Defects Per Unit (DPU)

DPU is the most intuitive defect metric: "on average, how many things are wrong with each unit we produce?" Unlike DPMO, it does not require defining the number of opportunities — which can be subjective and contentious.

For multi-defect products (PCBs, assembled systems, painted surfaces), DPU is often more practical than defect rate because a single unit can have multiple defects. If 100 units have a total of 15 defects, DPU = 0.15. Some units may have zero defects, some may have two or three. The Poisson distribution connects DPU to yield: FPY ≈ e^(-DPU). At DPU = 0.15, approximately 86% of units are defect-free.

DPU tracks improvement over time more transparently than percentage-based metrics. Reducing DPU from 0.50 to 0.25 is a 50% improvement in defects, regardless of the defect type distribution. It also enables cost analysis: if each defect costs $12 to repair, a DPU of 0.50 costs $6 per unit in rework.

EntropyStat contributes to DPU reduction by providing more accurate process characterization for each defect type. When process distributions are accurately modeled with the EGDF, engineers can predict which dimensional characteristics are most likely to generate defects and prioritize improvement efforts accordingly.

For variable data (measurements), the EGDF computes the probability of nonconformance for each characteristic, and summing these probabilities gives a predicted DPU. This prediction is only accurate if the individual probabilities are accurate — which requires correct distributional modeling. The EGDF's assumption-free approach ensures that the predicted DPU matches observed DPU more closely than normal-based predictions.

When DPU is driven by a few dominant defect types, the ELDF's cluster detection helps identify whether those defects come from a general process spread issue or from a specific subpopulation. A DPU of 0.30 driven by two distinct clusters (one from shift A, one from worn tooling) requires different corrective actions than a DPU of 0.30 from a single broad distribution. EntropyStat separates these scenarios automatically.