control Supplier->Input->Process->Output->Customer resource

Fixed ratio measurement scale: weight, height, absolute 0

Fixed distance measurement scale: temperature

Ordinal measurement scale: top, middle, bottom

Categorical (nominal) measurement scales: gender, occupation, place of birth

mean, median, mode, 1/4 quantile, 3/4 quantile

Standard deviation, variance, range, interquartile range

Degree of asymmetry: Skewness

Steepness: kurtosis

1. Look at the P value on the probability graph, P>0.05, normal 2. Then use normal to calculate Cp/Cpk/Pp/Ppk

Intra-group fluctuations Cp, Cpk-where Cp=USL-LSL/6σ, Cpk=min{USL-μ/3σ, μ-LSL/3σ} Inter-group fluctuations Pp, Ppk, Cp, Cpk, Pp, Ppk belong to the Wangda index ≤ 1 process capability is insufficient, 1 ≤ Pp ≤ 1.33 process capability is acceptable, 1.33 ≤ Pp ≤ 1.67 process capability is sufficient eg: Cp is large, Cpk is small, the difference between Cp and Cpk is large, the mean needs to be improved Cp is small, Cpk is small, the difference between Cp and Cpk is small, the fluctuation needs to be improved Cp is small, Cpk is small, the difference between Cp and Cpk is large, the mean and fluctuation need to be improved

1. Look at the P value on the probability plot, P<0.05, non-normal 2. Histogram, check the data formation mechanism, there is a double peak state, process the data and repeat again 3. For individual distribution identification, choose the one with smaller AD value and larger P value (multiple parameters are generally not selected, Box-cox and Johnson) and use non-normal calculations to calculate Pp and Ppk. 4. Prefer Box-cox, then Johnson 5. Non-parametric calculations

actual process fluctuations σ p²

Measurement system fluctuations σR&R²

Continuous

Discrete

1.FMEA factor screening

2.SOD scoring

3. Find the factor with the highest IPN score

Single Z (known σ) large sample ≥ 30

Single t (μ comparison, sample ≤ 30) 1 set of samples

Two-sample T (normal test first, equal variance test, double μ comparison) 2 sets of samples

Paired T (measure twice to confirm the difference in μ mean) 1 set of samples, different states of the same sample, test μ

One-way analysis of variance (normality first, then equal variance test) More than 2 groups of samples (multi-level test μ)

General linear model is also called GLM Multi-factor and multi-level, test μ (default data is normal and equal variances, if not normal and equal variances, use one-factor analysis of variance)

Single variance/equal variance/double variance Test σ

1P (known ratio and single group)

2P (Comparison of defective rates between the two groups)

Chi-square test (Known defective rate, unknown correlation)

Single factor linear regression

Multiple Regression

step

Binary logistic regression

1. Establish a hypothesis Null hypothesis: high probability events, equal, no difference, self-evident, irrelevant, independent, not significant Alternative hypothesis: small probability event, unequal, different, to be proven, related, not independent, significant

2. Set the significance level α (new regulations of 0.05 for stage A and 0.1 for stage I)

3. Assume H0 and calculate statistics

4. Calculate P value

5. Decide whether to reject H0 (P>0.05, the null hypothesis cannot be rejected, so pay attention to the expression)

Randomize the test sequence to evenly distribute test errors to all tests to prevent some unknown errors

Repeating the test multiple times under the same conditions is the only way to quantify the test error.

Homogeneous differentiation of uncontrollable factors to reduce experimental errors, such as: shift, date, manufacturer

Factor matching is "evenly dispersed, neatly comparable" and maintains the independence of factors

The impact of reducing the number of trials on Y is indistinguishable and only exists in some factor experiments.

Taguchi experimental design, number of factors ≥5

Partial factorial experimental design, number of factors ≥ 5

2-level full factor DOE, number of factors ≤ 5

General full-factor DOE, number of factors ≤ 5

Steps: 1. Determine X, Y 2. Create a Factorial Design 3. Test 4. Analyze Factorial Design 1) Whether model P <0.1 (the new version requires 0.1, the old one is still 0.05) 2)S,Rsq,Rsq(adj) 3) Residual diagnosis 4) Response optimizer (look at the eyes, look at the big, look at the small) 5) Get the optimal X 6) Test verification

best factor

continuity data

defect data (Poisson)

Defective product (two items)