Quality Improvement Using Advanced Process Control
Background
This food producer assures quality by periodically taking samples at the end of production and sending them to their in-house laboratory for analysis. The results are compiled to give an overall quality score and when plotted show a typical normal distribution.
Like most businesses, they wish to improve their quality but also equally important, to reduce the variability so that the consumer is always presented with a consistent product.
Earlier in-house attempts to identify the reasons for this variability had been unsuccessful. Their automation system included a process historian which had a significant history. Thus they were rich with historical data but poor in terms of data insight.
The Solution
From the earlier work done by the in-house team, there was a list of input parameters, known as features in data science language, which may impact quality. A neural network was used to map these inputs to the resulting quality. This class of algorithms learns from historical data, which is known as training. Thus, historical data must be cleaned, with incomplete and outlier data removed, before training can commence. Once training is complete, the network is tested using out of sample data to show that the network generalises well and has not just learnt the data by rote. The trained network can now be used to perform sensitivity analysis. Effectively asking the question, which of the inputs are major contributors to quality. This showed that the two major contributors were zone 1 temperature, followed by the prover water percentage. 
We now have a robust model, and we know which input parameters have a major impact on quality. The next question is how do we use this insight?
Often the initial stage is to deploy the model in an advisory capacity. In this case, the model is used to predict the quality. This is then compared with the results from the laboratory. Good agreement between the lab and predicted values brings confidence, which allows us to move to the next stage of the project.
We now start to consider what benefits an advanced process control (APC) algorithm might deliver by using an offline simulator. Previously it has been determined from the sensitivity analysis that the largest single contributor to quality is the zone 4 temperature. The initial question that we are trying to answer is, if we are only allowed to control this one input, what difference will it make to the final quality?
The question is answered by the two graphs. The first one shows what the results are with no optimisation, whilst the second one shows the results when optimising just the zone 4 temperature. It can be observed that the average quality value has been increased modestly from 71.8 to 72.18. However, the variability has been dramatically reduced.
What would happen if we optimised the process further. The next graph shows the results of optimising the two major contributors, zone 4 temperature and percentage water. Again we see modest increases in the quality value and a further reduction in variability.
The final graph shows the expected benefit of optimising all available adjustable inputs. In the original model, two input parameters, ambient temperature and humidity, are inputs that are thought to be relevant and therefore measured but are not considered to be adjustable. Nevertheless, by limiting ourselves to those inputs that are considered to be adjustable, we have been able to improve the quality value from 71.8 in the un-optimised state to 72.6 and almost eliminated variability.