Status:

CAP1 applies knowledge extraction techniques through the data collected. In this case, algorithms are being developed for data cleaning and identification and selection of relevant variables within the physical model of the plant. Besides, intelligent visualizations have been carried out. The mass balance of the plant can be visualized in the CAP platform, and the Thermal Balance is still in progress to accurate the model.

Status:

CAC1 algorithm has been developed based on a control algorithm where sensors and actuators that directly affect the drying process are used to calculate the optimum values for the different variables that run the drum, i.e., the burner and the rotary motors through the corresponding VFD’s. A dynamic modeling of the rotary drum has been developed using Deep Learning techniques. It is based on time series predictions using a long short-term memory (LSTM) Recurrent Neural Network (RNN) that has been trained using historical data. An LSTM network is a recurrent neural network (RNN) that processes input data by looping over time steps and updating the RNN state. The RNN state contains information remembered over all previous time steps. Thanks to this feature, subsequent values can be forecasted. At the same time, the control method to produce both, the optimum Burner Power Setpoint and the optimum Rotary Speed Setpoint is performed using using open loop forecasting methods with several time steps in the future. It finds the optimum value applying the fmincon algorithm which seeks for finding minimum of constrained nonlinear multivariable functions using as objective target the minimum error and the minimum energy consumption.

Status:

CAS1 is currently a laboratory prototype, almost in its final form.

Status:

CAS2, has been developed looking to two ways of work. The first, is to develop a new sensor, create new applications of sensors to obtain the flow of filler through the baghouse pipe. The second way of work is to use a commercial sensor that never have been use under these conditions of use. Both solutions are implemented in the use case, and are under the process of calibration and validation of results.

Status:

CAO1 solution is composed by two Deep Learning models

• The first model for predicting the health state of the baghouse is currently a laboratory prototype, almost in its final form. Currently, we are in the process of fine-tune and optimize the model.
• The second model for predicting the remaining useful life of a baghouse’s component is in progress.

Status:

CSS1 is done in two parts: the first connects the billet from the casting machine to the rolling mill by marking each billet with laser and later reading with visual cameras; the second connects the output from those billets, the bars, from the rolling mill to the finishing units, using laser marking and visual cameras too. In both cases, a close connection with the MES systems and Level 2 systems is necessary. The first part of the sensor is fully operative, whereas the second one has been operatively tested but faces some implementation difficulties due to a revamping in the rolling mill. However, a mitigation strategy has been developed based in one concrete type of orders.

Status:

In the CAPRI project, BFI develops a software-based solidification sensor for continuous casting of steel. Based on a 3-dimensional caster process simulation model and available process data, such as the initial temperature and steel grade in the tundish, the casting velocity, the cooling water flow rate and temperature increase in the mould as well as the spray water flow rates in the secondary cooling loops, the sensor calculates the temperature distribution and shell thickness in the casting strands. Upon cutting of a new billet, the sensor creates a digital twin of the semi-product. By feeding all the upstream data from secondary metallurgy and the casting process to the twin, the sensor provides a solid basis for the risk and anomalies detection, which will identify anomalous situations in the production early on and help the operator to make informed decisions about the further processing of the semi-products. The sensor has been calibrated by means of several temperature measurements on the strand surface at SIDENOR’s casting machine, where the sensor will be deployed in the CAPRI project.

Status:

The 2D- model is realized with FEM- analysis and uses Partial differential equations for the heat transfer. First step in modelling is to create the contours of the products. In Sidenor plant there are round profiles with diameters from 30 millimeters up to 100 millimeters. For each round contour, a FEM mesh is generated. The size of the mesh triangles have to be small enough to contain appropriate information, and big enough to guarantee a low calculation time. Additionally, boundary conditions, as material parameters are set. Second step of modelling is the creation of model calibration information. This concerns functions for the heat transfer coefficient. For later use and other environment conditions it is possible to calibrate this model input with real measured information.

Status:

The formation of scale on the surface of a steel product is a complex chemical reaction. It is a high temperature oxidation build-up of three kinds of iron oxid: FeO, Fe3O4 und Fe2O3, that is magnetite. The composition of the complex oxidation layer changes with the temperature. The formation of scale is high correlated to the steel composition. This information of the amount of scale is important for the following processes because scale can lead to quality issues. It is used in the CAPRI anomaly detector for the steel production. In this scale sensor the secondary scale is of interest. This scale grows after rolling, cooling by air on the cooling bed, in a temperature range from 1100°C down to 500° C. Below 500° C there is no increase of scale expected. There are two factors of influence for the scale: 1) The material composition of the steel, that is the steel grade and 2) the temperature, that is the cooling curve. The result of the scale model is a curve of the scale layer thickness over time as well as the final thickness, expected in the range [1, 100] µm.

Status:

The algorithm that we implemented for the anomaly detection so far is based on an autoencoder. Autoencoders are special types of neural networks, which are trained to simply reconstruct their input data. They consist of an encoder part and a decoder part, and the middle layer always has a lower number of neurons than the input layer. This implies that the autoencoder needs to learn a low-dimensional representation of the potentially high-dimensional input data, focusing on the features of most discriminative power and dropping strongly correlated features. The autoencoder may hence be thought of as a tool for dimensional reduction, like principal component analysis. There are at least two main approaches to using an autoencoder for anomaly detection. The first one is based on the reconstruction error: suppose we have a labelled dataset containing good cases only, without any outliers, and we train our autoencoder to learn a reduced representation for this dataset. When we then apply the autoencoder to an unknown good case, the expectation is that it will be able to reconstruct the data well. A bad case, on the other hand, should lead to an outlier in the latent representation, for which we do not expect the autoencoder to work very well, and consequently the reconstruction error should be large. We may then define a threshold in the reconstruction error for identifying anomalous input data. A second approach makes use of the latent representation to distinguish between good and bad cases. In this case we include the anomalous data or outliers in the model training and apply some learning algorithm to identify anomalies in the hidden layer. This may be done, for instance, by means of clustering, thresholds, or some regression algorithms.

Status:

For the steel use case DT‘s for billets and bars were realised. The billet DT is ‘born’ directly after the cut at the end of the casting machine. The billet DT is then equipped with information from steel works like the chemical analysis and casting, for example the casting speed. After the rolling of the billet the billet DT is the parent of several children, the bar DT’s. They inherit all the information from their parent and additionally information from the rolling process. The production of steel bars consists of several special conditions. The tracking is complicated due to the complexity of the optical detection of individuals e.g. in front of the rolling mill. (This is realised by the CSS1). Furthermore, the production of the bars may take weeks to months which makes it necessary to store the DT’s. For the persistence of the DT an object-oriented NoSQL database is used. This allows to store growing individuals containing different information located in their structure. New information can easily be added to this structure and stored again into the NoSQL database.