基于语义嵌入空间的离心泵未知复合故障识别

Unknown composite fault identification of centrifugal pump based on semantic embedding space

针对离心泵各类型故障相互耦合难以诊断的问题,提出了一种语义嵌入空间的未知复合故障识别框架。首先采用深度学习与加权交叉熵损失建模来提取数据不平衡分布下的离心泵已知单一故障特征。接着采用零样本学习结合语义编码将单一故障属性构建成复合故障语义空间。最后通过全连接神经网络将样本特征映射到语义空间,并通过相似度测量来识别未知复合故障。在自吸和单级单吸离心泵两个数据集上进行验证,结果表明:在复合样本缺失的前提下,该方法对四组数据不平衡分布下的未知复合故障平均识别率为73.63%和82.25%,验证了所提方法的有效性和泛化性。

Here,aiming at the problem of various types faults in centrifugal pump being coupled and difficult to diagnose,a semantic embedding space-based unknown composite fault recognition framework was proposed.Firstly,deep learning and weighted cross entropy loss modeling were used to extract known single fault’s features of centrifugal pump under imbalanced data distribution.Then,zero sample learning combined with semantic encoding was used to construct a composite fault semantic space for single fault attributes.Finally,sample features were mapped to the semantic space with fully connected neural network,and unknown composite faults were identified with similarity measurement.The verification was performed on two datasets of self-suction centrifugal pump and single-stage single-suction one.The results showed that under the premise of missing composite samples,the average recognition rates of unknown composite faults obtained using the proposed method for 4 sets of data under imbalanced distribution are 73.63%and 82.25%to verify the effectiveness and generalization of the proposed method.