⚡ electrical engineering

Low Latency Stand Alone Compute-Efficient Forecasting of Marine Engine Time Series Data

This paper proposes a robust, low-latency mathematical framework for forecasting marine engine time series data that utilizes an adaptive-window multi-particle stochastic differential equation approach with Girsanov-transformed probability weighting to outperform traditional linear models in capturing non-linear dynamics and ensuring real-time operational reliability.

Original authors: Y. Harsha Vardhana Reddy, Soumyendu Raha

Published 2026-05-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Y. Harsha Vardhana Reddy, Soumyendu Raha

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a massive, high-tech ship. Your engine is the heart of the vessel, but it's not a perfect machine. It gets old, the weather changes, and the load on the ship shifts. Sometimes the engine runs smoothly; other times, it sputters or surges unexpectedly.

The problem is: How do you predict what the engine will do next?

If you try to predict the future using old-school math (like the ARIMA models mentioned in the paper), it's like trying to predict the weather by drawing a straight line through yesterday's clouds. It works okay for a minute, but as soon as the wind changes direction (a "transient" event), that straight line becomes useless.

If you use complex AI (like deep learning), it's like hiring a genius who can guess the answer perfectly but refuses to explain why. In engineering, you need to know why something might break, not just that it might break.

This paper proposes a new way to forecast engine health called the Adaptive Multi-Particle Stochastic Framework. Here is how it works, using simple analogies:

1. The "Crowd of Explorers" (The Multi-Particle Method)

Instead of trying to guess the engine's future with a single number, the authors imagine sending out a crowd of 1,000 explorers (called "particles") into the future.

The Old Way: One explorer walks a straight path. If the path curves, the explorer falls off.
The New Way: The 1,000 explorers all start at the engine's current state. They walk forward, but they are allowed to wander slightly left or right to account for random noise (like a sudden gust of wind or a vibration).
The Result: Instead of one prediction, you get a "cloud" of possible futures. This gives you a safety envelope: "The engine will likely be somewhere in this area."

2. The "Smart Window" (Adaptive Windowing)

The paper introduces a clever trick called Adaptive Windowing. Imagine you are looking at the engine's history through a window.

When the engine is cruising smoothly: The window is wide. You look back at a long history to get a stable, calm average. This prevents you from overreacting to tiny, meaningless bumps.
When the engine accelerates or stutters: The window instantly shrinks. You only look at the very last few seconds of data. This makes the model "hyper-aware" and ready to react immediately to sudden changes.

It's like driving a car: you look far down the road when driving on a straight highway, but you look right in front of the bumper when you are turning a sharp corner.

3. The "Physics Filter" (Girsanov Transform)

Here is the most creative part. Sometimes, the crowd of 1,000 explorers will wander off into impossible territory (like an engine temperature dropping below absolute zero).

The authors use a mathematical tool called a Girsanov Transform. Think of this as a magnetic filter.

It looks at all 1,000 explorers.
It gives a "vote" to the explorers who are walking in a direction that makes physical sense (following the laws of thermodynamics).
It ignores or "discounts" the explorers who are wandering into nonsense.
The final prediction is a weighted average of the "sensible" explorers. This stops the model from "chasing noise" (reacting to random sensor glitches) and keeps it grounded in reality.

Why is this better than the old stuff?

Speed: The old methods (like VARIMA) are like a chef who has to taste every single ingredient and adjust the recipe from scratch every time a new customer walks in. It's slow. The new method is like a chef who knows the basic recipe and just adjusts the salt based on the current taste. It's instant.
Data Hunger: The old methods need a massive library of past data (years of logs) to learn. The new method is like a smart local guide; it only needs to remember the last few minutes (a small "sliding window") to know what to do next. This means it can start working immediately on a new ship without needing years of history first.
Reliability: In the tests, the new method stayed accurate even when predicting 10 or 20 steps into the future. The old methods tended to flatten out or go crazy when looking that far ahead.

The Bottom Line

The authors have built a "Grey-Box" solution. It's not a rigid, unchangeable rulebook (White Box), and it's not a mysterious black box that you can't understand. It's a hybrid: it respects the laws of physics (the engine must behave a certain way) but uses data to handle the messy, random reality of the ocean.

This allows ship operators to get real-time risk warnings instantly, even on small computers onboard the ship, without needing to send data to the cloud or wait for slow calculations. It's a faster, smarter, and more adaptable way to keep marine engines running safely.

Technical Summary: Low Latency Stand Alone Compute-Efficient Forecasting of Marine Engine Time Series Data

1. Problem Statement

The operational reliability of marine vessels is increasingly dependent on the health of their propulsion systems (Gas Turbines and Marine Diesel Engines), which operate under diverse loads and environmental stressors. Traditional approaches to monitoring and forecasting these systems face three primary limitations:

Physics-Based ("White Box") Models: While interpretable, they rely on ideal thermodynamic cycles (Brayton and Diesel) and fail to account for real-world stochastic deviations caused by component aging, fuel variations, and environmental noise.
Data-Driven ("Black Box") Models: Deep learning and regression techniques often lack physical interpretability, are prone to overfitting sensor noise, and require massive training datasets, making them unsuitable for "cold-start" scenarios or edge deployment.
Traditional Statistical Models (ARIMA/VARIMA): These rely on fixed historical windows and linear assumptions. They struggle with non-stationary regimes (e.g., engine acceleration), often leading to "flat-lining" or unrealistic extrapolations during multi-step forecasting. Furthermore, their iterative optimization processes introduce unacceptable latency for real-time, high-frequency monitoring.

