基于 BiLSTM-Attention 的迁移学习变工况故障识别方法研究

针对传统深度学习网络模型在变工况条件下的故障诊断泛化能力差的问题，提出一种基于迁移学习的双向长短时记忆网络和注意力机制(TLBA)融合的故障识别方法。将原始故障数据划分为源域及目标域；并构建融合注意力机制的双向长短时记忆网络(BiLSTM-Attention, BA)模型，之后使用此模型学习源域数据特征；最后利用迁移学习通过对目标域数据的学习，进一步优化调整BA模型的网络参数，最终得到目标域的故障分类辨识模型。以航空器翼梁故障为案例，结果表明，该方法与传统故障诊断方法BiLSTM-Attention相比，其综合评价指标F_1-score有3.4%的提高，故障平均诊断准确率在91%以上；同时针对变工况下的故障分类结果较为稳定。

Transfer learning based on BiLSTM-Attention research on fault identification methods for variable operating conditions

Aiming at the problem of poor generalization ability of fault diagnosis of traditional deep learning network model under variable working conditions, a fault identification method based on the fusion of transfer learning bidirectional long short memory network and attention mechanism ( TLBA) is proposed. Divide the original fault data into source domain and target domain, and construct a bidirectional long short-term memory network (BA) model that integrates attention mechanisms, and then use this model to learn source domain data features. Finally, transfer learning is used to further optimize and adjust the network parameters of the BA model by learning the data in the target domain, and finally the fault classification identification model in the target domain is obtained. Taking the aircraft wing beam fault as an example, the results show that compared with the traditional fault diagnosis method BiLSTM-Attention, the comprehensive evaluation index F1-score of this method is improved by 3. 4%, and the average fault diagnosis accuracy is above 91%. At the same time, the fault classification results under variable operating conditions are relatively stable.