Machine Learning in Maglev Transportation Systems: Review and Prospects

Abstract

Magnetic levitation (Maglev) technology has long garnered significant attention in the engineering community due to its inherent advantages, such as contactless operation, minimal friction losses, low noise, and high precision. Based on electromagnetic suspension (EMS) and electrodynamic principles, these systems are primarily developed for advanced transportation, while also inspiring emerging applications such as vibration isolation and flywheel energy storage. Despite progress, practical deployment faces critical challenges, including accurate modeling, robustness against nonlinear and uncertain dynamics, and control stability under complex conditions. Artificial intelligence (AI), particularly machine learning (ML) offers promising solutions. Studies show ML-based methods, i.e., improved particle swarm optimization (PSO) optimize proportional-integral-derivative (PID) to reduce controller output overshoot, deep reinforcement learning (DRL) reduces levitation gap fluctuation under complex conditions, ensemble learning achieves high fault diagnosis accuracy, and convolutional neural network-long short-term memory (CNN-LSTM) predictive maintenance cuts costs. This review summarizes recent AI-enabled advances in Maglev transportation system modeling, control, and optimization, highlighting representative algorithms, performance comparisons, technical challenges, and future directions toward intelligent, reliable, and energy-efficient transportation systems.

1. Introduction

Maglev levitation technology, renowned for its contactless operation, low friction losses, and environmental cleanliness, has found widespread applications across various engineering domains [1]. The integration of advanced electronic components and the continual evolution of control algorithms have driven significant progress in this field. Notably, the most prominent applications of Maglev technology are observed in magnetic bearing systems and high-speed Maglev transportation, indicating the potential of this technology to revolutionize modern engineering systems.

Maglev transportation systems hold significant potential to fulfill the growing demand for high-capacity transit solutions in both intercity travel and urban mobility. As an innovative mode of transportation, Maglev systems offer numerous advantages such as strong climbing ability, tight turning radii, low noise emissions, smoothness of operation, and robust acceleration as well as deceleration capabilities. These attributes make Maglev a promising candidate for future rail transportation development. Currently, the Maglev transportation technologies are categorized into three major classes based on their levitation principles: EMS, electrodynamic suspension (EDS) and hybrid suspension systems that combine the features of both EMS and EDS. Each class presents unique operational characteristics, offering diverse pathways for the advancement and development of Maglev infrastructure worldwide.

The core challenge in Maglev control lies in its inherent nonlinearity and open-loop instability. Traditional methods (e.g., PID) rely on Taylor series expansion around an operating point for linearization, which only works effectively within a narrow range. When confronted with real-world disturbances (e.g., payload changes, track irregularities, or high-temperature superconducting (HTS) thermal-vibration coupling), linear approximation fails, leading to levitation instability or even failure. To overcome these challenges, approaches based on AI have emerged as a promising solution, addressing key limitations of theoretical modeling and improving stability in levitation control.

Artificial Intelligence is an emerging field of computer science that focuses on the study, development, and systematic application of intelligent behavior inspired by natural intelligence systems [2,3]. As the core subfield of AI, ML offers a targeted solution by leveraging data-driven nonlinear modeling. It eliminates the need for precise physical models, adapts to wide operating conditions by learning from operational data (e.g., levitation gap, electromagnet current), and directly addresses traditional control pain points. For instance, supervised learning enables accurate fault classification via labeled sensor data, reinforcement learning optimizes real-time levitation strategies through reward-driven exploration, and deep learning captures temporal disturbances (e.g., track irregularity trends) via multi-layer networks, all of which outperform classical methods in handling Maglev’s complex dynamics.

Building on these capabilities, this study is strictly centered on three core application domains of AI in Maglev transportation. The first domain focuses on the optimization of traction, levitation, and guidance controllers, directly addressing the nonlinearity and open-loop instability challenges. The second domain emphasizes the enhancement of vehicle design, operation, and maintenance, including tasks such as sensor temperature compensation and fault diagnosis. The third domain covers the intelligent management of infrastructure, such as tracks and ground coils including predictive maintenance and structural health monitoring (SHM). To clarify the research scope, this manuscript focuses exclusively on Maglev transportation systems, excluding other applications such as Maglev bearings, vibration isolation, and flywheel energy storage, which are only briefly mentioned. Among transportation technologies, EMS and EDS systems are prioritized, while hybrid suspension systems are excluded due to their current confinement to theoretical research and lack of large-scale engineering applications. Traditional Maglev mechanical design (e.g., guideway material selection) and non-AI control methods (e.g., basic PID without ML optimization) are also outside the scope of this review.

The rest of this paper is structured as follows: Section 2 introduces the fundamental principles of Maglev technology, outlining its significant characteristics and their different characteristics. Section 3 provides an overview of fundamental ML techniques, highlighting six widely used methods with practical examples. Section 4 examines the application of AI methods in levitation control, operational maintenance, and track maintenance, emphasizing their technical contributions and real-world impact. Section 5 analyzes current challenges in Maglev systems and AI technology, offering insights into future research and development directions. Finally, Section 6 summarizes the advantages of AI in Maglev transportation systems, highlights its importance and prospects, and calls for expanded research beyond levitation control to enhance practicality and accelerate engineering implementation.

2. Representative Maglev Transportation Systems

Among Maglev technologies, EMS and EDS are the most mature and commercially deployed, covering the full range of operating speeds and embodying the core technical challenges of Maglev engineering. Hybrid suspension systems remain largely limited to theoretical studies and small-scale prototypes and are therefore not the focus of this paper.

2.1. EMS-Type Maglev Transportation Systems

2.1.1. System Classification and Motor Configuration

The EMS-type Maglev trains use attractive electromagnetic forces between onboard electromagnets and ferromagnetic guideways to achieve stable levitation. Based on the configuration of the linear motor, EMS systems can be classified into two categories: those employing short-stator linear induction motors and those using long-stator linear synchronous motors. The long-stator configuration enables higher operating speeds and is thus more suitable for high-speed applications [4]. Short-stator EMS systems, primarily used in urban rail transit applications, integrate the stator of the linear motor beneath the suspension frame of the vehicle. In contrast, long-stator systems feature the stator embedded along the track structure, allowing for continuous propulsion and improved speed capabilities. The two types also differ in suspension frame design: short-stator systems generally use a square-shaped frame, whereas long-stator systems employ an I-shaped frame optimized for high-speed dynamics.

2.1.2. Subsystem Composition and Functional Overview

The EMS-type Maglev system includes four major subsystems: guideway, vehicle, traction power supply, operation control, and communication. Each subsystem plays a critical role in ensuring the safe, stable and efficient operation of the Maglev train. The guideway system includes line, track, turnouts, and auxiliary facilities. The track provides suspension and traction support via steel or reinforced concrete beams and piers. Turnout operated hydraulically or electrically enables train routing and switching. Auxiliary facilities cover lighting systems, drainage infrastructure, and service roads to main support, maintenance and accessibility. The vehicle system comprises a suspension frame, secondary suspension system, car body, and electrical equipment. Electromagnets on the suspension frame enable smooth travel over curves and slopes. The secondary suspension system uses air springs and dampers to reduce vibration and improve passenger comfort. The car body focuses on structural strength, aerodynamics, and passenger safety. Electrical equipment adopts redundancy and modularity to enhance operational reliability and ease of maintenance. The traction power supply system includes on-board power supply units, substations, power cables, and switching stations. Power is drawn from the main electricity grid, then transformed, rectified, and inverted to power the long stator windings that generate propulsion forces. Substations supply electricity to specific guideway sections; multiple three-phase sets are used for transformation, and switching stations enable segmented power supply to improve efficiency and reduce energy losses. The operation control and communication system ensures the safe and efficient operation of the train by performing real-time monitoring, executing commands, and fault diagnosis. It maintains continuous wireless communication between the train and ground-based control centers to ensure smooth and responsive operation.

2.1.3. Speed Grade-Related Characteristics and Core Challenges

EMS systems are further divided into medium-low speed (urban rail transit, short-stator linear induction motor) and high speed (inter-city transit, long-stator linear synchronous motor) based on speed grades, with distinct dynamic characteristics and challenges that directly affect ML requirements.

Medium-low speed EMS (operating speed < 120 km/h): Key conditions include urban commuting with frequent starts/stops, small levitation gap fluctuations (±1 mm), and high sensitivity to environmental temperature (sensor measurement deviation easily caused by −20~60 °C temperature variation). Core challenges focus on “cost-effective real-time control” and “sensor anti-interference”.

High-speed EMS (operating speed > 300 km/h): Key conditions include inter-city travel with long-term high-speed operation, significant external disturbances (aerodynamic resistance, track irregularity-induced vibration), and strict requirements for response latency. Core challenges focus on “high robustness against disturbances” and “predictive control to reduce latency”.

2.2. EDS-Type Maglev Transportation Systems

2.2.1. Development Overview and Operating Principles

The EDS-type Maglev trains have been steadily developed domestically and internationally. Japan has taken a pioneering role with the construction of a low-temperature superconducting (LTS) Maglev test vehicle and the corresponding Yamanashi Test Line. The suspension principle of EDS-type Maglev trains is based on superconducting technology, where superconducting materials are cooled to approximately −269 °C to reach a zero-resistance state. EDS Maglev vehicles maintain a levitation height of 80–150 mm, effectively minimizing the risk of undesired contact between the train and the guideway.

