基于量子自组织神经网络的汽车三元催化器故障诊断方法

A research on the fault diagnosis of the three-way catalytic converter, via extracting the optimized fault features from the big data of automobile exhaust and employing a quantum self-organization neural network with semi-supervised learning, is carried out to solve the shortcomings appeared in the model-based diagnostic methods under the OBD framework, such as single species, simplified model, difficult integration, poor commonality and so on. This research meets the actual needs of energy conservation, and is beneficial to the heritage and development of the automotive fault diagnosis technology. Four contents contribute to this research: to generate the optimized fault features of the three-way catalytic converter having high-dimensional feature space, to quantized the fault features in the real space, to design a quantum self-organizing neural network with flexible neighborhood radius, to design a fault classifier using quantum self-organizing neural network with semi-supervised learning. Through an estimation of the compression dimension and a definition of the projection matrix, the dimension of the high-dimensional fault features for the three-way catalytic converter is reduced by the multiple random projection method. The optimal fault features are searched via a heuristic strategy under the constraints of scatter matrix. An unidirectional conversion function is constructed to achieve the “two-qubits” description of the real fault features. The neurons of self-organizing neural network are replaced by qubit neurons for building the topology structure of the quantum neural network. The flexible neighborhood radius, in the proposed quantum competitive learning algorithm, is derived from the distance and similarity between the quantized weights and the quantum excited states. After identifying the fault feature which fault type has been known, the supervised learning policy is formulated by the constraints of a label attraction domain and a punish function. Then, the unidentified fault features, under the semi-supervised learning framework, are fed to the quantum neural network for implementing the fault classification.

提取尾气大数据中三元催化器的优化故障特征，采用量子自组织神经网络对其进行半监督学习的故障诊断，以解决OBD框架下基于模型的诊断方法中模型简化单一、难于有机整合、通用性较差等问题，符合节能减排的需要，有益于汽车故障诊断技术的传承与发展。课题围绕三元催化器高维故障特征的优化生成及其量子化描述、具有弹性邻域半径的量子自组织神经网络设计、半监督学习的故障分类器设计等内容开展研究。通过估算压缩维数与定义投影矩阵，主动地多次使用随机投影实现高维故障特征的降维，在散布矩阵约束下采用启发式策略搜索最优组合特征；构造单向转换函数，完成实向量故障特征的双量子比特描述；用量子比特神经元替换自组织网络结构中的神经元，由量子化权值与量子激发态之间的距离和相似度共同定义量子竞争学习规则的弹性邻域半径；基于已标识的故障特征，制定标签吸引域和惩罚函数约束下的监督学习策略，采用量子自组织神经网络实施未标识特征的故障分类。

采用量子自组织神经网络对汽车三元催化器进行尾气大数据下的故障诊断，以解决OBD框架下基于模型的诊断方法中模型简化单一、难于有机整合、通用性较差等问题，在汽车故障诊断技术领域可借鉴的相关成果较少，且涉及到化学、热力学、数理统计、量子力学、神经计算、模式识别等多学科内容，是一项具有挑战性和前瞻性的研究课题。项目围绕三元催化器以尾气表征的高维故障特征的优化生成、实空间优化故障特征量子态描述、具有弹性邻域半径的量子自组织神经网络设计、三元催化器在量子自组织神经网络下的半监督学习故障分类器设计等研究内容，面向汽车三元催化器故障诊断的基础理论、设计方法等科学问题与关键技术开展基础研究，取得了阶段性成果。1）从大数据的角度，对汽车三元催化器故障的高维特征优化生成与量子态演化机理进行了系统研究，在一定程度上解决了“等距效应”而使故障特征不易区分的问题，提出类间/类内散布矩阵的可分性评估方法，以及故障动态特征的时序稳定性和收敛验证方法。2）从量子神经计算的角度，设计汽车三元催化器故障分类器，探索了半监督学习机制在量子神经网络中的应用，解决了故障特征的量子态描述与弹性领域半径设计这两个难题，提出量子神经计算下的自组织分类器设计方法，以及量子态时序全局稳定性证明方法。3）从模式识别和机器学习的角度，探索了汽车三元催化器故障数据的概率空间生成模型，提出了特征优化和分类的一般方法，并将这些方法与GPU/CPU异构计算相结合，应用用于故障诊断相类似的模式识别领域，例如，语音识别，图像中的障碍物识别，自然语言处理，图像中的人脸识别等领域，取得良好效果。

内容获取失败，请点击重试