Predictive Maintenance of an External Gear Pump using Machine Learning Algorithms

The importance of Predictive Maintenance is critical for engineering industries, such as manufacturing, aerospace and energy. Unexpected failures cause unpredictable downtime, which can be disruptive and high costs due to reduced productivity. This forces industries to ensure the reliability of their equip-ment. In order to increase the reliability of equipment, maintenance actions, such as repairs, replacements, equipment updates, and corrective actions are employed. These actions affect the flexibility, quality of operation and manu-facturing time. It is therefore essential to plan maintenance before failure occurs.Traditional maintenance techniques rely on checks conducted routinely based on running hours of the machine. The drawback of this approach is that maintenance is sometimes performed before it is required. Therefore, conducting maintenance based on the actual condition of the equipment is the optimal solu-tion. This requires collecting real-time data on the condition of the equipment, using sensors (to detect events and send information to computer processor).Predictive Maintenance uses these types of techniques or analytics to inform about the current, and future state of the equipment. In the last decade, with the introduction of the Internet of Things (IoT), Machine Learning (ML), cloud computing and Big Data Analytics, manufacturing industry has moved forward towards implementing Predictive Maintenance, resulting in increased uptime and quality control, optimisation of maintenance routes, improved worker safety and greater productivity.The present thesis describes a novel computational strategy of Predictive Maintenance (fault diagnosis and fault prognosis) with ML and Deep Learning applications for an FG304 series external gear pump, also known as a domino pump. In the absence of a comprehensive set of experimental data, synthetic data generation techniques are implemented for Predictive Maintenance by perturbing the frequency content of time series generated using High-Fidelity computational techniques. In addition, various types of feature extraction methods considered to extract most discriminatory informations from the data. For fault diagnosis, three types of ML classification algorithms are employed, namely Multilayer Perceptron (MLP), Support Vector Machine (SVM) and Naive Bayes (NB) algorithms. For prognosis, ML regression algorithms, such as MLP and SVM, are utilised. Although significant work has been reported by previous authors, it remains difficult to optimise the choice of hyper-parameters (important parameters whose value is used to control the learning process) for each specific ML algorithm. For instance, the type of SVM kernel function or the selection of the MLP activation function and the optimum number of hidden layers (and neurons).It is widely understood that the reliability of ML algorithms is strongly depen-dent upon the existence of a sufficiently large quantity of high-quality training data. In the present thesis, due to the unavailability of experimental data, a novel high-fidelity in-silico dataset is generated via a Computational Fluid Dynamic (CFD) model, which has been used for the training of the underlying ML metamodel. In addition, a large number of scenarios are recreated, ranging from healthy to faulty ones (e.g. clogging, radial gap variations, axial gap variations, viscosity variations, speed variations). Furthermore, the high-fidelity dataset is re-enacted by using degradation functions to predict the remaining useful life (fault prognosis) of an external gear pump.The thesis explores and compares the performance of MLP, SVM and NB algo-rithms for fault diagnosis and MLP and SVM for fault prognosis. In order to enable fast training and reliable testing of the MLP algorithm, some predefined network architectures, like 2n neurons per hidden layer, are used to speed up the identification of the precise number of neurons (shown to be useful when the sample data set is sufficiently large). Finally, a series of benchmark tests are presented, enabling to conclude that for fault diagnosis, the use of wavelet features and a MLP algorithm can provide the best accuracy, and the MLP al-gorithm provides the best prediction results for fault prognosis. In addition, benchmark examples are simulated to demonstrate the mesh convergence for the CFD model whereas, quantification analysis and noise influence on training data are performed for ML algorithms