You already know: time-series features and spectral analysis (D3), anomaly detection and autoencoders (D5), streaming pipelines (D16) and, from the fraud and spam cases, that the people acting on alerts have a fixed capacity. This case adds industrial sensor data: high-rate signals, machines that behave differently at different loads, failures that are rare and poorly labelled, and an edge that must keep working when the network does not. The design questions: what to compute at the edge, how to detect trouble without failure labels, how to turn scores into work orders a team can handle, and how to keep machine learning out of the safety loop.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "stream all sensor data to the cloud and train an autoencoder". A staff candidate first asks what the machines and sensors are, what a failure costs and how much warning helps, what labels exist and, most importantly, who acts on an alert and how many they can handle. Those answers produce feature extraction at the edge, a lead-time target, normal-behaviour models before supervised ones, and an alert budget per site.
Assumptions you state: