基于自编码器和多尺度符号转移熵的FCC沉降器跑剂故障检测

doi:10.11949/0438-1157.20250169

摘要/Abstract

摘要：

催化裂化（fluid catalytic cracking，FCC）沉降器跑剂故障复杂，故障发生时多个参数会偏离正常状态，需迅速检测和缓解，以保障FCC装置长周期稳定运行。提出了一种集成自编码器（autoencoder，AE）和多尺度符号转移熵（multi-scale symbol transfer entropy，MSTE）的无监督跑剂故障检测方法（AEM），充分考虑了时间序列数据中复杂的时间依赖关系及故障后时间序列的非平稳性。将双重注意力（dual attention，DA）机制融入长短期记忆单元（long short-term memory，LSTM）的编码器-解码器结构中，其中特征注意力（feature attention，FA）机制通过为输入特征变量分配权重，动态强调关键特征并识别主要故障变量；时间注意力（temporal attention，TA）机制则通过为时间维度内的每个时间步分配权重，捕获其依赖信息，从而进一步提升检测效果。此外，利用MSTE方法构建了非平稳过程的因果关系图，揭示了变量之间的时间延迟并排除了间接因果关系。通过分析沉降器快分头封头穿孔过程，验证了AEM的有效性，且提升了决策过程的可解释性。

关键词: 过程控制, 安全, 无监督学习, 神经网络, 因果关系发现

Abstract:

Fluid catalytic cracking (FCC) disengager catalyst loss faults are complex, with multiple variables deviating from normal when occurring. These faults must be detected and mitigated promptly to ensure the long-term stable operation of the FCC unit. This study proposes an unsupervised run-off fault detection method (AEM) that integrates autoencoder (AE) and multi-scale symbol transfer entropy (MSTE), which fully considers the complex time dependency in time series data and the non-stationarity of the time series after the fault. The dual attention (DA) mechanism is incorporated into a long short-term memory (LSTM) network with an encoder-decoder structure. By applying the feature attention (FA) mechanism to assign weights to input feature variables, key features are dynamically emphasized, enabling the identification of the main fault variables. The temporal attention (TA) mechanism further enhances fault detection by assigning weights to capture dependency information at each time step along the time dimension. Additionally, a causal diagram is constructed for the nonstationary process using the MSTE method to reveal time delays between variables and eliminate indirect causal relationships. The effectiveness of AEM is verified by analyzing the perforation process of the fast-separating head of the settler, and the interpretability of the decision-making process is improved.

Key words: process control, safety, unsupervised learning, neural networks, causality discovery

中图分类号:

TQ 021

朱春梦, 李增, 柳楠, 赵云鹏, 石孝刚, 蓝兴英. 基于自编码器和多尺度符号转移熵的FCC沉降器跑剂故障检测[J]. 化工学报, 2025, 76(9): 4512-4523.

Chunmeng ZHU, Zeng LI, Nan LIU, Yunpeng ZHAO, Xiaogang SHI, Xingying LAN. Fault detection of catalyst loss in FCC disengager based on autoencoder and multi-scale symbolic transfer entropy[J]. CIESC Journal, 2025, 76(9): 4512-4523.

图/表 13

图1 集成AM和LSTM的编码器结构

Fig.1 Encoder architecture with integrated AM and LSTM

图2 集成AM和LSTM的解码器结构

Fig.2 Decoder architecture with integrated AM and LSTM

图3 集成CFD、AE和MSTE的故障检测流程图

Fig.3 Fault detection flowchart with integrated CFD, AE and MSTE

图4 FCC沉降器结构及变量分布

Fig.4 Structure and variable distribution of FCC disengager system

表1 FCC沉降器系统中的过程变量

Table 1 Process variables in the FCC disengager system

工业变量	变量描述	单位
F1	顶旋入口催化剂流量	kg·s^-1
F2	催化剂跑损量	kg·s^-1
P1	下部稀相空间压力	kPa
P2	上部稀相空间压力	kPa
P3	快分出口压力	kPa
P4	顶旋压降	kPa
D1	快分出口催化剂浓度(固含率)	—
D2	隔离罩槽口催化剂浓度(固含率)	—
D3	顶旋入口催化剂浓度(固含率)	—
U1	快分出口线速	m·s^-1
U2	顶旋入口线速	m·s^-1
T1	下部稀相空间温度	K
T2	沉降器顶部温度	K
T3	快分出口温度	K
T4	顶旋入口温度	K

