基于卷积变分自编码和多头自注意力机制的断路器剩余机械寿命预测

针对万能式断路器退化过程的不确定性,考虑到振动信号对机械性能退化的完善表征,提出了一种基于卷积变分自编 码(CVAE)和多头自注意力机制(MSA)的断路器分闸机械机构寿命预测方法。 首先依据断路器不同的事件区间提取参数特 征,再通过 CVAE 挖掘信号成分中的深度特征,将参数特征与深度特征融合得到完备退化特征,最后建立 GRU-MSA 的定量寿 命预测模型,引入了多头自注意力机制,在多个不同表征子空间中捕捉信号的不同依赖关系,对重要的时间步赋予更大的权重。 最后利用 3 台试品的振动信号测量数据对所提断路器分闸机械机构寿命预测方法进行测试,结果表明,所提出的方法在 3 个数 据集中寿命预测均方根误差(RMSE)分别为 141. 46、128. 75 和 134. 16,平均绝对误差(MAE)分别为 112. 17、101. 52 和 106. 22, 预测精度高且稳定性好,相对于其他混合预测模型更具优势。

Remaining mechanical useful life prediction for circuit breaker based on convolutional variational autoencoder and multi-head self-attention

Arming at the uncertainty of the degradation of conventional circuit breakers and the perfect mechanical degradation characterization by vibration signals, an opening mechanical mechanism life prediction method based on CVAE and MSA mechanism is proposed. Firstly, the parametric features are extracted based on the different event intervals of the circuit breaker. Then, the depth features in the signal components are mined by CVAE, and the parametric features are fused with the depth features to obtain the complete degradation features. Finally, the quantitative life prediction model of the GRU-MSA is formulated, which introduces MSA to capture the different dependencies of signals in several different representation subspaces and assign greater weights to the important time steps. Finally, the proposed method is tested by using the vibration signal measurement data of three test samples. The results show that the proposed method has life prediction RMSE of 141. 46, 128. 75, and 134. 16, and MAE of 112. 17, 101. 52, and 106. 22, respectively. The prediction accuracy is high and the stability is good, which has more advantages compared with other hybrid prediction models.