Hardware Failure Prediction on Imbalanced Times Series Data

Abstract

Magnetic resonance imaging (MRI) systems and their continuous, failure-free operation is crucial for high-quality diagnostics and seamless workflows. One important hardware component is coils as they detect the magnetic signal. Before every MRI scan, several image features are captured which represent the used coil’s condition. These image features recorded over time are used to train machine learning models for classification of coils into normal and broken coils for faster and easier maintenance. The state-of-the-art techniques for classification of time series involve different kinds of neural networks. We leveraged sequential data and trained three models, long short-term memory (LSTM), fully convolutional network (FCN), and the combination of those called LSTMFCN as reported by Karim et al. (IEEE access 6:1662–1669, 2017). We found LSTMFCN to combine the benefits of LSTM and FCN. Thus, we achieved the highest F1-score of 87.45% and the highest accuracy of 99.35% using LSTMFCN. Furthermore, we tackled the high data imbalance of only 2.1% data collected from broken coils by training a Gaussian process (GP) regressor and adding predicted sequences as artificial samples to our broken labelled data. Adding 40 synthetic samples increased the classification results of LSTMFCN to an F1-score of 92.30% and accuracy of 99.83%. Thus, MRI head/neck coils can be classified normal or broken by training a LSTMFCN on image features, successfully. Augmenting the data using GP-generated samples can improve the performance even further.