A defining feature of this system lies in the principle of dynamic electromagnetic induction: as train speed increases, the induced currents in the figure-eight coils embedded along the U-shaped track sidewalls also rise. When the train reaches its rated speed, the induced current becomes sufficient to sustain stable levitation. Designed based on the zero-flux principle, these coils serve as integral components of both the suspension and guidance subsystems. Through the interaction between the vehicle’s superconducting magnetic field and that of the track coils, the EDS system achieves inherently stable levitation and guidance without continuous power input to the guideway, a key distinction from EMS systems.

In addition to LTS-based systems, HTS-type Maglev vehicles utilizing the flux-pinning effect have also been proposed. In this configuration, superconducting materials interact with magnetic fields to realize stable levitation without active control. HTS-based systems demonstrate advantages such as simpler cooling requirements and enhanced levitation stability, offering promising prospects for future high-speed Maglev development.

2.2.2. Speed Grade-Related Characteristics and Core Challenges

EDS systems can be broadly associated with medium HTS and ultra-high-speed configurations, each exhibiting distinct multi-physics coupling effects that influence the direction of ML applications.

Medium-high speed EDS achieves levitation through the flux-pinning effect, characterized by a moderate suspension height and significant thermal vibration coupling, where temperature fluctuations in superconducting materials affect the levitation force. Core challenges include thermal vibration correlation and low-speed startup stability.

Ultra high-speed EDS, based on zero-resistance superconducting coils, features a larger suspension height and places extreme demands on system reliability, as a sudden superconducting quench can pose serious safety risks. Key challenges include multi-modal fault early warning and defect detection under limited data conditions.

From a quantitative perspective, clear distinctions exist between EMS and EDS systems. EMS covers medium-low to high-speed ranges, while EDS operates primarily in medium-high to ultra-high-speed domains. EMS systems exhibit smaller levitation gaps and require active power input, whereas EDS systems rely on passive levitation with larger levitation gaps. Structurally, EMS employs short/long-stator linear motors with square or I-shaped frames, while EDS utilizes HTS/LTS superconductors with figure-eight track coils.

2.3. Common Technical Features of EMS and EDS Systems

The absence of mechanical wear due to the contactless operation between Maglev trains and their guideways significantly reduces maintenance costs. This contactless operation is enabled by the electromagnetic coupling between the train and the guideway, which constitutes the most distinctive feature differentiating Maglev trains from conventional wheel-rail systems. For EMS-type Maglev systems, the coupled train guideway system mainly comprises the vehicle’s suspension frame, track beam, and levitation control system, with the latter playing a pivotal role in ensuring stable and safe operation. Since optimizing the levitation control system generally does not require hardware replacement, it serves as an economical and effective means of performance enhancement through software updates, a practice commonly adopted in engineering applications [5]. In this context, AI represents a promising direction for enhancing levitation control, motivating a closer examination of representative AI methods in the following section.

3. Representative AI Methods

Among the numerous branches of AI, ML stands out as a core subfield of AI that goes beyond traditional rule-based programming. Instead, it empowers computers to autonomously learn from vast datasets to make predictions or informed decisions without explicit instructions. This learning approach has been a driving force behind many modern technological innovations, including autonomous vehicles, intelligent voice assistants, and personalized recommendation systems. However, ML is considered a key technology in realizing the broader vision of AI by simulating aspects of human intelligence through data-driven approaches. This paper delves into the fundamental principles of ML, its close association with AI, and how ML techniques address real-world challenges in Maglev transportation systems. Based on variations in training data and feedback mechanisms, ML can be classified into six categories: supervised learning, unsupervised learning, semi-supervised learning, deep learning, reinforcement learning, and transfer learning.

Supervised learning is an ML process, also known as a classifier, that is trained on labeled samples by adjusting the parameters to achieve the desired performance. Based on the type of labels, it can be divided into classification and regression problems. Classification predicts discrete categories for samples, while regression is concerned with continuous real-number outputs for samples. Common algorithms used in supervised learning are decision trees, support vector machines (SVM), k-nearest neighbors (K-NN), random forests, and naive Bayes. Each offers unique strengths depending on the nature of the task and dataset. For Maglev systems, this method is widely used in sensor temperature compensation (e.g., radial basis function (RBF)+ least squares-support vector machine (LS-SVM) for gap sensor calibration) and fault classification (e.g., SVM for levitation system fault diagnosis).

In contrast, unsupervised learning tackles pattern recognition problems using unlabeled training samples, and representative examples are clustering algorithms and dimensionality reduction techniques. In Maglev applications, this method is used for unsupervised clustering of unlabeled track beam vibration data (e.g., detecting subtle structural irregularities without manual annotation) to identify potential infrastructure anomalies.

Semi-supervised learning is considered a bridge between supervised and unsupervised learning. It uses a small amount of labeled data along with a larger amount of unlabeled data to perform either supervised or unsupervised tasks. This technique is particularly valuable in real-world scenarios, where obtaining labeled data is costly or time-consuming, enabling models to learn more effectively from limited supervision [6]. For Maglev, it reduces the cost of labeling scarce ultra-high speed EDS data (e.g., superconducting quench fault samples) by combining a small number of labeled quench cases with large unlabeled EDS operational data.

Deep learning is a subfield of ML that uses labeled training samples to create models capable of learning high-level abstract representations from lower-level features [7]. This is obtained by numerous layers of nonlinear transformations, enabling the model to extract the difficult pattern automatically. It primarily involves three types of methods: convolutional neural networks (CNNs), autoencoder neural networks based on multilayer neurons, and deep belief networks. Deep learning has achieved notable success in areas like image recognition and natural language processing (NLP) and exhibits high effectiveness in addressing complex pattern recognition challenges, significantly advancing the application of AI technology. In Maglev transportation, deep learning powers high-speed EMS levitation gap prediction (e.g., CNN-AGCN model reducing gap prediction error) and track structure predictive maintenance.

Reinforcement learning utilizes an unlabeled training set to describe and solve problems. Unlike supervised learning, reinforcement learning does not depend on labeled input and output pairs. Instead, it explores actions to maximize rewards or achieve specific goals over time. Common algorithms in reinforcement learning include Q-learning, temporal difference methods, and multi-agent reinforcement learning, each tailored to specific control and optimization tasks. For Maglev control, reinforcement learning enables adaptive levitation adjustment, e.g., DRL narrows high-speed EMS levitation gap fluctuation under track irregularities, meeting response latency requirements.

Transfer learning improves the learning efficiency in a target domain by transferring knowledge obtained from a source domain that shares similar features. This technique is helpful when labeled data in the target domain is scarce and costly to obtain. Transfer learning can be categorized into four types: instance-based, parameter-based, feature representation-based, and relation knowledge-based transfer. In Maglev maintenance, it transfers fault diagnosis knowledge from mature lines (e.g., Shanghai Maglev Demonstration Line) to new medium-low speed EMS lines, reducing the need for labeled fault data in the target line.

These ML techniques offer complementary strengths for addressing complex control, monitoring, and maintenance challenges in Maglev systems. The relationships and distinctions among the various ML approaches, along with their potential applications in Maglev transportation, are illustrated in the overview diagram (Figure 1).

4. AI Applications in Maglev Transportation Systems

Specific applications of AI in Maglev transportation systems are summarized (

Table 1. The specific applications of AI in the field of Maglev transportation systems.

Area	Application of AI	Relevant Literature
Traction, Levitation, and Guidance Controller Design	Optimize control performance through PID optimization algorithms.	[8,9,10,11,12]
	Fuzzy logic control to address nonlinearity and uncertainty in magnetic suspension.	[13,14,15,16,17,18,19,20,21,22,23,24,25,26]
	Neural network control to handle parameter variations and disturbances in Maglev transportation systems.	[27,28,29,30,31,32,33,34,35,36,37,38,39]
	Reinforcement learning enabled adaptive levitation control in complex Maglev environments.	[40,41,42]
	Fault diagnosis algorithms in adaptive levitation control.	[43,44]
	Sliding mode control to enhance anti-interference capability and dynamic response in magnetic suspension.	[45]
	Koopman operator theory to effectively resolve nonlinear control issues in Maglev train levitation systems.	[46]
	Nonlinear predictive control to accomplish dynamic modeling and predictive control for magnetic suspension.	[47]
	A CNN-AGCN-based prediction model to improve levitation gap prediction accuracy for high-speed Maglev trains.	[48]
	Deep learning and semi-supervised control algorithms to enhance vertical operational safety and reliability.	[49]
Vehicle Design, Operation and Maintenance	RBF neural networks and LS-SVM provide temperature compensation for gap sensors.	[49]
	Fuzzy neural networks modeled gap sensors in high-speed Maglev trains.	[50]
	ML optimized parameters for multi-surface HTS Maglev transportation systems.	[51]
	FNN and backstepping control enhanced hybrid Maglev transportation systems.	[52]
	A data-driven approach diagnosed faults in high-speed Maglev train levitation systems.	[53]
	An adaptive PSO-based linear active disturbance rejection controller designed for traction systems.	[54]
	A PSO-optimized real-time PID control strategy applied to Maglev transportation systems.	[55]
	SVM and decision trees simulated circuit fault diagnosis.	[56]
	Ensemble learning algorithms to perform integrated fault assessment for Maglev trains.	[57]
	A data-driven approach to optimize train operational strategies based on onboard battery condition monitoring.	[58]
Infrastructure Manufacture, Monitoring, and Maintenance	Deep learning enabled predictive maintenance.	[59,60,61]
	A maintenance management system developed and implemented for the Shanghai Maglev Demonstration Line.	[62]
	Real-time detection enabled fault detection in Maglev controllers.	[63,64]
	RCM used in maintenance strategies for urban Maglev trains.	[65]

), encompassing traction, levitation, and guidance controller design; vehicle design; operation and maintenance; infrastructure manufacturing, monitoring, and maintenance.

