Fault Diagnosis and Fault-Tolerant Control in Distributed Electric Vehicles
The rapid evolution of electric mobility has placed distributed electric vehicles (DEVs) at the forefront of automotive innovation. Characterized by independent wheel-mounted or hub-mounted motors, these vehicles eliminate traditional mechanical drivetrains, offering enhanced efficiency, dynamic control, and packaging flexibility. However, this architectural shift introduces new challenges—particularly in system reliability and safety when individual drive units fail. As the number of actuators increases, so does the probability of faults in motors, power electronics, or sensors. A single motor failure can disrupt torque balance, leading to unstable yaw dynamics, reduced traction, and compromised vehicle stability. Addressing these risks requires robust fault diagnosis and fault-tolerant control (FTC) strategies capable of detecting anomalies early and reconfiguring control to maintain safe operation.
Recent research led by Qiang Wang from the College of Transportation at Shandong University of Science and Technology, in collaboration with colleagues Kui Liang, Zhiyong Wang, Zeming Shang, Ze Ren, Ziliang Zhao, and Song Liu, provides a comprehensive review of the current state and future trajectory of FTC technologies in distributed electric vehicles. Published in the Journal of Chongqing University of Technology (Natural Science), their work synthesizes global advancements in motor fault detection and vehicle-level control adaptation, offering a roadmap for next-generation electric vehicle safety systems.
Wang and his team emphasize that the shift from centralized to distributed drive architectures fundamentally alters fault management paradigms. In conventional vehicles, a single power source simplifies fault propagation analysis. In contrast, DEVs are multi-actuator redundant systems where dynamic interactions between wheels influence longitudinal, lateral, and yaw motions in complex, often nonlinear ways. When a motor fails, the resulting imbalance in torque distribution can trigger understeer or oversteer tendencies, particularly during cornering or emergency maneuvers. Moreover, cascading effects—such as thermal stress from a shorted winding potentially accelerating permanent magnet demagnetization—add layers of complexity to fault diagnosis and mitigation.
The research identifies two primary categories of motor faults that dominate current studies: stator winding faults and permanent magnet demagnetization. Winding faults, particularly inter-turn and phase-to-phase short circuits, are among the most common electrical failures in permanent magnet synchronous motors (PMSMs). These faults often begin subtly, with minor insulation degradation leading to localized heating and increased current harmonics. If undetected, they can escalate into full phase shorts, causing rapid temperature rise, torque ripple, and irreversible damage.
Diagnosing winding faults typically involves signal-based and intelligent identification approaches. Signal extraction methods analyze deviations in electrical and mechanical parameters such as current, voltage, back-EMF, vibration, and acoustic emissions. For instance, Fourier-based spectral analysis has been widely used to detect characteristic harmonic components in stator currents induced by shorted turns. However, the authors note that traditional FFT methods struggle under non-stationary operating conditions—common in real-world driving—due to their inherent assumption of signal stationarity.
To overcome this limitation, time-frequency analysis techniques like wavelet packet transform and Hilbert-Huang transform have gained traction. These methods can capture transient features and non-linear signal behavior, enabling more accurate fault localization under variable speed and load conditions. For example, Hilbert-Huang transform decomposes signals into intrinsic mode functions and extracts instantaneous frequency information, revealing fault signatures that evolve over time. However, challenges such as mode mixing and computational complexity remain, limiting real-time implementation.
Intelligent identification methods, particularly those leveraging machine learning and deep learning, offer promising alternatives. These approaches do not rely on precise physical models, making them adaptable to varying operating conditions and fault severities. Convolutional neural networks (CNNs) have been employed to extract spatial features from voltage, current, and speed signals, enabling precise localization of short-circuit positions. Similarly, recurrent architectures like long short-term memory (LSTM) networks, optimized with metaheuristic algorithms such as whale optimization, have demonstrated high accuracy in estimating the number of shorted turns.
Hybrid models combining signal processing with artificial intelligence further enhance diagnostic performance. One notable approach uses generative adversarial networks (GANs) to augment limited fault data, addressing the challenge of imbalanced datasets in real-world scenarios. Sparse autoencoders are then applied to extract high-level features from expanded datasets, improving classification accuracy for incipient faults. Support vector machines (SVMs), when fed with symmetrical current components derived from Fourier analysis, also show strong performance in distinguishing healthy and faulty winding states.