图5 高保真数据集的故障检测结果

Fig.5 Fault detection results for high-fidelity datasets

图6 数据表现形式

Fig.6 Forms of data presentation

图7 添加数据噪声的故障检测结果

Fig.7 Fault detection results with data noise added

图8 基于FA机制的故障贡献度图

Fig.8 Fault contribution diagram based on the FA mechanism

图9 基于MSTE的因果关系

Fig.9 The causality generated by MSTE

图10 基于MSTE的快分头穿孔故障时间滞后因果关系

Fig.10 Causal diagram with time delay of vortex quick-separation perforation fault generated by MSTE

图11 变量P3与P1以及变量P4与P1之间的MSTE随时间延迟τ的变化关系

Fig.11 MSTE changes with τ in FCC disengager system for variables P3 and P1 or for variables P4 and P1

图12 基于MTE、STE和MSTE的因果关系

Fig.12 The causality generated by MTE, STE and MSTE

参考文献 31

[1]	许友好, 李浩天, 刘昌呈. “双碳”目标下石油炼制技术研发策略与工业实践[J]. 石油炼制与化工, 2024, 55(1): 171-179.
	Xu Y H, Li H T, Liu C C. Development strategy and industrial practice of petroleum refining technology under the goal of “carbon peak and carbon neutrality"[J]. Petroleum Processing and Petrochemicals, 2024, 55(1): 171-179.
[2]	Liu N, Wang J, Sun S L, et al. Optimized principal component analysis and multi-state Bayesian network integrated method for chemical process monitoring and variable state prediction[J]. Chemical Engineering Journal, 2022, 430: 132617.
[3]	宋冰, 郭涛, 侍洪波, 等. 基于双子空间并行回归的化工过程质量相关故障检测方法[J]. 化工学报, 2023, 74(11): 4600-4610.
	Song B, Guo T, Shi H B, et al. A chemical process quality-related fault detection method based on twin-space parallel regression[J]. CIESC Journal, 2023, 74(11): 4600-4610.
[4]	Hinton G E, Salakhutdinov R R. Reducing the dimensionality of data with neural networks[J]. Science, 2006, 313(5786): 504-507.
[5]	Park D, Hoshi Y, Kemp C C. A multimodal anomaly detector for robot-assisted feeding using an LSTM-based variational autoencoder[J]. IEEE Robotics and Automation Letters, 2018, 3(3): 1544-1551.
[6]	Li W X, Shang Z W, Zhang J, et al. A novel unsupervised anomaly detection method for rotating machinery based on memory augmented temporal convolutional autoencoder[J]. Engineering Applications of Artificial Intelligence, 2023, 123: 106312.
[7]	Yao Y Y, Ma J H, Ye Y M. Regularizing autoencoders with wavelet transform for sequence anomaly detection[J]. Pattern Recognition, 2023, 134: 109084.
[8]	Zeng L, Jin Q W, Lin Z M, et al. Dual-attention LSTM autoencoder for fault detection in industrial complex dynamic processes[J]. Process Safety and Environmental Protection, 2024, 185: 1145-1159.
[9]	Zhang Z W, Wu L F. Graph neural network-based bearing fault diagnosis using Granger causality test[J]. Expert Systems with Applications, 2024, 242: 122827.
[10]	Zhao C H, Chen X. Data-driven nonstationary industrial process monitoring[M]//Reference Module in Materials Science and Materials Engineering. Amsterdam: Elsevier, 2024.
[11]	Staniek M, Lehnertz K. Symbolic transfer entropy[J]. Physical Review Letters, 2008, 100(15): 158101.
[12]	Alobaid F, Almohammed N, Massoudi Farid M, et al. Progress in CFD simulations of fluidized beds for chemical and energy process engineering[J]. Progress in Energy and Combustion Science, 2022, 91: 100930.
[13]	何洋, 王利民, 唐春丽, 等. H型翅片管湿烟气对流冷凝传热的数值模拟研究[J]. 化工学报, 2019, 70(12): 4556-4564.
	He Y, Wang L M, Tang C L, et al. Numerical simulation of convection condensation heat transfer of H-fixed tubes with wet flue gas[J]. CIESC Journal, 2019, 70(12): 4556-4564.
[14]	周国峰, 王利民, 王小伟, 等. 基于时驱硬球算法与格子玻尔兹曼方法的颗粒流体系统直接数值模拟[J]. 科学通报, 2011, 56(16): 1246-1256.
	Zhou G F, Wang L M, Wang X W, et al. Direct numerical simulation scheme for particle-fluid systems based on a time-driven hard-sphere and the lattice Boltzmann method[J]. Chinese Science Bulletin, 2011, 56(16): 1246-1256.
