Sensor-Driven ML Pipelines on Embedded Systems Development

Time-Series Classification

Mon, 01 Jan 0001 00:00:00 +0000

Time-Series Classification#

Time-series classification maps a fixed-length window of sensor data to a discrete class label — a gesture type, an activity state, a machine operating condition. The input is a sequence of sensor readings sampled at a known rate, and the output is a categorical decision: swipe left, swipe right, walking, running, bearing fault, normal operation. Unlike image classification where spatial structure is the dominant signal, time-series classification depends on temporal patterns — the shape, frequency content, and evolution of the signal over the window duration.

Anomaly Detection

Mon, 01 Jan 0001 00:00:00 +0000

Anomaly Detection#

Anomaly detection learns what “normal” looks like and flags deviations. Unlike classification, which requires labeled examples of every category, anomaly detection requires only normal data for training — a form of one-class learning. The model builds a representation of normal operating conditions, and anything that falls outside that representation is flagged as anomalous. This property makes anomaly detection particularly valuable for industrial and embedded applications where collecting labeled fault data is expensive, dangerous, or impossible (a machine cannot be intentionally run to failure just to collect training data).

Predictive Maintenance

Mon, 01 Jan 0001 00:00:00 +0000

Predictive Maintenance#

Predictive maintenance estimates when a machine component will fail, enabling maintenance scheduling after most useful life is consumed but before failure occurs. The distinction from reactive maintenance (fix after failure) and preventive maintenance (replace on a fixed schedule regardless of condition) is that predictive maintenance uses actual sensor measurements to forecast the remaining useful life (RUL) of a component, optimizing the trade-off between maintenance cost and downtime risk.

Sensor Preprocessing for ML

Mon, 01 Jan 0001 00:00:00 +0000

Sensor Preprocessing for ML#

Sensor preprocessing is the bridge between raw sensor hardware and the ML model input tensor. The sensor produces a continuous stream of ADC samples or digital readings at a fixed rate. The model expects a specific fixed-size, normalized tensor at each inference cycle. Between these two endpoints lies a pipeline of buffering, windowing, normalization, and format conversion — and any mismatch between the preprocessing applied during training (in Python on a desktop) and the preprocessing applied during deployment (in C on an MCU) silently degrades model accuracy without producing any error message.