Return to Blog


PHM Logo


DECEMBER 9, 2021

Predictronics had a rewarding experience participating in the 2021 Virtual Prognostics and Health Management Society Conference.

Predictronics CTO Dr. David Siegel hosted a tutorial session during this year’s conference entitled, Methodology and Case Studies for Fielding PHM Systems – Successes, Challenges, and Lessons Learned. In addition, Dr. Siegel also helped organize the 2021 PHM Society Data Challenge and was part of the top three winning technical papers session.

Dr. Siegel shares his top 3 takeaways from his tutorial and the data challenge:

1. Data Quality

One of the common topics that received great interest and feedback during the tutorial session was the topic of data quality. Data quality can be a topic of great concern for industry, as many sources of historical data, control system data, maintenance data, and even sensor data may be available. However, this data may not be well maintained or consistently collected and therefore might not be appropriate or adequate for PHM. From our experience, real world data from fielded PHM systems is never perfect and some type of data quality check and pre-processing is almost always required.

One concern is whether these methods can be automated or if they require manual cleansing of the data. Fortunately, there are established routines for checking the data quality for high frequency measurements, such as vibration/acoustic. For low frequency signals, it requires more domain knowledge and offline inspection, but typically one can develop logical rules to have a routine for removing outliers and pre-processing the data.

All routines should be automated during the deployment and, in some cases, might require some fine-tuning during the offline analysis and configuration of the solution.

Other broader topics of interest were data quality with respect to acquiring enough training data and determining whether previously collected data is intended for a predictive maintenance system. When examining whether data is right for a predictive maintenance system, it is important to understand the original use of the data and capture any of the key contextual information that would be helpful when developing PHM systems with that data.

2. Practical Challenges for Fielded PHM Systems/Deployments

Within the field of PHM research, there is a drive to develop new analysis methods and algorithms, which can improve upon the existing capabilities for health monitoring, diagnosis, and failure prediction/prognosis. While these new techniques do generate excitement, practical challenges must also be considered when ultimately selecting a particular analysis method for a PHM solution.

Examples of these challenges include the amount of training data required, any computational requirements/limitations, whether the model can be easily refreshed/updated with new data, and the ability for the model to generalize for unforeseen data.

These practical challenges discussed during the tutorial really resonated with the audience and emphasized the importance of research for methods that address these unmet practical challenges for the development and deployment of solutions for PHM fielded systems.

3. Public Data Sets for Benchmarking and Updating Evaluation Methods for Data Challenges

The PHM Society encourages the use of public data sets for benchmarking and evaluating PHM analysis methods and algorithms. Their yearly data challenge is a great example of their promotion of this initiative.

This year’s data challenge focused on developing a model to estimate remaining useful life (RUL) for aircraft engines.

The data set provided for this competition is from a fleet of aircraft engines operating under variable conditions that experienced multiple modes of failure. These different failure modes showcased from real flight paths, with multiple regimes and flight phases, make working with the data a real challenge with many possible outcomes.

This particular industrial data set was chosen in part because the application of PHM in aerospace is big within the research community. Many aircraft engine manufacturers and airlines collect engine data, as well as other types of aircraft data. The topic of remaining useful life prediction is perhaps the most challenging part, since it requires a good deal of historical data, as well as good inputs from the health monitoring and diagnosis algorithms.

One point of note discussed during the data challenge evaluation was if additional aspects should be considered beyond prediction accuracy. The debate centered around whether it requires a subjective evaluation of a team’s approach, in order to see if it is logical or could be generally applied, or whether elements of explainability related to a team’s algorithm or computational requirements should also be used in the evaluation of data challenges. Explainability of AI/ML models is a topic of importance in the PHM field and can lead to greater understanding, trust, confidence, and adoption in any industrial deployment of such models.

The data challenge contest has concluded; however, if you’d like to evaluate your own methods and algorithms, the public data set utilized for the competition is still available to download here.

The goal of this conference is to provide an opportunity for industry professionals, technology providers, government officials, and academics who are experts in the field to share their knowledge and debate critical topics in PHM. Predictronics is proud to have been a part of the conversation with such distinguished colleagues and intellectuals, fostering a dialogue on the importance of PHM in the industrial space.

For more information and weekly updates, follow Predictronics:

Know What Happens Next with PDX!