, the model is deployed at the vehicle edge and can be periodically updated online using newly acquired data to maintain sensitivity to system changes and ensure that the predictor continuously provides accurate rolling predictions.

[0050] The Transformer prediction model uses offline historical data for supervised learning during the training phase and employs knowledge distillation to compress the model size, adapting to the limited computing resources of edge devices. After deployment, the edge computing platform can dynamically update and learn online, gradually adapting to dynamic system changes and further improving prediction accuracy.

[0051] In another implementation of the present invention, the prediction result is expressed as:

[0052] in, For the future The system state prediction sequence of the step, It is a nonlinear mapping. Indicates from arrive The input signal sequence at time t, This represents the corresponding control input sequence.

[0053] For example, select the most recent The multi-channel sensor features at each time point are used as the input sequence for the prediction model, including sensor measurements such as electromagnet current increment and levitation gap deviation. and the corresponding control inputs at the time The predictive model outputs the future. Step-by-step system state prediction sequence ， For example, predicting suspension acceleration or vibration amplitude over a given period. A sliding time window method is used to divide long-term series data into multiple training samples to fully extract local temporal patterns. The nonlinear mapping represented by the Transformer prediction model is denoted as... Then we have:

[0054] in, Indicates from arrive The input signal sequence at time t, This represents the corresponding control input sequence. For the future The prediction model outputs the predicted values ​​for each step. To improve multi-step prediction performance, the prediction model employs an autoregressive mechanism: when progressively predicting future sequences, the prediction output from the previous time step (e.g., the acceleration prediction from the previous time step) is used as the basis for prediction. Feedback, as one of the inputs for subsequent time steps, together with newly acquired observations, constitutes a new input sequence. Through output feedback, the model can use its own prediction results to continue making predictions during the inference phase, thereby improving the coherence and stability of the prediction sequence and reducing the phenomenon of error accumulation and amplification over time in multi-step predictions.

[0055] In another implementation of the present invention, the cost function of the model prediction controller is defined as follows:

[0056] in, For the future k+1 step (starting from the current time t) of the system prediction state, For the first k +1 step reference status, For the first k Step control input, To predict the terminal reference state at the end of the time domain (step N), Indicates a weighted quadratic form. Let be a given positive definite weight matrix.

[0057] For example, model predictive control (MPC) with finite predictive time is used to generate control commands for the maglev system. At each sampling time, MPC solves an optimal control problem online with the current state as the initial value, obtaining the future... Optimal control sequence of steps and actually implement the first step of control. For the system, the time domain is then rolled forward by one cycle and the above optimization is repeated. The MPC controller uses the state-space model obtained by the aforementioned linearization as the prediction model and incorporates the future disturbances estimated by the Transformer prediction model into the optimization to improve the decision-making of the control system. Its closed-loop architecture is as follows: Figure 6 As shown.

[0058] Edge computing platforms provide the hardware foundation for real-time optimization of MPC. Employing a high-efficiency quadratic programming (QP) solver and leveraging the parallel computing capabilities of edge devices, constrained optimization problems are solved rapidly within each control cycle. By localizing the prediction model and objective function at the edge, reliance on the central control unit is reduced, enhancing the system's autonomy and response speed.

[0059] The optimization formula of MPC is described in a formalized way, making Indicates the current system state, with the reference trajectory as follows. (For steady-state suspension control, the following can be taken) (This indicates that the suspension gap deviation is zero). In the time domain arrive To minimize trajectory error and control energy consumption, the cost function is defined as follows:

[0060] in, Indicates a weighted quadratic form. Given a positive definite weight matrix (or scalar), it is used to weigh terms such as state error, control increment, and terminal state deviation.

[0061] In another implementation of the present invention, the constrained optimal control problem solved by the model predictive controller is:

[0062] in, For the first k The system status of the step, The state matrix, For the control matrix, For the disturbance term, For the influence matrix, and Separately define the constraint boundaries for control variables and state variables.

[0063] For example, based on the cost function described above, MPC needs to solve the following constrained optimal control problem at each time step:

[0064] in, The state equations represent the current state and utilize the state matrix of the linearized model for constraints. Control matrix And add a perturbation term. and its influence matrix ; and These represent the constraint boundaries for control variables such as actuator current and levitation gap, as well as state variables. By solving the above convex quadratic programming problem, the optimal control sequence that satisfies the constraints can be obtained. MPC optimization can be numerically solved in real time using an efficient QP solver, due to the introduction of predictions for future disturbances. Therefore, MPC can proactively adjust the control strategy to counteract the effects of upcoming disturbances. For example, when it is predicted that the suspension gap will decrease due to a disturbance at the next moment (i.e., lead to (Decrease), optimization will tend to increase the current control input in advance. (Increase electromagnetic force) to preemptively suppress gap shrinkage, thereby playing a feedforward compensation role.

[0065] In each sampling period, apply the optimized first-step control. Then, based on sensor feedback, the state is updated to the next moment. The process is repeated iteratively within the optimization window. Since MPC considers input and state constraints during optimization, it ensures that the optimal solution will not lead the system into a dangerous state. However, in actual operation, inaccurate models or prediction errors may still cause the system to approach the constraint boundaries. Therefore, a layer of safety constraint control is superimposed on the MPC controller as a safety filter to further improve the safety of system operation. In actual implementation, the MPC control algorithm runs on an onboard edge computing platform, using a QP solver to ensure that the optimal solution is found within a millisecond-level control cycle, meeting real-time requirements.

[0066] By utilizing a lightweight Transformer network deployed on an onboard edge computer, short-term time-series predictions of the key states and disturbances of the maglev system are made. The prediction results are then integrated into the optimization process of MPC to compensate for the shortcomings of traditional model predictions in nonlinear and uncertain scenarios.

[0067] In another implementation of the present invention, the control barrier function is:

[0068] in, For gap deviation, h e This is the system's rated balance clearance. This is the minimum safe clearance.

[0069] For example, to further improve system security, a Control Barrier Function (CBF) is introduced as a security filter for MPC on the edge computing platform. Similarly, deploying it on the edge computing platform can significantly reduce latency and ensure the real-time performance of the security system. The CBF method constructs a function... To describe the safe set of system states and enforce it in control. The gap is not reduced, thus ensuring that the state remains within the safe range. For magnetic levitation systems, a natural safety set is that the levitation gap is not lower than a threshold to prevent collisions with the electromagnet by suspended objects. For example, setting a minimum safety gap... Then the barrier function can be defined as Safety requirements It holds true for all times.

[0070] According to CBF theory, if a continuously differentiable barrier function exists... At the dangerous boundary and Then you only need to control the input. satisfy:

[0071] For the evolution of a closed-loop system, the set It will remain unchanged. These are system drift term and control term pair. Lie derivative, for Class function (usually a linear function) The above inequality describes the linear constraint that the control input must satisfy. When When the state approaches 0 (close to the safety boundary), control needs to be selected. The rate of change is positive, which means that the state moves away from the danger boundary.

[0072] By incorporating control barrier function (CBF) constraints into the control framework, the control commands given by MPC are corrected in real time, ensuring that the suspension system state meets safety constraints such as suspension gap, and ensuring that the system remains safe and stable under disturbances.

[0073] In another implementation of the present invention, the optimal control sequence is:

[0074] in, These are the decision variables obtained through QP optimization. For candidate control inputs, These are system drift term and control term pair. Lie derivative, for Function-like.

[0075] For example, CBF constraints are combined with MPC control. At each sampling time, when MPC provides candidate control inputs... Then, a QP module is added for security filtering to calculate the distance. The most recent control input that meets safety conditions :

[0076] Since QP has only one linear inequality constraint, it can be solved within each cycle, ensuring computational speed and guaranteeing the obtained results. Safety conditions are met. If... If the safety constraints are already met, then The original control will not be changed; otherwise, the control parameters will be adjusted appropriately to avoid touching the safety boundary. It should be noted that when MPC and CBF conflict (i.e., MPC attempts to sacrifice safety to optimize performance), CBF constraints will become the last line of defense, ensuring safety even at the cost of some performance.

[0077] Figure 7 This diagram illustrates a control structure incorporating CBF (Safety Filtering) for safety. It shows that the CBF module monitors the system's critical states, and when control commands from the MPC (Master Control Controller) might violate safety limits, the CBF corrects the control input online. This correction is achieved by adjusting the parameters within the CBF. The value of can, to some extent, balance safety and conservatism with control performance: When the CBF is larger, it will move away from the constraint boundary more aggressively, but this may slightly reduce the tracking speed. Conversely, when the size is smaller. During experimental verification, select an appropriate... Ensure sufficient margin is maintained near the minimum suspension gap without compromising dynamic performance.

[0078] In summary, this suspension control framework deeply integrates IoT sensing, edge computing, and intelligent predictive control technologies to construct a closed-loop system that integrates data acquisition, local processing, real-time decision-making, and safety protection. By deploying the Transformer predictive model, MPC optimizer, and CBF safety filter on the vehicle edge platform, the system maintains excellent control performance and safety levels even under high-speed and high-disturbance conditions.

