Abstract:

Aiming at the problem that the existing fault prediction model does not have the ability to identify unknown faults， it is necessary to re-collect data to train the model or learn and identify unknown faults with the help of other components， an unknown fault prediction method based on the CapsNet model is proposed. This method can effectively process the multi-dimensional state sensing signals of complex equipment， realize the accurate sensing of equipment faults， and adaptively adjust the model and predict unknown faults when unknown faults occur. The conversion matrix is constructed to predict the existence and posture of the corresponding high-level features from the features of the low-level capsules. The process by which the dynamic routing algorithm integrates the prediction vector generated by the low-level capsule into the high-level capsule that agrees with it and forms the feature vector is described in detail. In the process of implementing fault feature classification in the last layer of CapsNets， a threshold judgment method is proposed. By reasonably selecting the value range of the threshold， the capsule network model can perfectly distinguish between known and unknown faults and realize accurate fault prediction. The proposed method is used to verify the performance of a series of well-trained CapsNets models. Through experiments， it can be found that the proposed method can better realize unknown fault prediction， which can prove the feasibility of the method.

Key words: CapsNet algorithm, unknown fault, fault prediction, accurate perception, complex equipment

摘要：

针对现有故障预报模型自身不具备未知故障识别能力，需重新收集数据对模型进行训练或借助其他组件对未知故障进行学习识别的问题，提出基于CapsNet模型的未知故障预报方法。该方法可有效处理复杂装备的多维状态感知信号，实现装备故障的精确感知，在出现未知故障时可自适应地调整模型并对未知故障进行预报。构建转换矩阵，由低层胶囊的特征预测出相对应的高层特征的存在及姿态。详细介绍了动态路由算法将低层胶囊生成的预测向量整合到对其表示同意的高层胶囊并形成特征向量的过程。在CapsNet最后一层胶囊实现故障特征分类的过程提出阈值判断法，通过合理选择阈值\lambda的取值范围，使胶囊网络模型能完美地区分已知与未知故障，实现故障的精确预报。使用提出的方法对经过良好训练的系列CapsNets模型进行性能验证。通过实验可以发现，提出的方法能较好地实现未知故障预报，可证实该方法的可行性。

关键词: CapsNet算法, 未知故障, 故障预报, 精确感知, 复杂装备