(340t) Process Data Analytics and Representation Learning to Predict Arc Loss in an Electric Furnace

Introduction

Electric Arc Furnace (EAF) is widely used for steelmaking and refining particulate ores into base metals. In mineral processes, ore undergoes several upstream processing (e.g., screening, crushing, drying, calcining, etc.) which produce a fine particulate EAF feed that is dried, reduced, and heated. The preprocessing step is crucial to maximizing the efficiency and conversion of energy-intensive twin electrode Direct Current (DC) EAF. As shown in Figure 1, a twin hollow graphite electrode is connected to the DC electrical power supply, which is converted into thermal energy employing open plasma arcs that span from the graphite electrodes to the surface of the molten scrap metal (i.e., slag). The DC EAF comprises
a refractory-lined vessel with two tapping Launders (one for slag and one for metal). The feed enters the furnace from multiple ports along the roof, whereas the slag and alloy are tapped intermittently from Launders (Kotze, 2002). To maintain safe operating conditions, both the overhead basket and the side walls are water-jacketed while the bottom side is air-cooled (Hurd and Kollar, 1991). To maximize the production efficiency, decrease the variability in the final product, and increase profitability, the EAF process needs to be stable. Precisely, loss of the plasma arc due to unknown causes is a recurring and unresolved fault that significantly impacts the production rate and the furnace's electrical efficiency. In this work, we aim to build a predictive inferential sensor from a historical process data to avoid high-risk operating regimes. The soft sensor's objective is to provide operators with an alarm five to ten minutes in advance of a faulty event with a 75% or higher probability.

Data Labeling

The open plasma arc occurs when the electric current jumps from the graphite electrodes to the slag. The arcs provide the thermal energy required to maintain the slag and alloy at the desired operating temperatures. Once any of the arcs is lost, the separation process of base metals and slag from the particulate ore stops. Identifying the leading cause of an arc loss is an open problem with two potential mechanisms, i.e., electrical disturbances from the DC power supply and feed disturbances from the upstream metallurgical processes (e.g., distributors, reducers and calciners). Figure 2 demonstrates the three conditions of the measured power of a single electrode that need to be satisfied to constitute an arc loss in that electrode. Specifically, the power must drop with a magnitude of 10 MW within the last 36 seconds, then the power must go back to within ± 5 MW of the stable power value within approximately 10 minutes, and finally, the power must stay within a standard deviation of 2 MW for around 11.25 minutes before the power drop. These conditions are used to indicate if there is an event of an arc loss in an electrode at each timestamp or not.

Data preparation

The one-year worth of daily exports from a real industrial process historian was collected and used in this work. The data consists of over one hundred process variables that include numerical features representing some physical process properties such as arc length and categorical features that correspond to valve openings. Industrial applications of any statistical machine learning techniques are often riddled with problematic artifacts due to the imperfect and inconsistent nature of the raw industrial data. Hence, it is not surprising that over 70% of the total time required to carry out an industrial process data analytics on the EAF data was spent on data preprocessing. Data preprocessing lays the groundwork for process data analytics and machine learning; therefore, the quality of data preprocessing is directly proportional to the generalization performance of any supervised ML method. The raw data contains errors such as missing values, wrong inputs, and not a number (NaN) values. Also, the sampling rate is inconsistent among the process features, as illustrated in the left-hand side of Figure 3. Systematically removing the corrupted data and structuring the data (i.e., unifying the time stamp) are among the first stages of preprocessing. Next, any observation or measurement that is unusually different with respect to the other measurement contained in a dataset is an outlier (Atkinson & Hawkins, 1981). Another step taken while preprocessing the data was to detect and remove the outliers in the data to enhance the classifier's predictive performance. In this work, we used both process knowledge and a model-based outlier detection method. As shown in Figure 3.b, a probabilistic model was fit for each process variable measurement, and the values with low probability or outside the process variable limits were identified as outliers. The treatment of the identified outliers is application dependent, and, in our application, we replaced each outlier by its closest corresponding process variable limit.

Methodology and Results

After structuring and transforming the raw EAF data into a form that is amenable for statistical machine learning algorithms, two categories of binary classification methods were applied and investigated, i.e., i) traditional linear representation learning and ii) deep learning methods. The binary classifier will predict whether the data segment contains fifty-five minutes' worth of data occurs five minutes before an arc loss or five minutes before a long period of normal operation. Firstly, basic linear classifiers such as Logistic Regression (LR) and Linear-Support Vector Classifier (L-SVC) with L2 regularization were used to get the baseline accuracy on the EAF data. Traditional dimensionality reduction techniques such as principal component analysis (PCA) and partial least squares (PLS) have been implemented to enable systematic elimination of irrelevant features. The number of principal components used for PCA models was chosen based on the total variance explained of the original dataset. On the other hand, the PLS model was trained and validated over the total number of features in the dataset, and the number of PLS components chosen was on the validation accuracy. 34 principal components explain 90 % of the total variance, while the highest validation accuracy was obtained with 9 components for PLS. In the deep learning category, two techniques investigated in this work. The first deep learning technique is a deep, fully connected artificial neural network (ANN), while the second method relies on a deep convolutional neural network (CNN). The testing accuracy obtained using each method implemented in this study is illustrated in Figure 4. Please note that since most of the methods applied in this work require tuning a combination of multiple hyperparameters, Figure 4 reports the preliminary results obtained by performing a random grid search.

Conclusion

The use of supervised learning algorithms for process data analytics is an emerging research area that offers significant benefits to the process industry. This work is a small part of a more significant movement to migrate advanced data analytics techniques from statistics and computing sciences to process industries. Successful completion of this work will yield a predictive alarm that can improve EAF operation and increase production. Moreover, this work will explore and develop state of the art methods for predictive classification and dimensionality reduction. A comprehensive evaluation and comparison of the various techniques will be provided.

References

Atkinson, A., & Hawkins, D. (1981). Identification of Outliers. Biometrics, 37(4), 860.

Hurd, D. and Kollar, J. (1991). Data for operating single electrode dc furnaces.

Kotze, I. (2002). Pilot plant production of ferronickel from nickel oxide ores and dusts in a dc arc furnace. Minerals Engineering, 15(11):1017–1022.