Distinguishing Sensor Faults from System Faults, and optimal placement of redundant sensors.

自动故障检测和诊断（FDD）系统依赖于传感器读数的可靠性。 在特定操作条件下传感器读数的不一致可能不一定是传感器本身的故障，而是被监控系统中更严重故障的症状，反之亦然。因此，系统和传感器故障可能会表现出相同的症状。识别准确故障源的能力在系统监控中至关重要，因为根据故障是来自传感器还是来自系统，需要采取不同的纠正措施。对于传感器和系统的故障检测和诊断，有大量的文献。尽管诊断方案的实际应用的重要性，传感器和系统故障之间的区分似乎并没有得到大量的关注，在监测和诊断文献。大多数当前的研究和检测故障的方法做出以下隐含假设之一： - 传感器功能齐全，因此故障可归因于系统。 - 系统功能齐全，故障可归因于传感器。在不了解系统特性的情况下，传感器读数与系统模型的差异可能会被错误地解释为监控传感器中的潜在故障。例如，如果在没有系统的一些知识的情况下使用极限检查来验证传感器测量，则诊断可能被错误地视为传感器故障。相反，一个简单的传感器故障可能被诊断为系统故障，并触发不必要的纠正措施。使用多个冗余传感器是改善这种情况的一种方法，并且它可以便于区分传感器故障和系统故障。然而，成本，物理和实际的约束限制了慷慨的冗余传感器的放置。虽然成本并不总是一个问题，但传感器的放置仍然应该是明智的，并基于科学原则，以避免复杂性。为了解决这个问题，通过旨在确定传感器冗余的最小程度，我们已经使用了先验知识的物理关系的监测变量来验证现有的传感器观测的可信度。随后，我们已经开发了一个冗余的传感器放置方法的系统，其变量可以建模为一个串联的因果网络。通过推导得出的结论是，如果传感器（基本的和冗余的）的数量大于被监测变量的数量的1.5倍，则可以确定地完成区分传感器和系统故障的任务。该研究旨在通过形式化手段证明所开发的方法，并将其扩展到具有多输入和多输出的一般网络（系统框图），而不限制互连形式。

Automated Fault Detection and Diagnosis (FDD) systems depend on the reliability of sensor readings. An inconsistency in a sensor's readings under specific operating conditions may not necessarily be a fault in the sensor itself, but a symptom of a more serious fault in the monitored system, and vice versa. Hence, system and sensor faults might manifest themselves with the same symptoms. The ability to identify the exact source of faults is crucial in the monitoring of a system because different corrective actions are required depending on whether the fault is from a sensor or from the system. There is an abundance of literature on fault detection and diagnosis for both sensors and systems individually. Despite the importance of the practical application of diagnostic schemes, distinguishing between sensor and system faults does not appear to have received substantial attention in the monitoring and diagnosis literature. Most current studies and methodologies of detecting faults make one of the following implicit assumptions: - Sensors are fully functional, so faults are attributable to the system. - The system is fully functional, and faults are attributable to the sensors.Without knowledge of the system characteristics, the discrepancy of sensor readings from the system model may erroneously be interpreted as potential faults in the monitoring sensors. For instance, if limit checking is used to validate the sensor measurement without some knowledge of the system, the diagnosis might wrongly be taken as a sensor fault. Conversely, a simple sensor fault might be diagnosed as a system fault and trigger unnecessary corrective actions. Using multiple redundant sensors is one way of improving the situation, and it can facilitate distinguishing sensor faults from the system faults. However, cost, physical, and practical constraints limit generous placement of redundant sensors. Although cost is not always an issue, the sensor placement should still be judicious and based on scientific principles to avoid complexity. To address this issue, by aiming to identify the minimum degree of sensor redundancy, we have used a priori knowledge of physical relationships between the monitored variables to verify the credibility of the existing sensor observations. Subsequently, we have developed a redundant sensor placement methodology for systems whose variables can be modeled as a serially connected causal network. The generalization of which by deduction has revealed that if the number of sensors (essential and redundant) is greater than 1.5 times the number of monitored variables the task of distinguishing between sensor and system faults can be accomplished with certainty. The proposed research aims to prove the developed method through formal means and extend it for general networks (system block diagrams) with multiple-input and multiple-output without any restriction on the form of interconnections.