The increasing complexity of production systems triggers the need for condition monitoring (CM) and condition-based maintenance (CBM) techniques adaptable to the requirements of various existing contexts. This research project is integrated with a bigger one, aimed at optimizing the application of CBM in industrial settings that operate under time-varying conditions. This expression refers to processes that have different standards over time. These processes in today's industrial scenario are widespread: multiphase processes, where the same machine carries out different operations without evidence of the shift between the phases; characteristics of the processed product not constant; an unfixed quantity of processed product during different production cycles. In this context, normal intrinsic changes in the production process can be mistaken for fault situations, causing several false alarms. Existing CM and CBM techniques are often limited to fixed operating conditions, making their implementation very challenging in such real settings. Besides, companies are increasingly focusing on the prognosis of their machines. The main aim is to obtain a sufficiently large forecasting period of the onset of a fault to allow the execution of maintenance tasks to prevent it from occurring. This research aims to study the application of multivariate statistical techniques, with attention to canonical variate analysis (CVA), mixing them with machine learning algorithms for fault detection, diagnosis, and prognosis in time-varying conditions processes. The role of CVA will be twofold: input for the diagnosis and prognosis phase of the machinery, based on information extracted in an initial training phase for fault detection; constant monitoring of the process for fault detection to make up for any inaccuracies in the prognosis model. A methodology will be structured to provide for the overall management of all areas of CBM for processes that operate under time-varying conditions.
As seen in the previous paragraph, where literature on the subject is presented, processes that operate under time-varying conditions are a highly debated topic, as well as an extremely relevant issue for current industrial settings. However, despite their importance on the world industrial scenario, many CM and CBM techniques are still not able to be applied in this context. Based on this, the innovations to be pursued are many. First of all, to guarantee in processes that operate under time-varying conditions the same maintenance performance currently achievable in more static and linear contexts. The proposed research will, therefore, be concerned with adapting or rather verifying the applicability and any necessary modifications or application methodologies of fault detection, diagnosis, and prognosis techniques currently performing in other contexts concerning the processes that operate under time-varying conditions. To do this, it is intended to analyze, as previously mentioned, a technique already tested in simpler contexts, that is CVA, to assess how to apply it effectively in the most variable operational contexts. Another point of interest is the integration in the same methodology of fault detection, diagnosis, and prognosis of machinery. Consequently, an improvement in the management of faulty variables, extracting information from the fault detection analysis to circumscribe the variables to be considered to carry out diagnostic and prognostic analyses is another of the objectives to be pursued. To do so, techniques currently used in other realities will be analyzed to check their performance in the context under consideration and, if necessary, adapt or modify them in this respect. Finally, another point of innovation to be explored is the use of CVA together with machine learning algorithms to implement the prediction of the onset of failures. In this way, the aim is to achieve a sufficiently wide forecasting period to be able to apply maintenance interventions aimed at avoiding the onset of the fault and the machine downtime. Considering the points just mentioned, the purpose is to combine them to structure a methodology able to perform fault detection, diagnosis, and prognosis at the same time. The objective is to provide prognosis analyses with a level of accuracy such that the process monitoring part of the fault detection process can be considered only an additional element to compensate for any inaccuracies in the prognosis model. In this scenario, the role of the CVA will give, as already said, twofold. Firstly, CVA will be the input for the prognostic and diagnostic analyses, based on the information extracted implementing fault detection in the training phase. Secondly, CVA will be applied for control monitoring in real-time to identify any unforeseen faults not predicted by the prognosis model. The overall novelty can, therefore, be examined from two points of view. The first is that it will increase the ability to maintain more realistic systems, such as the systems that operate under time-varying conditions. The second is for the methodological approach. In fact, a methodology will be structured in which the center is the prognostic analysis, while the fault detection becomes the step through which to identify unexpected failures not identified by the prognosis model. The objective is to move more and more towards the anticipation of the onset of the fault rather than its rapid identification. Finally, the techniques that will be analyzed are innovative in this context. CVA currently needs more study regarding processes operating under time-varying conditions and machine learning for predicting the useful life and time to failure of a machine is still an open challenge.