Crow-AMSAA analysis, also called reliability growth analysis, is a plotting technique that reveals whether a machine, fleet, or plant is becoming more reliable, staying flat, or quietly degrading. It plots cumulative failures against cumulative operating time on log-log axes; the slope of the resulting line, beta, is the verdict: below 1 means reliability is improving, near 1 means nothing is changing, above 1 means failures are accelerating.
The method descends from Duane’s observation that failure data from developing systems plots as a straight line on log-log paper, later formalized by Larry Crow at the US Army Materiel Systems Analysis Activity (AMSAA) as a statistically rigorous model. Its industrial appeal is robustness: it tolerates mixed failure modes and messy real-world data, exactly what plant maintenance records look like, and it asks only for failure dates and a usage measure (hours, cycles, tonnes).
A filling line logged 24 breakdowns in its first 4,000 operating hours, then a revised preventive program launched. Over the next 4,000 hours it logged 9 breakdowns. On the Crow-AMSAA plot the early data runs at beta of roughly 1.1, slightly degrading; after the program the slope breaks to about 0.6, strong growth. Management asked "is the PM program paying?" and the answer is visible in one picture: the same line that produced a failure every 167 hours now produces one every 440 and stretching. The plot also projects forward: staying on the current slope, the next failure is expected in roughly three weeks, which is a rational way to time spares and inspections. That projection assumes conditions stay as they are; the assumption should be stated, not hidden.
The two are complements, not rivals. Weibull analysis studies one failure mode across a population of identical components to characterize its life distribution, ideal for deciding replacement intervals. Crow-AMSAA studies a repairable system over time with all its failure modes mixed together, ideal for judging whether overall reliability management is succeeding. Weibull answers "when will this component type fail"; Crow-AMSAA answers "is this machine getting better or worse under our care," which also makes it a fair scoreboard for the strategies suggested by the six failure patterns research.
Only two things, but honestly: complete failure events and credible operating time. Both die in bad record-keeping: breakdowns fixed without work orders vanish from the numerator, and calendar time masquerading as run time corrupts the denominator. Clean, consistent failure records with real run hours are the entire entry fee, and metrics like MTBF inherit the same dependency.
Fabrico supplies the two ingredients the plot needs: every failure captured as a work order with timestamps and codes, and true operating time measured automatically by real-time production monitoring rather than estimated from the calendar. Export the pairs and the Crow-AMSAA chart takes minutes in any analysis tool; the hard part, trustworthy data, is already done. Fabrico does not run the statistical model itself; it makes the model worth running. EU-built, with EU data residency.
Meaningful slopes emerge from surprisingly few points; a dozen failure events already draw a usable line. That makes the method practical for single critical machines, not just fleets, though more events sharpen both the slope estimate and any change points.
Yes, that is its design strength. The plot tracks the system’s total failure intensity, mixed modes and all. When the line bends the wrong way, you then drill into modes and components with Pareto and Weibull tools to find out why.
They share a symbol and a flavor (below 1 improving or infant, above 1 worsening or wear-out) but belong to different models: Weibull’s shape describes a component life distribution; Crow-AMSAA’s beta describes the failure intensity trend of a repairable system. Conflating them is a common and confusing mistake.
Want failure history and true run hours clean enough to plot? Book a Fabrico demo to see automatic OEE capture and CMMS records build the reliability dataset your analyses deserve.
Login
Blog