Table 1 summarizes AI applications across three core domains of Maglev transportation systems: traction, levitation, and guidance controller design; vehicle design, operation, and maintenance; and infrastructure manufacture, monitoring, and maintenance. It maps over 12 AI techniques such as PID optimization, fuzzy logic control (FLC), neural network control, and deep learning to specific engineering tasks, based on more than 65 studies, reflecting comprehensive literature coverage. From a technical perspective, AI applications in Maglev are strongly scenario oriented.

In traction, levitation, and guidance control, AI primarily addresses system nonlinearity and open-loop instability, e.g., FLC for handling uncertainties and Koopman operator theory for nonlinear control challenges. In vehicle design, operation, and maintenance, the focus is on improving precision and cost-effectiveness, e.g., RBF neural networks combined with LS-SVM for levitation gap sensor temperature compensation. Finally, infrastructure management emphasizes predictive and intelligent monitoring, e.g., deep learning for predictive maintenance and BIM for SHM, forming a targeted technical solution system for diverse operational demands of Maglev trains.

4.1. Applications in Traction, Levitation, and Guidance Controller Design

4.1.1. Challenges and Limitations of Traditional Control

The levitation system of Maglev trains is an inherently nonlinear and open-loop unstable system, posing significant challenges for control design. Traditional control algorithm design relies on Taylor series expansion around the operating point to linearize the system. Control algorithms are then designed based on this linearized model, which only works effectively within a small range near the operation point. However, the Maglev train is subjected to significant external disturbances such as payload changes, track irregularity, vertical train vibrations, and sensor hysteresis. The system may deviate greatly from the operating point. Under such conditions, the linear approximation becomes invalid, and the system’s nonlinear characteristics cannot be addressed effectively, leading to instability or even suspension failure. These limitations highlight a significant challenge in the practical implementation of linearization-based control algorithms for Maglev traction, levitation, and guidance systems. The sensitivity of traditional controllers to nonlinear disturbances restricts their robustness in real-world scenarios. To resolve this issue, recent work has studied intelligent control techniques capable of adapting to nonlinearities and uncertainties in Maglev trains. Literature [8] demonstrates a detailed review of these techniques, evaluating their advantages and drawbacks, offering valuable recommendations for the future directions of levitation control strategies.

4.1.2. AI-Based Methods for Traction, Levitation, and Guidance Systems

Artificial intelligence technologies provide effective solutions to address the shortcomings of traditional control methods in the Maglev transportation system. Specific applications of AI-enhanced control can be categorized as follows:

PID Controller Optimization: AI algorithms have been successfully used to increase the performance of PID controllers, which are widely employed in Maglev levitation and traction systems. Literature [9] introduced an improved PSO algorithm to augment PID controllers in Maglev trains. This method significantly reduced controller output overshoot and addressed the limitations of the present PID and fuzzy control methods in terms of weak disturbance rejection and slow response. Literature [10] introduced a further parameter optimization method for PID controllers based on an improved PSO algorithm, which was applied to Maglev transportation systems. By improving the global search capability of the PSO algorithm, the proportional, integral and derivative parameters of the PID controller were more effectively optimized, leading to enhanced control accuracy and robustness. Literature [11] introduced an ant colony optimization algorithm to optimize the fractional-order PID controller for the reference design and control of Maglev transportation systems, reducing both the stable time and oscillation of the levitated object. This approach demonstrated superior adaptability in dynamic conditions. Literature [12] introduced a novel control method combining model predictive control (MPC), fuzzy control, and PID control for permanent-magnet Maglev trains. This hybrid system improves the dynamic performance and stability of the traction system by optimizing control parameters online.

Fuzzy Logic Control: This technique has proven to be a powerful approach to address the nonlinear, uncertain, and dynamic features of Maglev systems. Literature [13] introduced an adaptive fuzzy levitation control method based on the dynamic coupling model of multiple electromagnets. This approach simultaneously addresses dead-band and saturation problems while ensuring finite-time convergence of the air gap tracking error for multiple electromagnets. Literature [14] introduced a control method for Maglev transportation systems based on FLC to enable precise control of position and velocity through the stages of fuzzification, inference, and defuzzification. Simulation and experimental results show that the fuzzy control method has better performance than traditional PID controllers in terms of system stability, response speed, and anti-interference. This method effectively solves the nonlinear problem of the Maglev system and improves the control performance. Literature [15] introduced a nonlinear dynamic modeling method for the EMS system of a low-speed Maglev train, along with a fuzzy sliding mode control (SMC) strategy. This method combines fuzzy control and SMC to improve the system’s robustness to nonlinearities and external disturbances while maintaining stability and tracking performance.

To further improve control performance under external disturbance and nonlinear dynamics, advanced fuzzy-based and neuro-fuzzy control strategies have been developed for Maglev systems. Literature [16] introduced a fuzzy H_∞ robust control strategy based on the T-S fuzzy model for the magnetic levitation system of Maglev vehicles. This method integrates H_∞ robust control theory with fuzzy logic inference under the T-S fuzzy framework, effectively addressing the nonlinear and time-varying characteristics of the electromagnetic levitation system. By combining fuzzy approximation and H_∞ optimization, the proposed controller enhances the system’s robustness against parameter uncertainties and external disturbances while maintaining high-precision levitation stability. Simulation and experimental results verify that this method improves the dynamic response and reduces overshoot compared with conventional linear control schemes. Literature [17] introduced a control method based on a fuzzy neural network for precise control of Maglev transportation systems. This method combines the robustness of fuzzy control and the self-learning ability of neural networks. It successfully deals with the nonlinear characteristics and external disturbances of the system, allowing the system to maintain stable operation even in complex dynamic environments. Literature [18] introduced a networked levitation control strategy based on the Takagi-Sugeno (T-S) fuzzy model to address uncertainties and external disturbances in the networked Maglev control systems. This method combines the T-S fuzzy model with robust control strategies to enhance the robustness and stability of the system, particularly under conditions of communication delays and faults.

To address the uncertainty, nonlinearity, and perturbations in Maglev levitation systems, a variety of hybrid fuzzy control techniques have been established by combining fuzzy logic with optimization, adaptive, and robust control approaches. Literature [19] introduced a control method combining the T-S fuzzy model, fuzzy adaptive control, and PSO for the levitation system of Maglev trains. This method improves the stability, response speed, and robustness of the levitation system by optimizing the control parameters. Literature [20] introduced a fuzzy adaptive robust control method based on the T-S fuzzy model. By combining fuzzy logic with robust control theory, the system can effectively deal with uncertainty and external disturbances. Simulation results show that this method is superior to traditional control methods in terms of response speed, steady-state error, and anti-interference ability. Literature [21] introduced a fuzzy PID control algorithm for Maglev transportation systems to improve their stability and response speed. Literature [22] introduced a fuzzy-logic-based controller to address nonlinearity and uncertainties in Maglev transportation systems. Both simulation and experimental results validated its effectiveness in enhancing system stability and response speed, demonstrating more robustness than traditional PID control in handling disturbances and parameter changes. Literature [23] introduced an adaptive control method based on fuzzy-inverter control. In this method, the control strategy is dynamically adjusted through fuzzy control rules to improve the stability of the Maglev system under different conditions.

Recent developments in Maglev control systems have extended beyond classical fuzzy control by including robust control methods and vibration suppression approaches to further increase the system’s stability and performance. Literature [24] introduced a fuzzy H_∞ control method based on the T-S fuzzy model for levitation systems. By integrating fuzzy control and H_∞ control, this method enhances system stability and disturbance rejection performance, making it suitable for uncertain and noise-prone environments. Literature [25] introduced a synergistic integration of the Internet of Things (IoT) and an improved adaptive fuzzy control framework for medium- and low-speed Maglev transportation systems. Real-time monitoring capabilities provided by IoT enable dynamic feedback to the fuzzy controller, allowing for real-time optimization of control parameters and ensuring operational stability for different environmental conditions. Literature [26] introduced the vibration control problem caused by mechanical coupling in Maglev trains. It developed and optimized corresponding vibration control strategies and verified the effectiveness of the proposed methods in suppressing mechanical coupling vibration and improving the overall stability of Maglev trains through experimental tests.

Neural Network Control: Literature [27] introduced the application of adaptive neural network control to EMS-type levitation systems with the goal of both system stability and dynamic responsiveness. Literature [28] introduced a more advanced adaptive neural network control method to improve Maglev train stability under time-varying mass and external disturbances. Experimental results justified that this technique effectively mitigates system disturbances and uncertainties, leading to improved overall control performance, faster response efficiency, reduced tracking error, and enhanced robustness.

