Development and Application of a New Fault Diagnosis Method of Circuits by Neural Nets

本研究的目的是开发和应用一种新的电子逻辑电路故障诊断方法。神经网络不仅可以检测电路的故障，而且具有很强的抗故障能力。我们已将这一性质应用于故障诊断。该方法简单，对大多数类型的故障普遍适用。该方法的独特之处在于，它既不需要专门准备的测试数据集，也不需要任何故障诊断字典，而这些性质在其他故障诊断方法中是常见的；以前的故障诊断方法坚持可诊断性的完备性，即要求要么生成能够检测单个故障存在的测试模式，要么证明不可能诊断故障。另一方面，我们的方法逐个列举故障纠正的候选者，尽管诊断可能是错误的。虽然不能保证完全的可诊断性，但该方法允许同时进行简单的故障检测和诊断。因此，我们检查了我们的方法可以捕获哪种故障模型。实验结果表明，可以成功地检测/诊断单个静态故障(包括0/1退化故障和延迟故障)。我们对这一结果感到鼓舞，目前正在调查其原因。因此，为了开发神经网络方法的诊断能力，我们想要开发一种算法来诊断动态故障。我们还希望通过使用多值逻辑来扩展我们方法的目标故障模型，与二值逻辑相比，多值逻辑具有更多的通用表达能力。

The purpose of this research is to develop and apply a new fault diagnosis method of electronic logical circuits. Neural net not only can detect faults of circuits but also it is fault-resistive. We have applied this property to fault diagnosis. The method is simple and found universally applicable to most types of faults. Our method is unique in that it requires neither a specially prepared test data set nor any fault diagnosis dictionary, of which property used to be observed commonly in other fault diagnosis method.Previous approach of fault diagnosis adheres to completeness of the diagnosibility, i.e., it requires either to generate a test pattern that can detect existence of a single fault or to prove that it is not possible to diagnose a fault. On the other hand, our method enumerates candidates of fault corrections one by one, though the diagnosis may be erroneous. Though no complete diagnosibility is guaranteed, our method allows a simple detection and diagnosis of fault simultaneously.Previously we had designed a fault diagnosis algorithm and had obtained an excellent simulation result on small sized circuits. So we have examined what kind of fault model our method can capture. As a result we have observed by experiment that a single static fault (including 0/1 degenerate fault and delay fault) can be detected/diagnosed successfully. We are encouraged by this result and we are currently investigating its reason. As a natural consequence, to exploit the diagnosibility ability in our neural net approach, we want to develop an algorithm to diagnose dynamic fault. We also want to extend target fault models of our method somehow by employing multiple-valued logic, which possesses much versatile expressibility compared to the binary case.