The Sensor Stack: From Physics to Machine Learning

Engineers from different disciplines rarely disagree about facts. They disagree because they are standing on different layers of the same stack and using the same words for different objects. When a mechatronics engineer says "the signal is noisy", a machine-learning engineer hears "the labels are unreliable", and a firmware engineer hears "the ground return is shared with the motor". Learning the layer map is how you become useful in that room.

1. The seven layers

2. Phenomenon: the layer you cannot debug later

Before instrumentation you choose what physical story you believe. Crust formation is a coupled heat-and-mass-transfer process with a chemistry layer on top: water leaves the surface, surface temperature rises past the point where Maillard reactions accelerate, volatile compounds are produced and carried away by airflow. Each clause in that sentence suggests a channel — thermal, mass, gas, acoustic. If your physical story is wrong, your sensor selection is wrong, and no later layer will tell you why.

3. Sensor and analog front end: the information ceiling

Layers two and three set an upper bound on everything above. If the microphone's useful band stops at 8 kHz, the sizzle harmonics above it are gone. If the amplifier saturates during the first thirty seconds of searing, that interval is not merely noisy — it is absent. Machine learning cannot invent information that the front end discarded; it can only redistribute what remains.

4. Acquisition: timestamps are data

Multimodal systems live or die on time alignment. A camera at 30 frames per second, a probe at 1 Hz, an inertial unit at 400 Hz and a gas array with a 20-second response time do not share a clock or a latency. Fusion requires a common timebase, an explicit resampling policy, and honesty about each channel's group delay.

import pandas as pd

# Align three channels onto a common 1 Hz grid, respecting each one's nature.
grid = pd.date_range(start, end, freq="1s")

probe   = probe_df.reindex(grid, method="nearest", tolerance="2s")   # slow, smooth
mass    = mass_df.resample("1s").mean()                              # average out fan vibration
sizzle  = audio_energy_df.resample("1s").max()                       # transients matter, keep the peak

# A gas sensor with ~20 s response is shifted, not just resampled.
voc = voc_df.shift(freq="-20s").reindex(grid, method="nearest")

You will not write this code, but you should recognise it. The three different aggregation choices — nearest, mean, max — encode three physical claims. Reviewing those claims is a design conversation, not a coding one.

5. Preprocessing: the layer that silently breaks deployments

Preprocessing is deterministic and unglamorous: detrend, remove baseline, normalise, window, resample. Its danger is that it usually exists twice — once in the training notebook and once in the deployed firmware. Any divergence between the two produces a model that scored well offline and behaves strangely in the kitchen. Mature teams define preprocessing once, as shared, versioned code, and treat any duplication as a defect.

6. Features and embeddings: two ways to compress

A feature is physics you wrote down. Spectral centroid for brightness of a sizzle. Rate of mass loss per minute. Slope of the VOC baseline over the last ninety seconds. Features are interpretable, sample-efficient, and only as good as your hypothesis.

An embedding is physics the network inferred. You give a convolutional or transformer encoder raw spectrograms and it produces a vector whose geometry encodes similarity. Embeddings capture structure you never articulated — and also structure you never wanted, like the acoustic signature of your particular extraction fan.

7. Where to fuse

Fusion depth is one of the few genuinely architectural decisions in a multimodal system.

1. 01Signal-level (early) fusion: combine synchronised raw streams into one tensor. Maximum information, maximum sensitivity to missing channels and clock error.
2. 02Feature-level fusion: compute per-modality features or embeddings, then concatenate. The common industrial compromise; degrades gracefully when a channel drops.
3. 03Decision-level (late) fusion: each modality produces a score, then a rule or small model arbitrates. Most robust, most interpretable, lowest ceiling.

8. Model: calibration matters more than accuracy

For control systems, a model's confidence is as consequential as its prediction. An 85%-accurate model that knows when it is unsure is far more useful than a 90%-accurate model that is uniformly certain, because the first can defer to a human or a conservative default and the second cannot. Ask for reliability diagrams, not just accuracy numbers.

9. Decision, control and the feedback arrow

The final layer converts belief into action: reduce fan speed, hold the setpoint, alert the operator. Two properties dominate. First, latency budget — the sum of acquisition, transport, inference and actuation delay, which must be short relative to the process time constant. Second, hysteresis in the decision itself, so the system does not oscillate between two actions at the boundary.

10. A diagnostic order of operations

1. 01Reproduce on recorded data. If it disappears, the problem is in acquisition or timing, not the model.
2. 02Compare training and production distributions per channel. Look for a sensor that moved, aged or was replaced.
3. 03Verify preprocessing parity between notebook and device, byte for byte.
4. 04Check calibration and drift records for every channel.
5. 05Only then question the model architecture.

Recap

The stack runs from phenomenon to action, and information only ever decreases as you climb it. Layers one to four decide what is knowable; layers five and six decide what is expressed; layer seven decides what is believed; layer eight decides what is done — and then quietly changes layer one. Hold the whole picture and you can locate almost any problem to a layer within a few questions. That single skill is what lets a non-specialist founder hold a serious conversation with the specialists building the machine.