ORION-AE: Multisensor acoustic emission datasets reflecting supervised untightening of bolts in a jointed vibrating structure

Experiments were designed to reproduce the loosening phenomenon observed in aeronautics, automotive or civil engineering structures where parts are assembled together by means of bolted joints. The bolts can indeed be subject to self-loosening under vibrations. Therefore, it is of paramount importance to develop sensing strategies and algorithms for early loosening estimation. The test rig was specifically designed to make the vibration tests as repeatable as possible. The dataset ORION-AE is made of a set of time-series measurements obtained by untightening a bolt with seven different levels. The data have been sampled at 5 MHz on four different sensors, including three permanently attached acoustic emission sensors in contact with the structure, and one laser (contactless) measurement apparatus. This dataset can thus be used for performance benchmarking of supervised, semi-supervised or unsupervised learning algorithms, including deep and transfer learning for time-series data, with possibly seven classes. This dataset may also be useful to challenge denoising methods or wave-picking algorithms, for which the vibrometer measurements can be used for validation. ORION is a jointed structure made of two plates manufactured in a 2024 aluminium alloy, linked together by three bolts. The contact between the plates is done through machined overlays. The contact patches has an area of 12x12 mm^2 and is 1 mm thick. The structure was submitted to a 100 Hz harmonic excitation force during about 10 seconds. The load was applied using a Tyra electromagnetic shaker, which can deliver a 200 N force. The force was measured using a PCB piezoelectric load cell and the vibration level was determined next to the end of the specimen using a Polytec laser vibrometer. The ORION-AE dataset is composed of five directories collected in five campaigns denoted as B, C, D, E and F in the sequel. Seven tightening levels were applied on the upper bolt. The tightening was first set to 60 cNm with a torque screwdriver. After a 10 seconds vibration test, the shaker was stopped and this vibration test was repeated after a torque modification at 50 cNm. Then torque modifications at 40, 30, 20, 10 and 5 cNm were applied. Note that, for campaign C, the level 40 cNm is missing. During each cycle of the vibration test for a given tightening level, different AE sources can generate signals and those sources may be activated or not, depending on the tribological conditions within the contact between the beams which are not controlled. The tightening levels can be used to represent a reference against which clustering or classification results can be compared with. In that case, the main assumption is that the torque remained close to the level which was set at the beginning of every period of 10 s. This assumption can not be checked in the current configuration of the tests. For each campaign, four sensors were used: a laser vibrometer and three different AE sensors (micro-200-HF, micro-80 and the F50A from Euro-Physical Acoustics) with various frequency bands were attached onto the lower plate (5 cm above the end of the plate). All data were sampled at 5 MHz using a Picoscope 4824 and a preamplifier (from Euro-Physical Acoustics) set to 60 dB. The velocimeter is used for different purposes, in particular to control the amplitude of the displacement of the top of the upper beam so that it remains constant whatever the tightening level. The sensors are expected to detect the stick-slip transitions or shocks in the interface that are known to generate small AE events during vibrations. The acoustic waves generated by these events are highly dependent on bolt tightening. These sources of AE signals have to be detected and identified from the data stream which constitute the challenge. Details of the folders and files There is 1 folder per campaign, each composed of 7 subfolders corresponding to 7 tightening levels: 5 cNm, 10 cNm, 20 cNm, 30 cNm, 40 cNm, 50 cNm, 60 cNm. So, 7 levels are available per campaign, except for campaign C for which 40 cNm is missing. There is about 10 seconds of continuous recording of data per level (the exact value can be found according to the number of files in each subfolder). The sampling frequency was set to 5 MHZ on all channels of a picoscope 4824 and a preamplifer of 60 dB (model 2/4/6 preamplifier made by Europhysical acoustics). The characteristics of both the picoscope and preamplifier are provided in the enclosed documentation. Each subfolder is made of .mat files. There is about 1 file per second (depending on the buffering, it can vary a little). The files in a subfolder are named according to the timestamps (time of recording). Each file is composed of vectors of data named: A = micro80 sensor. B = F50A sensor. C = micro200HF sensor. D = velocimeter. Note that the measurements are stored in mV. Sample Matlab codes are provided to read the files provided. The characteristics of the sensors are provided in the enclosed documentation.

