基于AMWPSO-LSTM的多阶段间歇过程故障预测

过程安全对于间歇过程生产具有重要意义，为提高间歇过程生产安全性，提出一种基于改进粒子群算法(AMWPSO)优化长短期记忆网络(LSTM)的间歇过程故障预测模型AMWPSO-LSTM。针对LSTM中的神经元个数、迭代次数、学习率等参数需要人为设置的问题，采用AMWPSO对这些参数进行自动寻优。AMWPSO在原有粒子群优化算法(PSO)中融入了自适应变异和非线性递减惯性权重，提高了PSO的参数寻优能力。由于间歇过程具有多阶段性，因此先根据模糊C均值聚类(FCM)方法对间歇过程进行阶段划分，再利用Pearson相关系数对各阶段实验数据进行相关性分析，以降低系统变量的维数，并建立各阶段T²统计量控制限作为系统是否发生故障的指标。实验以青霉素发酵过程数据为例，建立基于AMWPSO-LSTM 的多阶段故障预测模型，并将该模型的预测结果与基于LSTM的多阶段预测模型、基于PSO-LSTM的多阶段预测模型的预测结果进行比较，结果表明，基于AMWPSO-LSTM 的多阶段故障预测模型可取得较高的预测准确度。

Multi-stage intermittent process fault prediction based on AMWPSO-LSTM

Process safety is of great significance in batch process production. In order to improve the production safety of batch processing, a fault prediction model AMWPSO-LSTM based on improved particle swarm optimization (AMWPSO) optimized long short-term memory network (LSTM) has been proposed. Given that parameters such as the number of neurons, the number of iterations, and the learning rate in LSTM need to be set manually, AMWPSO is used to optimize these parameters automatically. AMWPSO integrates adaptive mutation and nonlinear decreasing inertia weight into the original particle swarm optimization algorithm (PSO), which improves the parameter optimization ability of PSO. Due to the multi-stage nature of the batch process, it is first divided into stages according to the fuzzy C-means clustering (FCM) method. The Pearson correlation coefficient is then used to carry out correlation analysis on the experimental data for each stage to reduce the dimension of the system variables. The T² statistical control limit for each stage is taken as an indicator of whether the system fails. Using experimental data for penicillin fermentation as an example, we have established a multi-stage fault prediction model based on AMWPSO-LSTM and compared the predictions of the model with the predictions of the multi-stage prediction model based on LSTM and the multi-stage test model based on PSO-LSTM. The results show that the multi-stage fault prediction model based on AMWPSO-LSTM improves the prediction accuracy.