ANFIS control sharply damped half-car motion across bump and random road simulations
Passive suspension systems still do the same thing they have always done: absorb road shock with fixed springs and dampers, then accept whatever compromises follow. In this simulation study, a data-driven ANFIS controller was designed to push beyond that fixed response and actively reshape the vehicle’s vertical and pitch dynamics in a four-degree-of-freedom longitudinal half-car model.
The contrast with passive suspension was immediate in the double-bump scenario. At 10 m/s, body vertical displacement fell from 0.0435 m to 0.0084 m, a reduction of 80.6%, while vertical acceleration dropped from 2.40 to 0.77 m/s2, or 67.9%. Front suspension travel also fell steeply, from 0.0652 m to 0.0090 m, a reduction of 86.2% relative to passive suspension. Those are not cosmetic differences; they describe a vehicle body that settles faster, moves less, and transmits less disturbance to occupants.
The learned controller outperformed PID while remaining comparatively compact
The controller design used two inputs, five Gaussian membership functions for each input, 25 Takagi-Sugeno rules, and 150 training epochs. Training data came from PID-generated input-output behavior, which gave the ANFIS system a supervised starting point rather than forcing it to learn from scratch. In practice, that means the controller inherited a familiar control structure but translated it into a nonlinear mapping that better matched the simulated suspension dynamics.
Compared with PID-controlled active suspension, ANFIS still delivered clear gains in the same double-bump test. Body vertical displacement decreased by 53.6%, vertical acceleration by 49.3%, and front suspension travel by 43.8%. Settling time also improved, with body vertical displacement reduced by 38.8% and pitch acceleration by 38.6% relative to passive suspension. In other words, the system did not merely reduce peak motion; it also returned the chassis to equilibrium more quickly after the disturbance passed.
The pitch response matters because a half-car model captures the longitudinal coupling between the front and rear axle. A speed bump or similar input does not just move the body upward; it rotates the chassis and shifts load between axles. A controller that suppresses heave but leaves pitch uncontrolled would still leave the vehicle unstable in real driving. Here, the modeled active suspension addressed both motions together, which is the technical reason the results read as more than a simple vibration reduction exercise.
Road profile complexity exposed the value of active force generation
The study tested the 4-DOF model under a deterministic double-bump input and under ISO-classified random road profiles, specifically Type A and Type C, at 10 and 20 m/s. This is a useful spread of disturbances because it probes both sudden excitation and stochastic road roughness. The passive system, limited by fixed spring and damping characteristics, has no way to adapt when road input changes in frequency content or severity. Active control, by contrast, can generate additional force in real time and counter the motion as it develops.
At a mechanistic level, the ANFIS controller acts like an adaptive rule-based amplifier of the suspension’s counterforce. The more precisely the controller identifies the relationship between error and error rate, the more effectively it can oppose the sprung mass’s vertical and pitch movement before large oscillations accumulate. That is especially relevant in a longitudinal half-car model, where the front and rear unsprung masses do not respond identically and where pitch can magnify the occupant-perceived disturbance.
The paper states that the ANFIS result met all qualifications for riding comfort under ISO 2631. In simulation terms, that is the strongest claim the model can make: the controller moved the system into the comfort envelope defined by the standard, at least under the modeled excitation conditions.
Why the simulation result still stops short of vehicle deployment
The strongest gains came from a simulated environment, and the limitations are explicit. The model is linear, the training was supervised by PID-generated data, and actuator energy use was not assessed. Each of those choices narrows the claim. A controller that performs well in a clean numerical model can lose much of that advantage once sensor noise, delay, actuator saturation, and parameter drift enter the picture.
Experimental validation is therefore the real next step. Without it, the study cannot establish robustness to uncertainty or quantify the energy cost of delivering the control force that produced the better ride metrics. A controller that lowers body acceleration but demands excessive actuator power would still face a serious implementation barrier in a real vehicle platform.
The authors also frame their approach as intentionally simpler than more complex hybrid or robust control schemes. That simplicity is part of the appeal. ANFIS offers a compact nonlinear mapping from tracking error and error rate to force output, without requiring the type of online constrained optimization used in some model predictive or multi-agent chassis strategies. Simplicity, however, does not automatically mean superiority across all road conditions or hardware setups.
What the half-car model can and cannot prove
The vehicle representation used here is not a full-car model. It captures heave and pitch, along with front and rear unsprung motion, but it does not account for every chassis degree of freedom that a real vehicle will experience on the road. Even so, that half-car structure is valuable because it preserves the front-rear coupling that is central to pitch control during bumps, braking, and uneven terrain.
The study also reflects a broader pattern in suspension research: data-driven control can outperform fixed passive settings in simulation, but the performance margin depends heavily on the road excitation, the quality of the training data, and the underlying assumptions in the model. Here, the ANFIS controller clearly beat passive suspension and improved on PID control in the modeled scenarios. The open question is not whether the simulation improved; it did. The unresolved issue is whether the same control law can hold up when the suspension is exposed to the messy, delayed, and imperfect signals of real driving hardware.