本实验旨在复现航空、汽车或土木工程领域中采用螺栓连接（bolt joint）装配的结构所观测到的松动现象——此类结构中的螺栓在振动载荷下会发生自松动，因此开发用于早期松动检测的传感策略与算法至关重要。
为尽可能提升振动测试的可重复性，研究团队专门设计了测试台。ORION-AE数据集包含通过七个不同拧紧梯度松开螺栓所采集的多组时间序列测量数据。
所有数据以5 MHz的采样率通过四个传感器完成采集：其中三个为永久附着在结构上的声发射（Acoustic Emission, AE）传感器，另一个为非接触式激光测量装置。该数据集可用于监督、半监督或无监督学习算法（包括面向时间序列数据的深度学习与迁移学习）的性能基准测试，共包含7个类别。此外，该数据集还可用于挑战降噪方法或波拾取算法，此时可借助激光测振仪的测量结果进行验证。
ORION结构由两块采用2024铝合金制造的板材通过三颗螺栓连接而成，板材间通过机加工的搭接面实现接触。接触区域面积为12×12 mm²，厚度为1 mm。结构受到持续约10秒的100 Hz谐波激励载荷，该载荷由Tyra电磁激振器（electromagnetic shaker）施加，最大可输出200 N的力。载荷通过PCB压电式力传感器（piezoelectric load cell）进行测量，同时使用Polytec激光测振仪（laser vibrometer）在试样端部附近测定振动水平。
ORION-AE数据集包含五个测试批次（后续记为B、C、D、E、F）对应的五个目录。实验在上部螺栓上设置了七个拧紧梯度：首先使用扭矩螺丝刀将拧紧扭矩设置为60 cNm，完成10秒振动测试后关闭激振器，调整扭矩至50 cNm后重复振动测试；随后依次调整扭矩至40、30、20、10和5 cNm并重复测试。需注意：批次C缺少40 cNm这一梯度。
在给定拧紧梯度的振动测试周期中，不同的声发射源会产生信号，而这些信号的激活与否取决于梁接触界面间未受控的摩擦学条件。拧紧梯度可作为参考基准，用于对比聚类或分类结果。本场景的核心假设为：在每一轮10秒测试周期内，拧紧扭矩始终保持在初始设定值附近，但在当前的测试配置下，该假设无法得到验证。
每个测试批次均使用四个传感器：一台激光测振仪，以及三台不同的声发射传感器（Euro-Physical Acoustics公司的micro-200-HF、micro-80与F50A），三台传感器均安装在下部板材上（距离板材端部5 cm处）。所有数据均通过Picoscope 4824采集设备与增益设置为60 dB的前置放大器（来自Euro-Physical Acoustics）以5 MHz的采样率完成采样。激光测振仪还可用于额外用途：具体而言，用于控制上部梁顶端的位移幅值，使其在不同拧紧梯度下均保持恒定。
上述传感器可检测界面间的粘滑转变或冲击，这类现象在振动过程中会产生小型声发射事件。此类事件产生的声波与螺栓拧紧状态高度相关，从数据流中检测并识别这类声发射信号源正是本数据集所要应对的核心挑战。
文件夹与文件详情
每个测试批次对应一个文件夹，每个文件夹包含7个子文件夹，分别对应7个拧紧梯度：5 cNm、10 cNm、20 cNm、30 cNm、40 cNm、50 cNm、60 cNm。因此除批次C缺少40 cNm梯度外，每个批次均包含7个梯度。每个梯度对应约10秒的连续数据记录，精确时长可通过每个子文件夹中的文件数量计算得出。所有通道的采样率均设置为5 MHz，采集设备为Picoscope 4824，搭配Europhysical acoustics公司生产的2/4/6型前置放大器，增益设置为60 dB。采集设备与前置放大器的详细参数可随附文档中查阅。每个子文件夹均由.mat格式的文件组成，每秒约生成1个文件（受缓存机制影响可能存在小幅波动）。子文件夹中的文件以录制时间戳命名。每个文件包含四个数据向量，分别命名为：A = micro80传感器，B = F50A传感器，C = micro200HF传感器，D = 激光测振仪。所有测量数据均以毫伏（mV）为单位存储。随附提供了用于读取数据文件的MATLAB示例代码。传感器的详细参数可随附文档中查阅。