Time-series analysis (industrial)
- Almost all industrial ML is time-series ML: models built for ordered, multivariate, regime-switching sensor data.
- The working methods are multivariate anomaly models, drift/changepoint detection, and short-horizon forecasting — not general-purpose LLMs.
- High-frequency signals (vibration, electrical waveforms) are their own specialty with dedicated platforms.
- You evaluate platforms on behavior (attribution, regime handling, latency), not on which algorithm family they cite.
Factory data is time-series data: ordered, dense, multivariate, and constantly switching regimes as products change over. The ML that works on it is a specialized family, and knowing the shape of it inoculates you against buzzwords.
The three working methods
Multivariate anomaly models learn the joint distribution of many variables and score live data against it — the engine behind anomaly detection and root-cause attribution. Drift and changepoint detection finds slow creep and abrupt shifts in behavior — the precursor-hunting engine behind process-signal predictive maintenance. Short-horizon forecasting predicts a target (quality metric, energy draw, remaining useful life) minutes to days ahead — the engine behind predictive quality and maintenance scheduling. Every platform on the market is some packaging of these three, plus the attribution and UX layers that make them usable.
High-frequency is its own world
Vibration and electrical-waveform analysis runs at kilohertz where feature extraction and self-supervised methods dominate; it's why machine-health specialists and process-ML platforms remain different products even as they converge on the same maintenance budget.
Where LLMs fit
Nowhere in this layer. Large language models are not built for multivariate numerical inference, and the credible vendors say so plainly — they use LLMs to explain time-series model outputs, not to produce them. A vendor claiming their chatbot analyzes your sensor data directly has told you something important.
FAQ
- Do I need to know which algorithms a vendor uses?
- No — you need to test behavior: does it attribute causes, handle regime changes, run at your data rates, and explain itself to engineers? Two vendors with identical algorithm lists can differ 10× in usable output.
- What about foundation models for time series?
- Pretrained time-series models are an active research area and appear inside products as accelerators (faster cold starts on new lines). In 2026 they supplement plant-specific training rather than replacing it.
Every number on this site — in one sourced spreadsheet.
The $43.6B market math, use-case shares, adoption barriers and published deployment results, with the source column intact. Ready to paste straight into your business case.