[0079] This invention deploys a predictive control algorithm on an onboard edge computing unit, utilizing local low-latency computing resources and real-time train sensor data to construct a Transformer predictive model with spatiotemporal attention and an autoregressive mechanism. This model performs short-term, high-precision predictions of gaps / acceleration and disturbance trends at several future sampling times, providing feedforward compensation for the controller. Based on the prediction results, the Model Predictive Controller (MPC) handles hard constraints such as actuator current saturation and gap safety thresholds within a finite prediction domain, solving for optimized control sequences in real time to ensure constraint satisfaction while maintaining tracking performance and energy efficiency. A Control Barrier Function (CBF) is superimposed as an outer safety filter to ensure the state always remains within the feasible safety domain, forming a closed-loop chain of "prediction guidance—optimization decision—safety protection." Leveraging the processing power of onboard edge computing, this architecture offers the following advantages: First, the rapid response at the edge significantly reduces control loop latency and jitter, minimizing gap fluctuations and overshoot caused by phase lag; second, deep learning-based predictive compensation can adaptively compensate for the impact of model mismatch and environmental changes on control performance, improving system robustness and stability under conditions of strong disturbances and high uncertainty.

[0080] Example 1 To verify the performance of the proposed design, a high-speed maglev track-coupled vibration experimental system based on an Internet of Things (IoT) architecture and an onboard edge computing platform was built. The system consists of a high-speed maglev track-coupled vibration platform, a dSPACE real-time simulation platform serving as the onboard edge computing unit, and a suspension controller.

[0081] IoT Sensing Layer: The vibration platform simulates track excitation at different vehicle speeds through a vibrator. At the same time, a multi-source IoT sensor network (including gap sensors, current sensors, and accelerometers) collects key status data of the system in real time.

[0082] Edge computing layer: The dSPACE platform acts as the core of in-vehicle edge computing, interconnecting with sensors and control systems via a high-speed communication bus. It is responsible for running the aforementioned lightweight Transformer prediction model, MPC optimization solver, and CBF safety filtering algorithm, completing the entire "perception-prediction-decision-execution" process within a single control cycle.

[0083] Execution layer: The levitation controller receives control commands from the edge computing unit and outputs current to drive the electromagnet.

[0084] This verification platform can test the integrated "sensing-computing-control" IoT edge intelligent control process. Because it is deployed locally, it ensures that control commands generated by the edge computing unit can be sent to the actuator with extremely low latency, effectively verifying the real-time performance of edge computing in high-speed maglev systems. The experimental platform device is as follows: Figure 8 As shown.

[0085] Transformer prediction model input past Step-by-step sensing measurement and control inputs, output for future Step position prediction. MPC prediction time domain. Step, matching the predictor length; weight matrix taken The constraint conditions are: a rated suspension clearance of 10mm and a minimum suspension clearance of 5mm; clearances less than 5mm are considered unsafe. CBF parameters should be selected accordingly. The performance of the MPC controller is verified by comparing it with traditional PID control through three typical operating conditions.

[0086] Impulse Disturbance Performance Verification: First, the response of the suspension system to sudden disturbances is verified. There is an occasional pulse impact, simulating a sudden change in the orbit. Figure 9 The response curves of the suspension gap under different control strategies are presented. The traditional PID controller exhibits an instantaneous gap slip of approximately 5 mm after a disturbance, followed by a recovery to normal operation after about 4 seconds. In contrast, the method of this invention, due to the Transformer predictive model's anticipation of the disturbance's impact, allows the MPC to reduce control current in advance for compensation. Therefore, the gap experiences only a 2 mm fluctuation and recovers to normal within a very short time, with current and acceleration remaining essentially normal. Combined with the safety filtering effect of the CBF, the gap remains within the safe threshold range. Numerical results show that the MPC controller reduces the maximum deviation by more than 50% and shortens the recovery time by approximately 60%. This verifies the performance of the transformer network-based predictive compensation MPC in anti-impact conditions.

[0087] Verification of the disturbance performance of long-wave track irregularities: The disturbance of long-wave track irregularities can be regarded as the periodic vibration disturbance of the track on the vehicle. Using a vehicle-track coupled vibration table, the disturbance of long-wave track irregularities is simulated, and a sinusoidal disturbance with a vibration frequency of 3 Hz and an amplitude of 3 mm is set. Figure 10The curves for gap, acceleration, and current are shown for both traditional PID controllers and MPC controllers. It can be observed that, due to a lack of foresight, the traditional PID controller's control current lags behind the disturbance by approximately 0.3 seconds in phase, causing the output to deviate from the reference alignment. In contrast, the MPC controller provides pre-gain, appropriately adjusting the control current before the periodic disturbance arrives. This results in a significantly smaller gap fluctuation compared to the traditional controller, no phase lag, and smaller current fluctuations, indicating that predictive compensation avoids overly drastic control actions.

