Session: 18-01-02: AI Implementation in Industry - I

Paper Number: 150506

150506 - Explainable Machine Learning for Fault Diagnosis and Prognostics 

Physical systems’ state of health degrades under external stressors such as mechanical, electrical, thermal and chemical loading fluctuations. Calendar aging and utilization of any physical systems also contributes to the degradation of their constitutive materials. Depending on the application in which a physical system is utilized, monitoring the degrading system is of paramount importance in order to prevent the unexpected failure which can cause downtime or critical loss. Therefore, diagnostics and prognostics systems have been introduced over the past decades to monitor and forecast the degradation in wide range of physical systems.

Predictive algorithms are fundamental component of the diagnostics and prognostics systems to ensure the reliability of complex engineering systems. In critical applications where equipment failure have significant consequences, the algorithm monitors degradation of the system and predicts the formation and propagation of defects or classify faults in a wide range of industries such as aerospace, automotive, energy, manufacturing and medical device. The process include acquisition of signals from sensors or generating synthetic simulations by means of physics based models. The data flows to a signal processing module where noise and interference is removed. Depending on the specific algorithm requirements the high-fidelity data may be resampled to a lower sampling rate or adjustments to a desired sampling rate might be applied to unify the sampling rate of variety of sensors. If data is missing, interpolation and data augmentation is applied to fill the signals with interpolated or augmented data. Data Normalization is also often applied to raw data in order to bring the different sensor data to the same dimensional unit.

Once signal preprocessing and processing step is completed, the data flows to a feature extraction module where variety of features are computed from the preprocessed signals. Such features are computed from time, frequency and time-frequency (time-scale) representations of the signals. Feature engineering techniques is also applied to create new feature from the signals. This step requires field knowledge of the designer as the feature are the key element of the algorithm. Recent techniques of AutoML and deep learning-based techniques, have been developed to avoid the feature engineering and feature extraction that requires domain expertise.  However, the successful applicability of such techniques to diagnostics and prognostics domain problems need further research and investigation. Once the feature space is formed, candidates of machine learning algorithms are chosen for the training and performance evaluation. The design process is an iterative procedure, and the designer may repeat the steps several times until satisfactory evaluation metrics are obtained. Having a trained and evaluated model in hand, the next step is the model deployment. There are several strategies for model deployment such as Batch deployment  real-time deployment , edge-device deployment, server-based and server-less deployments. The deployment method is determined depending on the application of the algorithm and the computational resources of the deploying entity. Once the model is deployed, it is regularly monitored to ensure it is continuously performed well over time. The performance monitoring includes detecting any changes or drifts in the distribution of the data and retraining the model on new data to ensure it continues to perform well, either on a scheduled basis or in response to drift detection. An ensemble of models may be also utilized to improve the overall performance of the predictive machine learning algorithm. Performance dashboards are also designed to monitor the performance of the predictive machine learning algorithm.A diagnostics or prognostics algorithm in real world applications contain all or some of the above steps among which feature extraction and reduction is of paramount importance.

The premise of this presentation is that, if the feature space is designed based on physics of the failure, can result in accurate diagnosis of faults or high and accuracies of predictions. To this aim a case study is presented on vibration fault isolation and classification of rotating machinery such as ball bearings.

Presenting Author: Ali Kahirdeh The-Precursor Diagnostics

Presenting Author Biography: Ali Kahirdeh, is the founder of The-Precursor Diagnostics, a company at the intersection of degradation science and artificial intelligence. He has worked as a principal data scientist, algorithm developer, and researcher in various industries, including utilities, electric vehicles, research, semiconductors, aerospace and academia. He completed his Post-doctoral, Ph.D., and M.Sc. studies in Mechanical Engineering and Electrical Engineering at the University of Maryland, College Park, and Louisiana State University, Baton Rouge.

Authors:

Ali Kahirdeh The-Precursor Diagnostics