Literature [29] introduced an advanced intelligent control method for Maglev transportation systems combining backstepping, SMC, and recurrent neural networks (RNN). Targeting dynamic uncertainties in high-speed EMS levitation control, this hybrid approach employs a multi-input multi-output (MIMO) RNN (Figure 2), which receives real-time levitation gap and electromagnet current as inputs and outputs gap-deviation compensation and electromagnetic-force adjustment to enhance robustness and real-time performance. Experimental results demonstrate that it reduces dynamic tracking error by 35% compared with traditional SMC.

The dynamic neural network architecture adopted in this study is illustrated in the schematic diagram of the system (Figure 2). The model contains an input layer, a recurrent state layer, and an output layer. The input vector $Z_1\cdots Z_i$ represents external measurable signals of the levitation system (e.g., gap, current). These inputs are projected to the state layer, where each state neuron includes a nonlinear activation function S and a one-step delay operator $Z^{-1}$ which provides the previous state value $\mathsf{\Gamma }\left(t-1\right)$ for computing the current state $\mathsf{\Gamma }\left(t\right).$ The nodes labeled ${\mathsf{\Gamma }}_1\cdots {\mathsf{\Gamma }}_j$ denote the components of the recurrent state vector, forming the dynamic memory of the network.

Weighted connections $W_{ij}$ represent the parameterized mapping between inputs, recurrent states, and outputs. The summation nodes $\mathsf{\Sigma }$ aggregate these weighted inputs to generate the output vector $f_1\cdots f_n$ corresponding to the RNN-generated control actions, including gap correction and electromagnetic-force adjustment. By integrating $Z^{-1}$ based temporal feedback, nonlinear activation units, and MIMO weighted mappings, the network effectively models the nonlinear and time-varying dynamics of the EMS levitation process. These variables jointly define the core computational structure of the proposed intelligent levitation control strategy.

Neural network methods continue to gain traction in Maglev system control due to their capability to model complex, nonlinear dynamics and adapt to varying conditions in real time. Different innovative neural network-based control techniques have been established to increase the system’s performance. Literature [30] introduced a neural network-based modeling and control method for Maglev transportation systems. This method improves system accuracy and stability, particularly when handling nonlinear and complex dynamics. Literature [31] introduced an adaptive neural network control method for Maglev transportation systems, wherein the controller dynamically adjusts its parameters using a neural network model. This method effectively addresses nonlinearity and external disturbances. Simulation results show superior performance in stability, response speed, and robustness compared to traditional PID control. Literature [32] introduced the neural network event-triggered quantization control method based on an observer for active levitation systems. The method reduces computational and communication burdens through event-triggering mechanisms, improving system stability and control performance.

Literature [33] introduced a spiking neural network (SNN)-based fault prediction model for Maglev electromagnets. This method fuses vibration feature data from electromagnet sensors to identify incipient faults (e.g., coil insulation degradation), simplifying the difficulty of capturing early anomaly signals in traditional electromagnetic force models. With an accuracy of 95.6%, the model realizes real-time fault pre-warning during Maglev operation, avoiding the risk of sudden suspension failure caused by undetected electromagnet defects and providing an efficient tool for proactive maintenance of key Maglev components. Literature [34] introduced an adaptive neural fuzzy robust position control scheme for Maglev trains. Combining neural networks and fuzzy logic, this method maintains high-precision control under external perturbations and uncertainties. The structure of the neural fuzzy system, which highlights the integration of adaptive learning with fuzzy inference mechanisms, is shown in the system schematic (Figure 3). In this framework, each raw input $r_i\left(x\right)$ represents a specific sensing modality or system feature, which is subsequently transformed through expert-dependent feature mappings $I_{ij}\left(x,y^j\right)$. These mappings project the input signals into expert-specific latent spaces governed by the parameter sets $y^j$. The aggregated outputs ${agr}_j\left(x,y^j\right)$ of all experts encode localized dynamic representations, which are then fused by the summation operators $\mathsf{\Sigma }$ to obtain a unified control signal. Finally, the fused output is compared against the reference trajectory via a subtraction operator to produce the error signal that drives the adaptive update laws. This hierarchical structure enhances learning efficiency, improves robustness against multi-source uncertainties, and ensures reliable tracking performance in nonlinear Maglev levitation and guidance dynamics.

Recent advancements in intelligent control for Maglev systems have focused on a hybrid technique that integrates neural networks with advanced control theories. Literature [35] introduced a control method that combines neural networks and MPC for Maglev trains. In this method, neural networks are used to address system nonlinearity, while MPC optimizes real-time control. Experimental results demonstrate improved response speed and system stability, validating the effectiveness of this hybrid framework in dynamic operating environments. Literature [36] introduced a control method integrating Lyapunov’s stability theory and ML for nonlinear Maglev transportation systems. By dynamically adjusting control strategies in response to real-time system changes, this method ensures stability under nonlinearity and uncertainty. Literature [37] introduced a method combining robust dynamic SMC and adaptive extended neural networks (RENN) for Maglev transportation systems. The RENN network structure, illustrating the adaptive learning and compensation mechanisms incorporated into the controller, is depicted in the proposed control framework (Figure 4). As depicted in the figure, the RENN model adopts a recurrent neural network architecture in which the input layer receives real-time signals $x_i\left(N\right)$, and the nonlinear activation of hidden neurons is represented by SSS The forward connections from the input layer to the hidden layer are weighted by $a_{ij}$, while the recurrent feedback from the context layer, implemented through unit-delay operators $Z^{-1}$, is governed by the adaptive weights $b_{ij}$, enabling temporal memory and dynamic feature extraction. The hidden states are further aggregated in the output layer using weights $w_j$ to generate the control output $y\left(N\right)$. By integrating these adaptive learning, feedback memory, and nonlinear compensation mechanisms, the proposed RENN-based robust control method significantly enhances system stability and resilience under parametric uncertainties and external disturbances.

Radial basis function neural networks method is highly effective for adaptive control of nonlinear systems due to their universal approximation capability and fast convergence. Recent research on RBF network enhances the performance and robustness of Maglev transportation systems. Literature [38] introduced an adaptive SMC method based on the minimum parameter learning approach of RBF neural networks for Maglev transportation systems. The combination of RBF neural networks and SMC improves system stability and anti-interference capabilities. Literature [39] introduced a supervision and control method based on RBF neural networks to address network delays in Maglev trains running on elastic tracks. This method compensates for delays, significantly improving system stability and trajectory tracking accuracy.

Reinforcement Learning Control: Literature [40] introduced a levitation regulation method based on DRL for medium- and low-speed Maglev trains. This method optimizes levitation control strategies to improve stability and comfort. The DRL framework, which comprises two neural networks (the actor and the critic), is depicted in the model architecture (Figure 5).

The critic network (upper diagram) takes the system state as input and passes it through multiple hidden layers to output the Q-value, representing the expected cumulative reward of a given action. The actor network (lower diagram) has a similar multilayer structure but includes neuron and bias nodes (NB) to generate the optimal control command based on the current state. By alternately updating the actor and critic networks, the DRL controller continuously refines the levitation control strategy through interaction with the environment. Simulation results show that the proposed DRL-based method effectively handles system uncertainty and nonlinear dynamics, achieving superior stability and comfort compared with traditional control methods.

Literature [41] introduced a control method based on DRL for levitation systems. Using deep Q-networks (DQN), this method optimizes control strategies, improving stability and response speed. The reinforcement learning framework is illustrated in the schematic diagram (Figure 6).

Literature [42] introduced an optimal tracking control method based on reinforcement learning. The method significantly addresses input delays, enabling accurate trajectory tracking and improving response speed and stability.

Fault Diagnosis and Fault-Tolerant Control: Ensuring system reliability and maintaining safe operation under fault conditions are important for the success of Maglev transportation system. Literature [43] introduced the fuzzy multi-attribute decision-making methods to effectively analyze and prioritize the key fault risks of Maglev train levitation systems. This method increases the strength of safety management which helps to examine the crucial areas for intervention and prevention. Literature [44] introduced the Fault-tolerant control strategies for electro Maglev transportation systems. This study focused on developing advanced fault detection, diagnosis, and adaptive control methods that ensure stable operation in the presence of system faults. These techniques were verified through both simulations and experimental tests, manifesting their effectiveness in maintaining system stability and performance during fault conditions.

Nonlinear Modeling and Control: Exact nonlinear modeling and robust control design are significant to analyze the complex dynamics of multi-electromagnet levitation systems in Maglev trains. Literature [45] introduced the adaptive fuzzy super-twisting control (AFSTC) strategy that combines fuzzy compensation with a super-twisting algorithm to mitigate chattering effects for maglev trains’ multi-electromagnet system. This method was implemented on a portable field-programmable gate array (FPGA) controller and experimentally validated on Shanghai Maglev Line. Results determined that the AFSTC significantly outperforms PID controllers, providing superior stability and robustness under electromagnetic coupling and nonlinear conditions.

Literature [46] introduced the date-driven control strategies that were considered a useful method to address the difficult nonlinear dynamics of Maglev systems without depending on explicit physical models. A data-driven control method based on the Koopman operator for levitation systems. The Koopman operator transforms nonlinear systems into high-dimensional linear systems, simplifying control design. This method improves stability, reduces vibration, and enhances adaptability to external disturbances.

