ML model

The task of the Machine Learning(ML) model is to predict as accurately as possible the remaining lifetime of the catalyst. To achieve this goal, both the sensor technology and the domain knowledge of the process expert had to be integrated into the ML model. Specifically, this meant converting the domain knowledge of the process expert into a function and using this function as the output variable on which to train the ML model.

The following points explain the structure and operation of the ML model:

Quantification of process knowledge

Selection of relevant measured variables and course

Description of the measured data

Architecture of the ML model

Train / validate the model

Estimation of the remaining life of the catalyst used depends on the factors described below:

1. The initial catalyst lifetime is 100 days.
2. The function follows a descending straight line without temperature changes.
3. The sudden increase of the reactor temperature, the so-called reactivation of the catalyst, reduces the lifetime of the catalyst.
4. Reducing the furnace temperature does not increase the life of the catalyst.

Reactivation means that the catalyst can convert more methane into carbon dioxide again. However, this is at the expense of the lifetime. The resulting function for the remaining lifetime of the catalyst investigated with three reactivation steps is shown in the figure on the left.

On the left, the course of the relevant measured data is shown. In it, the principle operation of the catalyst for the conversion of methane into carbon dioxide can be seen on the basis of the underlying chemical equation CH4 + O2 → CO2 + H2O. The catalyst itself loses its ability to convert CH4 to CO2 over time. This is shown by the fact that the CO2 content decreases over time, but at the same time the CH4 content increases. If the temperature in the furnace is increased, the catalyst regains the ability to convert more CH4 to CO2. When the temperature is decreased, the catalyst changes back to its old operating state. The H2O is a by-product in the reaction and therefore follows the behavior of the CO2.

For the ML method, learning the correlation between measured variables and remaining lifetime is possible. Individual plant conditions can be learned on the basis of the temperature. For example, that the decreasing straight line at constant temperature follows the behavior of the measured variables CO2 and H2O.

As so often in the development of ML models in the production area, a relatively simple structure of the learning procedure is sufficient. As a preprocessing step, a Zscore standardization is applied. In total, the artificial neural network has only one hidden layer with four nodes. As an activation function, the ReLU function provided the best results.

At the Fraunhofer UMSICHT bench, tests were carried out with several catalysts in order to have a sufficiently large database for training and validating the ML model. After each experiment, the process expert described the course of the remaining life of the catalyst and translated it into a function. For the model, this function ultimately served as the output variable to which training was performed. During operation (see section on system architecture), the model was finally used to continuously predict the remaining lifetime of the catalyst.

To get an impression of the quality of the ML model, both the course of the expert knowledge and the predicted course of the ML model are shown in the figure on the left. Jumps in the remaining lifetime mean, as already mentioned, that a reactivation of the catalyst has taken place. In general, it can be seen that the model follows the course of the expert knowledge very well. Only in the areas where the catalyst changes its state (i.e. reactivation occurs due to an increase in temperature) and the residual lifetime is reduced abruptly does the model deliver greater deviations from the expert knowledge.