Abstract:

Fault prediction can indicate abnormal changes in variables and predict fault conditions in advance. Existing fault prediction methods primarily consider the global temporal dependencies of the complete sequence, which neglecting the dependencies between variables and the distinct local temporal features in the sampled subsequences. To address the above issues, a fault prediction architecture based on multi-sampled sequence feature extraction network (MSFEN) is proposed. First, a batch joint embedding mechanism is designed to better express the dependencies between variables while considering batch periodicity. Then, a sequence sampling mechanism is developed to divide the complete time series into sampled subsequences of different scales. Subsequently, the invert smoothing Transformer and the convolutional interactive extraction module are designed to comprehensively extract multi-scale temporal dependencies and variable dependencies. Finally, the multi-sampled sequence features are fused to obtain the final encoding features, and fault prediction is achieved through the feed forward layer. Experiments are conducted using the penicillin fermentation process, and the results show that this method has good fault prediction performance.

Key words: batchwise, fault prediction, sequence sampling, neural network, multi-scale

摘要：

故障预测可以指示变量的异常变化，提前预测故障情况。现有故障预测方法仅考虑完整序列的全局时间依赖关系，忽略了变量间依赖关系及采样子序列中不同的局部时间依赖关系。针对上述问题，提出了一种基于多采样序列特征提取网络（multi-sampled sequence feature extraction network，MSFEN）的故障预测架构。首先设计了一种批次联合嵌入机制，在考虑批次周期性的同时更好地表达变量间依赖关系。然后，开发了一种序列采样机制划分完整时间序列与不同尺度的采样子序列。之后，分别设计了翻转平滑Transformer与卷积交互提取模块，以全面地提取多尺度时间依赖关系与变量间依赖关系。最后，融合多采样序列特征获得最终的编码特征，通过前馈层实现故障预测。利用青霉素发酵过程进行实验，结果表明该方法具有良好的故障预测性能。

关键词: 间歇式, 故障预测, 序列采样, 神经网络, 多尺度