Road owners are concerned with the state of the road surface and they try to reduce noise coming from the road as much as possible. Using sound level measuring equipment installed inside a car, we can indirectly measure the road pavement state. Noise inside a car is made up of rolling noise, engine noise and other confounding factors. Rolling noise is influenced by noise modifiers such as car speed, acceleration, temperature and road humidity. Engine noise is influenced by car speed, acceleration, and gear shifts. Techniques need to be developed which compensate for these modifying factors and filter out the confounding noise. This paper presents a hierarchical clustering method resulting in a mapping of the road pavement quality. We present the method using a dataset recorded in multiple cars under different circumstances. The data has been retrieved by placing a Raspberry Pi device within these cars and recording the sound and location during various trips at different times. The sound data of our dataset was then corrected for correlation with speed and acceleration. Furthermore, clustering techniques were used in order to estimate the type and condition of the pavement using this set of noise measurements. The algorithms were run on a small dataset and compared to a ground truth which was derived from visual observations. The results were best for a combination of Generalised Additive Model (GAM) correction on the data combined with hierarchical clustering. A connectivity matrix merging points close to each other further enhances the results for road pavement quality detection, and results in a road type detection rate around 90%.

中文翻译：

---

使用统计聚类方法检测路面质量

道路所有者非常关心路面状况，他们会尽量减少来自道路的噪音。使用安装在车内的声级测量设备，我们可以间接测量道路路面状态。车内噪声由滚动噪声、发动机噪声和其他混杂因素组成。滚动噪音受噪音调节剂的影响，例如车速、加速度、温度和道路湿度。发动机噪音受车速、加速度和换档的影响。需要开发技术来补偿这些修改因素并过滤掉混杂的噪音。本文提出了一种分层聚类方法，用于绘制道路路面质量。我们使用在不同情况下记录在多辆汽车中的数据集来介绍该方法。通过在这些汽车中放置 Raspberry Pi 设备并在不同时间的各种旅行中记录声音和位置，已检索到数据。然后我们数据集的声音数据被校正为与速度和加速度的相关性。此外，为了使用这组噪声测量来估计路面的类型和状况，使用了聚类技术。这些算法在一个小数据集上运行，并与从视觉观察中得出的基本事实进行比较。对数据的广义加性模型 (GAM) 校正与分层聚类相结合的结果是最好的。一个连接矩阵合并彼此靠近的点，进一步增强了道路路面质量检测的结果，使道路类型检测率达到 90% 左右。