VIBRATION-BASED METHODOLOGY FOR ONLINE HEALTH MONITORING OF ROTATING MACHINES

Vibration-based condition monitoring represents the most efficient technology for early prediction and detection of failures in rotating machines. Faults can be detected by extracting typical features of vibration signature and comparing them to known thresholds of acceptable behaviour. Defining ap-propriate limit thresholds independently of the operating conditions, in order to perform a real time monitoring of any faulty state in system operation, is often a task not easy to achieve. The paper aims at presenting in this sense an effective condition monitoring technique for rotating machines, relying on a black box modelling approach of system dynamic behaviour. The powerful capabilities of the methodology are highlighted by implementing the model for a typical problem of rotating machines fault diagnosis. The proposed approach involves first identifying a Nonlinear ARX model, trained using the data from the healthy (nominal) operation of the machine. The model is then used for sim-ulation of system known dynamics, to compute residuals by subtracting the model-produced outputs from the corresponding measured signals. Through an accurate monitoring of the properties of resid-uals, such as their mean, variance and root mean square, the method is able to successfully distinguish normal and faulty operations as well as to properly rank fault severity. The high mode-discrimination power of each considered residuals feature demonstrates the robustness of the technique and its at-tractiveness to face with rotating machines health monitoring problems.