Maritime researchers have engineered an adaptive predictive control framework that significantly advances the precision and operational stability of autonomous ships navigating along designated straight-line corridors. Developed to overcome the twin challenges of hydrodynamic model uncertainty and persistent ocean environmental disturbances, the computational architecture merges a dynamically trained feedforward neural network with a real-time sliding data window. The system continuously identifies changes in vessel maneuvering behavior without relying on rigid physical mechanism equations that degrade when cargo, speed, or sea states fluctuate. Validated through numerical simulations grounded in the hydrodynamic characteristics of the commercial car carrier vessel Hual Trooper, the control strategy demonstrates that intelligent data-driven adaptation can effectively stabilize underactuated vessels along planned rhumb lines, offering a practical foundation for advancing autonomous waterway transportation and relieving navigational watch burdens.
Standard nautical practice relies on pre-planned rhumb lines between waypoints to establish safe, fuel-efficient transits across open waters and congested maritime lanes. However, as international trade expands and modern commercial vessels grow increasingly specialized and massive, automated trajectory execution must handle severe dynamic complexities. Traditional steering systems relying on classical mechanism-based formulations or fixed transfer functions often experience severe controller mismatch because mathematical coefficients drift significantly under varying operational loading and vessel speeds.
To overcome the fundamental limitations of mechanistic models such as the Abkowitz, Norrbin, and Nomoto formulations, the researchers turned to non-parametric black-box system identification. While grey-box identification methods like extended Kalman filtering, support vector regression, and recursive least squares depend heavily on predefined structural assumptions, black-box modeling extracts dynamic input-output relationships directly from operational data. Conventional deterministic machine learning approaches, including backpropagation neural networks trained via gradient descent or the Levenberg-Marquardt algorithm, often suffer from extreme sensitivity to initial parameters and frequently become trapped in suboptimal local minima. Furthermore, deep recurrent architectures like long short-term memory networks demand prohibitive computational training budgets that exceed the strict real-time control intervals required for active maritime steering corrections.
The newly proposed framework addresses these computational bottlenecks by coupling a shallow multilayer perceptron feedforward neural network with an enhanced metaheuristic trainer known as the improved butterfly optimization algorithm. Drawing inspiration from the natural foraging and mating behaviors of butterflies that track chemical fragrance gradients, the standard butterfly optimization algorithm offers structural simplicity and minimal tuning parameters. However, basic butterfly optimization models frequently suffer from premature convergence when navigating intricate multi-dimensional search spaces. To eliminate stagnation, the research team integrated two distinct algorithmic mechanisms: a global search operator employing a t-distribution mutation guided by a dynamically varying balance factor, and an intensified local search mechanism governed by an optimal individual guidance strategy.
Within the global search phase of the improved butterfly optimization algorithm, the balance factor begins with values distributed across a wider unit interval, ensuring robust progression toward global optima while the t-distribution mutation prevents early population homogeneity. As algorithmic iterations advance, the balancing threshold contracts toward a narrower range, selectively increasing the probability of t-distribution mutation updates to preserve swarm diversity and avoid blind displacement from promising regions. In the complementary local search phase, the guidance mechanism pools population resource data to conduct directed exploratory walks relative to the best-performing vector. Applying the global convergence criteria established by Solis and Wets, the authors formally demonstrated through rigorous mathematical analysis that the improved butterfly optimization algorithm converges to the globally optimal parameter region with a probability of one.
To provide the neural network with fresh, representative operational measurements without accumulating outdated historical drag, the framework incorporates a sliding data window as a localized observer. Operating as a strict first-in, first-out queue with a fixed temporal width, the window discards older motion pairs synchronously as new measurements arrive, packaging mini-batches of immediate input-output data. This streamlined mechanism supplies the neural network with the precise operational history needed to track rapid shifts in maneuvering dynamics. Systematic parametric evaluations across five distinct window widths established that an optimal width of fifty sample steps delivers an ideal compromise between swift adaptation and tracking accuracy, completely avoiding the severe computational delays observed when testing larger batch registers.
