Fig 6 - uploaded by Chang Peng
Content may be subject to copyright.
Source publication
In the real industrial production process, some minor faults are difficult to be detected by multivariate statistical analysis methods with mean and variance as detection indicators due to the aging equipment and catalyst deactivation. With structural characteristics, deep neural networks can better extract data features to detect such faults. Howe...
Contexts in source publication
Context 1
... need to be carefully edited for better visibility. We note that labels used in Figure 6 and Figure 7 have been revised. However, the resolution is very low compared to labels shown in Figure 4 and Figure 5. ...
Context 2
... until we get F í µí±¦ í µí±¡í µí±í µí± í µí±¡ vectors. A point with a value of 0 for the í µí±¦ í µí±¡í µí±í µí± í µí±¡ vector means that the system is operating normally. A point with a value of í µí±¦ í µí±¡í µí±í µí± í µí±¡ for the y vector represents that a fault has occurred at this time. The flow chart of AP-BLS modeling is shown in Fig. 6. Remark 3. In offline modeling we need to build F monitoring models and each model has to be divided into C stages, which is equivalent to training í µí°¹ × í µí° ¶ models. Deep network training í µí°¹ × í µí° ¶ models takes a lot of time, but the BLS network completes training of all models in a short period of ...
Citations
... In order to adapt to the demand for self-updating capability of models preferably, Chang et al. [22] based on copious literature extended the broad learning system (BLS) model proposed by Chen and Liu [23] to the domain of batch process. Instead of stacking numerous layers as deep network structures, the BLS designed on the idea of random vector functional-link neural network (RVFLNN) [24] to increase the width of neurons, whose weight can be rapidly obtained through the ridge regression algorithm without devilishly complex matrix operations. ...
Most industrial processes feature high nonlinearity, non-Gaussianity, and time correlation. Models based on overcomplete broad learning system (OBLS) have been successfully applied in the fault monitoring realm, which may relatively deal with the nonlinear and non-Gaussian characteristics. However, these models barely take time correlation into full consideration, hindering the further improvement of the monitoring accuracy of the network. Therefore, an effective dynamic overcomplete broad learning system (DOBLS) based on matrix extension is proposed, which extends the raw data in the batch process with the idea of "time lag" in this article. Subsequently, the OBLS monitoring network is employed to continue the analysis of the extended dynamic input data. Finally, a monitoring model is established to tackle the coexistence of nonlinearity, non-Gaussianity, and time correlation in process data. To illustrate the superiority and feasibility, the proposed model is conducted on the penicillin fermentation simulation platform, the experimental result of which illustrates that the model can extract the feature of process data more comprehensively and be self-updated more efficiently. With shorter training time and higher monitoring accuracy, the proposed model can witness an improvement of average monitoring accuracy by 3.69% and 1.26% in 26 process fault types compared to the state-of-the-art fault monitoring methods BLS and OBLS, respectively.
... Fei Chu proposed a weighted broad learning system (WBLS) based on BLS to address noise and outlier problems in industrial processes 19 . A batch process fault detection method for multi-stage broad learning system was proposed by Chang Peng 20 . Miao Mou proposed a Gated broad learning system based on deep cascaded for soft sensor modeling of industrial process 21 . ...
To deal with the highly nonlinear and time-varying characteristics of batch process, a model named Moving Window Stacking Approximate Kernel-Based Broad Learning System (MW-Stacking-AKBLS) is proposed in this paper. This model innovatively introduces the AKBLS algorithm and the MW-Stacking framework, giving it strong nonlinear fitting ability, excellent generalization ability, and adaptive ability. The Broad Learning System (BLS) is known for its shorter training time for effective nonlinear processing, but the uncertainty brought by its double random mapping results in poor resistance to noisy data and unpredictable impact on performance. To address this issue, this paper proposes an AKBLS algorithm that reduces uncertainty, eliminates redundant features, and improves prediction accuracy by projecting feature nodes into the kernel space. It also significantly reduces the computation time of the kernel matrix by searching for approximate kernels to enhance its ability in industrial online applications. Extensive comparative experiments on various public datasets of different sizes validate this. The MW-Stacking framework adopts the Stacking ensemble learning method, integrating multiple ABKLS models to enhance the model's generalization ability. Additionally, by adopting the moving window method, the model has been equipped with adaptive ability to better adapt to slow changes in industrial batch process. Finally, comparative experimental results on a substantial dataset of penicillin simulations indicate a significant improvement in predictive accuracy for the proposed MW-Stacking AKBLS model compared to other commonly used algorithms.