To resolve the issues faced by the nonlinear and time-varying dynamics of Maglev systems, scientists have studied various optimal control, predictive modeling and ML-based techniques to enhance stability and accuracy. Literature [47] introduced the application of linear quadratic regulator (LQR) and state-dependent Riccati equation (SDRE) methods in Maglev transportation systems. Simulation results show that the SDRE method outperforms LQR in handling nonlinear dynamics, demonstrating superior system stability and adaptability in real-world scenarios.

Deep Learning and Semi-Supervised Control: Recent developments in deep learning and semi-supervised learning have given powerful methods to increase control accuracy and ensure safety in Maglev systems. Literature [48] introduced the Hopfield neural network for parameter identification in dynamic systems. Experimental results confirm the method’s effectiveness in improving modeling accuracy. Literature [49] introduced a deep learning-based semi-supervised control method to ensure vertical safety and bound air gaps in Maglev trains. Experimental results demonstrate superior performance compared to traditional methods, improving system stability, safety and overall performance.

4.1.3. Auxiliary Support Technologies and Application Enhancements

In addition to the core AI control methods mentioned above, supplementary technologies have been developed to further optimize system performance. These advancements increase the system efficiency, minimize the computational load and confirm robust operation under dynamic conditions. For example, reference [32] introduced an event-triggered mechanism in adaptive neural network control. This method reduces the computational and communication burdens of the system while ensuring control performance and efficiency in real-time applications. Reference [36] introduced a combined Lyapunov’s stability theory with ML. This technique ensures system stability under nonlinearities and uncertainties, providing robust control in unpredictable operating environments. These technologies effectively complement core AI-based control methods, contributing to a more comprehensive and efficient traction, levitation, and guidance control system.

The diverse AI approaches applied to traction, levitation, and guidance controller design are comprehensively summarized (

Table 2. Summary of AI applications in traction, levitation, and guidance controller design.

Area	Main Purpose	AI Method
PID Controller Optimization	Reduce overshoot, improve response speed and stability	Improved PSO
Fuzzy Logic Control	Handle nonlinearity and uncertainty, improve robustness	FLC, T-S Fuzzy Model
Neural Network Control	Handle parameter changes and disturbances, improve control accuracy and adaptability	RNN, RBF, CNN-AGCRN, etc.
Reinforcement Learning Control	Achieve adaptive control, optimize levitation strategy	DRL, DQN
Fault diagnosis and tolerant control	Real-time monitoring and fault response, improving system reliability	Fuzzy Diagnosis, Adaptive Control
Nonlinear modeling	Enhance anti-disturbance ability, improve dynamic response performance	RBF-ARX Model
Koopman operator	Transform nonlinear systems into high-dimensional linear systems, simplify control design	Data-driven Linearization

). The table distills these approaches into seven distinct technical directions, explicitly linking each control task’s primary objective to its corresponding AI method. The focus is on addressing the inherent nonlinearity and open-loop instability of Maglev levitation systems. From a technical perspective, AI tackles these core challenges through two complementary pathways. The first is the optimization of traditional control frameworks, for example, improved PSO algorithms tune PID parameters, reducing controller output overshoot in practical applications. The second is the innovation of nonlinear control techniques, where reinforcement learning enables adaptive levitation control, narrowing levitation gap fluctuations under complex conditions such as track irregularities. Additionally, Koopman operator theory provides a data-driven means for linearizing nonlinear systems, simplifying control design. Neural networks including RNNs, RBF networks, and CNN-AGCRN models serve as versatile tools across multiple directions, effectively handling parameter variations and external disturbances in Maglev systems.

4.2. Applications in Vehicle Design, Operation and Maintenance

4.2.1. Challenges and Limitations of Traditional Methods

Despite significant advances in Maglev transportation technology, traditional methods face several critical limitations across vehicle design, operation, and maintenance. In levitation system control, conventional strategies lack predictive capabilities, preventing pre-adjustment of parameters based on future system states and leading to delayed responses to sudden changes or external disturbances. Temperature compensation for gap sensors is often inaccurate, and sensor measurements are easily influenced by environmental variations, causing deviations in levitation control. Fault diagnosis methods rely on manual experience and fixed-threshold criteria, resulting in low diagnostic accuracy and difficulty in identifying complex fault types.

In parameter optimization for HTS-type Maglev systems, trial-and-error or purely analytical approaches are inefficient and cannot fully explore optimal parameter combinations. Similarly, accurate analysis of thermal-vibration coupling is essential for system health monitoring and early fault detection, but traditional methods lack intelligent data-processing capabilities, making it challenging to establish precise correlations between thermal and vibration signals.

Furthermore, conventional vehicle operational status monitoring depends on wired sensors and centralized data processing, which are limited by high latency, low scalability, and poor adaptability in complex vibration environments. These limitations highlight the need for intelligent data-driven approaches to enhance the performance, reliability, and adaptability of Maglev transportation systems.

4.2.2. AI-Based Methods for Vehicle Design, Operation and Maintenance

Artificial intelligence has been increasingly applied to address the limitations of traditional temperature compensation and fault diagnosis methods in Maglev systems. AI-driven approaches provide higher accuracy and adaptability than classical techniques. Key applications are as follows.

Sensor Temperature Compensation: Literature [49] introduced a hybrid model combining RBF neural networks and LS-SVM to predict degradation trends of Maglev gap sensors. This method not only compensates for temperature-induced measurement deviations but also realizes 30-day life prediction of gap sensors by learning the correlation between sensor output drift and internal component degradation. It addresses the limitation of traditional temperature compensation methods that only correct errors without predicting sensor failure, laying a foundation for predictive replacement of key sensing components.

Literature [50] introduced a two-level inverse model compensation scheme for high-speed Maglev train levitation gap sensors, utilizing RBF neural networks, T-S fuzzy neural networks, and LS-SVM. The combined models demonstrated smaller compensation errors compared to individual models, providing a foundation for further calibration to eliminate environmental interference.

Fault Diagnosis: Literature [51] introduced a data-driven fault diagnosis method for high-speed Maglev train levitation systems. By analyzing sensor data and employing ML algorithms such as SVM and neural networks, the method accurately identifies system failure types, improving diagnostic precision and reducing the misclassified faults. Literature [52] introduced a multi-modal fault diagnosis method for Maglev circuits based on SVM-decision tree fusion. The method integrates edge-sensed data (e.g., circuit current fluctuations, temperature rise, and voltage ripple) to classify fault types (e.g., capacitor breakdown, resistor aging). Deployed on a Maglev fault diagnosis platform, it reduces the false alarm rate to 1.8% and shortens fault localization time to <3 min significantly outperforming traditional single-modal SVM/decision tree methods and adapting to complex on-site electromagnetic interference.

Literature [53] introduced a fuzzy comprehensive evaluation method based on ensemble learning for Maglev levitation system fault diagnosis. To solve the low accuracy of traditional threshold-based diagnosis (e.g., misjudging electromagnet faults), this method combines fuzzy reasoning (for sensor noise handling) and ensemble algorithms (aggregating 5 base learners). Its optimization process (Figure 7) includes sensor data preprocessing (gap/current), feature selection, and parameter tuning experimental results show that it achieves 98.5% diagnostic accuracy for levitation system faults (e.g., coil short circuits) and outperforms single-classifier methods.

Literature [54] introduced a speed-tracking control strategy based on particle swarm optimization and adaptive linear active disturbance rejection control (ALADRC). This method integrates PSO-optimized controller parameters with an adaptive disturbance-rejection mechanism to enhance velocity regulation under uncertainties and external perturbations. By compensating for modeling errors and dynamic variations in real time, the proposed approach achieves faster convergence, reduced steady-state error, and improved robustness during high-speed train operation. Simulation results demonstrate that the optimized ALADRC significantly outperforms conventional PID and fixed-parameter ADRC in suppressing disturbances and maintaining precise speed tracking across varying operating conditions.

Parameter Optimization and System Modeling: Studies [45,55] introduced ML techniques, including supervised learning, reinforcement learning, and ensemble learning to optimize parameters of HTS-type Maglev transportation systems. Experimental results confirmed that ML-based optimization outperforms traditional methods, offering new insights for Maglev and other complex systems. This approach also yields significant aspects applicable to other higher-order nonlinear engineering systems. Literature [55] introduced an edge-deployed adaptive fuzzy neural network control (AFNNC) scheme to simultaneously address chattering in SMC and realize early fault warning for Maglev drive systems. The lightweight model is deployed on onboard edge devices, with a response latency of <40 ms, which not only supplies real-time control outputs to electromagnets and linear induction motors but also monitors current/torque anomalies to pre-warn drive system faults. Experimental validation on the Shanghai Maglev Demonstration Line showed AFNNC reduces chattering amplitude and improves fault detection rate compared to traditional SMC. Literature [56] introduced a federated PSO-optimized predictive maintenance strategy for Maglev traction systems. This method integrates operational data from 3 inter-city Maglev lines via federated learning (avoiding raw data sharing to protect privacy) and uses PSO to optimize the weight of the predictive model. It not only improves the response speed of traction systems but also realizes 72 h early warning for traction inverter faults. With a fault prediction accuracy of 97.2%, it reduces unplanned downtime of traction systems by 50% compared to traditional PID control which lacks predictive capability.

