基于循环神经网络的半监督动态软测量建模方法

Semisupervised dynamic soft sensing approaches based on recurrent neural network

数据驱动的软测量技术被广泛应用于难测关键变量的在线实时预报。然而,在工业过程中,有标签样本通常十分稀少,且动态特性显著,导致传统有监督、静态的软测量建模方法性能不佳。为此,提出一种基于循环神经网络的建模方法,首先将传统带有长短时记忆单元(LSTM)的循环神经网络(RNN)扩展为半监督模式,然后针对LSTM的不足,进一步提出一种基于注意力机制的改进方案。通过一个实际工业案例验证半监督LSTM-RNN在软测量应用中的有效性,以及所提出的改进方案的有效性。

Data driven soft sensing techniques have been widely accepted for predicting important yet difficult-to-measure variables in real-time. However, in industrial processes, labeled samples are usually very scarce, and dynamics widely exist, which result in that traditional supervised and static soft sensing approaches do not perform satisfactorily. To this end, this paper proposes a soft sensing method based on the recurrent neural network(RNN). Firstly, the traditional long-short time memory-aided RNN(LSTM-RNN) is extended to the semisupervised version(S^2LSTM-RNN). Subsequently, to deal with the drawback of the LSTM, a further improved attention-based S^2LSTM-RNN(attention-S^2LSTM-RNN) is proposed. A case study is conducted on one real-world industrial process, through which both the S^2LSTM-RNN and attention-S^2LSTM-RNN can be verified.