The development of low-cost smart vibration sensors (less than $100 per node) will enable wider deployment of machine health monitoring to military platforms, especially for less costly assets such as light tactical and utility vehicles where the high cost of existing health monitoring systems (typically in excess of $1000 per node) would be difficult to justify. There are hundreds of thousands of light utility vehicles in service throughout the military and the Navy and Marine Corps has more than 50,000 such vehicles in service. Predictive maintenance and reliable estimates of the remaining useful service lifetime of individual vehicles will improve overall fleet reliability and reduce total maintenance costs. The goal of the proposed program is to create low-cost (less than $100 per node), machine health monitoring smart vibration sensors. The smart sensor will include the following hardware elements: a vibration sensor element with bandwidth and resolution required to monitor drive-train component (transmissions, gearboxes, differentials), low power sensor interface electronics and signal preconditioning, and an ultra-low power computational engine for data acquisition, data logging, analytics, and communications. The smart sensor will support the implementation of trained neural networks for various applications such as gearbox and transmission condition monitoring. The complete system also will include an application running on an external device to communicate with smart sensors to allow users to configure sensors, download trained models, and upload and aggregate data for off-line model training. The goal of reducing the hardware cost to below $100 per node requires a reassessment of existing approaches to smart vibration sensor design. We are exploring the tradeoffs in power, performance, and cost of solutions that employ commercial off the shelf hardware versus custom designs, including the sensing element itself, to data processing and analysis on custom application specific integrated circuits (ASICs). The outcome of the Phase 1 effort will be a complete conceptual system design, from the sensor element to the user interface. The platform will support the implementation of a novelty detection autoencoding neural network. This method does not require labeled fault data for training, which makes it a promising and practical approach for the detection of machine faults. The platform will be open and will support user-defined machine learning models that are consistent with data and program memory constraints of the platform, the target is 1 Mbyte. The availability of an inexpensive, flexible, and easily re-programmable platform will enable and encourage much wider adoption of machine health monitoring for both DoD and commercial applications.
Benefit: Reduction of the cost of machine health monitoring to less than $100 per node, roughly a factor of 10 reduction from current costs, will enable a transformation of how the military maintains its more ubiquitous, lower cost assets such as the fleet of several hundred thousand light utility vehicles deployed throughout the Services. Employing machine health monitoring to avoid periodic maintenance when not required and to alert users of faults before they progress to failure will save the military significant maintenance time and costs and will more than justify the relatively modest investment in low-cost smart sensors. The Navy and Marine Corps alone have more than 50,000 light utility vehicles in service and the wide deployment of machine health monitoring across this vehicle fleet will improve readiness in addition to reducing maintenance costs. Making machine health monitoring more affordable for the military will be enabled by the large established commercial market for vibration monitoring, which was more than $1B world-wide in 2019 and is estimated to reach $2B by 2026. The commercial opportunities for vibration-based machine health monitoring primarily are monitoring of pumps and rotating machinery for gas and oil production, energy production, mining, chemical production, and a growing automotive application segment. Low-cost autonomous monitoring will reduce the need for expensive walk-around programs in which a technician with hand-held instruments periodically takes machine vibration readings throughout a plant, and it will enable machine monitoring to be employed on lower cost assets where an expensive condition-based maintenance installation would not be economically justified. An important feature of the proposed research is that the deep learning novelty detection models that will be integrated into the proposed smart sensor platform do not require labeled fault data for training, which is difficult to obtain in sufficient quantities. Rather, to train the novelty detection models, only data from normally operating machines is required, and machine signatures that deviate from normal are flagged for closer inspection. Simulated data from digital twins of the machines may be used initially to deploy such sensors and the deep learning models may be refined as data is collected by the deployed sensors. This feature of the proposed platform, in addition to its affordability, will enable broader adoption of the machine health monitoring in both military and commercial sectors. Finally, the proposed research will advance the use of machine learning in autonomous edge devices designed to provide users with actionable information, not simply raw data. The value of deploying deep learning to low-cost, low power, embedded platforms will extend far beyond machine health monitoring applications to sensors, smart speakers, and other autonomous and interactive devices in the built environment.
Keywords: Machine Learning, Machine Learning, Vibration Sensing, Novelty Detection, Machine Health Monitoring
