Predictronics Chief Technology Officer David Siegel attended the 2017 Prognostics and Health Management (PHM) Society conference Oct. 1 - 5 in St. Petersburg, Florida, where he spoke on a variety of topics, including smart manufacturing standards and intelligent system technologies.

Below are Dr. Siegel’s top five takeaways:

1. Diverse Set of Industries, Common Challenges in PHM Applications

While the industrial segment is extremely diverse, challenges such as fully labeled data sets and data scarcity from degraded or failed systems were a reoccurring theme throughout the conference. There also seems to be an underlying need for models that are scalable, update over time, and can learn new failure patterns and operating modes.

Despite all of these challenges, there are still many success stories to be told. It just goes to show that with the right technology, knowledge and problem selection, effective solutions are possible.

2. Excellent Technical Work and Emerging Technologies

The conference’s technology demonstrations, technical papers and poster sessions were extremely innovative, but deep-learning was perhaps one of the most interesting topics during the technical sessions. Deep-learning’s ability to learn features, either in an unsupervised manner or for anomaly detection, and build models that could better generalize for PHM applications is impressive.

Researchers should continue to explore deep-learning's ability to learn features and generalize models, as opposed to looking at a single machine test-lab data and showing that its classification accuracy or prediction accuracy is slightly better than other machine learning methods. While incremental improvement is great, it is thinking outside of the box that drives technology forward.

The component level of PHM for gears, bearings and other mechanisms was also an interesting topic that came up during the conference. It is exciting that work on the direct estimation of the spall size from the vibration signal is showing great promise. Most current condition indicators are indirect assessments of the fault size. A direct estimation would be a big game-changer in a fielded system.

3. Addressing the Right Problem is Half the Battle

The success of a PHM project depends on the problem at hand. The right problem is technically feasible and can be narrowed down to a specific objective, such as reducing unplanned downtime, optimizing maintenance, improving product quality or gaining an edge over the competition.

It is important to note that developing a PHM system is not the only way to solve a problem. If an engineering or design countermeasure could solve an issue, that option should be considered as well.

4. The Balance of Data-Driven and Domain Expertise

There is too much emphasis on whether a solution is data-driven or physics-based. It is perhaps more insightful to consider how much each approach plays into a solution, and whether that solution is logical and fully addresses the problem by incorporating both data and domain knowledge.

More often than not, solutions include domain expertise to constrain and build better machine learning models for PHM. Similarly, physical models are frequently turned with data to calibrate model coefficients. That distinction gets especially difficult to make when, to train various data-driven machine learning models, solutions use physical models that stimulate different fault levels and failure modes.

5. Use Success Stories to Your Advantage

There are numerous defense, aerospace, automotive, manufacturing and heavy industry examples of successfully deployed PHM systems, which have saved companies millions of dollars. Highlighting these success stories could convince those hesitant to adopt this type of technology that it is, in fact, effective. However, while it is important to leverage these good stories as much as possible, it is also crucial to learn from past mistakes and experiences others have had in this field.