量子–经典混合神经网络及其故障诊断应用

量子神经网络由于结合了量子计算和神经网络的优点, 近年来受到了广泛的关注. 然而由于目前量子计算 资源受限(如量子比特数、量子逻辑门的保真度等)以及贫瘠高原现象(量子神经网络优化过程中解空间变得平坦时 出现的训练困难)的存在, 量子神经网络当前还难以大规模训练. 针对上述问题, 本文面向量子–经典混合神经网络 模型提出了一种基于无监督学习的特征提取方法. 所采用的无监督学习方法结合了量子自编码器和K-medoids聚类 方法, 可用于多层次结构的特征学习. 该方法创新地利用了K-mediods方法对训练得到的量子自编码器进行聚类, 以 最大化量子自编码器性质的差异. 进一步, 本文在轴承异常检测问题上, 通过实验验证了所提出的无监督特征提取 方法的有效性和实用性, 测试集准确率在二分类、四分类和十分类分别达到100%, 89.6%和81.6%.

Quantum-classical hybrid neural network and its application in fault diagnosis

Combining the merits of quantum computation and neural networks, quantum neural networks (QNNs) have gained considerable attention in recent years. However, because of the limitation of quantum resource (the number of qubits, quantum logic gate fidelity, et al.) and the barren plateau phenomenon (the trainability problem that occurs in quantum neural networks when the landspace turns flat as the optimization is run), it is costly to train QNNs. In this paper, an unsupervised feature learning is proposed for quantum-classical hybrid neural networks to alleviate the problems. The unsupervised feature learning method, which can learn a hierarchy of feature extractors, is introduced by combining quantum autoencoders and K-medoids clustering algorithm. Key to our approach is to use K-medoids clustering to maximize the difference of properties of the trained quantum autoencoders. The effectiveness and practicability of the proposed approach is verified on bearing anomaly detection using numerical simulation on binary, four-class and ten-class classification, achieving 100%, 89.6%, 81.6% on test set, respectively.