... Penicillin simulation software PenSim 2.0 [38] was utilized in this paper to simulate an actual fermentation production process. According to Chang et al [39], 10 process variables, listed in table 6, were selected for fault detection. The simulation conditions of normal and fault states are listed in table 7. The simulated data are supplied in the supplementary material. ...
Process monitoring is crucial to ensure the safety of industrial processes. Generally, the monitoring process involves all measured variables; however, large industrial processes contain many redundant variables. For a method based on describing the intrinsic correlation relationships among variables, vine copula-based dependence description (VCDD) method shows significant advantages for describing nonlinear and non-Gaussian processes. However, redundant and irrelevant variables adversely affect the correlation between variables containing the most important information, reducing model performance. The lack of research in this area may substantially weaken the advantages of VCDD for fault monitoring. Therefore, this article introduces a variable selection vine copula dependence description (VSVCDD) monitoring model. It utilizes known faults as validation data to select the relevant variables for constructing the VCDD model, specifically designed for monitoring known faults. Furthermore, to prevent information loss, the remaining unselected variables are also employed to create a separate VCDD model, dedicated to monitoring unknown faults. The performance of the proposed method is verified by a numerical example, the Tennessee-Eastman (TE) process and the Penicillin fermentation process (PFP).
... Although deep learning offers performance benefits, most deep-learning algorithms are affected by excessive training time and a large number of model parameters in deep architecture [16], [17]. Moreover, the structural complexity and parameter diversity of deep learning negatively impact model analysis and interpretation. ...
Federated learning (FL) guaranteeing data privacy is of great interest in decentralized fault diagnosis. However, limited research attention has been paid to the dynamic domain-shift issue due to varying working conditions. This paper proposes an active federated transfer algorithm based on broad learning to address the domain shift issue in FL. First, a central server dispatches a global model to the source clients for collaborative modeling. Subsequently, the global model is initialized with a federated averaging strategy. Next, the initialized global model is used to annotate emerging signals from the target clients based on an active sampling strategy proposed. Finally, an asynchronous update scheme is designed to adapt the global model to the target domain. The performance of the AFTBL algorithm is validated with three datasets, including 24 centralized- and decentralized-modeling tasks. The computational results indicate that the proposed algorithm is more accurate and efficient than the prevalent algorithms.
... In the past decade, deep learning-based fault diagnosis (DLFD) has achieved remarkable success and produced numerous variants [4]. With structural characteristics, deep architectures can effectively extract data features and detect faults [5]. However, most deep learning models are affected by a timeconsuming training process due to the involvement of numerous parameters and complicated structures [6]. ...
Deep learning has led to tremendous success in machine maintenance and fault diagnosis. However, this success is predicated on the correctly annotated datasets. Labels in large industrial datasets can be noisy and thus degrade the performance of fault diagnosis models. The emerging concept of broad learning shows the potential to address the label noise problem. Compared with existing deep learning algorithms, broad learning has a simple architecture and high training efficiency. An active label denoising algorithm based on broad learning (ALDBL) is proposed. First, ALDBL captures the embedded representation from the time-frequency features by a recurrent memory cell. Second, it augments wide features with a sparse autoencoder and projects the sparse features into an orthogonal space. A proposed corrector then iteratively changes the weights of source examples during the training and corrects the labels by using a label adaptation matrix. Finally, ALDBL finetunes the model parameters with actively sampled target data with reliable pseudo labels. The performance of ALDBL is validated with three benchmark datasets, including 30 label denoising tasks. Computational results demonstrate the effectiveness and advantages of the proposed algorithm over the other label denoising algorithms.
