Analysis of batch asymmetry and batch line position for the decision support in the glass melting process

Glass furnaces are a crucial part of the production line in every glass factory. The operating parameters of furnaces have a decisive influence on the efficiency of the process and pollution (especially NO_x emission), and on the quality of the final product. Therefore, its correct operation is extremely important. To a large extent, the parameters of the glass are estimated by the operator on the basis of the image from a camera installed inside the furnace. In particular, the operator should ensure the optimal positioning of the batch line (i.e. the distance to which the unmelted glass pieces reach) and similar distribution of the unmelted glass on both sides of the furnace. If the batch line moves too close to the working end of the furnace, the glass quality dramatically decreases, while the batch line too close to the front of the tank results in high emissions of NO_x, and CO₂ [1] and unnecessary energy consumption, which is a significant part of production costs [2].

The subject of the research is a glass furnace with a capacity of approx. 150 tons a day. A schematic drawing of the furnace is shown in Fig. 1. It is a 6 × 10 m furnace with one hopper located on the left side.¹ Two gas burners (left and right) are used to melt the glass in the furnace, working alternately for 30 min. The process of changing the working burner (called the reversal) lasts one minute. During this time, neither of the burners are working. Every 30 min it is therefore possible to obtain an image of the glass pane inside the furnace, not obscured by the flame of a burner. This determines the maximum frequency at which the camera on the wall of the furnace opposite the burners can analyse the image. Two sets of electric electrodes are used to heat the glass from the bottom—one in the melting part of the furnace and the other at the break wall.

Computer vision systems are becoming common elements of production lines. They are used for the inspection and control of industrial processes, including glass production. Various vision systems have been developed for monitoring different glass manufacturing stages, from the hot end of the production line, for example vision systems for gob inspection [7, 8] to the inspection of semi-products and the final product [9, 10]. Vision systems have also been applied for the inspection of furnace structure [11] or vision-based analysis of flames [12].

In the literature, there is also some research on computer vision for the automatic detection of batches in glass furnaces [13, 14, 19] the distribution of which is crucial for the correct control of the furnace. Vision systems can be used not only for full automation of the process, but also to help the operator to make correct decisions by making suggestions. They may also help by visually presenting the furnace interior, as in [15], where the image from a single furnace camera is used to create an enhanced 3D image of the furnace, which can be viewed from a virtual camera located at an arbitrary point in the furnace or using stereoscopic devices.

This study is a continuation of our previous research on the vision system supervising the operation of a glass furnace [3, 4] and control systems in the glass industry [5]. In the above-mentioned works, we proposed an image processing method for automatic analysis of the symmetry of the glass melting process based on a series of images captured by a camera inside a glass furnace. We recommended a set of asymmetry indicators that are not influenced by asymmetries that result from glass melting technology (i.e. consecutive heating of first one, then the other side of the furnace). As a result of this research we proposed rule-based system which makes inferences about the asymmetry of chargers and burners from the distribution of the batch on the glass surface and the distribution of correlated colour temperature. The proposed method indicates the reasons for any asymmetry, for example whether heating of the left and right hand sides of the furnace is unbalanced or whether chargers are not working properly.

The above works resulted in the SCADA-like industrial vision system for glass furnaces supervision, which is briefly described in [6]. The system was implemented in the glassworks and the results of its operation are presented in the following chapters. In particular, we focused on analysis of historical data from the period of over a month of operation of the furnace, concerning the asymmetry of the batch and the position of the batch line. Based on the data, selected statistics of furnace operation were calculated. An attempt was also made to calculate the position of the batch line using regression and neural models, based on historical process data from the SCADA system.

The main contribution of this research is thorough analysis of the furnace process variables and indicators provided by the vision system over a long time. These studies gave deeper insight into the melting process and showed that it is possible to model certain aspects of the furnace operation by the use of machine learning technics based on historical data. Analysis performed here is, according to our knowledge, the first study of this type based on real data obtained during the glass melting process.

The research presented in this paper is based on the vision system created by the authors, whose prototype version was described in [3]. It was initially used for the extraction of relevant glass melting parameters based on pairwise comparisons [4]. The system was later developed [6], and the current version is equipped with tools for remote connection to the glassworks, making it possible to collect and process data remotely in real time. It also includes tools for statistical analysis of data, such as statistics describing the variability of the distribution of the batch over time, including symmetry analysis. The interface of the system is presented in Fig. 2.

