Institute of Acoustics

The accurate classification of infrasound events is significant in natural disaster warning, verification of nuclear test bans and geophysical research. Current deep learning-based classification methods mostly focus on denoised and filtered signals. To simplify the process, avoid information loss, and address the issues of incomplete feature extraction by single-scale convolution kernels and the potential loss of physical information by single models, this paper directly utilizes raw infrasound signals and proposes two fusion classification models based on multi-scale Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). Experiments were conducted on a typical infrasound signal dataset (comprising four signal types: mountain-associated waves, auroral infrasound waves, volcanic eruptions, and microbaroms). The performances of the two models were compared in terms of accuracy, convergence speed, and stability. The results indicate that both models achieve classification accuracies exceeding 99% with optimal parameter combinations. The dual-branch multi-scale CNN-LSTM model generally outperforms the multi-scale CNN-LSTM model in classification accuracy, while also demonstrating faster convergence speed and better stability. Addressing the class imbalance in the dataset, evaluations using precision, recall, and F1-score further validated the effectiveness of the proposed models. This study demonstrates that the proposed methods can effectively achieve end-to-end classification of raw infrasound signals and are competitive with existing techniques.

Reliable nondestructive evaluation of fine weave pierced carbon/carbon (C/C) composites is essential because these materials are increasingly used in critical components, yet ultrasonic inspection is often compromised by dispersion and frequency-selective filtering that distort waveforms and complicate imaging. This study aimed to experimentally characterize the anisotropic acoustic dispersion and frequency-filtering behavior of up-to-date fine weave pierced C/C composites with a pitch-based matrix. Phase velocities along the three principal directions (x, y, z) were measured over a frequency range of 0.5–5.0 MHz. Along the z-direction, phase velocity increases from 7250 m/s to 13,500 m/s with rising frequency, revealing four selective passbands. This indicates pronounced geometric dispersion and a wave-filtering effect due to the larger-scale fibers aligned in this direction. In contrast, the x- and y-directions exhibit only a single low-frequency passband dominated by the strong viscoelasticity of the matrix, with phase velocities of 8100 m/s at 0.5 MHz and 7100 m/s at 0.3 MHz, respectively. Furthermore, temperature-dependent measurements in the z-direction demonstrate a transition from viscoelastic-dominated to geometric-dominated dispersion as temperature increases. These results provide frequency-selection guidance for reliable ultrasonic nondestructive evaluation of advanced C/C composite components.

In recent years, the abuse of antibiotics has led to the emergence of numerous drug-resistant bacteria. Given that single antimicrobial materials can hardly achieve satisfactory inhibition against pathogenic bacteria, it is urgent to develop composite antimicrobial materials and strategies. In this work, inorganic piezoelectric ZnO nanoparticles were used as the core, modified to improve hydrophilicity, coated with polyethylene glycol (PEG), and then conjugated with the sonosensitizer chlorin e6 (Ce6) via amide bonds to form ZnO@PEG@Ce6 (ZGC) nanocomposites. The nanocomposites were characterized by Transmission Electron Microscopy (TEM), Fourier Transform Infrared Spectroscopy (FTIR), cavitation intensity, absorbance, band gap, zeta potential, and particle size analysis. ZGC exhibits an octahedral structure with a uniform particle size distribution, and FTIR verified the stable chemical linkage among components. Synergy between ZnO piezoelectricity and Ce6 sonosensitivity greatly enhances ROS generation and results in superior sonodynamic antimicrobial activity. Gram-negative bacterium Pseudomonas aeruginosa, Gram-positive bacterium Staphylococcus aureus, and Candida albicans were selected as experimental subjects to study the antimicrobial activity and mechanism of the composite material ZGC. The experiment showed that under the action of ultrasound, the inhibition rates of ZGC against P. aeruginosa, S. aureus, and C. albicans were 58.7%, 68.29%, and 94.55%, respectively, which overcomes the disadvantage of the narrow antimicrobial spectrum of traditional sonosensitizers. Scanning Electron Microscope (SEM) images revealed that the synergistic effect of ZGC and ultrasound destroyed the surface structure of pathogenic strains, allowing reactive oxygen species to penetrate and damage the pathogens. ZGC-mediated sonodynamic therapy shows promising applications in chronic infected wound healing and implant-related infections, enabling efficient elimination of drug-resistant bacteria and precise antimicrobial treatment. This study provides insights and theoretical support for the development of broad-spectrum and highly efficient sonodynamic antimicrobial therapy.