In the overall control loop, straight-line tracking errors relative to waypoint line segments are transformed into an equivalent heading control problem through a proportional line-of-sight guidance law. By establishing a forward-looking distance scaled to three ship lengths between perpendiculars, the line-of-sight algorithm generates continuous course commands, with uniform semi-global asymptotic stability confirmed via Lyapunov stability analysis. The dynamically updated feedforward neural network acts as an online multi-step predictor, mapping past heading angles, yaw rates, and rudder deflections to forecast future vessel states over a finite prediction horizon. A sequential quadratic programming algorithm then solves the constrained finite-time domain optimization problem at each control cycle, determining an optimal sequence of incremental rudder movements while explicitly respecting steering gear mechanical constraints.
To establish rigorous theoretical guarantees for maritime safety, the researchers conducted a comprehensive stability analysis using the principle of input-to-state stability. Nonlinear ship maneuvering was formulated as a discrete-time constrained dynamic system subject to bounded disturbances and parameter modeling discrepancies. By treating the closed-loop objective cost function as an input-to-state stability Lyapunov function and demonstrating that the multi-step neural prediction error remains strictly bounded under uniform continuity and Lipschitz constraints, the analysis proved that the overall path-following control system achieves robust input-to-state stability within the invariant state set, ensuring that external environmental disturbances produce bounded, controlled deviations rather than catastrophic instability.
Before deployment in maritime motion control, the optimization capabilities of the improved butterfly optimization algorithm were thoroughly evaluated across six recognized numerical benchmark suites, encompassing unimodal functions designed to assess exploitation and multimodal surfaces testing global exploration. On multimodal test functions, the coordinated balance factor mutation and guidance recombination prevented entrapment in deceptive local valleys, maintaining search diversity without inflating overall computational complexity beyond the standard initialization and fitness evaluation bounds.
Operating at a cruising speed of 20.8 knots with a sampling period of one second, the ship model was subjected to continuous twenty percent random parameter perturbations in its turning and lag indices to reproduce severe hydrodynamic uncertainty. In online model identification trials, the improved neural framework achieved a mean squared error of 0.1136 and a root mean squared error of 0.3370, representing error reductions of 54.40 percent and 78.91 percent compared to conventional Levenberg-Marquardt backpropagation and baseline butterfly-trained networks, respectively, with Wilcoxon rank-sum testing confirming statistical significance at the five percent level.
Full closed-loop path-following evaluations further confirmed the superior performance of the predictive control architecture across multi-waypoint straight-line transits under white-noise environmental perturbations. The control system driven by the improved butterfly-trained neural predictor achieved an average path-following error of 7.8238 meters and an average rudder angle amplitude of 6.7037 degrees, outperforming alternative architectures based on radial basis function networks, grey wolf optimizers, and standard butterfly predictors by up to 65.57 percent in trajectory error and up to 29.72 percent in steering mechanical effort. Crucially, the average per-step computational runtime was restricted to 0.0872 seconds, comfortably beneath the one-second sampling threshold, proving that the proposed framework delivers high-precision, energy-efficient path-following within the strict timing constraints required for practical maritime implementation.
Subject of Research: Adaptive straight-line path-following predictive control of underactuated ships using online neural network identification
Article Title: Adaptive straight-line path-following of underactuated ships via online identification with an enhanced stochastic trainer-based neural network and sliding data window
Article References: Xu, D., Yan, J., & Yin, J. (2026). Adaptive straight-line path-following of underactuated ships via online identification with an enhanced stochastic trainer-based neural network and sliding data window. Discover Computing, 29(1), Article 921. https://doi.org/10.1007/s10791-026-10856-w
Image Credits: AI Generated
DOI: 10.1007/s10791-026-10856-w
Keywords: underactuated ships, path following, model predictive control, feedforward neural network, butterfly optimization algorithm, sliding data window, online system identification, line of sight guidance, sequential quadratic programming, input to state stability, Hual Trooper, Nomoto model, marine navigation, autonomous ships, stochastic optimization, rudder control
News Source: Cassandra Pierce. (November 22, 2026). Neural Network With Sliding Data Window Enhances Autonomous Ship Path Following. Scienmag.



