贫数据中基于模型自训练的空气处理设备故障诊断

Fault Detection and Diagnosis of Air Handling Unit via Model Self-training in Poor-data Scenario

摘要:

针对空气处理设备（AHU）故障贫数据，基于深度置信网络（DBN）模型对4种特征选择算法进行对比研究，结果表明最大相关最小冗余算法的特征子集在诊断准确率及子集元素稳定性上表现最优。提出将DBN 嵌入自训练框架的故障诊断模型，发现DBN自训练的诊断准确率较单纯DBN最高可提升19.5%。提出均匀抽样及按比例抽样2种自训练伪标签抽样策略，二者的诊断准确率均随抽样数减小而增大，在不同抽样数中的最大差异为3.42%；在所有贫数据样本中，均匀抽样策略始终优于按比例抽样，诊断准确率最大相差1.39%，表明在故障标签匮乏时，采用均匀抽样策略及较小的抽样数有利于提升DBN自训练的诊断性能。

Abstract:

A comparative analysis was conducted to evaluate four feature-selection algorithms in the context of diagnosing air handling unit （AHU） faults using deep belief network （DBN） with poor-data. The results indicate that the feature subset filtered by the maximum correlation minimum redundancy algorithm exhibits superior performance in terms of diagnostic accuracy and stability. Subsequently， a fault diagnosis model was developed by integrating DBN into a self-training framework， and a case study was performed to validate its efficacy. The findings demonstrate that the diagnosis accuracy of DBN self-training model can be improved by up to 19.5% than that of pure DBN. Furthermore， two self-training pseudo-label sampling strategies， namely uniform sampling and proportional sampling， were proposed. While both strategies contribute to increased diagnostic accuracy with a reduction in sampling number， the maximum difference observed among different sampling numbers is 3.42%. Notably， the uniform sampling strategy consistently outperforms the proportional sampling strategy， with a maximum accuracy difference of 1.39% across all scenarios with poor-data， which indicates that， in situations where the fault labels are seriously lacking， the uniform sampling strategy with the smaller sampling number is beneficial to improve the diagnosis performance of DBN self-training model.