Product Yield
Background
The manufacturer produces a discrete product. Unexplained variations in the production process result in a variable final product. This variation is represented by a composite quality metric which, as shown, approximates a normal type distribution.
This composite quality metric is then used together with their quality specification to classify the final product. This results in the product being classified as either Grade A, Grade B, Grade C etc. In extreme cases the product may be designated as waste.
Higher graded products attract a premium price and therefore the business wishes to maximise the yield of grade A products whilst minimising waste. Previous attempts by their in-house Process Improvement Team had been unsuccessful.
Thus the business decided to turn to Data Science to provide the necessary insight.
The Solution
Although previous attempts by the Process Improvement Team had been unsuccessful it had yielded a list of factors, often known as features in data science, which may contribute to product quality. However, their contribution, if any, to each feature is unknown. Fortunately, the customer has a significant amount of history which records the values of the various features and the resulting quality. Typically such data must be cleaned before use to address issues such as:
1. Incomplete records.
2. Start-up and shutdown conditions which may be described as atypical.
3. Outliers that may be attributed to faulty instrumentation. 
Essentially, after cleaning, all that is required is to extract the insight from the data to determine those features which contribute to quality, and how to convert this insight into a competitive advantage.
In this example, the results of the analysis were presented as a decision tree and for discussion purposes, a simplified version is shown. Starting at the top of the tree it shows that only 43.8% of all product is grade 1. However, if you follow the red path to get to the bottom branch it can be shown that the tree branches at both forming pressure and material supplier only. Thus although initially it was thought that ten features may be significant, the analysis suggests that the two dominant features are pressure and the raw material supplier.
Furthermore, the decision tree indicates that if the pressure is kept within the range 367.0 to 370.0 and the raw material supplier is limited to suppliers 2 and 3, then the yield of grade 1 product will rise from 43.8% to 100%.
In this particular case, the deployment strategy was trivial, no advanced control algorithms were required. It was only necessary to ensure that all raw materials came from suppliers 2 and 3, and the forming pressure was carefully controlled between the two limits specified above.