VECTOR VXP2: A Physics-Informed Neural Network-Driven Predictive Maintenance System for Aero-Engines

The VECTOR VXP2 platform developed by Orion Spacetech aims to address the certification bottlenecks and false positive issues in aero-engine predictive maintenance. This system integrates a dual-layer LSTM architecture and thermodynamic constraint mechanism, adopts the Physics-Informed Neural Network (PINN) approach, implements uncertainty quantification via Monte Carlo Dropout, and achieves over 80% accuracy in Remaining Useful Life (RUL) prediction on the NASA CMAPSS FD004 dataset, providing a feasible path for AI deployment in safety-critical domains.

Aero-engine predictive maintenance faces two major challenges: First, airworthiness certification requires AI models to be safe and reliable, but traditional black-box AI lacks physical interpretability and is difficult to pass regulatory reviews; Second, the false positive issue—sensor noise, environmental fluctuations, etc., easily lead to misjudgments, causing high costs such as flight delays and scheduling chaos. The core innovation of VXP2 lies in shifting to a physics-constrained self-certification architecture to address these problems.

VXP2 uses a dual-layer LSTM architecture: the bottom layer captures short-term fluctuations at the sensor level, and the top layer integrates degradation trends; the key innovation is the outer thermodynamic constraint mechanism—verifying the real-time P30/T30 ratio (compressor outlet pressure to temperature ratio) before inference. If it violates physical laws, the prediction is rejected, eliminating false positives caused by sensor noise at the source. This 'physics first, AI second' model is a practical application of PINN in industry, where AI exerts its predictive capabilities within a physics-constrained framework.

VXP2 uses Monte Carlo Dropout technology, keeping the dropout layer active during inference, and performing multiple forward propagations to obtain the prediction distribution, providing the standard deviation of RUL prediction instead of a single point estimate. This is crucial for aviation decision-making: high confidence (narrow interval) supports extending maintenance intervals; low confidence (wide interval) triggers conservative strategies; multi-modal distribution implies multiple degradation modes. Probabilistic output enhances the system's auditability, meeting the requirements of safety-critical systems.

VXP2 was trained and validated on the NASA CMAPSS FD004 dataset, which is a standard benchmark for aero predictive maintenance. It features multiple operating conditions (different flight altitudes, Mach numbers, etc.), multiple failure modes (compressor, fan degradation), high-dimensional sensors (24 channels), and complete degradation trajectories, making it the most challenging subset in the CMAPSS series. VXP2 achieved over 80% RUL prediction accuracy on this dataset.

The value of VXP2 lies in: 1. Transforming black-box AI into an auditable white-box system, suitable for safety-critical domains such as aviation and healthcare; 2. The self-certification architecture lowers the certification threshold, accelerating AI deployment from the laboratory to production environments; 3. Supporting condition-based maintenance, maximizing engine on-wing time, and optimizing operational costs; 4. The thermodynamic constraint concept can be extended to other physical systems such as gas turbines and wind turbine gearboxes.

As a prototype system (version 1.0B), VXP2 still needs improvements: 1. Adapt to multi-fleet operations and research cross-model transfer learning strategies; 2. Optimize edge deployment, compress the model to adapt to limited computing resources in aviation environments; 3. Integrate onboard ACARS or 5G real-time data links; 4. Add fault root cause diagnosis functionality to identify the type of degraded components.