Co-author: Peeyush Pankaj Peeyush Pankaj is an application engineer at MathWorks with a strong Aerospace background and deep domain expertise in vibrations and mechanical systems (engine lubrication circuit design & analysis). In this blog post, he tells us about his experience at the 2025 Prognostics and Health Management Society Conference Data Challenge.

For the last several years, MathWorks has attended the Prognostics and Health Management (PHM) Society Conference – a global conference focused on predictive maintenance applications. This year, the team participated in the 2025 PHM Data Challenge – and took home first place. The MathWorks team comprised Peeyush Pankaj, Reece Teramoto, Shyam Joshi, Xiaomeng Peng, and Taylor Hearn.

 

 

A Real-World Data Challenge

The 2025 PHM Data Challenge asked participants to predict when two components of a commercial jet engine would need maintenance. Using sensor data from four engines across thousands of flights, teams had to create models to estimate how many flights remained before three different maintenance events: High Pressure Turbine (HPT) shop visit, High Pressure Compressor (HPC) shop visit, and HPC water wash. The three maintenance predictions made this a more complex problem than previous PHM data challenges. The dataset was intentionally realistic and messy, with missing sensors, noise, and flights out of order. And predictions that came too late were penalized more heavily. The teams faced a big challenge to generalize well and avoid overfitting the data. This was no simple toy example.

Engineering Expertise + MATLAB = Win

This year, MathWorks had two advantages in the data challenge: deep engineering domain expertise and a knowledge of AI and Predictive Maintenance tools in MATLAB. The team was led by Peeyush Pankaj, co-author of this post, whose prior career in aerospace engineering helped him lead the team to interpret degradation patterns, identify practical domain-specific features, and understand realistic engine behavior. Next, Peeyush describes how MathWorks did it.

The Solution

Read the full Paper: Maintenance Service Events Prediction Modeling of Aircraft Gas Turbine Engines | Annual Conference of the PHM Society

What made this year’s PHM Society Data Challenge particularly interesting was how closely it resembled the messiness of real engine prognostics. We were given only four engines for training, each with multiple failure-linked events that interact with one another the same way they do on a real fleet. The test engines operated in envelopes not seen in training, several sensors were entirely missing, and the individual flight files were shuffled out of sequence. This is very typical of fielded assets: a model must learn from whatever limited history is available and still generalize to engines of the same make and model that operate differently, age differently, and undergo maintenance at different intervals. The challenge captured that reality extremely well.

 

 

 

To handle the missing sensing channels in the test and validation sets, we built virtual sensor models to reconstruct key pressures and temperatures. Once the physical variable set was complete, we introduced domain-informed features such as compressor efficiency drop, turbine temperature gradient change, pressure ratio decay across the engine core, and several thermodynamic health markers that typically reflect aging. In addition, the raw dataset consisted of up to eight snapshots per cycle, sometimes fewer, so we converted everything into cycle-level statistics. This preserved the underlying physics and made it possible to analyze degradation on a cycle-by-cycle basis rather than dealing with snapshot irregularities.

A major breakthrough came when we built global health indicators using the Health Indicator Designer app in Predictive Maintenance Toolbox. When plotted, the HPC health indicator showed clear upward recoveries after Water Wash events, confirming exactly what we see in real field operations: regular washing restores compressor health and extends the life of downstream modules. That insight influenced the modeling strategy. We used neural networks for HPT shop visit predictions, and for HPC and Water Wash predictions we used sequence-to-sequence LSTM models with custom loss functions that penalized incorrect predictions more severely when the maintenance event was near. This helped prevent late forecasts for safety-critical events. MATLAB’s Deep Network Designer and Experiment Manager were extremely helpful here, allowing us to iterate quickly through model architectures and training strategies without writing everything from scratch.

 

 

Since the challenge deliberately scrambled the order of flight files in the test set, we finally introduced a profile-based registration method that aligns the files chronologically by matching the shape of the predicted health indicators. This step ensured that the degradation sequence remained physically consistent, which improved our confidence in the predictions.  

 

 

Overall, combining aircraft engine domain knowledge with productivity tools in MATLAB allowed us to move rapidly from noisy, inconsistent raw data to a complete, physics-aligned prognostics pipeline, ultimately resulting in first place in this year’s competition.

Congratulations to all the participants in the challenge and we hope to see you next year!