基于反事实推理的柴油机故障诊断

Fault Diagnosis in Diesel Engines Based onCounterfactual Reasoning

摘要:

针对柴油机故障因果关系的隐蔽性、网络模型向诊断模型转换过程中存在的不确定性，以及传统反事实推理框架在处理复杂系统时推理效率低的问题，提出了一种基于Leaky Noisy-OR假设和孪生网络的因果推理模型，以明确定量因果关系。首先，构建故障因果网络，并融入Leaky Noisy-OR假设和隐变量节点，以表示不确定性因素，构建了一个适用于反事实推理的因果推理模型；其次，将该模型转换成孪生网络，并采取反事实干预措施进行简化，提高了推理效率；最后，定义了故障充分因和故障必要因这两个指标，以定量衡量故障原因和故障征兆之间的因果关系强度，并且在推理的准确性上构成了双重保障；实验结果表明，该方法克服了传统相关性分析的局限性，为复杂系统故障诊断提供了可解释的因果推理范式。

Abstract:

A causal inference model based on Leaky Noisy-OR assumptions and twin networks is proposed to address the uncertainties in the process of the hidden causality of diesel engine fault, the network model to diagnostic model conversion, and the low inference efficiency of the traditional counterfactual inference framework when dealing with complex systems. Firstly, the fault causal network is constructed, and a causal inference model suitable for counterfactual reasoning is constructed by incorporating the Leaky Noisy-OR assumption and hidden variable nodes into the model to represent the uncertainty factors. Secondly, the model is converted into a twin network and simplified with counterfactual interventions, which improves the reasoning efficiency. Lastly, the definition of the two metrics, fault sufficient cause and fault necessary cause, to quantitatively measure the strength of the causal relationship between the cause of the fault and the signs of the fault, and constitutes a double guarantee in terms of the accuracy of the reasoning. Experimental results demonstrate that this method overcomes the limitations of conventional correlation analysis and provides an interpretable causal reasoning paradigm for complex system fault diagnosis.