Research on a Safety Dynamic Early Warning Model for Hazardous Chemical Tankers under Multi-Parameter Spatio-Temporal Conditions
摘要
针对现有危化品运输罐车预警系统依赖固定阈值、难以刻画多源时空参数动态耦合关系而导致误报与漏报的问题,研究了基于长短期记忆自编码器与多元状态估计技术(long short-term memory autoencoder-multi-variate state estimation technique,LSTM-AE-MSET)的时空多参数安全动态预警模型.该模型以危化品罐车运行安全动态预警为目标,融合温度、压力、液位、倾角和振动等具有时空属性的多源传感器数据,并引入GNSS、IMU和API获取的道路空间信息,利用LSTM-AE提取时序特征与表征变量间的长期依赖关系,将高维输入映射至低维潜在空间;基于历史健康状态数据构建MSET记忆矩阵,通过状态估计与残差分析实现预警阈值的动态调整.采用田纳西-伊斯曼过程(Tennessee Eastman process,TEP)数据集验证模型有效性,数据集包含41个过程测量变量、12个操纵变量与21种故障模式,涵盖阶跃扰动、随机变化、阀门黏滞等典型异常.模型在TEP数据集上完成训练后迁移至罐车运行场景,用于识别不同时空条件下的温度突变、压力漂移、液位振荡等异常.结果表明,LSTM-AE-MSET模型的平均决定系数 R2 值为0.621 8,均方误差值为4.301 4;其中位数分别达到0.998 5与0.000 1,表明模型在多数正常运行工况下具有优异的拟合能力;异常检测召回率由57%提升至88%,F1 分数从60%提升至74%.与单一MSET方法相比,LSTM-AE-MSET模型整体准确率提高2%,漏报率降低72%.本文方法能够增强危化品运输罐车异常状态的识别能力,为复杂运行环境下的实时动态预警提供技术支持.
Existing early-warning systems for tank trucks transporting hazardous chemicals commonly rely on fixed thresholds and fail to characterize the dynamic coupling among multi-source spatio-temporal parameters,which leads to false and missed alarms.To address this limitation,a spatio-temporal multi-parameter dynamic safety ear-ly-warning model was proposed based on a long short-term memory autoencoder and the multivariate state estima-tion technique(LSTM-AE-MSET).The model integrates multi-source sensor data with spatio-temporal attributes,including temperature,pressure,liquid level,inclination,and vibration,and incorporates road spatial information ac-quired via the global navigation satellite system(GNSS),the inertial measurement unit(IMU),and map application programming interfaces(API).The LSTM-AE extracts temporal features,captures long-term dependencies among variables,and maps high-dimensional inputs into a low-dimensional latent space;an MSET memory matrix is con-structed based on historical healthy-state data,and the early-warning thresholds are dynamically adjusted through state estimation and residual analysis.The Tennessee Eastman Process(TEP)dataset is used to evaluate the model.The dataset comprises 41 process measurement variables,12 manipulated variables,and 21 fault modes,covering typical anomalies such as step disturbances,random variations,and valve stiction.After being trained on the TEP da-taset,the model is transferred to a tank-truck operating scenario to identify anomalies such as abrupt temperature changes,pressure drift,and liquid-level oscillations under different spatio-temporal conditions.The results show that the model achieves an average coefficient of determination(R²)of 0.621 8 and a mean squared error of 4.301 4,with medians of 0.998 5 and 0.000 1,respectively,indicating excellent fitting performance under most normal op-erating conditions.Compared with the standalone MSET method,the proposed model increases the anomaly-detec-tion recall from 57%to 88%and the F1-score from 60%to 74%,improves the overall accuracy by 2 percentage points,and reduces the missed-alarm rate by 72%.The proposed method enhances the identification of abnormal op-erating states in hazardous-chemical tank trucks and provides technical support for real-time dynamic early warning under complex operating conditions.