EC-20.7 Sensor Fusion, and Machine Learning on the Device

Work through these

• Why one sensor is never enough: bias, drift, noise and dropout

  Each sensing technology fails in its own characteristic way, and the failures are usually uncorrelated. Combining two whose errors are independent produces an estimate better than either, which is the whole argument for fusion.

  NPTEL: Introduction to Internet of Things (IIT Kharagpur) · Course

• The complementary filter, and when it is all you need

  Trusting one sensor at low frequency and another at high frequency is a two-line calculation that solves a surprising number of real problems. Reaching for something heavier before trying it is a common waste of effort.

• The Kalman filter as a prediction corrected by a measurement

  The filter keeps a model of the state, predicts forward, and blends the prediction with the next measurement in proportion to how much each is trusted. Reading it that way makes the equations follow from the idea rather than the reverse.

  NPTEL: Introduction to Internet of Things (IIT Kharagpur) · Course

• Fitting a model into a few hundred kilobytes: quantisation and pruning

  Cutting weights to eight-bit integers and removing the ones that contribute little can shrink a model by an order of magnitude for a small loss of accuracy. Whether that loss is acceptable is a question about the application, not the model.

  NPTEL: Embedded System Design with ARM (IIT Kharagpur) · Course

• Deciding where inference runs, and what each choice costs

  Running on the device costs memory and energy, running on a server costs a radio transmission and a round trip, and running nearby costs a second box to maintain. The energy budget from the previous topic usually settles it.

  NPTEL: Design for Internet of Things (IISc Bangalore) · Course

• Collecting the data honestly, and knowing when the model has gone stale

  A model trained on data from one installation often fails at the next, and nothing on the device will announce it. Deciding in advance what evidence would show the model has drifted is part of shipping one.