Machine learning in smart factories and failure analysis

This thesis is related to the concept of “new world and new opportunities” that have been appearing due to the increasing power of virtual world. Since the improvement of computer technology, all type of data can be collected and moreover, it can be treated in many ways. Getting data, transforming them and solving problems are the main objectives with the irruption of this new world. If all of this could be done, some problems of the companies would be solved in a simple way and they could take advantage of it. On the other hand, solving problems is a tricky question, because each problem is different and it could accept more than one solution. Finding the best solution in a minimum lack of time is the key point to avoid losses in your company and it could be done with subsets of artificial intelligence. The most useful artificial intelligence in this context is machine learning. It allows to learn and improve the model that is built from an initial dataset and it could also recognize or predict some values thanks to training methods. Some years ago, predicting values and obtaining results was not as easy as nowadays. Since the barrage of machine learning, some business are trying to adapt its infrastructures to this new virtual concept in order to achieve better results than before and consequently, increasing their efficiency. This kind of business are called Smart factories because they combine physic world with robotics world. The main aim is to join forces and to get the best result they can reach. Machine learning, in terms of artificial intelligence, is based on getting data from some kinds of processes, then select the best algorithms to work with these data, train one attribute that will be the bottleneck of the process, and finally analyse the predictable results obtained. If the predictable results are similar or equal to the real values, then it means that all the models built are working well. Otherwise, if the results are not similar to the real values, it means that something is going wrong and the models need more training to recognize better the data and the attributes given. Furthermore, machine learning has two types of learning methods. The first one is supervised learning and the other one is unsupervised. While the first one needs some inputs to be able to produce an output, the second one learns by itself from the moment it receives data. Figure 1.1 shows the evolution of data minig and machine learning mentioned.Outgoin