Our vision system calculates the distribution of the batch at each reversal, based on an image from a camera placed in the furnace. Then, on the basis of the distribution, the average asymmetry value of the batch is calculated based on the following procedure:

Since the nature of the asymmetry differs significantly depending on which burner preceded the given reversal (which will be further elaborated in the next chapter), the trend is drawn separately for the left burner (more precisely—for the series of every second reversal preceded by operation of the left burner) and for the right burner.

Additionally, on the basis of the batch coverage, the location of the batch line is calculated. The location in this context is the position of the farthest (from the burner wall), unmelted part of the batch (yellow line in the image of the glass pane). Figure 3 shows different examples of recognising this line. The first three examples are correct, and the last is incorrect. In practice, incorrect identification is rare. It is usually caused by misclassifying stubble on the camera as a batch. During the analysed period, incorrect line recognition occurred 47 times out of 1678 reversals, which is 2.8% of all cases.

For correct operation, the vision system needs information from the furnace control system, such as the beginning of the reversal (initiates analysis of the camera image), time to the next reversal (for information purposes) and information about which burner is currently operating (to divide the asymmetry into the left and right burner). This information is obtained from the SCADA system using the OPC protocol (for security reasons, the system is not connected directly to the PLC).

Data from 35 days of continuous operation of the installation (about 847 h) were taken for analysis. The data cover 1678 reversals.

We used machine learning methods to create a number of models predicting the position of the batch line on the basis of selected process variables. The best of the several dozen tested models turned out to be the regression model, calculated using the GPR method (Gaussian Process Regression—a non-parametric kernel-based probabilistic model) [18]. The input variables are listed in Table 5.

The mean square error of the model fit to the actual data (calculated using the five-fold cross-validation method) is 38.6 (RMSE), and the fit factor R² is 0.56. These results show that the model is not very accurate.

Since the classical methods did not give satisfactory results, we used artificial neural networks. We arbitrarily assumed a shallow network with three hidden layers (see Fig. 11). The inputs and outputs were the same as for the regression model. The data were randomly divided into a training set (70%), used to train the network, a validation set (15%), used to assess the effectiveness of the learning process, and a test set (15%), on which the final accuracy of the model was verified.

Since the learning method and the number of neurons in individual hidden layers are hyperparameters that cannot be determined optimally in advance, a simple method was used to search the space of parameters. For each of 11 back propagation-based learning methods (such as BFGS quasi-Newton, Conjugate gradient, Gradient descent with adaptive learning rate, Gradient descent with momentum, Levenberg–Marquardt, and Bayesian regularisation), all combinations of architectures with three hidden layers, where the number of neurons in each layer varied from 10 to 50 with a step of 10, were created and trained. As a result, 1375 networks were obtained, the best of which was the network size [h1 = 30, h2 = 20, h3 = 30] and the learning method based on back-propagation with Bayesian regularisation [19].

When analysing the results shown in Fig. 12, it can be seen that the model follows the trend in the data, although it does not reflect all the details. The histogram of the model residuals (the differences between the model and the actual data) and the auto-correlation waveform of the residuals indicates that the residuals are practically uncorrelated white noise. This means that the residuals do not contain any useful information, so the model is adequate. The mean square error of fitting the model output to the real data is 31.3, and the fit coefficient R² is 0.84. The results are much better than for the regression model. Such a model could be used to predict the position of the batch line depending on other process variables in the glass furnace.

The information obtained from our vision system allowed observation of phenomena occurring in glass furnaces, which were previously difficult to notice. Historical data on batch asymmetry show that the furnace distributes the batch asymmetrically—10% more batch is accumulated on the right side of the furnace. Moreover, data analysis showed that the asymmetry was greatly influenced by which burner was running before the reversal, and that asymmetry was much greater after operation of the left burner. The large asymmetry of the batch in the furnace is disadvantageous, as it makes it difficult to melt the batch evenly and may lead to the penetration of the unmelted batch into the working end.

Historical data of the batch line position allows an assessment of whether the operator is controlling the process optimally. Typically, operators tend to shorten the batch line unnecessarily to avoid the risk of unmelted glass entering the working end. However, doing so leads to much higher energy consumption and process inefficiency.

Summarising this research, it can be stated that:

In future research, additional process variables logged in SCADA can be included in analysis. These variables may include gas calorific value or the level of a glass cullet in the batch (as it acts as a fluxing agent and decreases the melting energy). Perhaps it would make sense to build a separate model for each burner, because, as research has shown, the process is slightly different depending on which burner is working.

From the point of view of the process operator, running the process as efficiently as possible is important. Therefore, operators should observe the following recommendations: