September 23, 2019
By Kai Goebel, Principal Scientist, System Sciences Lab, PARC, a Xerox Company
Companies that depend on large, expensive equipment have a fundamental responsibility to maintain the integrity and safety of their assets, not only to protect their employees and customers, but also to hold down their operating costs.
This responsibility is universal across all industrial sectors. Utilities need to maintain large power plants and extensive transmission and distribution lines. Infrastructure operators are responsible for ensuring the safety and operability of large, dispersed assets such as roads, bridges, tunnels, waterways, sewers and water supplies. Transportation and delivery companies must continually maintain and repair far-flung fleets of trains, buses and trucks. Aircraft companies and airlines are required to reliably ensure that their jets are 100 percent safe to fly on every flight. Manufacturing plants and process plants need to make sure that machines are in prime working order to prevent them from failing unexpectedly and shutting down whole production lines.
Such downtime of valuable equipment can cause expensive disruptions to businesses and their supply chains. The most cost-effective solution to this complex problem involves the adoption of predictive maintenance systems. Predictive maintenance (PdM) is an approach to monitoring the condition and performance of high-value equipment during operations to understand their state of health, predict time of failure, and to take action to mitigate the impact of impending failure through life extension mechanisms, planning, and performing maintenance, thereby reducing the likelihood of unforeseen failures.
Also known as condition-based maintenance (CBM), predictive maintenance has been contemplated in the industrial world since the 1990s. However, there have been barriers to achieving the full benefits due to the challenges of reliably scaling these systems up. The difficulty comes in part from the need to instrument equipment with sensors that can observe events of interest. Additionally, it is required to collect, store, and analyze sensor data from spread-out assets such as large fleets of equipment. One way to tackle the analytics problem involves the development of libraries of mathematical models and algorithms for each industrial group. These models can be then be applied in standardized architectures for components of interest to each particular site.
The Art of Analytics in the Industrial Internet of Things (IIoT)
Recent advances in communications and deployment of sensors through the Industrial Internet of Things (IIoT) have made predictive maintenance more feasible for far-reaching industrial portfolios. A powerful mix of physics-based models and machine learning systems that are designed and trained to recognize equipment faults can also make an accurate prognosis to predict the time at which specific components will fail.
In this way, industrial companies no longer need to go through costly procedures to perform regular periodic inspections and upkeep through schedule-based maintenance, which in fact may not always be necessary. Instead, predictive maintenance systems can logically prioritize any needed repairs to prevent failures before they occur, thus greatly saving precious time, money and resources.
The first step in this process involves anomaly detection, in which smart sensors on the network’s edge detect something that doesn’t seem right – without knowing exactly what is going wrong yet. The next step is to perform a diagnosis of any flagged equipment to determine the root causes of the problem. Finally, prognostic algorithms can make an accurate estimation of when the components will fail.
A new IIoT System Analytics technology platform called MOXI enables engineers, operators and maintenance professionals to remotely monitor and proactively manage unexpected system failures and maintenance problems. MOXI is designed for large-scale industrial assets in sectors such as manufacturing, power grids, railways, bridges and other heavy infrastructure. The IIoT suite combines embedded sensing, complex system models and artificial intelligence technologies to predict adverse system conditions with high accuracy, negligible false alarm rates and near-zero missed detections.
Deploying with Predictive Maintenance to Improve Railway Efficiency
In one case, PARC is pursuing a pilot project with the East Japan Railway Company (also known as JR-East). The railway has been facing rising costs due to its aging infrastructure amid shrinking budgets for regular maintenance.
PARC partnered with Nomura Research Institute (NRI) to interview JR-East engineers, R&D teams and maintenance technicians to fully understand the problem and determine how the PdM technology should be used. PARC then created dashboard mock-ups, gathered feedback from JR-East end users, and began rapid iterations of algorithm and software development.
The team developed customized fault detection and diagnosis pilot software by applying advanced machine-learning and a model-based system analyses approach. Dashboards were also developed for JR-East engineers to visualize and better understand the obtained data. The initial pilot focused on maintenance of train doors and railway tracks, with a goal to expand the program over time to include other assets such as rail cars, stations, bridges and tunnels.
Initial tests for both the railway and train door fault detections indicated very high true positive rates and very low false alarm/missed detection rates. The dashboards allow the JR-East technicians to pinpoint and repair upcoming issues before they escalate into asset failures/downtime. Further tests are being conducted to validate the findings and roll out a comprehensive field implementation.
As new IIoT systems get rolled out and begin to mature, the JR-East pilot project is just one example of the practical benefits of predictive maintenance to enable self-aware, self-adaptive systems. By applying the principles of physics to enhance AI-based predictive models, industrial companies can generate actionable insights about the condition, safety, and performance of their equipment to keep it running smoothly and optimally with minimal downtime.
Kai Goebel is a Principal Scientist in the System Sciences Lab at PARC, a Xerox Company. He is focused on leading research at the intersection of AI and cyber-physical systems for integration into both Xerox’s core business as well as for external partners. His particular interest is systems health management and autonomy for a broad spectrum of cyber-physical systems in the manufacturing, energy, transportation, aerospace, and defense sectors. Prior to joining PARC, Kai worked at NASA Ames Research Center on machine learning, physics modeling, quantum computing, and prognostics and systems health management as applied to aeronautics and space applications.