4 November 2024
Abstract: This article presents the results published in the paper "System Condition Monitoring Based on a Standardized Latent Space and the Nataf Transform" published in IEEE Access Reliability Society Section in February 2024. This work introduced a condition-monitoring approach suitable for complex engineering systems such as aircraft, wind turbines and automated manufacturing plants. Optimal operation and maintenance of these systems require sophisticated strategies based on continuous health monitoring. The proposed approach analyzes measured data to assess the health condition of the systems. It combines an Artificial Intelligence (AI) model with the Nataf transform. The resulting model includes a built-in visualization tool for enhanced interpretability, which is key for high-risk and safety-critical systems. We demonstrated the proposed approach in two real-life case studies—aircraft engine run-to-failure data and wind turbine operation data—where it detected degraded conditions earlier than methods using only an AI model.
Keywords: Complex engineering systems, condition monitoring, data-based approaches, detection, artificial intelligence, Nataf transform, variational autoencoder.
Introduction
What do an aircraft, a wind turbine and an automated manufacturing plant have in common? These modern engineering systems are extraordinarily complex and must meet high reliability and safety standards. To ensure these systems accomplish their mission during their lifetime, operators adhere to sophisticated operation and maintenance strategies based on health condition-monitoring. To that end, modern engineering systems embed dozens, if not hundreds, of sensors that continuously capture data describing their physical state.
The ever-increasing databases from modern engineering systems are challenging to analyze for two reasons. First, because of the multiphysics coupling within these systems. Consider, for example, a wind turbine: its operation involves mechanical, thermal, and electrical variables, all influenced by the stochastic nature of wind and environmental conditions. Second, the behaviour can be highly non-linear, and the relationships between variables can be subtle. Classical analytical methods are often too limited for modeling such systems.
Artificial Intelligence — A Most Interesting Avenue!
AI models (particularly Deep Neural Networks) are well suited for capturing complex information from high-dimensional and complex system descriptions. AI-based condition monitoring is a very active research domain. An extensive corpus of articles on AI-based condition monitoring in diverse application domains brings solutions to enhance complex systems' reliability, security and profitability. However, AI models' "black-box" nature can hinder their adoption in high-risk or high-cost applications. In our research, we designed an interpretable AI-based approach featuring a visualization tool.
Artificial Intelligence — Four Takeaways for Your Next Implementation
- Keep it simple: If a simpler model, like polynomial regression, can solve the problem, stick with it. More than an overkill solution, AI modeling may introduce unnecessary complexity and waste numerical resources. Checking for more straightforward solutions while you grasp the system characteristics is a good practice.
- “Garbage in, garbage out”: Sometimes the data you have is simply not correct. No (AI) modelling could provide a good description of such a case. High-quality input data is essential. Pre-processing the database should include validation and pre-processing of the input data. These steps comprise filtering out incorrect data, handling gaps or biases, and normalizing measured values.
- Model design: Consider the train-validation loss functions to choose appropriate architecture and hyperparameters. Seek convergence and a model that generalizes accurately on both train and validation datasets—you do not want to overtrain or undertrain. Look for guidance and previous implementations to choose your model, activation functions, layers and hyperparameters wisely.
- Last (but not least!): Test. Test. Test. The testing dataset should be carefully designed. Include cases your model should be able to model. Use metrics such as accuracy, precision, F1 score, and recall to evaluate your model. Go back to steps 2 and 3 if needed.
Our AI Model for Enhanced Interpretability
Our paper introduced a new method suited for systems described by multiple measures. We assume datasets corresponding to the different conditions can be created with a suitable labelling strategy.
We developed a novel approach for condition monitoring of engineering systems using a combination of an AI model, the Variational Autoencoder (VAE), and a conventional tool from reliability theory, the isoprobabilistic Nataf transform. Here’s how it works:
Variational Autoencoder: The VAE captures essential information from high-dimensional data. Why the VAE? We chose this model for its ability to compress complex, high-dimensional data into a lower-dimensional latent space while preserving essential features. Also, the VAE offers a built-in visualization tool, enhancing model explainability. Nevertheless, the latent space distribution changes whenever the model is training again due to randomness in the VAE model definition and the training algorithm. That's where the Nataf Transform comes into play!
- Nataf Transform: The Nataf transform maps the healthy condition cluster—with an arbitrary distribution—in the latent space into a multivariate standard normal distribution in the Nataf space. See Fig. 1. Having the same distribution for the healthy condition across different training instances is quite an advantage. This consistency of the healthy condition distribution allowed for the definition of two complementary health indexes: IM and IN.
- IM is an outlier detector. It is sensitive to sudden changes and detects anomalies that evolve quickly outside the ±3σ range of the standard normal distribution.
- IN is an out-of-distribution detector. It monitors gradual changes that manifest as changes in the distribution itself.
These indexes capture different degradation patterns, making the CM more capable of detecting abnormal conditions as early as possible.
Figure 1. The NT maps the latent space z into the standard Nataf space s. In the latter, the health condition follows a 2D standard normal distribution.
After the model is properly trained, it uses recent data to assess the condition status—whether the system is operating in normal or in degraded condition—with two complementary health indexes, IM (outlier detection) and IN (out-of-distribution detection), as depicted in Fig. 2.
Figure 2 Proposed condition monitoring approach. The physical high-dimensional dataset space is VAE-encoded into the 2D latent space, then Nataf-transformed into the standard Nataf space. The condition status combines health indexes IM and IN.
Testing with Real-World Case Studies
We tested our method with data from two engineering systems: aircraft engines and wind turbines.
1. Aircraft Engines. The Commercial Modular Aero-Propulsion System Simulation (CMAPSS) database, created by NASA, describes multiple physical variables from commercial aircraft engines for run-to-failure cases. Our method detected failures earlier than a concurrent technique based on a regressive VAE.
Figure 3 Aircraft engine analyzed in the CMAPSS database. The IN health index anticipated the detection of failure initiation in 13 cycles.
2. Wind Turbines. Our research used data measured from a wind farm operating in North America (data from the Supervisory Control and Data Acquisition system, or simply SCADA system). The proposed approach successfully detected overheating of critical components, such as the main bearing and the generator.
Figure 4 Trajectory of the wind turbine from healthy condition (blue cluster) to a degraded condition (red cluster) in the latent space and the standard Nataf space. Cyan shows points at the beginning of the time series, and magenta the most recent data points.
Conclusion
The original contribution of our paper was to combine the VAE model with the Nataf transform, a conventional tool from the reliability theory. Why It Matters? While AI has gained popularity, its implementation is still limited because of its lack of interpretability. The VAE-Nataf combination includes a built-in visualization tool that confers interpretability to the proposed condition monitoring approach. Our method was demonstrated with real-life case studies from aircraft engine and wind turbine operations. It can be applied to other complex systems with lots of sensors. What's next? We are currently working on the diagnosis and prognosis of wind turbines based on measured data, and further results will be published soon. Interested in research and development for a sustainable energy future? Get in touch with our team.