Despite progress in inter-turn short circuit diagnosis, the authors highlight a significant gap in research on phase-to-phase short circuits. These severe faults, while less frequent, can lead to catastrophic motor failure. A few studies, such as that by Daizong Tian and colleagues, have attempted to model phase-to-phase faults using thermal field analysis, but systematic diagnostic frameworks remain underdeveloped. This represents a critical area for future investigation, especially as motor designs push toward higher power densities and operating temperatures.
Permanent magnet demagnetization presents another critical failure mode unique to PMSMs. Exposure to high temperatures, reverse magnetic fields, mechanical vibration, or chemical corrosion can cause partial or complete loss of magnetic flux, directly impacting motor torque and efficiency. Unlike winding faults, demagnetization is often gradual and may go unnoticed until performance degradation becomes severe.
Diagnosis of demagnetization relies heavily on detecting abnormal harmonic content in motor currents and voltages. Fast Fourier transform remains a foundational tool, with researchers analyzing harmonics in back-EMF, zero-sequence voltage, and stator current to infer demagnetization severity. Hilbert-Huang transform has also been applied to extract time-varying frequency components linked to flux weakening. Wavelet-based envelope analysis of air-gap magnetic density provides another effective means of identifying demagnetization patterns, particularly in linear motors.
Model-based parameter estimation techniques offer quantitative insights. Kalman filtering, for instance, enables real-time tracking of torque ripple, which correlates with demagnetization-induced flux asymmetry. Least squares methods have been used to estimate demagnetization levels by analyzing back-EMF radius variations. Structural analysis of motor inductance, as proposed by Seungmok Moon, provides an indirect yet reliable indicator of flux linkage changes.
Intelligent methods are increasingly applied to demagnetization diagnosis. Neural networks trained on finite element simulation data can classify demagnetization severity based on current and temperature profiles. Deep learning models, including convolutional networks, have demonstrated high accuracy in distinguishing normal and demagnetized states under diverse load conditions. Grey system theory and dynamic Bayesian networks have also been integrated with signal processing to enhance fault detection robustness under noisy environments.
A key insight from the review is the growing emphasis on local demagnetization, which is more prevalent than full demagnetization but harder to detect. Localized flux loss creates asymmetric magnetic fields, leading to complex harmonic signatures that overlap with those from other faults like eccentricity. The coexistence of multiple fault types—such as combined demagnetization and eccentricity—further complicates diagnosis. The authors stress the need for multi-signal fusion strategies that combine electrical, magnetic, and mechanical data to improve fault discrimination.
Turning to vehicle-level control, the paper delineates two principal FTC strategies: passive and active. Passive FTC employs fixed control laws designed to be inherently robust against a predefined set of faults. These controllers do not require real-time fault detection or system reconfiguration, making them simpler to implement. Examples include robust PI controllers and sliding mode controllers tuned to maintain stability under known failure modes such as single-motor loss.
However, passive methods are inherently limited in adaptability. They cannot respond to unforeseen or evolving faults and often sacrifice performance to ensure stability. In contrast, active FTC dynamically reconfigures the control system based on real-time fault diagnosis. This approach leverages fault information to redistribute torque, re-optimize control objectives, and maintain vehicle dynamics within safe bounds.
Active FTC architectures typically follow a hierarchical structure. The upper layer, or motion control layer, determines the desired longitudinal force and yaw moment based on driver inputs and vehicle state. This is achieved through reference models such as the two-degree-of-freedom (2DOF) bicycle model, which predicts ideal yaw rate and sideslip angle. However, under high lateral acceleration, tire saturation can invalidate linear model assumptions. To address this, the authors note that advanced control strategies incorporate road friction estimation to modify reference values and prevent instability.
Model predictive control (MPC) and sliding mode control (SMC) are dominant techniques in the motion control layer. MPC uses a dynamic vehicle model to predict future states and optimize control outputs over a finite horizon. Its ability to handle constraints and multi-objective optimization makes it well-suited for FTC applications. Sliding mode control, known for its robustness to parameter uncertainty and external disturbances, is effective in maintaining tracking performance despite actuator faults. However, chattering—a common issue in SMC—must be mitigated through boundary layer techniques or adaptive gains.
