Industrial Internet of Things embedded devices fault detection and classification. A case study

Industries transition to the Industry 4.0 paradigm requires solutions based on devices attached to machines that allow monitoring and control of industrial equipment. Monitoring is essential to ensure devices' proper operation against different aggressions. We propose a novel approach to detect and classify faults, that are typical in these devices, based on machine learning techniques that use as features the energy, the processing, and the time consumed by device main application functionality. The proposal was validated using a dataset collected from a testbed executing a typical equipment monitoring application. The proposal machine learning pipeline uses a decision tree-based model for fault detection (99.4% accuracy, 99.7% precision, 99.6% recall, 75.2% specificity, and 99.7% F1) followed by a Semi-Supervised Graph-Based model (99.3% accuracy, 96.4% precision, 96.1% recall, 99.6% specificity, and 96.2% F1) for further fault classification. Those results demonstrate that machine learning techniques, based on easily obtainable metrics, help coping with common device faults.

产业向第四工业革命范式转型，亟需基于连接至机械设备的装置所提供的监控与控制解决方案。监控对于确保设备在面对各种攻击性环境下的正常运作至关重要。本研究提出一种新颖的方法，基于机器学习技术检测和分类这些设备中典型的故障，该技术以设备主要应用功能消耗的能量、处理时间和时间作为特征。该提议通过从执行典型设备监控应用的测试平台收集的数据集进行了验证。提议的机器学习流程采用基于决策树的模型进行故障检测（准确率99.4%，精确率99.7%，召回率99.6%，特异性75.2%，F1分数99.7%），随后采用基于半监督图模型的模型（准确率99.3%，精确率96.4%，召回率96.1%，特异性99.6%，F1分数96.2%）进行进一步的故障分类。这些结果证明了基于易于获取的指标的机器学习技术有助于应对常见的设备故障。