Autonomous robots are one of the critical components in modern manufacturing systems. For this reason, the uninterrupted operation of robots in manufacturing is important for the sustainability of autonomy. Detecting possible fault symptoms that may cause failures within a work environment will help to eliminate interrupted operations. When supervised learning methods are considered, obtaining and storing labeled, historical training data in a manufacturing environment with faults is a challenging task. In addition, sensors in mobile devices such as robots are exposed to different noisy external conditions in production environments affecting data labels and fault mapping. Furthermore, relying on a single sensor data for fault detection often causes false alarms for equipment monitoring. Our study takes requirements into consideration and proposes a new unsupervised machine-learning algorithm to detect possible operational faults encountered by autonomous mobile robots. The method suggests using an ensemble of multi-sensor information fusion at the decision level by voting to enhance decision reliability. The proposed technique relies on dissimilarity-based sensor data segmentation with an adaptive threshold control. It has been tested experimentally on an autonomous mobile robot. The experimental results show that the proposed method is effective for detecting operational anomalies. Furthermore, the proposed voting mechanism is also capable of eliminating false positives in case of a single source of information is utilized.

自主机器人是现代制造系统的关键组件之一。因此，制造过程中机器人的不间断运行对于自主的可持续性非常重要。检测可能在工作环境中导致故障的故障症状将有助于消除操作中断。当考虑监督学习方法时，在有故障的制造环境中获取和存储带标签的历史训练数据是一项具有挑战性的任务。此外，机器人等移动设备中的传感器在生产环境中暴露于不同的噪声外部条件下，影响数据标签和故障映射。此外，依靠单个传感器数据进行故障检测往往会导致设备监控出现误报。我们的研究考虑了需求，并提出了一种新的无监督机器学习算法来检测自主移动机器人可能遇到的操作故障。该方法建议通过投票在决策层面使用多传感器信息融合的集合来增强决策的可靠性。所提出的技术依赖于基于相异性的传感器数据分割和自适应阈值控制。它已在自主移动机器人上进行了实验测试。实验结果表明，该方法对于检测运行异常是有效的。此外，所提出的投票机制还能够在使用单一信息源的情况下消除误报。