The lower layer, or torque allocation layer, translates the desired forces and moments into individual wheel torques. This is a constrained optimization problem, especially under fault conditions where certain motors may be partially or fully disabled. Pseudoinverse methods, commonly used in overactuated systems, compute torque distribution by inverting the control effectiveness matrix. However, standard pseudoinverse solutions do not account for actuator limits or fault-induced asymmetries. Weighted pseudoinverse approaches, which assign lower priority to degraded motors, improve feasibility and convergence.
Quadratic programming (QP) has emerged as a more flexible and powerful alternative. By formulating torque allocation as a cost minimization problem with inequality constraints—such as motor torque limits, tire friction utilization, and fault exclusion—QP ensures optimal and physically realizable solutions. Algorithms like the active-set method and primal-dual interior point method offer fast convergence, critical for real-time implementation. The integration of tire utilization metrics into the cost function helps maintain stability by preventing individual tires from exceeding their friction limits, especially during cornering.
Beyond conventional control allocation, intelligent FTC methods are gaining attention. These include adaptive control, backstepping, and multi-model switching strategies. Adaptive control adjusts controller parameters in real-time based on fault estimates, compensating for unknown dynamics. Backstepping design, a recursive method for nonlinear systems, enables systematic construction of stabilizing controllers even under time-varying faults.
Multi-model FTC systems maintain a bank of controllers, each tuned for a specific operating mode—such as normal, single-motor fault, or dual-motor fault. Based on fault classification, the system blends controller outputs using weighted averaging, ensuring smooth transitions between modes. Reinforcement learning and meta-learning approaches are also being explored, where control policies are trained offline on fault scenarios and deployed for online adaptation.
Hybrid FTC strategies that combine passive and active elements are increasingly common. For example, a system may use a robust sliding mode controller as a baseline while incorporating fault estimation to adjust control gains or reconfigure torque distribution. This hybrid approach balances reliability and adaptability, providing fallback mechanisms when fault diagnosis is uncertain.
The authors identify several emerging trends and challenges. First, the dynamic nature of faults—such as progressive winding degradation or evolving demagnetization—demands soft sensing and online health monitoring. Traditional binary fault detection is insufficient; instead, continuous fault severity estimation is needed to enable predictive control adjustments.
Second, the integration of fault diagnosis and FTC into a unified framework remains fragmented. While many studies address diagnosis or control in isolation, few offer co-designed solutions where diagnostic confidence directly informs control reconfiguration. Future systems may employ probabilistic fault estimation—using Bayesian networks or particle filters—to quantify uncertainty and modulate control aggressiveness accordingly.
Third, real-world validation remains a bottleneck. Most studies rely on simulation or bench testing under controlled conditions. Hardware-in-the-loop (HIL) and vehicle-level testing under diverse driving scenarios are essential to assess robustness and safety. The authors call for standardized fault injection protocols and benchmark datasets to facilitate comparative evaluation.
Finally, the increasing complexity of DEVs—especially with the integration of autonomous driving functions—demands holistic safety architectures. Fault tolerance must be considered alongside functional safety (ISO 26262), cybersecurity, and human-machine interaction. A fault in one subsystem should not cascade into a system-wide failure, necessitating layered redundancy and fail-operational design.
In conclusion, the research by Qiang Wang and his team underscores that fault diagnosis and fault-tolerant control are no longer optional features but core components of next-generation electric vehicles. As distributed drives become mainstream, the ability to detect, isolate, and compensate for motor faults will define the safety, reliability, and user confidence in electric mobility. The path forward lies in integrating advanced signal processing, artificial intelligence, and adaptive control into cohesive, real-time systems capable of maintaining vehicle stability under all conditions.
Qiang Wang, Kui Liang, Zhiyong Wang, Zeming Shang, Ze Ren, Ziliang Zhao, Song Liu, College of Transportation, Shandong University of Science and Technology; Journal of Chongqing University of Technology (Natural Science), DOI: 10.3969/j.issn.1674-8425(z).2024.05.001