Design of ANN-Optimized Second Order Sliding Mode Control of Switched Reluctance Motor for Aircraft Actuators

A modern passenger plane, like a Boeing 747 or an Airbus A380 consists of several million different parts. And many of them are moving. Control surfaces need to go up and down, flaps need to deploy when landing and retract when cruising. Landing gears need to extend and lock. All of these motions need some kind of force (or power) behind them. For most planes, this has been the domain of hydraulic actuators. Little cylinders and servos that convert pressurized fluid into mechanical motion. They deliver high force, they can respond quickly, and they have decades of operational history. The problem is, they're also heavy, maintenance-intensive, and prone to leaks. Plus, they need a system of tubes and plumbing to run throughout the aircraft.

So in the last few decades, engineers have started trying out alternatives. One contender is called Switched Reluctance Motors (SRMs). They're small electric motors that generate torque by pulling a rotor toward positions of lower magnetic reluctance, using energized stator windings and a magnet-free rotor. There are no permanent magnets, no rotor windings, and each phase is electrically isolated. This makes them simple, robust, and considerably fault-tolerant. But, while they do have a ton of benefits, we haven't quite figured out how to control them smoothly and reliably under real operating conditions yet. Today, we can make SRMs run and track a target speed, but only with noticeable torque ripple and chattering. We also need to use fairly conservative tuning that sacrifices smoothness for stability. What we need instead is fast, precise, low-ripple motion across changing loads and operating conditions.

How do we get there?

That's where today's paper comes in. In it, the authors propose a control architecture (for SRMs) that combines sliding-mode control with neural networks and automated gain optimization, to tame those nonlinearities. On today's episode we'll walk through why SRMs are difficult to control, what's been tried before, and what this new strategy does differently. Let's dive in.

First a little clarification. Yes, we are talking about motors. No, we are not talking about propulsion motors. This paper is not about "electric aircraft" in general, it's about a field called MEA: More-Electric Aircraft. Planes that use electric motors and power electronics for everything they can, but still use jet engines and turbofans to actually make the plane go.

Also worth noting: we're not talking about switching all of a plane's electric motors to SRMs. There are electric motors for pumps, fans, compressors, environmental control systems, avionics cooling, cabin systems, and so on, that are all fine as-is. Those motors work well, and those systems don't have needs that are unmet. This paper is about using SRMs for a specific class of problems: high-load, safety-critical electromechanical actuators. The type that used to be hydraulic, where the motor has to deliver precise, repeatable motion under changing loads. Where it has to survive faults and high temperatures, and continue operating even when other parts of the system degrade. This is the use-case for SRMs.

But SRMs have a reputation problem. The rotational force they generate causes their assembly to pulse and shake. These kinds of vibrations can propagate through the airframe and accelerate fatigue in surrounding components. And the acoustic noise is, well, not a particularly comforting sound for passengers to hear. But why do SRMs do this?

Well, the fundamental issue is that they operate on magnetic reluctance rather than the interaction between magnets and fields.

• In a traditional electric motor, the rotor carries magnets and the stator produces a rotating magnetic field. When you drive the stator windings with phased currents, the rotor is continuously pulled forward, producing smooth, continuous torque.
• An SRM is different. The motor generates torque by energizing stator poles, which pulls rotor poles toward alignment.

Reluctance refers to the tendency of magnetic flux to follow the path of lowest magnetic resistance, so the rotor is repeatedly pulled into preferred alignment positions. The issue is, this process is inherently discontinuous. The motor doesn't produce smooth, continuous rotation so much as a series of magnetic tugs. Layer on top of this the nonlinear magnetic characteristics and variable inductance from the double-salient stator-rotor geometry, and you end up with a serious control issue.

Researchers have tried a range of strategies to deal with that.

• Classical linear controllers (like PID) can stabilize SRMs around a fixed operating point, but they break down once loads change or nonlinear effects dominate.
• More advanced approaches, like model reference adaptive control and torque sharing functions, try to explicitly shape current waveforms to smooth out torque production, but they depend heavily on accurate motor models and careful calibration.
• Sliding-mode control has been especially attractive because of its robustness to uncertainty and disturbances, but in practice it introduces its own problems: high-frequency switching leads to chattering, which shows up directly as vibration and noise in the actuator. Second-order variants reduce some of that chattering, but they still rely on fixed gain choices that are hard to tune across the full operating envelope.

