Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Explainable AI applied to Malfanction Parameter Determination on Rotating Machines using Bayesian Neural Networks and Sobol Index
Submission Author:
Olympio Belli , SP
Co-Authors:
Olympio Belli, Helio Fiori de Castro
Presenter: Olympio Belli
doi://10.26678/ABCM.COBEM2023.COB2023-0943
Abstract
It is of great importance in mechanical systems to identify possible deviations from standard operation. In the rotating machinery branch, this issue is crucial due to its extensive use in various types of equipment. Not rarely, rotating components perform critical functions in mechanical systems that require high levels of reliability. These systems may be employed in environments and operating conditions that are vulnerable to failure due to fatigue, wear, corrosion, and assembly errors. As an emerging line of research in recent years, fault detection in rotating machinery is increasingly being carried out using artificial intelligence (AI) techniques fed by data from vibration, temperature, sound, and image sensors. These types of algorithms have been relatively successful due to their ability to handle a large set of input data and provide high accuracy in machine diagnostics. However, these statistical techniques still lack explainability and are often referred to as black boxes. Furthermore, these types of models, with a few exceptions, do not possess extrapolation capabilities and do not provide early warning to the user when the model is extrapolating and when the model is interpolating predictions between training data. To address this problem, this paper aims to combine the ability of Bayesian Neural Networks (BNNs) to express uncertainty with global sensitivity analysis using Sobol Index to determine which inputs are more determinant for the meta-models formulated. In this sense, BNNs were trained to regress malfunction parameters of a double rotor with a coupling. The regressed faults were: coupling misalignment angle and distance, and crack size. The training of the AI models was done using a data set generated by numerical simulations whose bearing displacement data were processed using the full spectrum transform. The input parameters of the finite element model were modeled with probabilistic distributions and sampled using Monte Carlo. After training, the models were evaluated in their uncertainty expression, verifying that the higher the uncertainty, the higher the error performed in the predictions. Finally, the Sobol index analysis was applied and validated, identifying the most influential inputs for model prediction.
Keywords
Explainable Artificial Intelligence, Bayesian neural networks, condition based maintenance, Rotor Dynamics, Sobol Index
