Ship cabin noise management is shifting from passive static solutions toward intelligent dynamic autonomous management, for which accurate multi-source vibration state sensing is a prerequisite. Under limited observation channels, the core challenge is decoupling physically distinct noise sources — narrowband periodic vibrations, broadband transient pulses, and quasi-periodic human-activity impacts — from multi-source aliased signals, where a unified model readily induces cross-task negative transfer. We propose a function-decoupled heterogeneous multi-source state sensing framework with source-specific feature pathways and independent sub-models for device classification, speed regression, and transient recognition, together with task-wise feature compression reducing dimensionality by over 75% while preserving core performance. A key empirical finding — termed the weak-signal advantage — is that base-side points, despite much weaker signals, significantly outperform raft-side points on oil-pump speed regression and drop/walking detection (paired tests), with motor speed favoring the raft side, establishing that information separability rather than absolute signal strength governs sensing performance. The compression process further reveals nearly non-overlapping core discriminative bands across tasks, supporting the function-decoupled design.
Federated Learning (FL) is a privacy-preserving distributed machine learning paradigm that enables collaborative model training without exchanging local data. Based on data distribution characteristics, FL is typically categorized into horizontal, vertical, and federated transfer learning. As a crucial preliminary step, Private Sample Alignment (PSA) aims to identify common samples across participants without revealing non-intersecting information. However, achieving efficient alignment while preserving privacy remains a key challenge for FL deployment. Existing schemes vary significantly in security assumptions, performance, and applicability, which complicates practical implementation. This paper systematically reviews PSA techniques for FL. First, it clarifies the fundamental concepts and threat models of sample alignment. Next, using Private Set Intersection (PSI) protocols as the main thread, it outlines the evolution of PSA schemes from basic PSI to its various extensions, and explains how techniques such as Differential Privacy (DP) and threshold control are integrated to enhance security, applicability, and scalability. Furthermore, it analyzes the strengths and limitations of existing schemes. Finally, it summarizes the current shortcomings in research and proposes future directions to promote the further application and optimization of privacy-preserving alignment technologies in FL.
The parallel multi-channel structure of the Casing Once-through Steam Generator (COTSG) is prone to inducing flow instability under boiling two-phase flow conditions, posing a threat to equipment safety. In this study, the subchannel analysis method is applied to investigate flow instability in the COTSG. A multi-region subchannel model of the COTSG is developed using the subchannel code CTF, and its reliability is validated through comparison with experimental data and the RELAP5 code. The effects of primary-side flow distribution, transverse mixing, tube bundle scale, and local blockage on the simulation results are systematically examined. The results indicate that more reliable outcomes are achieved by allocating primary-side flow according to the number of heat transfer casings while accounting for transverse mixing. A simplified subchannel model with 19 casing tubes provides a favorable balance between computational accuracy and efficiency. System instability is initially triggered in the edge channels with the highest thermal load, where out-of-phase oscillations first appear between edge channels and subsequently induce flow oscillations in the inner channels. In the multi-channel model, a blockage in a single secondary-side channel enhances system stability, whereas a primary-side blockage reduces system stability. The extent of this effect is related to the location and axial position of the blockage. This study provides a new approach for refined thermal-hydraulic design and safety analysis of COTSGs.
The introduction of composite materials into a Pump-Jet Propulsor (PJP) rotor alters its line-spectrum thrust fluctuation in a frequency-dependent manner. We combine a transient Computational Fluid Dynamics-Finite Element Method (CFD-FEM) fluid-structure interaction algorithm with theoretical analysis to investigate the mechanisms under spatially fixed flow-field perturbations (caused by the rudder wake and stator) and the rotating flow-field perturbation (caused by the rotor itself), considering both single-blade and whole-rotor behaviour. The theoretical predictions agree with the numerical results. For a single blade under spatially fixed perturbations, the fluid-structure interaction deformation and the non-uniform flow field are in the same direction and thus cancel, reducing the thrust fluctuation by 99.61% at the wake blade passing frequency (64 Hz) and 97.62% at the stator blade passing frequency (144 Hz); this cancellation effect scales inversely with the square of the circumferential period number of the fixed perturbation. Under the rotating perturbation, however, the deformation and the flow field are vectorially orthogonal and superpose constructively, increasing the fluctuation by 27.5% at the rotor blade passing frequency (112 Hz). For the whole rotor, under spatially fixed perturbations the amplitude is governed by the single-blade amplitude and remains of the same order, whereas under the rotating perturbation it becomes seven times the single-blade value. Overall, the composite rotor strongly suppresses the wake blade passing frequency line spectrum, where the energy of the metallic PJP is most concentrated, which is beneficial for reducing the low-frequency radiated noise of the propulsor.
To address the lack of interpretability in Convolutional Neural Networks (CNNs) for ship-radiated noise classification, this paper proposes a visual analysis framework integrating Guided Backpropagation (GBP) and Gaussian Mixture Models (GMMs). Firstly, a convolutional kernel visualization process based on GBP is constructed, combined with a multi-frame feature alignment algorithm to eliminate time-frequency interference from harmonic line components, enabling the precise reconstruction of deep convolutional kernel response features. Secondly, a GMM-based pattern decomposition method is proposed, which models activation locations to decompose a single convolutional kernel into multiple sub-patterns with distinct frequency distributions. Experiments were conducted on a CNN model with CQT spectrogram inputs, validated on the DeepShip dataset. The visualization results reveal that deep convolutional kernels contain two primary patterns: line spectra and background, detecting harmonic line components and continuous spectrum features in LOFAR spectrograms, respectively. After GMM decomposition, the response regions of the sub-patterns become more concentrated, feature expression becomes more consistent, and an association mechanism linking “convolutional kernel - sub-pattern - ship class” is established. This research provides a quantitative basis for understanding CNN decision logic, filling a gap in interpretability studies for ship-radiated noise classification models.
