Abstract
This paper illustrates a data-driven approach adopted to address the PHME2020 Data Challenge competition. The aim of the challenge was to estimate the Remaining Useful Lifetime (RUL) of an experimental filtration device analyzing its clogging status by means of static (e.g. data sheets, fluid type, sensors) and dynamic (e.g. sensing data) information. We address the problem employing different state-of-the-art feature extraction, feature selection, and machine learning techniques. In this paper, we describe the approach followed to assess the problem and to generate robust and adaptable prediction models together with a corresponding performance assessment and robustness evaluation. The performance of the proposed solution is calculated in term of penalty score. The final penalty score 57.24 ranked 2nd in the above-mentioned data challenge competition.
Introduction
In industrial environments, filtration systems are essential components whose failure can lead to significant downtime and maintenance costs. Prognostics and Health Management (PHM) aims to mitigate these risks by accurately predicting the Remaining Useful Life (RUL) of such components. The PHME2020 Data Challenge provided a platform to develop and evaluate models for predicting filter clogging in an experimental rig, where failure is defined as a pressure drop reaching 20 psi. The primary challenge lies in the variability of operating conditions, including different particle sizes and concentrations, requiring models that are both accurate and robust to unseen scenarios.
Experimental Setup and Dataset
The dataset was generated from a filtration rig using Polyetheretherketone (PEEK) particles. It consists of 48 run-to-failure experiments. The training set includes 24 experiments covering "small" (45-53 micron) and "large" (63-75 micron) particle sizes across three solid ratio profiles (0.4%, 0.425%, and 0.45%). The validation set comprises 8 experiments with a fourth solid ratio profile (0.475%), specifically designed to test the model's ability to generalize to conditions not encountered during training.
High-frequency (10Hz) sensor data provided measurements for flow rate, upstream pressure, and downstream pressure.
Methodology
Feature Engineering and Selection
A multi-stage pipeline was implemented for feature extraction. Statistical features, including kurtosis, root mean square (RMS), and skewness, were calculated using a 10-second rolling window. A key innovation was the introduction of a physics-inspired "hybrid" feature, clog_am. This feature represents the cumulative amount of dirt intercepted by the filter, calculated as the time integral of the product of flow rate, particle size, and solid ratio.
Feature selection prioritized monotonicity—the tendency of a feature to trend consistently in one direction as degradation progresses. Techniques such as bivariate distribution analysis, correlation matrices, and Recursive Feature Elimination (RFE) were used to identify the most predictive and stable features, filtering out noise that could lead to overfitting.
Model Architecture
The final model is a four-layer sequential neural network. During development, various architectures and activation functions were evaluated. While models using ReLU activation showed lower Mean Absolute Error (MAE) on training data, a model utilizing Sigmoid activation was ultimately selected for its superior stability and robustness when applied to the validation set's unseen conditions.
Results and Discussion
The proposed approach achieved a final penalty score of 57.24, securing 2nd place in the PHME2020 Data Challenge. This score reflects a balance between prediction accuracy (MAE) and the model's performance across training and validation sets.
As shown in the loss curves, the model demonstrates rapid convergence and maintains a stable gap between training and validation loss, indicating effective generalization. Furthermore, the model's performance remained consistent even when the training data volume was significantly reduced, meeting the challenge's requirements for data efficiency and robustness.
Conclusion
This study demonstrates that combining domain-specific knowledge (the clog_am feature) with robust feature selection and stable neural network architectures can effectively predict RUL in complex filtration systems. Prioritizing model stability over raw training accuracy proved crucial for handling unforeseen experimental conditions. Future work will explore the integration of more detailed physics-based models and the inference of static parameters directly from sensor streams to further enhance prediction accuracy in real-world industrial applications.
Summary updated