Hi @TIM,
Worth being precise here, because "predictive monitoring" in TrendMiner isn't one feature; there are three different mechanisms, and the strongest setup combines them. None of them is a trained ML forecast they're precursor detection and a linear soft sensor, but for "warn before a cooking cycle overruns" that's exactly what you want.
Route 1: Validate the precursor first: Diagnose → Cross-correlations. Before you alert on anything, confirm the early signals actually lead cycle duration. Cross-correlation recommends upstream parameters that correlate with your target and detects the time shift — up to 24 hours, split into 100 equidistant intervals — so it tells you not just that slow temp-rise predicts a long cycle, but by how many minutes it leads. This is offline root-cause work, not the alert itself, and the usual caveats apply (correlation ≠ causation; very short windows can score spuriously high). Use it to pick which precursor to trust and to set your thresholds.
Route 2: The real-time early warning (the one I'd build): VBS on the early indicator → Monitor. Scope a search to the early part of the cycle and alert on the precursor there e.g. within the first N minutes of a cooking cycle, temp-rise-rate below X and/or steam pressure below Y. Promote it to a Monitor and it fires within ~2 minutes of the precursor, before the cycle finishes (setting up a monitor). Two practical notes: there's no native "only look at the early window" switch, so you compose it a cycle-start condition AND the anomaly condition — reusing the cycle-start search from your cycle-identification work. And if you want "any of several precursors" (an OR across different tags), the VBS UI only ANDs conditions, so bake the logic into one flag tag EarlyWarn = if(or(rise_rate < X, steam_p < Y), 1, 0) and search on = 1.
Route 3: Optional soft sensor: the Prediction tag. This is multi-variable linear regression: you give it candidate tags, it ranks them by influence score and builds a formula with an accuracy score (your data as a solid line, the model dotted) (prediction tags). You can point it at early-cycle metrics as candidates to get a single predictive score, but two honest caveats: it's linear-only (3–10 inputs), and the docs don't show a worked cycle-duration example so check the accuracy score before trusting it. For a lightweight "where is this heading / time remaining" estimate, the linear-extrapolation pattern is simpler and often enough.
What I'd do: cross-correlation to confirm and time the precursor → VBS-on-early-indicator → Monitor as the live alert → add the Prediction tag only if you want a single soft-sensor score and the relationship turns out roughly linear. The closest community example to this whole pattern (early indicators + value-based search + monitor + dashboard) is the Catalytic Reactor Monitoring and Prediction Dashboard different process, but the methodology transfers directly.
Kind regards
Frederik