[15]	Tang Y N, Liu C, Kuang S B, et al. CFD modeling investigation of oxy-fuel combustion application in an industrial-scale FCC regenerator[J]. Journal of the Energy Institute, 2024, 117: 101796.
[16]	Zhao Y P, Li C L, Shi X G, et al. Simulation analysis of CO₂ in situ enrichment technology of fluidized catalytic cracking regenerator[J]. Powder Technology, 2024, 434: 119386.
[17]	Lan X Y, Shi X G, Wang C X, et al. Process intensification of multiphase flow and reaction system: perspectives[J]. Chemical Engineering and Processing-Process Intensification, 2024, 204: 109938.
[18]	Zhang H L, Zhu A Q, Xu J, et al. Gas-solid reactor optimization based on EMMS-DPM simulation and machine learning[J]. Particuology, 2024, 89: 131-143.
[19]	Aminmahalati A, Fazlali A, Safikhani H. Multi-objective optimization of CO boiler combustion chamber in the RFCC unit using NSGA Ⅱ algorithm[J]. Energy, 2021, 221: 119859.
[20]	高学金, 姚玉卓, 韩华云, 等. 基于注意力动态卷积自编码器的发酵过程故障监测[J]. 化工学报, 2023, 74(6): 2503-2521.
	Gao X J, Yao Y Z, Han H Y, et al. Fault monitoring of fermentation process based on attention dynamic convolutional autoencoder[J]. CIESC Journal, 2023, 74(6): 2503-2521.
[21]	Sun R R, Wang Y Q. Key-performance-indicator-related fault detection based on deep orthonormal subspace analysis[J]. IEEE Transactions on Industrial Informatics, 2024, 20(5): 7249-7258.
[22]	Yu J H, Gao X, Li B F, et al. A filter-augmented auto-encoder with learnable normalization for robust multivariate time series anomaly detection[J]. Neural Networks, 2024, 170: 478-493.
[23]	邵远哲, 赵忠盖, 刘飞. 基于共同趋势模型的非平稳过程质量相关故障检测方法[J]. 化工学报, 2023, 74(6): 2522-2537.
	Shao Y Z, Zhao Z G, Liu F. Quality-related non-stationary process fault detection method by common trends model[J]. CIESC Journal, 2023, 74(6): 2522-2537.
[24]	Runge J, Bathiany S, Bollt E, et al. Inferring causation from time series in Earth system sciences[J]. Nature Communications, 2019, 10(1): 2553.
[25]	刘梦溪, 卢春喜, 时铭显. 催化裂化后反应系统快分的研究进展[J]. 化工学报, 2016, 67(8): 3133-3145.
	Liu M X, Lu C X, Shi M X. Advances in quick separators of post-riser system in FRCC unit[J]. CIESC Journal, 2016, 67(8): 3133-3145.
[26]	马明亮, 卢朝鹏, 赵静. 催化裂化装置旋风分离器运行情况分析及对策[J]. 石油与天然气化工, 2020, 49(6): 34-39.
	Ma M L, Lu Z P, Zhao J. Operation situation analysis and countermeasures of the cyclone separator for catalytic cracking unit[J]. Chemical Engineering of Oil & Gas, 2020, 49(6): 34-39.
[27]	Hu X, Zhang T Y, Geng Z Q, et al. Noise adaptive filtering model integrating spatio-temporal feature for soft sensor[J]. Expert Systems with Applications, 2024, 239: 122453.
[28]	Hao X C, Wang Y Z, Zhang Z P, et al. Root cause analysis of faults in cement pre-decomposition system using kernel principal component analysis and multi-scale symbolic transfer entropy[J]. Chemical Engineering Science, 2024, 286: 119681.
[29]	Duan S Y, Zhao C H, Wu M. Multiscale partial symbolic transfer entropy for time-delay root cause diagnosis in nonstationary industrial processes[J]. IEEE Transactions on Industrial Electronics, 2023, 70(2): 2015-2025.
[30]	Faes L, Nollo G, Porta A. Compensated transfer entropy as a tool for reliably estimating information transfer in physiological time series[J]. Entropy, 2013, 15(1): 198-219.
[31]	Zhao X J, Sun Y P, Li X M, et al. Multiscale transfer entropy: measuring information transfer on multiple time scales[J]. Communications in Nonlinear Science and Numerical Simulation, 2018, 62: 202-212.