The paper identifies a gap for a "Grey Box" solution that combines the interpretability of physical trends with the adaptability of data-driven methods, specifically tailored for low-latency, stand-alone computation on marine vessels.

2. Methodology

The authors propose a dual-layered framework based on Stochastic Differential Equations (SDEs), treating engine states as particles moving through a state-space. The framework consists of two distinct methods:

A. Single Particle Method (SPM)

Designed for short-term, univariate prediction, SPM models engine parameters as a Geometric Brownian Motion (GBM) process.

Local Parameter Learning: It estimates local drift ( $a$ ) and diffusion ( $b$ ) coefficients using a sliding window of historical data.
Prediction: Future states are predicted using the Euler-Maruyama discretization, providing a closed-form solution for the expected mean and variance-based confidence intervals.
Advantage: It avoids iterative optimization, computing coefficients via closed-form algebraic solutions (means and variances) in $O(N)$ time.

B. Adaptive Multi-Particle Method (MPM)

Designed for non-linear, multivariate time-series prediction, MPM evolves an ensemble of particles to capture complex system dynamics.

Adaptive Windowing: Unlike fixed-window models, MPM dynamically adjusts the learning window size ( $W$ ) based on the instantaneous drift magnitude ( $D_t$ ). The window shrinks during rapid transients (e.g., acceleration) for responsiveness and expands during steady-state for stability.
Ensemble Evolution: It generates a large ensemble of particles (e.g., 1,000) representing potential future states via Euler-Maruyama discretization of a multi-dimensional SDE.
Girsanov Transform & Weighting: To mitigate "noise-chasing" behavior, a Girsanov transform induces a change of probability measure. Particles are assigned weights based on their proximity to the expected physical drift. The final prediction is a weighted probability distribution rather than a simple arithmetic mean.
Drift Estimation: Utilizes a fourth-order central finite difference scheme for higher numerical accuracy in estimating the drift vector.

3. Key Contributions

Adaptive Windowing Mechanism: A novel approach that dynamically scales the learning window based on system volatility, ensuring responsiveness during non-stationary regimes while maintaining statistical stability during steady-state operation.
Girsanov-Induced Weighting: The implementation of a Girsanov transform to re-weight particle ensembles. This filters out stochastic noise by assigning higher probabilistic weights to trajectories aligning with physical drift, significantly improving prediction stability compared to raw ensemble averages.
Compute-Efficient "Grey Box" Framework: The models utilize closed-form algebraic solutions for drift and diffusion estimation, eliminating the need for iterative grid searches (like AIC in ARIMA) or massive training corpora (like Transformers). This enables execution at frequencies of 1 Hz+ on low-power edge devices.
Stand-Alone Capability: The framework is designed to function as a self-contained unit onboard a vessel, requiring no cloud connectivity or large historical datasets for initialization.

4. Experimental Results

The proposed methods were benchmarked against ARIMA (univariate) and VARIMA (multivariate) baselines using marine engine sensor data (e.g., Lubricating Oil Temperature, Fuel Oil Pressure, RPM).

Multi-Step Forecasting Stability:
- SPM vs. ARIMA: SPM demonstrated superior stability over a 10-step horizon. While ARIMA error escalated rapidly due to its inability to model stochastic drift (leading to unrealistic linear extrapolations), SPM maintained a predictable error envelope governed by the diffusion term. For example, SPM achieved an RMSE of 0.0393 for Lubricating Oil Temperature (LOT) compared to ARIMA's 0.0567.
- MPM vs. VARIMA: In multi-step (20-step) forecasting, MPM significantly outperformed VARIMA. VARIMA tended to "flat-line" or diverge, whereas MPM provided a realistic range of outcomes bounded by physical constraints.
Noise Robustness: The MPM's weighting mechanism prevented the model from overfitting to high-frequency sensor noise, a common failure mode for VARIMA which often reacts sharply to negligible fluctuations.
Computational Efficiency:
- Latency: The SPM/MPM methods are orders of magnitude faster than VARIMA/ARIMA. The latter require iterative Maximum Likelihood Estimation (MLE) and grid searches for every new data window, introducing unacceptable latency for real-time monitoring.
- Data Dependency: Unlike VARIMA, which requires a large training corpus (e.g., 85% of the dataset) for global parameter optimization, MPM operates effectively with a limited local history (e.g., 50–300 samples), enabling rapid "cold-start" deployment.

5. Significance and Claims

The paper claims that the proposed SDE-based framework offers a significant departure from standard time-series forecasting for marine engineering. Its primary significance lies in providing a scalable, low-latency, and physically consistent solution for real-time risk quantification.

Operational Relevance: By treating fluctuations as diffusion rather than structural patterns, the model provides a "probabilistic envelope" that maintains physical consistency, which is critical for high-stakes marine engineering decisions.
Maintenance 4.0 Alignment: The approach supports the "Maintenance 4.0" paradigm by enabling proactive maintenance and real-time risk quantification without the computational overhead or data dependency of deep learning or rigid physics-based models.
Deployment Viability: The ability to run on edge devices with minimal data requirements makes the model uniquely suited for the harsh, connectivity-constrained environments of modern marine vessels.

The authors conclude that this hybrid approach successfully bridges the gap between the interpretability of physics-based models and the adaptability of data-driven methods, offering a robust tool for the non-linear, stochastic nature of marine propulsion systems.