Can equipment faults be detected without labelled failure data?

What unlabelled methods can and cannot say

A label-free approach extracts normal operation from historical data. Several methods do this: statistical limits, Mahalanobis distance, isolation trees, and reconstruction error from principal component analysis (PCA). Each asks the same question: does this observation fit the normal behaviour the model has learned?

What the method knows is that something departs from normal. It does not know the fault type, the part or the cause, because it never saw those labelled. A flag therefore means “review this period”. The same departure could come from a bearing fault, a sensor drift, a shift handover or a new product grade, and the model cannot tell these apart.

Validating by replaying past incidents: a hypothetical

Suppose a compressor has a two-year maintenance log with four bearing failures, and the model flagged 12 periods over those two years. Operators confirm 5 of the 12 flagged periods as real abnormal conditions.

• Operator confirmation rate: 5 / 12 ≈ 0.42, the share of flags found to be real.
• Failure coverage: 3 / 4 = 0.75, so three of the four failures had a flag beforehand.
• Lead time: for those three, the first flag came 9, 6 and 3 days before failure, a median of 6 days.

The maintenance log is itself an imperfect label. Failures may be recorded late or against the wrong asset, so these numbers guide the review rather than prove accuracy.

Limits and common mistakes

The most common mistake is treating the number of alarms as success. Alarm count is not a measure of performance, and a model that flags constantly looks busy while proving nothing. A second mistake is training on a window that includes unrecorded faults; the model then learns the fault as normal and misses it. A third is validating with a random split, which lets future data leak into training.

Operator feedback is valuable. Each verdict becomes a label, and enough of them can support a supervised model later. In the early stage, however, the labels are biased toward the periods operators chose to inspect, so they reflect attention as much as reality.

Follow-up questions