Data driven study for optimization of industrial production

The fourth industrial revolution has already happened. In industry 4.0, factories and machines are instrumented with sensors able to acquire data along the production process and the produced product. Fault prediction is one of the solution provided by industry 4.0 to reduce wastes in terms of time and resources due to problems in the production line. Until now, in order to deal with production errors, a domain expert has been charged to analyze all the parameters and processes related to the production line to figure out what caused manufacturing defects. The goal of this thesis is to improve this process by reducing the time spent on identifying which are the factors responsible for failures and remove the bias introduced by the domain expert who is involved in the analysis. Technological progress has lead to the use of more sophisticated machines, which are now able to provide a lot of data about the operations they carry out. This allows to use machine learning algorithms to generate classification models which are capable of: deal with the great amount of data generated by the production line, predict and eventually anticipate the occurrence of manufacturing defects. This thesis analyzes the production line of a smart meter produced by BITRON. The first part of the work is related to the study of the three datasets with the aim of reconstructing the path related to each board (main board, radio frequency module and display) that compose the smart meter, gather all the relevant information related to each phase and clean the data. This first analysis allowed us to understand that there are two testing stages which have a great occurrence of errors: phase 45 and phase 170. Consequently, based on the type of classification performed (binary or multiclass) and the variable's type, some statistical tests are applied in order to select only the most important features. At this point, for each combination of module and test, we build different Catboost classifiers with different sets of features for both the binary ans multiclass problem. This allow us to investigate the predictive capability of the data. Each model is then used to predict defects on the produced boards along the production line. The result is that the different models cannot discriminate between defective and working boards in the binary case, nor to identify the type of error for the multiclass classification. In some cases, by checking the feature importance output of the model, some value of certain features, like for example batches, are more likely to generate defective modules. This could be due to the nature of the datasets that contain information about the traceability during the production line of the different boards which compose the smart meter. Thus those datasets contain high level characteristics about when a board crossed a specific production step without the information that characterize the operations performed at each step. In conclusion, it is not possible to predict manufacturing errors, for the given production line, with data related to traceability. For this reason, it may be beneficial to collect also data regarding the operations performed at each step, since those information are likely to be more correlated with faults in the production line.