Acoustic and deep features extracted from acoustic emission signals for the characterization of biocomposites and glass fiber epoxy composites

INTRODUCTION
The dataset contains 12 Excel files with extracted features for the characterization of loaded materials and classification of the source material as biocomposite and glass fiber epoxy composite. Each file contains features for 22323 samples, extracted from acoustic emission measurements on loaded composites. The dataset was collected to examine various configurations of convolutional autoencoders for deep feature extraction, and different machine learning methods for the classification of the source material.

INPUTS
The inputs of this dataset are composed of two types of extracted features:
1. Standard features
2. Deep features

Standard features are extracted from acoustic emission signals as follows:
c1: Peak amplitude [nm] – burst signal linear peak amplitude,
c2: Burst signal rise-time [µs] – elapsed time after the first threshold crossing and until the burst signal maximum amplitude,
c3: Burst signal duration [µs] – elapsed time after the first and until the last threshold crossing of a burst signal,
c4: Burst signal energy [au],
c5: RMS Background noise [µV],
c6: Counts – Number of positive threshold crossings [/],
c7: Spectral centroid [Hz],
c8: Frequency of the max. amplitude of Fourier transform spectrum [Hz],
c9: Frequency of the max. amplitude of continuous wavelet transformation (using the complex Morlet wavelet) [Hz],
c10: Partial power of Fourier spectrum between 0 and 75 kHz [/],
c11: Partial power of Fourier spectrum between 75 and 150 kHz [/],
c12: Partial power of Fourier spectrum between 150 and 300 kHz [/],
c13: Partial power of Fourier spectrum between 300 and 475 kHz [/].

Deep features are extracted by using the convolutional autoencoders (CAE) as described in the accompanying paper.
Deep features are extracted automatically in an unsupervised manner and therefore don’t have physical meaning but are designed to minimize information loss of the input-output mapping of the CAE. Deep features are denoted as: d1, d2, d3, etc.

OUTPUT
The outputs in this dataset are collected in the last column of each file and denote the class (labels_bio-0_cfe-1) as:
Class = 0: biocomposite
Class = 1: fiber epoxy composite

FILE NAME DESCRIPTION
File names are composed of CAE1 or CA2 prefixes denoting two types of used CAEs, and the remaining part of the file name defines the CAE hyperparameters as “s1-s2-s3-s4-s5-s6-s7”, which stand for:
s1: Number of kernels used in 1st convolutional layers,
s2: Number of kernels used in 2nd convolutional, and 1st transposed convolutional layers,
s3: Number of kernels used in convolutional and transposed convolutional layers of the latent section,
s4: Number of neurons in 2nd and 4th FC layer,
s5: Number of neurons in the bottleneck (3rd FC) layer. This number also represents the number of extracted deep features per input image.
s6: Number of training epochs,
s7: The batch size of training samples.

引言
本数据集包含12个Excel文件，其中提取了用于表征负载材料及将原材料分类为生物复合材料和玻璃纤维环氧树脂复合材料的特征。每个文件包含22323个样本的特征，这些特征是通过在负载复合材料上进行的声发射测量提取的。数据集的收集旨在考察卷积自编码器（CAE）的各种配置以进行深度特征提取，以及不同机器学习方法在原材料分类中的应用。

输入
本数据集的输入由两种类型的提取特征组成：
1. 标准特征
2. 深度特征

标准特征如下提取自声发射信号：
c1: 峰值幅度 [nm] – 爆发信号的线性峰值幅度，
c2: 爆发信号上升时间 [µs] – 首次阈值穿越后至爆发信号最大幅度的经过时间，
c3: 爆发信号持续时间 [µs] – 首次至最后一次阈值穿越的爆发信号的经过时间，
c4: 爆发信号能量 [au]，
c5: 均方根背景噪声 [µV]，
c6: 计数 – 正阈值穿越次数 [/]，
c7: 频谱中心 [Hz]，
c8: 傅里叶变换频谱最大幅度频率 [Hz]，
c9: 连续小波变换（使用复Morlet小波）频谱最大幅度频率 [Hz]，
c10: 0至75 kHz之间的傅里叶频谱部分功率 [/]，
c11: 75至150 kHz之间的傅里叶频谱部分功率 [/]，
c12: 150至300 kHz之间的傅里叶频谱部分功率 [/]，
c13: 300至475 kHz之间的傅里叶频谱部分功率 [/]。

深度特征通过使用在附带论文中描述的卷积自编码器（CAE）提取。
深度特征以无监督方式自动提取，因此不具有物理意义，但旨在最小化CAE输入-输出映射的信息损失。深度特征表示为：d1, d2, d3等。

输出
本数据集中的输出收集在每个文件的最后一列，表示类别（标签_bio-0_cfe-1）如下：
类别 = 0：生物复合材料
类别 = 1：纤维环氧树脂复合材料

文件名描述
文件名由表示两种不同CAE类型的CAE1或CA2前缀组成，文件名剩余部分定义了CAE超参数，格式为“s1-s2-s3-s4-s5-s6-s7”，其中：
s1: 第一卷积层的核数量，
s2: 第二卷积层和第一转置卷积层的核数量，
s3: 潜在部分的卷积和转置卷积层的核数量，
s4: 第二和第四全连接层的神经元数量，
s5: 瓶颈（第三全连接）层的神经元数量。此数字还代表每个输入图像提取的深度特征数量，
s6: 训练周期数，
s7: 训练样本的批量大小。