Nonlinear Controller Design: Literature [57] introduced an adaptive particle swarm optimization-based nonlinear active disturbance rejection controller (APSO-NLADRC) to address deficiencies in automatic train operation control algorithms. The NLADRC parameter adjustment process, highlighting how APSO searches the parameter space to improve controller performance under certain constraints, is illustrated in the optimization diagram (Figure 8). The APSO algorithm optimizes the key NLADRC parameters, namely the proportional gain $K_p$, derivative gain $K_d$, observer bandwidth $\omega$, and control gain estimate b, thereby tuning both the response speed and disturbance compensation capability of the controller. Specifically, $K_p$ and $K_d$, determine the feedback strength and dynamic stability,$\omega$ defines the convergence speed of the extended state observer (ESO), and b represents the system’s estimated control gain used for nonlinear compensation. This optimization process addresses premature convergence issues in traditional PSO algorithms and significantly enhances tracking accuracy, disturbance rejection, and overall control robustness in highly dynamic Maglev systems.

Energy Management and Battery Health Monitoring: Literature [58] introduced an operation control strategy based on on-board battery state monitoring for Maglev trains. This strategy employs multi-sensor fusion technology to collect key parameters of onboard batteries in real time, including charge discharge current, terminal voltage, and core temperature, thereby overcoming the limitations of traditional fixed-threshold management methods that fail to adapt to dynamic battery state variations.

4.2.3. Auxiliary Support and Simulation Validation Technologies

In addition to core AI algorithms, a range of supplementary technologies play a key role in ensuring the practical accuracy and engineering applicability of AI methods in Maglev vehicle design, operation and maintenance. For example, reference [56] not only developed an SVM-based decision tree but also implemented it on a dedicated fault diagnosis platform, realizing the engineering application of fault diagnosis algorithms. In study [52], numerical simulations were used to verify the superiority of AFNNC over traditional control strategies, providing a reliable validation method for the promotion of AI control algorithms. These supporting technologies, such as signal preprocessing, platform implementation, and simulation validation, ensure the practicality and effectiveness of AI methods in engineering conditions.

A summary of AI applications in vehicle design, operation, and maintenance is presented (

Table 3. Summary of AI applications in vehicle design, operation, and maintenance.

Area	Main Purpose	AI Method
Sensor Temperature Compensation	Improve sensor measurement accuracy under temperature variations	RBF Neural Network + LS-SVM
Fault Diagnosis	Identify system failure types, improve diagnostic accuracy	SVM, Decision Tree, Ensemble Learning
Parameter Optimization and System Modeling	Optimize HTS Maglev system parameters, improve system performance	ML (Supervised/Reinforcement/Ensemble)
Energy Management and Health Monitoring	Dynamically adjust energy use, extend battery life	Data-driven Strategy
Thermal-Vibration Correlation Analysis	Enable intelligent monitoring and early fault warning for HTS Maglev transportation systems	BP Neural Network

). The table highlights five critical applications, including sensor temperature compensation, fault diagnosis, parameter optimization and system modeling, energy management and health monitoring, and thermal-vibration correlation analysis, explicitly linking each scenario to its functional purpose and corresponding AI method. The technical design of these applications emphasizes precision improvement and cost-effectiveness. For sensor temperature compensation, RBF neural networks combined with LS-SVM effectively mitigate measurement deviations of gap sensors caused by −20 to 60 °C temperature variations. In fault diagnosis, ensemble learning achieves high accuracy in assessing levitation system faults, substantially outperforming traditional manual judgment and threshold-based approaches. For parameter optimization and energy management, machine learning methods, including supervised learning, reinforcement learning, and ensemble learning, along with data-driven strategies, enhance system performance without hardware replacement. This reduces engineering modification costs while ensuring operational stability, which is particularly critical for HTS Maglev systems with high technical complexity.

4.3. Applications in Infrastructure Manufacturing, Monitoring, and Maintenance

4.3.1. Challenges and Limitations of Traditional Methods

Maglev infrastructure, including tracks, track beams, and ground coils forms the structural backbone of Maglev systems. Ensuring the integrity and functionality of this infrastructure is important to preserve reliability and system safety. However, traditional manufacturing, monitoring, and maintenance methods have notable limitations. Infrastructure maintenance is still largely manual-dominated operations suffer from low efficiency, high labor costs, and poor real-time comprehensive monitoring. This often leads to delayed fault detection and raising safety risks due to undiagnosed wear or structural fatigue. For track gap detection, contact methods are inaccurate and environmentally sensitive. Meanwhile non-contact methods lack high-precision image processing capabilities, limiting their ability to reliably detect subtle anomalies in dynamic environments. Traditional SHM systems are often passive and static, lacking dynamic visualization and intelligent early warning capabilities. This hinders the ability to form intuitive state grasp and failure prediction. Ground coil maintenance relies on error-prone manual recording, a process that is time-consuming, inconsistent and prone to human error. Track defect detection methods have weak image processing abilities, especially for small/obscure defects, leading to low detection rates. Additionally, maintenance strategies lack scientific failure mode analysis, resulting in unreasonable cycles (over- or under-maintenance) that increase costs or reduce reliability.

4.3.2. AI-Based Methods for Infrastructure Manufacturing, Monitoring, and Maintenance

Artificial intelligence technology has brought significant innovations to Maglev infrastructure manufacturing, monitoring, and maintenance, with specific applications as follows:

Predictive Maintenance: Real-time monitoring and intelligent maintenance of Maglev track structures are crucial for ensuring the safety and reliability of the system. However, existing maintenance practices remain predominantly manual, leading to enhance operational cost, low efficiency and delayed fault response. This poses an essential barrier to the realization of intelligent and autonomous Maglev infrastructure management.

Track Gap Detection: It is a crucial task to ensure the safe and stable operation of Maglev systems. Traditional contact-based methods are often inaccurate and environmentally sensitive, while non-contact methods lack the image processing precision needed to detect small defects. To resolve these issues, study [59] introduced a vision-based gap detection method for Maglev transportation systems. The approach used the YOLOv3 system to perform real-time object detection and identify the essential structural regions in track image. An image edge feature-based gap detection is then used to extract critical geometrical information. Sub-pixel edge detection using the Zernike moment method improved accuracy. The architecture of the improved deep convolutional generative adversarial network (DCGAN) with an integrated attention mechanism is shown in the schematic diagram (Figure 9).

Intelligent Inspection System: Literature [60] introduced an intelligent inspection vehicle combining AI image recognition, chord measurement track inspection, and big data-based lifecycle management for medium- and low-speed Maglev transportation systems. The vehicle offers high-precision, efficiency, and standards-compliant inspections tailored to Maglev characteristics. It is specifically engineered to accommodate the unique structural and operational features of Maglev infrastructure, thereby advancing the development of AI ecosystems in future transportation work.

Track Defect Detection: Literature [61] introduced a deep learning-based method for stator-surface defect identification in high-speed Maglev tracks. To compensate for the limited availability of real defect samples, the authors constructed a targeted image dataset and applied data augmentation strategies to enhance feature diversity. A multi-scale CNN architecture was introduced to capture both global stator-surface patterns and fine local textures, improving model robustness under varying illumination and reflective conditions. Experiments demonstrate that the proposed approach effectively identifies abrasion, corrosion, and surface-peeling defects, providing reliable localization performance and forming a technical basis for automated Maglev track inspection and intelligent maintenance.

Intelligent Maintenance Technologies and Equipment: Literature [62] introduced a digital twin-enabled edge maintenance system for medium and low-speed Maglev lines. The system integrates three core components: edge sensors along the track for real-time beam deflection and stress data collection; a digital twin platform to visualize the structural status of the physical line and intelligent inspection vehicles equipped with AI-based image recognition. Validation on the Qingdao Medium-Low-Speed Maglev Line demonstrated a reduction in fault localization time from 30 min to 4 min and a 60% decrease in manual inspection workload, achieving enhanced visualization, predictability, and operational efficiency in Maglev infrastructure maintenance.

Real-Time Fault Detection: Literature [63] introduced real-time malfunction detection of Maglev suspension controllers. A data-driven diagnostic framework and multi-signal feature extraction module were established for the EMS system of a medium-low-speed Maglev line. By integrating real-time current deviation identification, controller state estimation, and anomaly pattern recognition, the method enables fast localization of malfunction types such as coil current drift and amplifier degradation. Experimental validation on an operational Maglev suspension platform demonstrates that the proposed detection scheme significantly improves fault-response speed and enhances operational safety, providing an effective technical foundation for intelligent health monitoring of Maglev suspension controllers. Literature [64] introduced a data-driven method for evaluating EMS control systems. Different traditional methods that rely on mechanism models or experience, it establishes a multi-dimensional index system with real-time data, and realizes objective, quantitative evaluation of performance, reliability and robustness through data mining and statistics.

Reliability-Centered Maintenance (RCM): Literature [65] introduced an AI-driven RCM strategy for urban Maglev trains. This approach employs LSTM networks to predict failure modes of critical components, such as air springs and levitation electromagnets, and integrates reliability analysis to optimize maintenance intervals. For instance, the predicted degradation trends allow extending the maintenance cycle of air springs from six to nine months while maintaining a failure rate below 0.1%. Compared to conventional RCM based on fixed schedules, this AI-based method reduces overall maintenance costs and increases train availability.

