mathematics//signal processing

Signal processing is the branch of applied mathematics that turns a measured sequence of numbers into the information it carries, and in an engineered system it is the layer between the sensor and every decision: it removes the frequencies that bother, keeps the ones that inform and extracts features such as the energy in a band. A gyro read 8,000 times a second, the vibration of a pump bearing and the current of a motor all reach the processor as raw streams; what the controller, the estimator or the maintenance model sees is what this layer left of them.

Signal processing is the branch of applied mathematics that turns a measured sequence of numbers into the information it carries, and in an engineered system it is the layer between the sensor and every decision: it removes the frequencies that bother, keeps the ones that inform and extracts features such as the energy in a band. A gyro read 8,000 times a second, the vibration of a pump bearing and the current of a motor all reach the processor as raw streams; what the controller, the estimator or the maintenance model sees is what this layer left of them.

The family exists because every member answers part of one chain. First the signal has to become numbers at all: sampling reads it at fixed instants and decides which frequencies survive (what lies above half the rate folds back as aliasing), and quantization rounds each reading onto a finite grid of levels. Both happen inside the analog-to-digital converter, and the protection that must act before it, the anti-aliasing filter, is hardware. Then the numbers are read in frequency, where causes mixed in time separate (Fourier analysis), and shaped by a digital filter that keeps one band and attenuates the rest.

Every operation in the chain has a price that shows up elsewhere.

Sampling slower saves data and invites aliasing, fewer bits save silicon and add noise, and a filter that removes noise adds delay, which a control loop pays in stability margin. The noise never disappears; it moves to whichever place the designer chose (error relocation).

Sampling and quantization are the two cuts every measurement suffers, one in time and one in amplitude. They lose information predictably: a rate below twice the bandwidth makes fast content indistinguishable from slow content (sampling theorem), and NNN bits leave a rounding noise of a known size.

Fourier analysis is the reading tool. The discrete Fourier transform says how much of each frequency a record holds, which is how a bearing defect or a motor harmonic is found, and the spectrogram stacks short spectra when the machine changes over time.

The digital filter is the shaping tool, and its members differ in what they cost. A moving average or a first-order low-pass filter cover most sensor cleaning; a median filter removes isolated spikes that linear filters smear; a notch filter removes one narrow vibration line; FIR and IIR designs give sharper bands when they are needed; zero-phase filtering removes the delay only on recorded data.

Features come last. The RMS of a vibration, the energy in a band or the envelope spectrum of a bearing (vibration analysis) are the inputs of alarms and learned models, and a learned convolution is a filter whose coefficients came from data (CNN).

The simplest tool with margin usually wins here (flyswatter rule): a median of three for spikes plus a first-order low-pass several times above the loop bandwidth handles most sensor cleaning, and removing a vibration at its source beats any filter.