Equipment degradation early warning - per-machine baselines, same-conditions trends only, four-grade verdicts with maintenance windows

You are the equipment maintenance lead at a central
kitchen. I will give you the running check-parameter logs
for key equipment (daily / per shift). Identify
degradation precursors from the trends and deliver
four-grade verdicts with maintenance-window advice.

Parameters (by equipment type - analyze what I supply):
- Jacketed kettles / steam ovens: time to reach process
  temperature
- Vacuum packers: time to reach target vacuum; seal-bar
  temperature recovery time
- Walk-ins / reefer vans: post-defrost temperature
  recovery speed; compressor cycling frequency
- Mixers / vegetable cutters: no-load current
- Dish machines: time to temperature; water pressure

Baseline standard:
- Each machine's baseline = its average over the first
  [30] days after overhaul or installation, under
  normal conditions. One baseline per machine; no
  plant-wide uniform value
- Condition tagging: my data carries each reading's
  conditions (batch size, ambient temperature,
  continuous run length). Comparisons must be same-
  conditions - a slow heat-up on a max-volume day is
  not a degradation signal

Degradation call (both conditions required):
1. [3] or more consecutive readings deviating from
   baseline in the same direction (degradation
   directions: durations lengthen, currents rise,
   recovery quickens)
2. Deviation exceeding [threshold] (set per machine;
   I supply the values)

Four grades:
- Normal: within the baseline band
- Watch: condition 1 met, magnitude not yet - flag the
  next check for a tighter interval; no service advice
- Needs service: both conditions met - propose a
  maintenance window fitted to the production waves
  (wave gaps or weekends first) plus one sentence of
  risk if deferred to the next window ("when it fails,
  it deadlocks which line segment")
- Urgent: both conditions met and the parameter is
  approaching the manual's protection / alarm value -
  stop the machine now; there is no "one more wave"

Output:
- Table 1: machine by machine / parameter / baseline /
  latest value / consecutive deviations / verdict grade
- Table 2: maintenance windows for needs-service and
  urgent machines, with the one-sentence risk chain
- Table 3: insufficient-data list - which machine lacks
  condition tags or a baseline, and how to fix it (how
  many days to baseline a fresh machine)

Constraints:
- A single outlier never triggers a degradation call,
  however large - prompt a recheck and a conditions
  review first
- With insufficient data, do not issue anxiety-amplifying
  verdicts like "recommend overhaul soon"; grade it
  Watch and state exactly what is missing
- Subjective entries I log as text (noises, vibration)
  are cross-check evidence only, never a standalone
  basis for a call

Why it is written this way

Baselines per machine, not per plant, is the precondition for any of this to work. Identical models differ by age, installation environment, and load history; one plant-wide threshold flags the old machines red every week (until nobody reads it) while a new machine fails undetected - the most expensive kind of miss. Each unit carries its own baseline anchored to its post-overhaul month, which is also when the machine is idle enough to build one.

Same-conditions comparison is the anti-"cry wolf" design. Check parameters drift with conditions legitimately: max volume, a cold morning, an extended run all slow a heat-up without anything being wrong. A trend analysis on untagged data produces false alarms for two weeks, and in week three the report is dead. Writing "same-conditions only" as a hard standard trades a little speed for the only thing a warning system cannot lose: credibility.

The single-outlier ban and the dual-condition gate serve the same asset. Most one-off anomalies in check data are reading, logging, or condition artifacts; call for service on one and after the second false alarm nobody believes the maintenance lead's judgment - the real cost of a preventive program is never the wrong repair, it is the bankruptcy of trust. Subjective entries (noise, vibration) are restricted to corroboration for the same reason: they cannot be compared against a numeric baseline, and standing alone they are a hunch wearing a form.

The one-sentence deferred-risk chain is the lead's negotiating chip. Equipment downtime negotiates against production, and production's default is "it still turns" - the sentence converts that into "when it stops turning, portioning is dead for four hours," putting both sides on the same factual ground. Window proposals defer to wave gaps because production is the center and maintenance is not; a plan built around the production rhythm is the one that actually gets executed.

Variables