Welcome to the Glass Age

120 Much more data is available than in the past. But, how to handle it? Near Infrared (NIR) furnace cameras can act as virtual thermocouples and see the temperature trends within the furnace over time (Figure 7.8). Indeed even temperature profiles can be explored. In the following, the capabilities of such cameras are considered first and the question of data accuracy later. First though we consider what artificial intelligence (AI) and its neural networks are? How can they help the glass industry? Most glass operators are familiar with DCS, a digital control system for a process or plant usually with many control loops and MPC. Before AI, irregular issues evaded their operators creating inefficiencies and low-quality glass production. AI anticipates and performs the tasks that layer of the analysis, to be formalized and then inserted into the neural network. We then teach this neural network to fill certain highs and constants inside different neurons, to learn (with lots of data on the input side) to predict what is produced as output —and to recognize it automatically. The key thing is that we don’t fully understand these neurons, and we don’t have to understand them. They are simply filled out by giving sufficient data and sufficient output for the neurons that are going to be filled with the mechanism that they recognize best. Figure 7.9 shows the data input, the data analysis and the process output. To illustrate these concepts, let’s look at an imaging technique which we use with an NIR furnace camera. The camera software is trained to recognise the images it sees, and after time can differentiate between batch, flame, glass surface, refractory, and camera build-up. So if buildup around a camera covering a thermocouple occurs, it can no longer be used reliably. Then, input data from this thermocouple should not be applied could not previously be resolved by hands-on techniques. It allows the computer to mimic human intelligence to solve a problem, using neural network decision trees trained by machine learning. Deep learning may appear magical, but is simply a multi-layered deep neural network that handles vast amounts of information. Actually, a daily search for something on Google uses the same technology. Google suggests an answer to what you are really searching for. So, this is already AI. What is a neural network? It was probably named after neurons in the human body which have similar characteristics. A data set needs first to be analyzed, and after analysis, the result is the outer layer which is its meaning. So first this data was born into the inner Figure 7.9. Neural networks, with deeper hidden layers. Source: Glass Service, a.s.