To reduce the emissions, the plant process controllers should be tuned well enough so that the chemical reactions produce as much end product as possible under all circumstances with minimal emissions. It is a very complicated and time-consuming task to tune all the controllers and for human it is hard to analyze a great amount of data and find the right dependencies for individual parts of process function as a whole that always produce maximal efficiency. For that reason, the plant should consider taking advantage of machine learning and, especially, deep learning methods, to tune the controllers.
Before developing a deep learning algorithm for controller tuning, some relevant data must be collected from the plant processes. This can be done, for example, by installing sensors that measure different things from the controlled process and send those measurements into a database, for instance, every 10 seconds. Data points to collect into the database are all the values that are available, such as set values, control values and measurement values of the controllers and binary variables for the run type (automatic/manual) and the access type (near/remote), as examples.
When data have been collected for a long time, for example, two years, it is probable that this time period includes some abnormal operations like annual maintenances, replacements of catalysts or instruments, or some other actions that require switching off a production process. To simplify the future steps and ensure smooth data analysis process, these operations and their time ranges should be marked down.
After the data collection phase, an explorative data analysis stage begins. Some data are exported from the database to visualize the variables that it contains. The data visualization usually brings out some of the abnormal operations, but it is easier and more likely to recognize all of them if they have been marked down already during the data collection process. In this phase, all the anomalies and data errors are tried to be identified. If it can be seen that some errors or missing data points are caused by data exportation, the data may have to be exported from the database again.
The next step is data preprocessing. All the incorrect and exceptional data points that have been identified are eliminated, or corrected if possible. If some operation, like a replacement of some catalyst, makes data points to behave in a different way than before, it might cause problems and distort the results. In the worst-case scenario, all the data points before that operation may need to be removed. Because of this risk, it is important that there are enough data in the database. In addition to these tasks, the preprocessing step usually contains several mathematical and statistical operations like variable encoding, variable rescaling and creating new variables. The aim of data preprocessing is to convert the data into a form in which it can be analyzed.
The next step is building a deep neural network. Read more about that on the next blog post Deep Learning in Process Optimization, Part 2: Building a Deep Neural Network.