4.3.3. Supplementary Technologies

The successful application of AI in Maglev system depends not only on intelligent algorithms but also on a suite of complementary technologies that enable real-world deployment, enhance data quality and facilitate decision-making. Supplementary technologies provide strong support for the practical application of AI in infrastructure management. For example, study [65] introduced several types of intelligent inspection, such as intelligent inspection vehicles, F-track operation vehicles, and dedicated operation platforms for Maglev infrastructure maintenance. These devices integrate AI image recognition and sensor technologies to realize the integration of detection-analysis-management. Consequently, they significantly improve maintenance efficiency, fault detection accuracy and safety across the operational environments. A 3D digital visualization platform based on the actual Jiading Maglev Test Line was developed. This platform enables engineering personnel to quickly interpret structural status, enhancing real-time situational awareness and decision-making accuracy in maintenance and control operations.

The mapping of AI applications in Maglev infrastructure management to seven core tasks, which includes software-driven techniques such as CNN and LSTM for predictive maintenance and data-driven algorithms for real-time fault detection as well as hardware and software integrated solutions including BIM technology with sensor networks for structural health monitoring and RFID for ground coil maintenance management, is summarized in the overview table (

Table 4. Summary of AI applications in infrastructure manufacturing, monitoring, and maintenance.

Area	Main Purpose	AI Method
Predictive Maintenance	Predict system failures, optimize maintenance cycles	CNN, LSTM
Track Gap Detection	Achieve high-precision, non-contact gap measurement	Machine Vision (e.g., YOLOv3, Zernike Edge Detection)
Intelligent Inspection System	Achieve automated, high-precision track inspection and lifecycle management	Artificial intelligence Image Recognition + Big Data Management
SHM	Build a 3D digital monitoring platform for dynamic visualization and intelligent early warning	BIM Technology + Sensor Network
Real-Time Fault Detection	Monitor controller status in real time, improve system safety and reliability	Data-driven Algorithms
RCM	Optimize maintenance strategies, improve system reliability and operational efficiency	Failure Mode Analysis and Prediction
Maintenance Management	Enable maintenance data management, target search, and status monitoring	RFID Technology

). Together, these approaches form a comprehensive intelligent infrastructure management system. These applications drive the intelligent transformation of traditional maintenance practices: CNN-LSTM-based predictive maintenance reduces maintenance costs and extends fault early-warning time to 48 h, addressing the inefficiencies and delays associated with manual scheduled maintenance. Machine vision methods, combining YOLOv3 with Zernike edge detection, enable high-precision, non-contact track gap measurement, overcoming environmental sensitivities such as vibration and dust that affect traditional contact-based methods. Furthermore, BIM technology integrated with sensor networks constructs a 3D digital twin of track structures, transforming static post-fault inspections into dynamic, real-time visualizations, thereby significantly enhancing the safety and reliability of Maglev infrastructure operations.

4.4. Validation Summary of AI Applications

To clarify the maturity of AI applications in Maglev systems, we summarize the validation types of 75 cited studies (Table 1, Table 2, Table 3 and Table 4) as follows.

Experimental validation: Most focus on core applications such as levitation control (e.g., [45] on Shanghai Maglev Line), fault diagnosis (e.g., [56] on fault diagnosis platforms), and infrastructure monitoring (e.g., [62] on Tongji Test Line). These studies use real Maglev lines, test benches, or environmental chambers to verify effectiveness, providing reliable data for engineering applications.

Numerical simulation: Mainly involve preliminary algorithm design, such as DRL-based levitation control [41], GAN-based data augmentation [61], and LQR/SDRE control comparison [47]. These studies lay the foundation for experimental research but lack real operational data validation.

Hybrid validation: Combine simulation and experiments, e.g., [40] (DRL simulation + medium-speed test line). This approach balances algorithm feasibility (simulation) and practicality (experiment), becoming a mainstream validation mode for complex AI-Maglev systems.

Overall, experimental validation dominates, reflecting the strong engineering relevance of AI methods in Maglev applications. At the same time, hybrid validation is emerging as an effective approach to ensure both algorithm reliability and real-world applicability.

5. Challenges and Future Perspectives

In the preceding chapters, this paper systematically reviewed the structural characteristics of Maglev transportation systems, representative AI methods, and their application achievements in levitation control, vehicle design and operation and maintenance, as well as infrastructure monitoring. These studies indicate that AI technologies have already shown significant potential in improving control precision, enhancing robustness, reducing maintenance costs, and enabling higher levels of intelligence. However, as Maglev systems evolve toward higher speeds, greater intelligence, and large-scale engineering deployment, existing technical solutions still face a series of practical obstacles in terms of data acquisition, model generalization, real-time performance, safety, and system integration. Therefore, it is necessary, on the basis of the existing achievements, to further analyze the key challenges that hinder the deep integration of AI and Maglev, and to explore potential future research directions and technical pathways.

5.1. Real-World Challenges

5.1.1. Data Collection and Processing

Effective AI deployment in Maglev transportation systems requires large amounts of operational data. Collecting and processing such data efficiently and accurately is challenging. Semi-supervised learning (Section 3) can alleviate the high cost of labeling by combining small amounts of labeled data with large volumes of unlabeled data, reducing reliance on fully annotated datasets. GAN-based data augmentation techniques (e.g., GAN with joint attention layer, Section 4.3.2) generate simulated data for minor track defects or extreme levitation conditions, supplementing scarce or low-quality real data and improving data integrity. Additionally, RBF neural networks (Section 4.2.2) and CNN feature extraction (Section 4.1.2) can automatically filter redundant sensor noise (e.g., vibration-induced disturbances), enhancing preprocessing efficiency and reducing manual intervention.

5.1.2. Real-Time Response Requirements

High-speed Maglev systems demand rapid decision-making from AI controllers. DRL-based levitation control (e.g., DQN with actor-critic network, Section 4.1.2) adapts control actions in real time without complex model derivation, narrowing levitation gap fluctuations. Event-triggered neural network control (Section 4.1.3) reduces computational load by performing updates only when system states deviate from predefined thresholds. Lightweight ML models deployed on edge AIoT devices (Section 4.2.2) process vehicle monitoring data locally, avoiding cloud transmission latency and ensuring timely responses for high-speed operations.

5.1.3. Safety and Fault Management

Safety is a critical concern in AI-driven Maglev systems. Machine learning addresses this through early warning, real-time diagnosis, and fault-tolerant control. CNN-LSTM-based predictive maintenance (Section 4.3.2) provides advance warnings for track structural fatigue or superconducting coil anomalies. Ensemble learning methods, such as fuzzy comprehensive evaluation combined with multiple base learners (Section 4.2.2), achieve high fault diagnosis accuracy. SVM and decision tree algorithms (Section 4.1.2) enable real-time identification of levitation system faults, reducing misclassification risk. Fault-tolerant control strategies, such as adaptive fuzzy PID adjustments, allow automatic switching to backup control strategies in response to detected faults, preventing system collapse.

5.1.4. Environmental Adaptability

Maglev systems are affected by environmental factors such as temperature, humidity, and aerodynamic disturbances. AI techniques mitigate these influences through environmental interference compensation and adaptive control. RBF neural network combined with LS-SVM (Section 4.2.2) reduces gap sensor errors across temperature variations from −20 °C to 60 °C. Fuzzy logic controllers (T-S fuzzy model, Section 4.1.2) dynamically adjust control parameters to handle uncertainties caused by humidity-induced friction changes. Adaptive neural networks learn environmental interference characteristics in real time, such as aerodynamic resistance changes or payload variations, and optimize levitation and traction strategies to maintain system stability.

5.1.5. System Integration

Integrating AI with existing Maglev infrastructure is challenging due to compatibility and modularity issues. Software-defined intelligent control (e.g., PSO-optimized PID, Section 4.1.2) allows embedding ML algorithms via software updates without replacing hardware components like sensors or electromagnets. Digital twin platforms combining BIM and SHM technologies (Section 4.3.2) act as an integration layer, connecting AI model outputs, such as track health status, with existing monitoring systems. Modular ML components, including independent fault diagnosis and temperature compensation modules (Section 4.2.2), can interface with traction power and other systems, reducing overall integration complexity.

5.1.6. Feasibility Gaps in Safety-Critical Levitation Control

Levitation control is the core safety-critical function of Maglev trains. AI controllers face three major unresolved challenges:

Black-Box Opacity vs. Safety Traceability: Deep learning models such as DRL and CNN-LSTM rely on hierarchical feature extraction, producing decisions that are difficult to interpret. For example, when a DRL-based levitation controller adjusts electromagnetic force, it is unclear which inputs triggered the action. Traditional PID controllers provide explicit logic for real-time fault localization. This lack of transparency violates traceability requirements for safety-critical systems (IEC 61508 SIL 4), complicating post-incident investigation.

Insufficient Robustness Under Out-of-Distribution (OOD) Scenarios: AI models trained on historical data may fail under rare or extreme conditions. In EDS systems, a sudden superconducting quench can invalidate flux pinning.