In this paper the authors believe that no single technique is sufficient on its own. That's why they're building a "cascading" control structure instead.

• At the outer level, a second-order sliding-mode controller handles speed tracking and disturbance rejection. This provides the robustness needed for safety-critical actuator operation.
• At the inner level, a neural network is used to model the relationship between phase currents and electromagnetic torque. The idea is to have it learning the behaviors that are difficult to capture analytically.

The gains of the controller are not chosen by hand, but tuned automatically using PSO: particle swarm optimization. This allows each component to do what it does best. Sliding-mode control for robustness, learning for nonlinearity, and an optimization algorithm for tuning. All combined into a single control loop. Let's talk about each of these components in a bit more detail.

First, what do we mean when we say "second-order sliding-mode controller"? At its core, this is a feedback strategy for deciding how hard to drive the motor based on how far the system is from where it should be, and how that error is changing over time. The "sliding-mode" part means the controller is designed to force the system's behavior onto a specific desired pattern, even when the system is nonlinear or disturbed. The "second-order" part means it does this smoothly, by controlling how the correction signal evolves, rather than abruptly switching it on and off. Unlike linear controllers, it doesn't rely on a precise motor model. It defines a sliding manifold based on the speed tracking error and its derivative, and drives the system toward that manifold in finite time. Once on the manifold, the dynamics become largely insensitive to parameter variations, load torque changes, or unmodeled effects. And by acting on the derivative of the sliding variable rather than directly switching on the control input, it significantly reduces the high-frequency switching that causes chattering. That's everything you need to know about the outer-level.

The inner level focuses on an entirely different problem: how "commanded-torque" or "commanded-speed" translates into actual electromagnetic torque. In SRMs, this relationship is highly nonlinear and strongly dependent on rotor position, saturation effects, and inductance variation. Analytical models do exist for this, but they're difficult to parameterize accurately across operating conditions, and they tend to break down under transient loads. Here the authors put a neural network in the inner loop to sidestep this entirely. It just learns the nonlinear mapping directly from data. It takes inputs like current and outputs references that better align with the motor's true behavior. In effect, the neural network acts as a nonlinear compensator, shaping the current commands so that torque production is smoother and more predictable, even when the underlying motor physics are...well...messy. This reduces torque ripple at the source rather than trying to correct it after the fact.

So you have two different systems both addressing the same underlying problems: vibration and instability. The outer layer keeps the actuator motion stable and on target, and the inner layer reduces the ripple itself by compensating for the motor's nonlinear torque. The final piece is the optimization layer, which addresses tuning. That is: how much of each control action should be applied at any given time. This is where the PSO algorithm comes in. Each candidate set of gains is evaluated against performance. The algorithm combs through the search space, and iteratively converges toward better and better solutions. This removes much of the trial-and-error, and ensures that the controller is tuned for the specific SRM and actuator it's connected to.

So how well did this system perform? Do the motors still shake, or are they smooth as butter now? A bit of both. In simulation, the system clearly improves on classical approaches. Speed tracking is tighter, recovery from load disturbances is faster, and torque ripple is reduced. But none of this makes the SRM behave like a naturally smooth motor. The vibration doesn't disappear, there's just less of it. The control problem doesn't go away, it's just less of a problem. What the authors show is a reduction in severity, not a cure for the motor's ills. And worth nothing: the results are demonstrated only in simulation, we still have yet to see what happens in practice.

So what can we learn from this paper?

• Narrowly: that there's still room to push SRMs much closer to actuator-quality performance. They'll likely never be perfectly smooth machines, but that doesn't mean we can't make them a lot better than they currently are.
• And broadly: that the goals of MEA likely won't be met with a single breakthrough technology. They'll be achieved only through careful system-level engineering that combines hardware choices, control architecture, and optimization to make these imperfect components more reliable and predictable.

If you want to see the authors' derivations of the control laws, their implementation of PSO, or their neural network architecture, then I'd highly recommend downloading the paper.