To develop and validate a machine learning model integrating ultrasound radiomics and clinicopathological parameters to predict intrahepatic recurrence in colorectal cancer liver metastases (CRLM) patients after curative hepatectomy. This retrospective study enrolled 278 eligible CRLM patients (age, 55 ± 12 years; male, 188) from two centers, including a main cohort (n = 224, July 2010-February 2021) and an external cohort (n = 54, February 2015-October 2020). Patients were stratified by recurrence status during a 2-year follow-up. Preoperative ultrasound images and clinicopathological parameters were collected. Radiomics features were extracted from liver metastases, peri-tumor areas, and disease-free liver parenchyma. Using least absolute shrinkage and selection operator (LASSO) analysis and support vector machine algorithms, three predictive models were developed: clinical, radiomics, and clinical-radiomics combined (cRadiomics) models. Model performance was assessed using five-fold cross-validation (main cohort) and external validation (external cohort), with metrics including receiver operating characteristic (ROC) curve, the area under the ROC curve (AUC), accuracy, sensitivity, and specificity. Six clinical parameters (pathological lymph node positivity, synchronous liver metastases, bilobar liver metastases, preoperative chemotherapy, use of targeted drugs, and preoperative CA19-9 > 200 U/mL) and seven radiomics features were identified as strong predictors. The cRadiomics model achieved AUC values of 0.811 (95% CI: 0.755-0.861) and 0.784 (95% CI: 0.644-0.880) during testing on the main cohort and external cohort data, respectively, significantly outperforming both radiomics (AUC 0.744 and 0.724; p < 0.01) and clinical models (AUC 0.706 and 0.696; p < 0.05). The cRadiomics model, integrating ultrasound radiomics and clinicopathological parameters, improved the prediction of intrahepatic recurrence within two years for colorectal liver metastases after curative hepatectomy. The cRadiomics model, with enhanced accuracy in predicting intrahepatic recurrence of colorectal liver metastases after curative hepatectomy, holds great potential to improve clinical decision-making and enable personalized management and risk-adapted follow-up for colorectal cancer liver metastases (CRLM) patients. The cRadiomics model can predict intrahepatic recurrence of colorectal cancer liver metastases (CRLM) after curative hepatectomy. Machine learning with data integration of ultrasound radiomics and clinicopathological parameters enables better prediction of intrahepatic recurrence for CRLM after curative hepatectomy. The developed model holds great potential to improve clinical decision-making and personalized management for CRLM patients.

Although lip-to-speech synthesis (L2S) has achieved significant progress in recent years, current state-of-the-art methods typically rely on intermediate representations such as mel-spectrograms or discrete self-supervised learning (SSL) tokens. The potential of latent diffusion models (LDMs) in this task remains largely unexplored. In this paper, we introduce SLD-L2S, a novel L2S framework built upon a hierarchical subspace latent diffusion model. Our method aims to directly map visual lip movements to the continuous latent space of a pre-trained neural audio codec, thereby avoiding the information loss inherent in traditional intermediate representations. The core of our method is a hierarchical architecture that processes visual representations through multiple parallel subspaces, initiated by a subspace decomposition module. To efficiently enhance interactions within and between these subspaces, we design the diffusion convolution block (DiCB) as our network backbone. Furthermore, we employ a reparameterized flow matching technique to directly generate the target latent vectors. This enables a principled inclusion of speech language model (SLM) and semantic losses during training, moving beyond conventional flow matching objectives and improving synthesized speech quality. Our experiments show that SLD-L2S achieves state-of-the-art generation quality on multiple benchmark datasets, surpassing existing methods in both objective and subjective evaluations.

Wide-range and high-sensitivity hydrogen sensors are critically important for hydrogen safety in aerospace and advanced transportation sectors. This work demonstrates a thermal-conductivity surface acoustic wave (SAW) based sensor to achieve high sensitivity hydrogen sensing. By integrating thermal balance and acoustic wave equations, a precise mechanistic model elucidating the structure-activity relationships among gas flow rate, operating temperature, and MEMS architecture in determining sensing sensitivity is constructed. Guided by this model, the SAW hydrogen sensor with on-chip microheater integration was developed. Furthermore, a highly integrated SAW hydrogen sensing system with ultra-low baseline noise (<30 µV) was constructed for performance evaluation. Leveraging the exceptional thermal sensitivity of the SAW device and system stability, the optimized sensor achieves wide detection range (up to 100% vol), low detection limit (~6 ppm), rapid response and recovery time (T90/T10: ~15 s), excellent repeatability (error<2.4%) at a relatively low operating temperature (120 °C). The prepared SAW sensor provides an effective solution for hydrogen leakage monitoring across unprecedented concentrations (ppm-100% vol), establishing a new paradigm for hydrogen safety applications.

