改进CNN和Bi-LSTM的集成化装备故障检测研究

Research on integrated Equipment Fault Detection of improved CNN and BI-LSTM

摘要:

集成化装备的故障检测和健康管理（PHM）已成为装备领域研究的重点，但是由于其集成度高，结构复杂，综合性强等特点，采用常规的检测方法常面临信息多源异构，体量浩大，且实时性难以保证的问题，不仅消耗大量的人力物力，而且需要极强的数据分析及管控能力。为保证准确性、实时性和有效性的统一，研究提出一种基于CNN和Bi-LSTM(双向长短记忆网络，Bidirectional short and long memory network)及其优化算法的故障检测算法，构建了Bi-LSTM-CNN-FCM模型，并通过田纳西-伊斯曼化工过程数据集进行验证。在实验过程中通过观察不同激活函数对模型精度和效果的影响选择合适的激活函数，最终确定在卷积层使用tanh激活函数，在全连接层使用relu激活函数。在确定激活函数后对模型不断优化，在模型末端加入FCM聚类算法，提高了故障检测分类的准确率，最后以准确率和损失值为依据，通过与单一的LSTM模型，CNN模型和LSTM-CNN模型对比，证明该模型的优越性。该模型使得故障检测的准确率提升至98.25%，损失值减少至0.0104，在性能上明显优于其他模型。

Abstract:

Integrated equipment fault detection and health management (PHM) has become the focus of the researches on the equipment, but because of its high integration, complicated structure, the characteristics of comprehensive, the conventional detection method often face a multi-source heterogeneous information, size, and difficult problems to ensure real-time performance, not only consume large amounts of resources, It also requires strong data analysis and control skills. A fault detection algorithm based on CNN and BI-LSTM (Bidirectional Short and Long Memory Network) and its optimization algorithm is proposed to ensure the unity of accuracy, real-time and validity. Bi-lstm-cnn-fcm model was constructed and verified by tennessee-Eastman chemical process dataset. During the experiment, appropriate activation functions were selected by observing the influence of different activation functions on the model accuracy and effect, and finally tanH activation function was determined to be used in the convolution layer and Relu activation function was used in the full connection layer. After the activation function was determined, the model was continuously optimized, and FCM clustering algorithm was added at the end of the model to improve the accuracy of fault detection and classification. Finally, based on the accuracy and loss value, the superiority of the model was proved by comparing with the single LSTM model, CNN model and LSTM-CNN model. This model improves the accuracy of fault detection to 98.25% and reduces the loss value to 0.0104, which is obviously superior to other models in performance.