Edge Computing and Machine Learning in Machinery Condition Monitoring

Thumbnail Image
Aloise, Chris
Machine Condition Monitoring , Machine Learning , Edge Computing
Reactive maintenance in industry is associated with high downtime costs and the risk of catastrophic failure; predictive maintenance has the potential to mitigate these costs and risks. Currently, to analyze tool wear in CNC machines, rolling element bearing health, and other vibration-sensitive environments, accelerometers are used to gather vibration data. It is common in industry to use feature extraction (i.e., Fourier Transform) to identify fault frequencies in the vibration signals. Machine learning can also be used to identify patterns in the faulty equipment’s vibrations. Machine learning can provide accurate predictions on bearing and tool health from raw vibration data given enough data, training time, and computational power. However, connection to powerful computational resources (the cloud) is not always possible in remote locations or for security reasons. In many applications, it would be useful to leverage the power of machine learning in a local device. This practice is called edge computing, and its feasibility is investigated in this research. A small device is used to demonstrate a low-cost, easy to install solution. Data preprocessing, model training, and data analysis are all possible on a small local device. However, it is possible to save time and computational power by training the model elsewhere and simply using it for inference (to make predictions) on the small device. The prediction accuracy and time to compute of eight machine learning models was compared between a high-end CPU and a small device CPU over four data sets (two bearing health and two tool wear). From all the models, a convolutional neural network was shown to be optimal using compute time and accuracy as metrics. The small device CPU was able to make predictions in only 15 times the amount of time as the high-end CPU. In this way, the tests verified the feasibility of using an edge computing device with machine learning for machine condition monitoring.
External DOI