What’s happening with wheel-speed noise and LSTM denoising here is really a microcosm of a much broader pattern in modern sensor and signal-processing systems. And hence, the approach here can be extrapolated extensively.

Let us first try to understand what the researchers have done here. https://lnkd.in/gztqjmv8

At its core, wheel speed is simply how fast a wheel is rotating. It’s usually expressed as:

A. Rotations per minute (RPM) or
B. Angular velocity in radians per second (ω = dθ/dt)

Wheel-speed sensors (often called ABS sensors or vehicle speed sensors) measure this rotational motion and send a signal to the car’s control units (ABS, traction control, stability control, etc.) so they know how fast each wheel is turning.

But the sensor doesn’t “see” rotation directly. It detects a periodic signal that represents rotation. And that is where noise tries to ruin the party. Each wheel has a toothed ring (called a tone wheel or reluctor ring) that spins with the wheel.

The sensor sits close to the ring and measures changes in magnetic flux as the teeth (metal) and gaps (air) pass by. Every tooth-gap transition produces a voltage pulse. So, as you can assume, counting pulses corresponds to the number of rotations, and measuring the time between pulses gives us rotational speed.

So, the sensor outputs a periodic electrical waveform, which is something like a square or sine wave whose frequency corresponds to the wheel’s angular speed. Now, because this is a tiny voltage signal sitting in a harsh mechanical environment (rotating metal, vibration, electromagnetic fields, etc.), a plethora of noise types, ranging from mechanical to electromagnetic, can distort it.

This paper addresses the problem of noise interference in wheel-speed sensors (used in vehicle safety systems like ABS, TCS, ESP), in particular, high-amplitude or non-stationary noise that degrades sensor accuracy.

The authors propose an improved deep-learning model: an enhanced Long Short-Term Memory (LSTM) network augmented with an attention mechanism.

The enhanced LSTM model significantly outperformed the other methods across multiple metrics: Mean Squared Error (MSE), Root MSE (RMSE), Signal-to-Noise Ratio (SNR), Mean Absolute Error (MAE), Peak SNR (PSNR), and correlation coefficient (R) in recovering a clean wheel-speed signal.

So, if we zoom out and think about it, at its core, the wheel-speed problem is about a true physical process (rotation), which is converted to a measurable electrical signal. This signal is corrupted by non-stationary noise, so there is a need for the signal to be reconstructed digitally.

That chain exists everywhere in sensing.

The only things that change are the domain, frequency, sensor modality, and consequence of error. You see where I am going with this?