[0088] Performance verification under model mismatch: One magnetic pole module of the actual system is removed, and a load is added simultaneously. This affects the mass of the levitation system. and Simultaneous changes cause a mismatch between the real system and the mathematical model. Figure 11 The results show the differences in gap, current, and acceleration responses between the traditional PID controller and the MPC controller. The results indicate that the traditional PID controller takes nearly 4 seconds to reach stability from initial floating, while the MPC controller reaches stability in just 2 seconds, and the amplitude of the gap oscillation is smaller than that of the traditional PID controller. This is because the predictor learns the actual dynamic characteristics of the system from real-time data, compensating for model mismatch to some extent. Furthermore, the CBF appropriately increases the control strength when it detects a slight safety risk in the early stages of oscillation, quickly pulling the system back to the safe region. Overall, the method of this invention exhibits a certain degree of robustness under model mismatch conditions.

[0089] In summary, the proposed MPC+CBF control framework based on onboard edge Transformer prediction enhancement exhibits superior performance compared to traditional methods under various typical operating conditions, ensuring the stability of the suspension system while guaranteeing operational reliability through safety constraints.

[0090] This invention addresses the stability control problem of maglev suspension systems, proposing a real-time closed-loop control method integrating onboard IoT edge intelligence. By deploying a Transformer timing predictor on an edge computing platform, future system state changes are detected in advance, and this is combined with model predictive control to achieve rapid and precise adjustment of the suspension system. Simultaneously, a control barrier function is introduced to safely filter the control input, ensuring the system always operates within a safe range. The invention is validated through an experimentally constructed hardware system comprising a suspension system dynamics model, predictor, optimized controller, and safety filter. Results show that the proposed method improves dynamic response speed, effectively resists transient and periodic disturbances, significantly improves suspension stability and tracking performance, and exhibits robustness. Compared to traditional PID controllers, this method reduces overshoot and phase lag in the suspension gap, effectively improving response speed while maintaining suspension stability, and is significantly superior to the baseline control scheme without predictive compensation.

[0091] Another aspect of the present invention provides a control system for a high-speed maglev suspension system based on edge computing and Transformer prediction, comprising: Model building module: Based on the acquired local low-latency computing resources and real-time train sensing data, a Transformer prediction model with spatiotemporal attention mechanism and autoregressive mechanism is constructed, and the Transformer prediction model is deployed on the vehicle edge computing unit.

[0092] Parameter prediction module: The Transformer prediction model is used to make short-term, high-precision predictions of the gaps, accelerations and disturbance trends at several future sampling times, and the prediction results are obtained.

[0093] Sequence optimization module: Utilizes a model predictive controller to handle actuator current saturation and hard constraints such as gap safety threshold within a finite prediction domain, and solves for and optimizes the control sequence in real time by combining the prediction results.

[0094] Safety filtering module: The superimposed control barrier function serves as a safety filter for the model predictive controller, correcting the optimized control sequence to obtain the optimal control sequence.

[0095] Control module: Uses the optimal control sequence to control the high-speed maglev suspension system.

[0096] The high-speed maglev suspension system control system based on edge computing and Transformer prediction of this invention forms a closed-loop chain of "predictive guidance - optimized decision-making - safety protection". The fast response at the edge end significantly reduces control loop delay and jitter, and reduces gap fluctuation and overshoot caused by phase lag. The prediction compensation based on deep learning can adaptively compensate for the impact of model mismatch and environmental changes on control performance, and improve the robustness and stability of the system under strong disturbance and high uncertainty conditions.

[0097] In another aspect of the present invention, the electronic device includes: a processor, a memory, and a communication bus and a communication interface.

[0098] in: The processor, memory, and communication interface communicate with each other via a communication bus.

[0099] A communication interface is used to communicate with other electronic devices or servers.

[0100] The processor is used to execute programs, specifically, to execute any of the steps of the high-speed maglev suspension system control method based on edge computing and Transformer prediction in the above embodiments.

[0101] Specifically, the program may include program code, which includes computer operation instructions.

[0102] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0103] Memory is used to store programs. Memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device.

