Five failures your threshold alarms catch too late.

A threshold is a good instrument for a bad event. It is a poor instrument for a slow one.

Set a limit, cross it, get an alarm. That works when the failure arrives as a step change: a trip, a lockout, a phase loss. It works badly when the failure arrives as a slope, because a limit only ever sees the current reading. It cannot see that the reading has doubled every month for four months and is about to cross whether you like it or not.

Five degradation modes where the trend carries the information and the limit carries almost none. In every one, the signal is the rate of change, and a threshold is structurally incapable of reporting a rate.

Rate, not level

what these five failures signal with

Weeks to months

typical warning available in the trend

One reading

what a threshold gets to work with

1. Partial discharge in switchgear and cable terminations

Partial discharge erodes insulation from the inside. Small voids ionise, the discharge widens the void, and the widened void discharges more readily. The mechanism accelerates itself, which is why the shape over time is a curve rather than a line.

A magnitude threshold set where flashover risk becomes serious is set at the top of that curve. It fires late by construction. What matters far earlier is that discharge activity is present and rising, and that its pattern relative to the power cycle has a consistent phase signature, which separates a defect from switching noise. Neither of those is a level.

What to watch instead

• Whether discharge activity exists at all where it previously did not.
• The rate of increase in magnitude and repetition, not the magnitude itself.
• Phase-resolved pattern stability, which separates a defect from intermittent interference.

2. Dissolved gas evolution in transformer oil

Dissolved gas analysis is the most established trend-first diagnostic in the industry, and it is still routinely operated as a limit check.

The absolute concentration of a given gas depends on oil volume, load history, temperature and how long since the last degassing. A transformer sitting well below every limit while its acetylene generation rate climbs is in worse condition than one sitting stable near a limit it has held for a decade. The first is an arcing fault developing; the second is a unit with a known, stable characteristic.

Ratio methods exist because the useful information is in the relationships between gases and in their rate of production. A limit on a single gas discards both.

3. Bearing degradation in rotating plant

Bearing failure has a well-documented progression, and it is unusually unhelpful to a threshold. Early damage appears in the ultrasonic and high-frequency envelope. It then moves down into the detectable defect frequencies. Broadband vibration, which is what most overall-level alarms measure, rises meaningfully only in the later stages.

There is a further trap. As spalling progresses and the damaged surface becomes uniformly rough, the discrete defect frequencies can subside, and a level-based reading may briefly improve on its way to failure. A threshold watching that number sees recovery. The asset is closer to the end than it was.

4. Capacitor and DC-link ageing in power electronics

Electrolytic capacitors dry out. Capacitance falls and equivalent series resistance rises, both gradually and both temperature-accelerated. The converter keeps meeting its output specification for most of that journey, because the control loop compensates.

So the observable symptom is not a threshold crossing. It is that the unit is working harder to produce the same result: ripple climbing, operating temperature drifting up at unchanged load, efficiency edging down. The compensation hides the degradation right up until the compensation runs out, and then the failure looks sudden despite having been visible for months to anything measuring drift.

5. Thermal path degradation in cooling systems

Blocked filters, fouled heat exchangers, failing fans and silted cooling passages all produce the same signature: the equipment reaches a slightly higher temperature for the same load and ambient conditions than it used to.

A temperature alarm cannot see this, because it is not watching the relationship. Ninety degrees on a hot afternoon at full load is normal. Ninety degrees on a mild morning at half load is a cooling system that has lost a large part of its capacity. It will read as perfectly fine right up to the first hot day, which is also the day everything else is at peak load.

The measurement that matters is temperature normalized for load and ambient, tracked over time. That is a derived quantity. No threshold on a raw sensor will ever produce it.

The common shape

All five share a structure. The degradation is gradual, self-accelerating or compensated, and it announces itself as a change in rate long before it reaches any defensible limit. Setting the limit lower does not fix this. It converts a late alarm into a noisy one, and a noisy alarm gets acknowledged without being read.

Thresholds fail in two directions at once, and tuning trades one failure for the other. What breaks the trade is a baseline: what this asset, in these conditions, normally does, and how far today departs from it.

Our analysis of 14,898 support cases from monitored sites found that on 94.4% of equipment events, that same site had raised that same kind of event before. The pattern was already in the history.

What this changes in practice

Alarm on rate as well as level.

A value rising faster than it has ever risen needs a look even when it is nowhere near a limit.

Normalize before you compare.

Temperature, efficiency and ripple mean nothing without load and ambient alongside them. Compare like conditions, not raw readings.

Keep the baseline per asset.

Class limits are a starting point. A specific unit's own history is a far better predictor of that unit than a specification written for the class.

Treat recurrence as a finding.

The third instance of the same event at the same site is not three independent alarms. It is one unresolved problem, and it should be presented that way.

None of this requires new sensors. Every one of these five signals is already present in data most operators are collecting and discarding. The data is already there.