Predicting the failure of any structure is a difficult task in a mechanical system. However complicated and difficult the prediction might be, the first step is to know the actual condition of the system. Given the complexity of any machine tool, where a number of subsystems of electro-mechanical structures interact to perform the machining operation, failure diagnostics become more challenging due to the high demand for performance and reliability. In a production environment, this results in maintenance costs that the management always strives to reduce. Condition-based machine maintenance (CBM) is considered to be the maintenance strategy that can lead to failure prediction and reducing the maintenance cost by knowing the actual condition of the asset and planning the maintenance activities in advance.

Grinding machines and grinding processes have come a long way since the inception of the centuries old grinding technique. However, we still have a number of challenges to overcome before a completely monitored and controlled machine and process can be claimed. One such challenge is to achieve a machine level CBM and predictive maintenance (PdM) setup which is addressed in this thesis. A CBM implementation framework has been proposed which combines the information sampled from sensors installed for the purpose of the process as well as condition monitoring. Accessing the machine's controller information allows the data to be processed with respect to different machine states and process stages. The successful implementation is achieved through a real-time and synchronized data acquisition setup that allows data from multiple sources to be acquired, stored, and consolidated. The dataset thus generated is used in a significant part of this project and is also published in Swedish National Data Service (SND).

The thesis also presents the failure diagnostic model based on two step classification approach using benchmarked random forest models. The binary classifier predicts if there is a fault present in the machine based on crucial sensors data from the Idle segment of the grinding cycle. Multi-class random forest classifier diagnosis the fault condition. PdM, knowing when to trigger maintenance action, is achieved through predicting the overall quality of the produced parts from the feature set extracted from sensor data of the Spark-out segment of the grinding cycle. Combining fault diagnosis with the predicted quality information resulted in reliable and actionable maintenance decisions for the bearing ring grinder. The demonstrated setup, based on a production bearing ring grinder, is adaptable to similar machines in production.