This paper presents a deconvolution algorithm for high resolution acoustic imaging, specifically tailored for small-element acoustic vector-sensor (AVS) arrays. While existing fourth-order cumulant (FOC) based methods can improve spatial resolution, their localization accuracy remains fundamentally limited by the inadequate spatial sampling of small arrays. To address this limitation, the deconvolution method is developed based on a forward propagation model for sound sources received by AVS arrays. The proposed method is designed to achieve high-precision localization under challenging conditions, specifically at low signal-to-noise ratios (SNR) and in the presence of spatial aliasing. The FOC-based vector beamforming is initially performed to obtain the “dirty map”, with the FOC matrix iteratively updated to progressively remove sidelobe components merged in the main lobe. By incorporating coherence information, the proposed cumulant-based vector CLEAN-SC (CUM-VCLEAN-SC) method effectively mitigates spatial resolution degradation caused by the sidelobes of coherent sources in small-element arrays. Meanwhile, the clean source intensity power is estimated based on the main lobe peak, and iterations continue until convergence, ultimately yielding the high-resolution acoustic imaging results. Simulation results demonstrate that the proposed method significantly enhances imaging resolution and localization accuracy compared to traditional algorithms, while maintaining robustness in low SNR environments and in the presence of spatial aliasing caused by high frequencies. Furthermore, under minimal planar array configurations, the method exhibits superior robustness over traditional acoustic imaging algorithms. Experimental results in a semi-anechoic chamber further confirm the effectiveness of the proposed algorithm with small-element arrays.

Information-centric networking (ICN) uses namebased forwarding and in-network caching to reduce latency and backbone traffic and is therefore a key candidate for future networks. ICN architectures using a Name Resolution Service (NRS) improve scalability and flexibility, but also create new distributed denial-of-service (DDoS) threats against resolution and forwarding resources, such as cache pollution and forwarding exhaustion. We propose a trust-based defense for NRS-based ICN that binds entities to decentralized identifiers, derives multifactor credit scores from resolution and caching behaviors, and uses lightweight credentials to steer rate limiting, caching, and forwarding policies. A federated-learning-based detector uses local observations to identify low-rate attacks that span multiple resolution domains without sharing raw traffic and with only modest extra coordination traffic. ns-3-based simulations under cache pollution, forwarding exhaustion, and hybrid attacks show that our scheme improves detection F1-score by up to about 10 percentage points and maintains request success and cache hit ratios around <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$88-90 \%$</tex> and 70 % during attacks, whereas representative ICN/NDN defenses can drop below 80 % and 65 %.

Underwater ad-hoc networks (UANETs) suffer from slow propagation, high error rates, and sluggish control-plane reaction, making them highly vulnerable to sudden multi-node failures. Stale routes during convergence send packets along obsolete next hops, creating blind zones with cascading losses and severe energy waste, while existing recovery schemes rely on slow signaling, heavyweight computation, or non-trivial overhead that does not fit dynamic underwater conditions. This paper presents UW-LFRR, a lightweight, data-plane-driven fast-recovery mechanism. UW-LFRR constructs a purified local topology view, models failure impact via a severed-branch abstraction, and uses a two-stage relay selection strategy that first searches near the original shortest path and then performs a constrained fallback search inside the affected region. A minimal-overhead firstpacket paving procedure installs temporary forwarding entries along the detour, enabling subsequent packets to bypass the damaged area without additional computation. Results from NS-3 simulations indicate that UW-LFRR achieves loop-free recovery using only local connectivity information with localitybounded computation cost, while maintaining high packetdelivery ratios and low end-to-end delay with substantially reduced control overhead across diverse multi-failure scenarios, demonstrating its practicality for resilient underwater networks.

Butenolides are important features of many bioactive compounds. Current petrochemical-based methods for asymmetric butenolide synthesis are unsustainable. Biocatalysis could enable cheaper, greener, and more sustainable alternatives. However, known butenolide biosynthetic pathways are complex and challenging to harness. Here we show that avenolide, a 4-alkylbutenolide that regulates avermectin production in Streptomyces avermitilis, is assembled from a fatty acyl thioester by a multifunctional flavoenzyme and an iterative cytochrome P450. The flavoenzyme catalyzes homochiral 4-alkylbutenolide formation via successive desaturation, hydroxylation, and lactonization reactions, using dioxygen as the sole stoichiometric reagent. This lays the foundation for development of sustainable biocatalytic asymmetric butenolide synthesis.