A classification-driven neuron-grouped SAE for feature representation and its application to fault classification in chemical processes

Pan, Zhuofu; Wang, Yalin; Yuan, Xiaofeng; Yang, Chunhua; Gui, Weihua

doi:10.1016/j.knosys.2021.107350

Cited by 24 publications

(6 citation statements)

References 41 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…An SAE‐based deep learning algorithm was developed to extract deep features for predicting TSS and TA in grapes. The SAE is composed of multiple layers of auto‐encoder (AE) (Pan et al ., 2021). The structure of AE can be viewed as the combination of an encoder and a decoder, and there are typically three layers within an AE (i.e.…”

Section: Methodsmentioning

confidence: 99%

Non‐destructive prediction of total soluble solids and titratable acidity in Kyoho grape using hyperspectral imaging and deep learning algorithm

Sun

Cheng

et al. 2022

Int J of Food Sci Tech

View full text Add to dashboard Cite

Summary Total soluble solids (TSS) and titratable acidity (TA) are essential quality properties for postharvest commercialisation of grapes. This study aimed to estimate the TSS and TA in grapes using hyperspectral imaging (HSI) technique in the range of 400–1001 nm. A deep learning‐based stacked auto‐encoders (SAE) algorithm was developed to extract deep spectral features from pixel‐level spectra. Then, these features with a compensation factor (i.e. size of fruits) were fed into partial least squares (PLS) and least squares support vector machine (LSSVM) for predicting TSS and TA in grapes. Additionally, competitive adaptive reweighed sampling and successive projections algorithm as conventional wavelength selection approaches were also investigated for comparison. The optimal prediction accuracy was achieved by the SAE‐LSSVM model with size compensation, where RP2=0.9237, RMSEP = 0.5041% and RPD=3.25 for TSS; Rp2=0.9216, RMSEP = 0.1091 g L−1 and RPD=3.21 for TA. The results suggested that SAE has great potential for extracting features from pixel‐level hyperspectral image data; the well‐performed deep learning model SAE‐LSSVM with size compensation can be used for rapid and non‐destructive predicting TSS and TA in grapes, which may provide a valuable reference for internal quality evaluation of postharvest fruits via HSI technique.

show abstract

Section: Methodsmentioning

confidence: 99%

Non‐destructive prediction of total soluble solids and titratable acidity in Kyoho grape using hyperspectral imaging and deep learning algorithm

Sun

Cheng

et al. 2022

Int J of Food Sci Tech

View full text Add to dashboard Cite

show abstract

“…The deterministic or probabilistic characteristics of collected samples can be captured using different unsupervised NNs. By training the network parameters, the reconstructed outputs will be approximately identical to its inputs [24].…”

Section: Preliminaries and Problem Formulation A Unsupervised Neural ...mentioning

confidence: 99%

A New Perspective on AE- and VAE-based Process Monitoring

Pan¹,

Chen²,

Wang³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

<p>Unsupervised neural networks (NNs) specialize in mining potential patterns from unlabeled data in a self-organizing manner. Recently, they have also been employed as observers for process monitoring using the generated residual signals. However, few studies have explained the model behavior and analyzed the monitoring performance of unsupervised NNs awing to their multilayer nonlinear structure. The interpretability of NN covers the analysis of its working principle to seek better network design and achieve improved performance. Thus, this paper develops explainable residual generators based on unsupervised NNs, which are applicable to deep autoencoder (DAE) and variational autoencoder (VAE). Through Taylor expansion, the residual deviation caused by the fault signal, a.k.a. the fault-affected term, is proven not to disappear in the presence of a non-zero Hessian matrix. Then, the consistency of achieving the optimal monitoring performance and the training loss of NNs is presented. A new indicator function is established based on the sum martingale, a representative of a weakly dependent stochastic process. Freedman’s inequality is first applied to describe the reliability of the learned thresholds during fault evaluation, which requires a smaller sample size for training. Finally, simulations on the continuous stirred tank reactor verify the effectiveness of the proposed methods.</p>

show abstract

“…In 2006, Hinton, the famous machine learning expert, first proposed deep learning to break the barrier of insufficient training of shallow learning [19] and started the wave of deep learning in academia and industry. Common deep learning methods include stacked autoencoder (SAE) [20][21][22], convolutional neural network (CNN) [23][24][25], deep belief network (DBN) [26][27][28], long short-term memory network (LSTM) [29][30][31], among others. To extract features related to fault types, Yu et al [32] proposed a supervised convolutional autoencoder, which reconstructs process samples and corresponding labels by performing multi-layer encoding and decoding on original samples, and the obtained fault features can clearly distinguish different types of faults.…”

Section: Introductionmentioning

confidence: 99%

A latent representation dual manifold regularization broad learning system with incremental learning capability for fault diagnosis

Mou

Zhao

et al. 2023

Meas. Sci. Technol.

View full text Add to dashboard Cite

When new fault occurs, the parameters and structure of fault diagnosis model based on deep learning need to be adjusted for retraining, which is often very time-consuming. For the above problem, a latent representation dual manifold regularization broad learning system (LRDMR-BLS) with incremental learning capability is proposed for process fault diagnosis. This model embeds latent representation learning into feature selection and utilizes the link information between data to guide feature selection. Meanwhile, manifold regularization is introduced in the objective function to preserve the local manifold structure of the original data space. Further, a manifold regularization term is added to the objective function of the broad learning system to preserve the local structure of the features. Finally, the incremental learning capability of the proposed model is given, which enables the proposed model can be updated quickly when new fault occurs. The superiority of the proposed model is demonstrated by two Chemical processes.

show abstract

A classification-driven neuron-grouped SAE for feature representation and its application to fault classification in chemical processes

Cited by 24 publications

References 41 publications

Non‐destructive prediction of total soluble solids and titratable acidity in Kyoho grape using hyperspectral imaging and deep learning algorithm

Non‐destructive prediction of total soluble solids and titratable acidity in Kyoho grape using hyperspectral imaging and deep learning algorithm

A New Perspective on AE- and VAE-based Process Monitoring

A latent representation dual manifold regularization broad learning system with incremental learning capability for fault diagnosis

Contact Info

Product

Resources

About