[0104] Specifically, the program can be used to cause the processor to execute the steps of any of the high-speed maglev suspension system control methods based on edge computing and Transformer prediction described in the embodiments. The specific implementation of each step in the program can be found in the corresponding descriptions of the steps and units executed in any of the aforementioned high-speed maglev suspension system control methods based on edge computing and Transformer prediction, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments.

[0105] An exemplary embodiment of this application also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods of various embodiments of this application.

[0106] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0107] Specific embodiments of the present invention have now been described. Other embodiments are within the scope of the appended claims. In some cases, the actions described in the claims can be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result.

[0108] It should be noted that all directional indications (such as up, down, left, right, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship between the components in a certain order (as shown in the figure). If the specific order changes, the directional indication will also change accordingly.

[0109] In the description of this invention, the terms "first" and "second" are used only for convenience in describing different components or names, and should not be construed as indicating or implying a sequential relationship, relative importance, or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include at least one of that feature.

[0110] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0111] It should be noted that although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of the present invention.

[0112] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention, so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be an improper limitation of the embodiments of the present invention.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for a high-speed maglev suspension system based on edge computing and Transformer prediction, characterized in that, include: Based on the acquired local low-latency computing resources and real-time train sensing data, a Transformer prediction model with spatiotemporal attention mechanism and autoregressive mechanism is constructed, and the Transformer prediction model is deployed on the vehicle edge computing unit. The Transformer prediction model is used to make short-term, high-precision predictions of the gaps, acceleration, and disturbance trends at several future sampling times, and the prediction results are obtained. A model predictive controller is used to handle actuator current saturation and hard constraints of gap safety threshold within a finite prediction domain, and the optimized control sequence is solved in real time by combining the prediction results. The superimposed control barrier function serves as a safety filter for the model predictive controller, correcting the optimized control sequence to obtain the optimal control sequence. The optimal control sequence is used to control the high-speed maglev suspension system.

2. The method according to claim 1, characterized in that, The spliced ​​outputs of the spatiotemporal attention mechanism, after being linearly mapped, yield the multi-head attention result as follows: in, These represent the query, key, and value matrices, respectively. For the first The projection matrix of each attention head. This is for outputting the projection matrix.

3. The method according to claim 1, characterized in that, The prediction result is expressed as follows: in, For the future The system state prediction sequence of the step, It is a nonlinear mapping. Indicates from arrive The input signal sequence at time t, This represents the corresponding control input sequence.

4. The method according to claim 3, characterized in that, The cost function defined for the model predictive controller is: in, For the future k +1 step system prediction state, For the first k +1 step reference status, For the first k Step control input, To predict the terminal reference state at the Nth step in the time domain, Indicates a weighted quadratic form. Let be a given positive definite weight matrix.

5. The method according to claim 4, characterized in that, The constrained optimal control problem solved by the model predictive controller is: in, For the first k The system status of the step, The state matrix, For the control matrix, For the disturbance term, For the influence matrix, and Separately define the constraint boundaries for control variables and state variables.

6. The method according to claim 1, characterized in that, The control barrier function is: in, For gap deviation, h e This is the system's rated balance clearance. This is the minimum safe clearance.

7. The method according to claim 6, characterized in that, The optimal control sequence is: in, These are the decision variables obtained through QP optimization. For candidate control inputs, These are system drift term and control term pair. Lie derivative, for Function-like.

8. A control system for a high-speed maglev suspension system based on edge computing and Transformer prediction, characterized in that, include: Model building module: Based on the acquired local low-latency computing resources and real-time train sensing data, a Transformer prediction model with spatiotemporal attention mechanism and autoregressive mechanism is constructed, and the Transformer prediction model is deployed on the vehicle edge computing unit. Parameter prediction module: The Transformer prediction model is used to make short-term, high-precision predictions of the gaps, accelerations and disturbance trends at several future sampling times, and the prediction results are obtained. Sequence optimization module: Utilizes a model predictive controller to handle actuator current saturation and hard constraints such as gap safety threshold within a finite prediction domain, and solves for the optimized control sequence in real time by combining the prediction results; Safety filtering module: The superimposed control barrier function serves as a safety filter for the model predictive controller, correcting the optimized control sequence to obtain the optimal control sequence; Control module: Uses the optimal control sequence to control the high-speed maglev suspension system.

9. An electronic device, characterized in that, include: The system comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the high-speed maglev suspension system control method based on edge computing and Transformer prediction as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps in the high-speed maglev suspension system control method based on edge computing and Transformer prediction as described in any one of claims 1 to 7.

Images