A batch-wise LSTM-encoder decoder network for batch process monitoring

Ren, Jiayang; Ni, Dong

doi:10.1016/j.cherd.2020.09.019

Cited by 39 publications

(20 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Then, the weight gradient updates are calculated as the network is feedback . However, RNNs trained with this algorithm can present vanishing or explosion to infinity gradient issues, introducing a necessity to improve this algorithm and create new structures of RNN, such as the echo state network (ESN) …”

Section: Neural Network Modelingmentioning

confidence: 99%

Modeling and Predictive Control of Cooling Crystallization of Potassium Sulfate by Dynamic Image Analysis: Exploring Phenomenological and Machine Learning Approaches

Moraes

Lima

Lage

et al. 2023

Ind. Eng. Chem. Res.

View full text Add to dashboard Cite

Representative mathematical modeling is essential for understanding the batch cooling crystallization processes. Efficient process design and operation are relevant to achieving high-quality criteria and minimizing variation between batches. This work first presents the modeling of batch cooling crystallization based on online dynamic image analysis. A flow-through microscope was used to track the temporal evolution of the crystal population. A population balance modeling (PBM) approach, parameter estimation, and validation were obtained for the batch cooling crystallization of potassium sulfate in water. The performed experiments provided new experimental data, giving dynamic information about the crystal size throughout each run. The kinetic model parameters for crystal nucleation and growth were estimated using a hybrid optimization algorithm, followed by the confidence region construction using a more exploratory particle swarm algorithm. In the parameter estimation framework, in addition to solute concentration, the first fourth-order moments computed throughout all experiments were included in the objective function. A linear size-dependent growth rate was found to capture well the dynamics of the potassium sulfate crystal size distribution. The experimental results evidenced that the crystal shape of potassium sulfate is predominantly constant, allowing the adequacy of the developed model. The validated PBM was also employed as a digital twin of the crystallization process to develop a machine-learning-based control for the process. Then, a surrogate model based on a recurrent neural network, called an echo state network (ESN), was applied in a nonlinear model predictive controller approach (ESN-NMPC). The ESN model could predict the moments of the population balance model up to five steps (5 min) forward. The ESN-NMPC achieved the desired control scenarios for the crystal size and its coefficient of variation. Its performance was comparable to the controller that uses the PBM as the internal model (PB-NMPC).

show abstract

Section: Neural Network Modelingmentioning

confidence: 99%

Modeling and Predictive Control of Cooling Crystallization of Potassium Sulfate by Dynamic Image Analysis: Exploring Phenomenological and Machine Learning Approaches

Moraes

Lima

Lage

et al. 2023

Ind. Eng. Chem. Res.

View full text Add to dashboard Cite

show abstract

“…There have been numerous studies on the use of deep learning models for process monitoring of nonlinear processes. A variety of deep learning architectures have been employed, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), , and autoencoder neural networks (AEs). − …”

Section: Introductionmentioning

confidence: 99%

Unsupervised Fault Detection of Pharmaceutical Processes Using Long Short-Term Memory Autoencoders

Aghaee

Krau

Tamer

et al. 2023

Ind. Eng. Chem. Res.

View full text Add to dashboard Cite

Unsupervised multilayer long short-term memory autoencoder (LSTM-AE) models are proposed for monitoring nonlinear batch processes. The methodology is demonstrated for a simulation-based study of an industrialscale penicillin process and for an industrial vaccine manufacturing process, using production data. The LSTM-AE model was trained with two different loss functions: minimizing mean square error (MSE) between the input and reconstructed data and maximizing the average fault detection rate (FDR) in the training data set. Two algorithms are also proposed for obtaining contribution plots for the diagnosis of faults. For the industrial case study, where the faults are not known a priori, the contribution plots are found to be a valuable tool for identifying possible sources of faults. Furthermore, a semisupervised procedure has been proposed to select the normal process region for training the model. Two metrics are also presented to evaluate the performance of the proposed methodology: one for the simulator case study in which fault knowledge is available and one for the industrial case study in which fault knowledge is not available a priori. The proposed unsupervised algorithms exhibit a clear improvement in accuracy over linear methods or nonlinear techniques that do not explicitly account for dynamic behavior.

show abstract

“…The lower dimensional representations obtained from these methods can be better used for the detection of process abnormalities than representations using the entire dimensionality. [ 14 ] Ren and Ni [ 15 ] used a multi‐layer recurrent neural network in the encoder—decoder structure for the fault detection of batch process. Lee et al [ 16 ] proposed an on‐line monitoring method based on probability‐based synchronization for uneven multiphase batch process.…”

Section: Introductionmentioning

confidence: 99%

Slow‐varying batch process monitoring based on canonical variate analysis

Zhang

Bao

2022

Can J Chem Eng

View full text Add to dashboard Cite

Most industrial batch processes display normal slow changes over batches because of the influence of external factors. The question of how to take this slow-varying behaviour into account during dynamic process monitoring remains a challenge for researchers. To address the above issue, a slow-varying batch process monitoring method is proposed for multiphase batch processes. It makes subtraction between two different batches with a certain interval for pretreating the original data and subsequently utilizes a canonical variate analysis (CVA) algorithm with kernel density estimation (KDE) for fault detection. For each subphase, the proposed method can extract a variable-wise dynamic characteristic with the time evolution for every single batch and capture statistical features of slow variations along batch direction simultaneously. It is less sensitive to normal gradual changes with a low false-positive rate. The performance of the proposed method for batch process monitoring was tested on a numerical simulation system and penicillin fermentation production process with comparison to multiphase CVA, multiphase slow feature analysis (SFA), multiphase principal component analysis (PCA), and traditional CVA without phase division. The achieved results clearly demonstrate the effectiveness of the proposed method, which is a remarkable and promising tool for modelling and monitoring batch processes with regular slow-varying characteristics.

show abstract

A batch-wise LSTM-encoder decoder network for batch process monitoring

Cited by 39 publications

References 22 publications

Modeling and Predictive Control of Cooling Crystallization of Potassium Sulfate by Dynamic Image Analysis: Exploring Phenomenological and Machine Learning Approaches

Modeling and Predictive Control of Cooling Crystallization of Potassium Sulfate by Dynamic Image Analysis: Exploring Phenomenological and Machine Learning Approaches

Unsupervised Fault Detection of Pharmaceutical Processes Using Long Short-Term Memory Autoencoders

Slow‐varying batch process monitoring based on canonical variate analysis

Contact Info

Product

Resources

About