Automated Random Sampling Control of Data Sets

Due to the increasing number of articles to be classified in customs tariffs, the topics around artificial intelligence and machine learning are becoming more and more important. It is hoped that the automation of processes in the area of goods classification can reduce manual effort and save resources. But how is the quality of the classification of goods ensured, and who actually determines what “good quality” means?
The tried and tested four-eyes principle guarantees high quality, but often no longer meets today’s demands. For example, if you decide to add another product to your range, you will soon realise that your resources will quickly reach their limits. It takes time to classify and check goods correctly.
One efficient solution for quality control is an automated, random check of the goods classification.

In concrete terms, random sampling means that a certain number of classifications are checked again manually, which allows conclusions to be drawn about the quality of the entire group of articles. But how is this sample selected, and who sets the standard for quality? Manual selection of the sample makes no sense, as the objectivity of the articles to be selected is called into question with this approach.

What is needed, therefore, is an objective, transparent and comprehensible method. We therefore recommend following the sampling procedure according to DIN ISO 2859-1 AQL (Acceptance Quality Level). These standards provide orientation and represent a common objective measure.

The ISO2859-1 AQL standard is based on an assumption sampling test. This means that a random sample is taken on the basis of a total quantity, i.e. a “batch”. In practice, this means, for example, 2,500 articles from one quarter.

The AQL process takes place in four steps:

1. Set Inspection Level

First you have to determine how strict the sampling will be interpreted. The standard provides for a total of seven different levels. There are three standard levels and four special levels. Level III in the standard is very strict and is used, for example, for braking systems in cars.

2. Set acceptance quality level

Then the acceptance quality level must be determined. The AQL tables specify how many items of the sample may be “defective” such that the sample is still accepted and the entire batch may therefore be awarded an agreed quality level.

3. Select sampling plan

Then the sampling plan is selected. You can choose between a Simple, Double or Multiple Sampling Plan. With the Simple Plan, you get the largest tested sample. In a Double Sampling Plan, you first get a medium sample. If the result is in the uncertainty range, a further sample is created, which is then tested.

4. Determine the size of the sample

The last step is to determine the sample size. A sample is defined here according to a fixed plan, which is then checked in terms of content according to common guidelines.

If you have 30,000 items and want to apply Normal Inspection Level II and a maximum error probability of 2.5% (AQL) in the data sets, then the sample is 315 items (Simple sampling). The number 315 is derived from the DIN standard, based on the error probability and batch size. Of these 315 articles, a maximum of 22 may be classified “incorrectly”. If the number of defects is higher than this, the batch is rejected and the specified quality level is not reached.

Now, if there are no rejections of the sample when the quality of the sample checks is continuously good, the standard provides that the sample sizes also become smaller. In the event of defects, however, it is the other way round: Then the sample sizes are expanded again.

It is important to ensure a uniform policy on when data sets are judged to be “false”. The people who made the original determination of the master data are also not allowed to evaluate the sample.  At this point, an IT-supported implementation would also be advisable, as the standard entails some complexity.

Since the entire standard is based on statistical procedures and therefore follows its own logic, the plans and procedures of the standard can also be mapped in a system.

A typical test might go like this: The data sets are uploaded to the system with the item master data. A Duplicate check during import into the system ensures that only data from the respective quarter enters the “batch”. The system then randomly selects the appropriate sample based on batch size and inspection level.

These items are then manually checked in the “blind classification” process. Blind classification means that the previous classification result is not displayed in order to be able to classify the goods again objectively. Subsequently, you receive a report on the inspection carried out and therefore have an objective view of the quality.

Quality testing according to AQL opens up many possibilities. It can in principle be applied to many processes. The great advantage of this is that it allows both the internally or externally supplied or determined data to be checked and an objective and standardised assessment of the quality to be made.  The automated results can also be thoroughly checked again.