I suppose you can try to rely on the ‘fact’ that your Customer hasn’t seen any defects but that won’t last long.
Below is the 35,000 level view of acceptance sampling. I don’t explain a lot except the bare mechanics. Quality Engineering is not a paint by numbers exercise. There are several approaches that might be acceptable to your auditor (note I didn’t say that they were effective just that an auditor would accept them)
Please start reading the collective works of Donald Wheeler. You can also take a look at my Essential References for Quality Engineering list as it has many links to free articles and a set of books that are pretty useful.
You could determine a plan based on AQL and lot size just use general inspection and don’t get ‘fancy’. There are plenty of tables and calculators out there on the interwebs for this. They are not truly statistically valid as they are based on lot size and negotiations but since they’ve been there for decades most people just accept them.
The AQL is the Acceptable Quality Level. This means that the AQL defect rate will be accepted 95% of the time it is present. It really isn’t a very good protection method for today’s quality levels, but as I said, history gives it a pass.
Now here are the parts you won’t like:
The sample is taken after the lot is produced and is RANDOM. It isn’t from the top of the box or the closest upper corner of the pallet. (Those are known as ‘convenience samples’.). RANDOM means RANDOM. IF you fail the accept number (which is often 0 as it provides the smallest sample size) then you must inspect the entire lot.
After the lot is produced and Random also mean that you cannot simply spread the samples out across the lot as it is being produced. (The math relies on truly random samples form the finished lot). Personally I find that taking samples throughout the lot at predetermined levels provides far more information and protection but that is a different type of sampling plan and some auditors don’t understand it at all or accept it. It isn’t statistical in the traditionally accepted sense…
A better approach is to determine the Rejectable Quality Level (RQL). This is the minimum level of defects you want to detect and reject most of the time. Use the c=o tab in the spreadsheet. You will also have to determine what probability (aka Confidence) you want to have in detecting the RQL defect rate. It is the simplest plan. Typical probability levels (that won’t make your auditor’s eyebrows shoot up his forehead) are 95% and for critical characteristics it may 99%. Going below 95% is not recommended unless you are making inconsequential parts.
Now the caution I’m going to give is that categorical (pass/fail) data requires fairly large sample sizes.
A Good approach is to utilize Statistical Process Control (SPC) to ensure low defect rates as it can be used quite effectively with continuous data (measured data, actual values). Basically you sample the process at each known change point: first pieces, last pieces and every material, setting and operator change in between. A default for this in high volume manufacturing is every 2 hours. (There are way better ways to determine the frequency but that is beyond the scope of this post). The sample size for each sample point can be as small as 1 and maybe no more than 3-5. (If you are machining more than 1 part at a time the sample plan is different.). You start by plotting each value against the specification limits on a run chart; data in sequential order of manufacture…if you get a failed part you 100% sample back to the last known good measurement. SPC is way more complicated than this but it’s a start.
Below is the 35,000 level view of acceptance sampling. I don’t explain a lot except the bare mechanics. Quality Engineering is not a paint by numbers exercise. There are several approaches that might be acceptable to your auditor (note I didn’t say that they were effective just that an auditor would accept them)
Please start reading the collective works of Donald Wheeler. You can also take a look at my Essential References for Quality Engineering list as it has many links to free articles and a set of books that are pretty useful.
You could determine a plan based on AQL and lot size just use general inspection and don’t get ‘fancy’. There are plenty of tables and calculators out there on the interwebs for this. They are not truly statistically valid as they are based on lot size and negotiations but since they’ve been there for decades most people just accept them.
The AQL is the Acceptable Quality Level. This means that the AQL defect rate will be accepted 95% of the time it is present. It really isn’t a very good protection method for today’s quality levels, but as I said, history gives it a pass.
Now here are the parts you won’t like:
The sample is taken after the lot is produced and is RANDOM. It isn’t from the top of the box or the closest upper corner of the pallet. (Those are known as ‘convenience samples’.). RANDOM means RANDOM. IF you fail the accept number (which is often 0 as it provides the smallest sample size) then you must inspect the entire lot.
After the lot is produced and Random also mean that you cannot simply spread the samples out across the lot as it is being produced. (The math relies on truly random samples form the finished lot). Personally I find that taking samples throughout the lot at predetermined levels provides far more information and protection but that is a different type of sampling plan and some auditors don’t understand it at all or accept it. It isn’t statistical in the traditionally accepted sense…
A better approach is to determine the Rejectable Quality Level (RQL). This is the minimum level of defects you want to detect and reject most of the time. Use the c=o tab in the spreadsheet. You will also have to determine what probability (aka Confidence) you want to have in detecting the RQL defect rate. It is the simplest plan. Typical probability levels (that won’t make your auditor’s eyebrows shoot up his forehead) are 95% and for critical characteristics it may 99%. Going below 95% is not recommended unless you are making inconsequential parts.
Now the caution I’m going to give is that categorical (pass/fail) data requires fairly large sample sizes.
A Good approach is to utilize Statistical Process Control (SPC) to ensure low defect rates as it can be used quite effectively with continuous data (measured data, actual values). Basically you sample the process at each known change point: first pieces, last pieces and every material, setting and operator change in between. A default for this in high volume manufacturing is every 2 hours. (There are way better ways to determine the frequency but that is beyond the scope of this post). The sample size for each sample point can be as small as 1 and maybe no more than 3-5. (If you are machining more than 1 part at a time the sample plan is different.). You start by plotting each value against the specification limits on a run chart; data in sequential order of manufacture…if you get a failed part you 100% sample back to the last known good measurement. SPC is way more complicated than this but it’s a start.
