MOIRA – Uncertainty-aware diagnostics for operational condition monitoring

Reliability of diagnoses in AI-based condition monitoring systems

Experts at Fraunhofer LBF are investigating the reliability of diagnoses in multisensory, machine learning-based condition monitoring systems. The goal is to quantify uncertainties in diagnosis and thereby improve decision-making processes in machine and plant operations. A strategy was developed that systematically captures the uncertainties of probabilistic methods and various information flows and integrates them into the evaluation of results.

Use of vibration measurement data as time series for AI analyses

Vibration data from a small wind turbine under real and artificially generated fault conditions were used as the basis. A methodology for AI-based monitoring was developed and tested using real vibration measurement data. What sets this methodology apart is that it performs a reliability assessment of each individual processing step and passes the results on to the subsequent analysis steps. This enables the AI diagnostics to output classification results along with metrics indicating their reliability.

The dataset comprises over 6 hours of time series data from 28 sensor channels. In addition to normal operation, six damage conditions are recorded and validated.

Uncertainty quantification for reliable condition monitoring

At the heart of the methodology is the application of Bayesian convolutional neural networks to individual sensor channels. Uncertainties are decomposed into epistemic and aleatory components through uncertainty quantification. To fuse the sensor channels, the researchers used Bayesian model averaging to systematically combine the uncertainty metrics. The introduction of three-valued logic allows uncertain diagnoses to be assigned to a neutral class, thereby avoiding erroneous decisions.

10 percent improvement in diagnostic quality

The results show that the chosen fusion of information sources improves diagnostic quality by just under 10 percent compared to individual sensors. Uncertainty metrics assist experts in evaluating and further developing the system. The extensive dataset is available to researchers and offers potential for further applications.

The method enables more reliable condition monitoring and well-informed operational decisions.