Health monitoring in turbofan engines represents one of the most complex challenges within aeronautical predictive maintenance. Traditional models based on recurrent neural networks, such as GRUs, often offer high accuracy in predicting remaining useful life, but suffer from a lack of interpretability: their internal states do not clearly reflect a coherent degradation process. This limits their adoption in environments where certification and explainability are as critical as numerical accuracy. An emerging line of research proposes the use of liquid neural networks as latent dynamics models, where a hidden state evolves through differential equations that capture the temporal evolution of degradation, separating it from variations inherent to operating conditions.
This approach introduces a factorization of the latent state into two components: one dedicated to irreversible degradation and another that captures environmental fluctuations. The network is trained with supervised losses on remaining useful life, monotonic risk, and temporal consistency, while penalizing any leakage of operational information into the degradation component. Results on the C-MAPSS benchmark show that the liquid model improves sensor prediction RMSE under multiple conditions (subsets FD002 and FD004), and that the learned degradation state forms a much clearer temporal axis, achieving a Spearman correlation of 0.5960 in state velocity. However, direct regression of remaining useful life remains superior with the GRU, indicating that this representation is more useful as an interpretable world model for understanding failure dynamics than as a calibrated residual life regressor.
From a business perspective, this technology opens the door to artificial intelligence solutions for companies that not only predict when a failure will occur, but also explain how the asset is degrading. This capability is vital in regulated sectors such as aviation, energy, or advanced manufacturing. Implementing these models in production also requires a robust data and deployment architecture. That is why companies like Q2BSTUDIO offer custom applications that integrate sensor ingestion pipelines, liquid model training, and latent state visualization in interactive dashboards. The use of AWS and Azure cloud services allows scaling real-time inference computation, while business intelligence tools like Power BI facilitate fleet health monitoring.
One of the most promising aspects is the possibility of training AI agents that, based on the liquid latent state, make autonomous maintenance decisions: from rescheduling inspections to adjusting operating parameters to extend useful life. This requires cybersecurity to protect both sensor data and deployed models. Q2BSTUDIO integrates these capabilities within its custom software platform, combining the latest deep learning research with robust software engineering practices. Ultimately, the combination of liquid latent states and a well-designed digitalization strategy enables a shift from purely predictive maintenance to interpretable and proactive maintenance, where each decision is supported by a clear and verifiable representation of the degradation process.
