On the use of context information for an improved application of data-based algorithms in condition monitoring

The irruption of sensorized, connected and autonomous machinery has already marked a milestone in the manufacturing industry. This paradigm also known as Industry 4.0 consists in the digitalization of industrial processes and it is providing a new and previously underused source of knowledge: data.

The data streams that are massively being generated and stored are big sources of information. With the proper tools, the knowledge extracted from these data can enable monitoring and control, boost the decision making and customer satisfaction as well as the ef?ciency, the productivity, and the optimal use of facilities in manufacturing.

The process of knowledge extraction or data mining is carried out using algorithms that can handle large data volumes and are capable of extracting patterns that are hidden in the data. Once patterns are found, they can be captured in Machine Learning (ML) models to be later applied in a range of different ?elds with different purposes.

In the discipline of condition monitoring, the discipline that deals with the detection of the health status of the assets, the so-called data-based models have been around for a long time. Statistical methods as the quality control charts have been used to detect anomalies in operating machines since the last century.

Later on, more sophisticated classi?cation and regression algorithms such as arti?cial neural networks (ANNs) or support vector machines (SVMs) have been used to detect and predict failures with condition monitoring purposes. However, most of the works in the literature overlook one of the most important factors that are involved in the implementation of algorithms: the context. In short, we can say context is composed by the factors that have in?uence in the monitoring of a machine. Hence, a detailed understanding of the opportunities and limitations of the context of a particular application is needed to put algorithms in production. As they could limit the use of certain algorithms or could enable the use of some other algorithms that are not suited for the problem at a ?rst sight.

Most of the algorithms are designed with a certain context in mind, nevertheless, they tend to be applied without considering that the ?nal application context might change. In condition monitoring, this is the case of fault diagnosis algorithms that are trained and validated in test rigs, without on-site data that will enable retraining the algorithm in the real-life application. Or other algorithms developed in steady conditions that do not consider the real operation is varying over time.

This work discusses the role of the context in condition monitoring algorithms in three different situations dealing with the constraints and the chances that are related to each case of study. The applications discussed are wind turbine gearboxes operating under varying operating conditions; rotating machinery operating under steady conditions with a lack of knowledge regarding its degradation; and, an electromechanical actuator that has been diagnosed with the help of a physical model. The contexts are studied and afterwards, solutions are proposed to provide ad hoc designed algorithms that compose condition monitoring systems.

Although this thesis project deals with three special cases of study, the contexts that are addressed in this work can be found in many other condition monitoring problems, which makes the lessons obtained in this work be transferable to other real condition monitoring problems.