无人系统监听模式的模型开发音频及测试结果评估数据

本数据集是面向水下目标监测无人系统的“常开监听模式”（Always-on）研发与验证的基准数据集，旨在支持高能效AI处理器在水声信号二分类任务（有用信号与背景噪声区分）中的模型训练与性能评估。数据集内容涵盖了设计阶段的合成音频数据与测试阶段的实地水声测试数据。设计数据基于ModSignalGenerator程序模块生成，包含涵盖2FSK、OFDM等多种调制方式的信号以及船舶辐射噪声，通过受控注入高斯白噪声与设备背景噪声，生成了信噪比（SNR）覆盖0.5dB至70dB区间的标准化基带信号，采样率为44.1kHz。测试数据采集于南京市东南大学信息大楼实验室水池，包含通过水听器接收的实地测试音频及对应的系统推理结果。本数据集的数据结构依据算法开发与系统测试流程组织，主要包含五类关键文件：一是用于模型训练的信号类（多种水声调制信号）与噪声类（多种船舶辐射噪声）标准化音频文件；二是用于数据集划分的索引文件；三是用于二分类任务的浮点模型训练代码（Python）与定点模型（8-bit及1-bit量化）推理代码；四是系统测试运行的C代码；五是详细记录模型在浮点、8-bit及1-bit量化条件下准确率的测试结果文件。本数据集面向水声通信、边缘计算及低功耗芯片设计领域，支持用户在Linux或Windows环境下配合VSCode及Python等开发工具使用。通过使用本数据集，研究人员可评估端到端深度学习算法在水下环境的监测性能，验证低比特量化技术在高能效无人系统中的应用潜力。This benchmark dataset is developed for the research and validation of the "Always-on" listening mode of unmanned systems for underwater target monitoring, aiming to support model training and performance evaluation of high-efficiency AI processors for the binary classification task of underwater acoustic signals (distinguishing useful signals from background noise). The dataset includes synthesized audio data from the design phase and field underwater acoustic test data from the testing phase. The design data is generated based on the ModSignalGenerator program module, including signals with various modulation modes such as 2FSK and OFDM, as well as ship radiated noise. By controlled injection of additive white Gaussian noise and equipment background noise, standardized baseband signals with signal-to-noise ratio (SNR) ranging from 0.5 dB to 70 dB are generated, with a sampling rate of 44.1 kHz. The test data was collected in the laboratory pool of the Information Building of Southeast University in Nanjing, including field test audio received by hydrophones and corresponding system inference results. The data structure of this dataset is organized according to the algorithm development and system testing workflow, and mainly includes five categories of key files: 1. Standardized audio files of signal type (various underwater acoustic modulation signals) and noise type (various ship radiated noises) for model training; 2. Index files for dataset splitting; 3. Floating-point model training code (Python) and fixed-point model inference code (8-bit and 1-bit quantization) for the binary classification task; 4. C code for system test operation; 5. Test result files that detailedly record the accuracy of the model under floating-point, 8-bit and 1-bit quantization conditions. This dataset is targeted at the fields of underwater acoustic communication, edge computing and low-power chip design, and supports users to use it with development tools such as VSCode and Python in Linux or Windows environments. By using this dataset, researchers can evaluate the monitoring performance of end-to-end deep learning algorithms in underwater environments, and verify the application potential of low-bit quantization technology in high-efficiency unmanned systems.

搜集汇总

背景与挑战

背景概述

该数据集是为水下目标监测无人系统的'常开监听模式'研发与验证而设计的基准数据集，包含合成音频和实地水声测试数据，支持二分类任务的模型训练与性能评估。它涵盖了多种调制信号和噪声，信噪比范围广泛，并提供浮点与量化模型的代码及测试结果，适用于水声通信、边缘计算和低功耗芯片设计领域。