Sei sulla pagina 1di 13

Accuracy and Precision

Repeatedly measure a fixed quantity Compare the results with the correct answer Accurate (unbiased): The mean is on target Precise: Standard deviation is small (close together) Example: A person who weighs 75 kg is weighed repeatedly on a scale. Any of the following patterns could arise.

Using Statistics James Reilly

Fig 8.7

Using Statistics James Reilly

Gauge Linearity and Bias Studies


Estimates the average bias How bias changes with bigger objects (linearity) Instrument: Gauge Objects: Parts Select a wide range of parts Get the master measurement for each part Measure each part a number of times Example: A kitchen scale was tested by weighing five consumer products a number of times each.

Using Statistics James Reilly

Table 8.3
Part 1 2 3 4 5 Master 200 300 400 625 700 Measurement 260 270 375 340 500 505 725 740 780 790 Deviation 60 70 75 40 100 105 100 115 80 90 85.0 107.5 102.5 57.5 Average Deviation 65.0

Average bias = 83.5 Regression equation of average deviation on master: Average Deviation = 53.5 + 0.0673 Master % Linearity = +6.73%
Using Statistics James Reilly

Components of Variance
Total variation consists of: Material variation (the parts are not all the same) Measurement system variation Model 8.1
2 2 2 Total = Mat + MS

Measurement system variation consists of: Reproducibility (rpd): Explained variation, i.e. different sessions (operators, instruments, days) Repeatability (rpt): Unexplained variation Model 8.2
2 2 2 MS = rpd + rpt

Model 8.3
2 2 2 2 Total = Mat + rpd + rpt
Using Statistics James Reilly

Gauge R&R Studies


A designed experiment Take a number of typical parts Measured them a number of times in succession Do this on a number of different sessions Example: A measurement system is used to measure the deflection of timber beams, for which LSL = 20 mm and USL = 100 mm. An R&R study involved measuring fifteen randomly chosen beams twice each, on each of four days. This involved taking 120 measurements; thirty on each session; eight on each part. Two-way ANOVA Response: Measurement Factors: Part, Session
Using Statistics James Reilly

Table 8.4 Two-way ANOVA: Measurement versus Part, Session


Source Part Session Error Total DF 14 3 102 119 SS 1982.98 6.52 752.14 2741.64 MS 141.641 2.174 7.374 F 19.21 0.29 P 0.000 0.829

Using Statistics James Reilly

Variance components
Error (repeatability): 7.374 Session (reproducibility): (2.174 7.374)/30 = -0.1733 Cannot be negative, adjust to zero Measurement system: 7.374 + 0 = 7.374 Part (Material): (141.641 7.374)/8 = 16.7834 The corresponding standard deviation estimates are: Repeatability: 2.71549 Reproducibility: 0 Measurement system: 2.71549 Material: 4.09676 Signal-to-noise ratio (SNR) Formula 8.6

Mat SNR = MS
Using Statistics James Reilly

Can it distinguish between typical parts? Threshold values for the SNR are as follows: >10 - good 3-10 - acceptable <3 - unacceptable In this case: SNR = 4.09676 / 2.71549 = 1.5 Can it distinguish between good and bad parts? Precision to tolerance ratio (%R&R) Formula 8.7

6 MS %R & R = 100 USL LSL


Using Statistics James Reilly

Threshold values for the %R&R are as follows: <10% - good 10-30% - acceptable >30% - unacceptable In the example above, USL LSL = 80 %R&R = (6 2.71549) / 80 100 = 20.4% Corrective action for measurement systems Identify main problem: rpd or rpt If rpd: May be training or calibration If rpt: May be no solution

Using Statistics James Reilly

Attribute Gauge R&R Studies


Parts are good or bad Measurement system accepts or rejects Go/no-go gauges or human inspectors Use samples of both good and bad parts Inspect at least twice Have at least two instruments Use colour-coded parts for automated test systems Disguise and randomise for human inspectors Example: Fifty biscuits (thirty-one good and nineteen bad) were presented twice, in random order, to two inspectors (Evan and Silvia) who were required to judge their colour acceptability. An extract from the results is shown in Table 8.5: 1 denotes good or accept, while 0 denotes bad or reject.

Using Statistics James Reilly

Table 8.5
Part 1 2 3 4 5 Master 1 0 1 1 1 Evan 1 1 0 1 0 0 Evan 2 1 0 1 0 0 Silvia 1 1 0 1 1 0 Silvia 2 1 0 0 1 0 Outcome Effective Effective Rpt Rpd Bias

Using Statistics James Reilly

Rows 1, 2: Measurement system is effective Row 3: Repeatability - inconsistent inspector Row 4: Reproducibility - inspectors disagreed Row 5: Bias - classification was wrong Effective: 80% (forty rows) Repeatability: 10% (five rows) Reproducibility: 8% (four rows) Bias: 2% (one row) Remedial actions: Training and communication (bias and rpd) Provision of specimens (for rpt) (200% inspection) Bias 1: False positive (rejected but good) Bias 2: False negative (accepted but bad) False positive is more frequent False negative is more costly

Using Statistics James Reilly

Potrebbero piacerti anche