基于深度学习的燃气发电机组剩余寿命预测

Residual life prediction of gas generator set based on deep learning

为了保障油气站场关键设备随运行年限增加的安全稳定运行,减少安全生产事故,同时,提高检维修效率降低运行成本,基于深度学习模型与遗传优化算法,针对油气站场的燃气发电机组提出一种新的剩余寿命预测模型。首先,对油气站场采集的燃气发电机组运行工况数据进行预处理,利用主成分分析(PCA)算法分析主成分多维传感器特征,提取出信号主成分作为特征输入;利用一维转置卷积(TransConv1D)与长短期记忆(LSTM)网络构建深度学习剩余寿命预测模型进行训练;针对模型训练过程中特征堆叠冗余的问题,利用量子遗传算法(QGA)优化模型的2个超参数;通过实例验证并与传统方法进行对比分析。结果表明:相比于传统模型,提出的剩余寿命预测模型预测结果的平均绝对误差(MAE)与均方根误差(RMSE)分别降低2.46%与7.85%。所提模型预测误差更小,能够更加准确地预测燃气发电机组的剩余寿命。

To ensure the safe and stable operation of critical equipment in oil and gas stations,reduce safety accidents and improve maintenance efficiency while reducing operating costs,this paper proposed a novel residual life prediction model for gas engine generator sets in oil and gas stations based on deep learning models and genetic optimization algorithms.Firstly,the operational data of gas engine generator sets collected from the oil and gas stations were preprocessed,and the Principal Component Analysis(PCA)algorithm was applied to extract the main components of the multi-dimensional sensor features as input features.A deep learning residual life prediction model was constructed using 1-Dimensional Transposed Convolution(TransConv1D)and Long Short-Term Memory(LSTM)for training.To solve the problem of feature stacking redundancy during model training,the QGA was employed to optimize two hyperparameters of the model.Through case verification and comparative analysis with traditional methods,the proposed model achieved at least a 2.46%reduction in MAE and a 7.85%reduction in RMSE in the prediction results.The research demonstrates that the proposed model outperforms the traditional models with smaller prediction errors and can predict the residual life of gas engine generator sets more accurately.