Lack of Safety Certification Standards: Traditional controllers like PID or SMC benefit from mature verification and validation protocols, including HIL simulations with over 10,000 fault injections and extended field testing. Currently, no unified standards exist for certifying AI controllers in Maglev levitation, limiting commercial deployment. Even high-performance AI systems cannot meet formal safety compliance, as EU CEN/TC 256 and China GB/T 24338 lack AI-specific criteria.

5.2. Future Prospects

Despite the challenges in applying AI to high-speed Maglev transportation, the field presents substantial opportunities. The following research directions are both actionable and tightly aligned with the technical characteristics of Maglev systems, including real-time control requirements, data privacy considerations, and hardware resource limitations. They emphasize the integration of lightweight edge machine learning and federated learning to address practical engineering bottlenecks while maintaining system safety and efficiency.

5.2.1. Advanced AI-Driven Maglev Controllers

Future AI controllers may move beyond basic adaptive control toward higher real-time performance and cross-scenario generalization. For high-speed EMS systems operating, lightweight DRL models, such as pruned DQN or quantized Actor-Critic networks, may be integrated with on-board FPGA modules. By compressing model parameters, cloud transmission delays may be significantly reduced or even eliminated, enabling adaptive re-al-time adjustments of levitation gaps. For instance, MIMO RNN structures for levitation control may undergo channel pruning to reduce computational overhead while maintaining anti-disturbance performance. In parallel, federated averaging algorithms can facilitate crossline model training without sharing raw operational data, which may enhance adaptability to various line types, such as urban medium-low speed versus high-speed intercity Maglev, and may help reduce new line controller debugging cycles.

5.2.2. AI-Assisted Vehicle Design and Manufacturing

Building upon controller advancements, AI may also help enhance vehicle design and manufacturing precision. Lightweight computer vision models, such as pruned YOLOv5 or Mobile Net-based segmentation networks, deployed with edge sensors (e.g., high-definition industrial cameras and laser rangefinders), may allow real-time monitoring of critical assembly metrics, including superconducting coil gaps and aerodynamic surface flatness. This approach may help reduce manual inspection errors. For new HTS Maglev vehicles with limited test data, federated transfer learning may enable leveraging design knowledge and fault data from mature systems, such as Shanghai Maglev or Japan’s L0 series, possibly decreasing aerodynamic simulation and structural optimization costs and potentially shortening R&D cycles.

5.2.3. Predictive and Real-Time Maintenance

Effective vehicle design and high-performance controllers complement predictive maintenance. Lightweight CNN-LSTM models, with parameter compression to reduce the original size, be deployed on track-side edge gateways to process real-time FBG sensor data, possibly providing early warnings for faults such as track cracks and support structure fatigue. This edge-based processing may reduce prediction latency compared to cloud methods while accommodating hardware limitations. For ultra-high-speed EDS systems, where superconducting quench samples are scarce, federated learning across multiple test lines, such as Japan’s Yamanashi and China’s Qingchengshan Lines, may improve fault prediction accuracy, potentially mitigating the risk of overfitting to single-line datasets.

5.2.4. AI-Driven Energy Management

Energy optimization is tightly linked to both real-time controllers and predictive maintenance systems. Lightweight PPO algorithms deployed on-board dynamically adjust traction power and regenerative braking based on operational conditions, such as passenger load, track gradient, and wind speed. For medium-low speed urban lines, this may help reduce energy consumption. Federated learning may further enable crossline energy strategy sharing, allowing, for instance, optimized superconducting system cooling based on regional temperature differences, potentially avoiding the limitations of single-line strategies and possibly enhancing system-wide efficiency.

5.2.5. Intelligent Network Scheduling

AI integration extends beyond vehicles to the operational network. Lightweight CNN-Transformer models at regional edge nodes may fuse multi-modal data including track status, real-time passenger flow, and weather conditions to potentially adjust train speeds and stop times dynamically. This may help increase inter-city Maglev network on-time rates. Federated learning may support cross-regional collaboration without exposing sensitive operational data. Jointly trained models may optimize service coordination, such as connections between Shanghai Suzhou and Suzhou Nanjing lines, possibly improving overall network transport efficiency.

5.2.6. Distributed Resilient Fault-Tolerance and Cybersecurity Protection

The integration of AI into intelligent agents exposes a series of critical and high-impact risks [66,67]. These problems are equally serious in high-speed maglev systems, including cascading failures at the physical layer and network threats such as DoS attacks and data tampering. Traditional centralized control of architecture alone may not be able to address these risks in a timely and adequate manner. It must be emphasized that these vulnerabilities are directly related to the two most fundamental safety requirements in EMS maglev operation: reliable vertical stability and strictly limited air gap. Any failure in these areas may trigger a chain of failures with system-level consequences. Recent research further shows that even advanced controllers based on deep learning cannot fully guarantee air gap safety under uncertain conditions [49], highlighting the urgency of building a more resilient and intelligent risk mitigation framework. However, in typical simulation environments, it is still possible to build targeted strategies based on existing technologies and given assumptions to mitigate the above risks to some extent.

For physical-layer actuator anomalies (e.g., partial electromagnet demagnetization, traction motor bias faults) and network-induced disturbances, an adaptive neural-network-based fault estimation framework can theoretically achieve bounded fault-estimation errors. Under these controlled conditions, distributed compensation mechanisms have been shown to reduce levitation-gap fluctuations. Likewise, when sensor attenuation occurs due to degraded transmitters, an NN-based adaptive observer is capable of reconstructing the latent true signals with sub-millimeter accuracy, thereby reducing the likelihood of issuing erroneous levitation commands that could destabilize the suspension system [49]. These results remain scenario-dependent and reflect performance attainable under the modeled assumptions rather than guaranteed physical-system metrics.

For cyber-layer threats such as false-data-injection attacks on levitation-gap measurements, multi-agent cross-validation mechanisms can, under the designed at-tack models, achieve high detection accuracy, outperforming simple CRC-based schemes whose effectiveness diminishes under time-varying tampering patterns. These outcomes rely on the adopted threat model and communication topology.

In the context of distributed control and resilience against DoS attacks, the Maglev system, comprising traction, levitation, guidance and track-monitoring subsystems which can be abstracted as a heterogeneous multi-agent system (MAS). With appropriate assumptions on communication availability, a data-driven resilient learning algorithm can converge to consistent control targets within tens of seconds in simulation [50]. Under the theoretical conditions ensuring bounded connectivity and admissible attack duration, a resilient estimator can dynamically adjust inter-agent communication weights, maintaining control-loop latency within tens of milliseconds and ensuring operational stability when DoS durations remain within a specified fraction of total operation time. These bounds reflect algorithmic guarantees rather than fixed engineering specifications, and final performance depends on implementation scenarios.

5.2.7. Implications for Future Research

These directions collectively may help address the main challenges in AI Maglev integration, including real-time performance, data privacy, and sample scarcity, potentially providing a coherent technical roadmap for future research. Lightweight ML models (e.g., pruned DRL for levitation control and edge-deployed CNN-LSTM for predictive maintenance) may ensure rapid decision-making, while federated learning may enable crossline generalization without exposing sensitive operational data. Digital twin platforms and sensor-integrated monitoring systems may enhance real-time visualization and predictive maintenance, possibly bridging the gap between experimental validation and practical deployment. By linking controller optimization, vehicle design, maintenance, energy management, and network scheduling, this integrated approach may help tackle unresolved issues such as scarce high-speed fault samples, strict EMS latency requirements, and adaptation across Maglev speed grades.

6. Conclusions

Artificial intelligence exhibits significant potential to advance Maglev transportation by mitigating the inherent nonlinearity, open-loop instability, and operational constraints of traditional systems. Although recent progress has demonstrated notable gains in levitation control, fault diagnosis, predictive maintenance, and sensor compensation, the practical application of AI in Maglev environments is still limited by insufficient data quality, stringent real-time requirements, model generalizability issues, and difficulties in integrating modern algorithms with existing infrastructure. Current studies nonetheless provide strong evidence of AI’s effectiveness: advanced controllers enhance levitation stability under complex disturbances, ensemble learning improves diagnostic reliability, CNN-LSTM architectures enable early fault prediction, and hybrid RBF-SVM methods strengthen sensor robustness. Together, these advancements indicate that AI has the capacity to improve control precision, enhance operational safety, and reduce lifecycle maintenance costs.

Despite these encouraging developments, several unresolved challenges must be addressed before AI can be reliably deployed in safety-critical Maglev operations. Rare or extreme events continue to create data scarcity, real-time latency constraints restrict the use of computationally intensive models, and performance consistency across multiple Maglev speed grades and system configurations remains difficult to achieve. Addressing these issues requires the development of lightweight, edge-deployable learning models that can meet real-time constraints, as well as closer coordination between algorithm design and hardware implementation to ensure that control responsiveness is not compromised. Future research should also explore broader application scenarios beyond levitation control, including planar motor coordination, vibration suppression, and high-power electromagnetic subsystems. Techniques such as transfer learning, federated learning, and digital-twin-based modeling may help improve model adaptability and maintain data privacy.

Ultimately, realizing AI-enabled Maglev systems will depend on sustained collaboration among AI researchers, control engineers, and transportation practitioners. Such interdisciplinary efforts are essential for translating theoretical advances into verifiable and deployable technologies, paving the way for safe, efficient, and intelligent Maglev transportation platforms.