... The results conducted on the TE process illustrated that the proposed method performs well. Chang et al. (Chang, Lu, Oliver & Wang, 2020) proposed a multi-stage BLS method for batch process monitoring. The experiment in penicillin fermentation process verified that this method could capture the multi-stage characteristics and quickly establish the monitoring model with low time consumption, acquiring satisfactory results. ...
Affected by the operation environment and uncertainties, batch processes have complex dynamic characteristics, presenting autocorrelation and mutual correlation among process variables. Many conventional methods tend to ignore this property when constructing monitoring models, resulting in inadequate process feature extraction and unsatisfactory monitoring performance. A novel monitoring method named Dynamic Hidden Variable Fuzzy Broad Neural Network (DHVFBN) monitoring model is constructed for batch processes to address the aforementioned issues. For the details, in order to capture dynamic feature and nonlinearity in batch process, Slow Feature Analysis (SFA) is first used to extract the slowly changing components and sorting them from the raw time series data, which has strong capability in dynamic feature processing. In the monitoring model, the incremental learning ability of Fuzzy Broad Learning System (FBLS) is adopted to complete the quick reconstruction and expeditiously updating of the monitoring model without having to retrain the entire network when new fault samples are added to the training set or the accuracy of the network barely meet the requirements, which hugely relieves the computation burden and thus accomplishes the online fault surveillance. In addition, to fully extract the feature of process data, the fuzzy mechanism of the FBLS is then selected to extract the full fuzzified feature information of the process data so as to identify the slight difference between abnormal and normal process data effectively, which can improve the performance of fault monitoring. Finally, this method is evaluated by conducting experiments on the penicillin fermentation platform and Real-world industrial application. Compared with common state-of-the-art methods involved, the monitoring results indicate that the DHVFBN outperforms them.
... Although deep learning offers performance benefits, most deep-learning algorithms are affected by excessive training time and a large number of model parameters in deep architecture [16], [17]. Moreover, the structural complexity and parameter diversity of deep learning negatively impact model analysis and interpretation. ...
Knowledge transfer with class-imbalanced data is a challenge in predictive maintenance and fault diagnosis. Deep learning algorithms have provided promising results in fault diagnosis. However, their prediction performance is affected by class-imbalanced data in cross-domain tasks. Broad learning algorithms present promising performance in handling class-imbalanced domain-adaptation problems. In the presence of a domain shift, ABTCI, an active broad-transfer learning algorithm for class-imbalanced domain adaptation, is proposed. First, the ABTCI algorithm extracts the time-frequency features and feeds them into a recurrent cell to capture spatial-temporal features. Subsequently, it augments the feature space using a sparse autoencoder and an orthogonal mapping projector. By solving the ridge regression problem, the classifier is initialized. Next, the algorithm samples the target data with reliable pseudo-labels and synthesizes new data using random intraclass interpolation among the minor classes containing source and target knowledge. Finally, the classifier is updated using an incremental continuous learning strategy. The performance of the ABTCI algorithm is validated using three datasets, which include 20 class-balanced and 27 class-imbalanced transfer tasks. The performance of the proposed algorithm benchmarked against other deep learning algorithms is promising.
... Chang et al. [34] first introduced the BLS method into the field of batch process monitoring, and proposed the method of Multi-Stage Broad Learning System (MS-BLS) to solve the problem of Multi-stage and rapid modeling. However, the BLS-based methods and deep neural network-based methods seldom consider the non-Gaussian characteristics of the data. ...
To measure whether the sewage treatment meets the standards, biochemical oxygen demand (BOD5) is often used to determine, but the measurement of this indicator often has a long time lag and difficult to observe the real-time changes of BOD5, which brings inconvenience to the industrial process. The soft measurement technology based on neural network can realize BOD5 prediction at every moment by means of auxiliary variables, which has attracted people’s attention. However, there are still two problems with soft measurement technology, neural network-based soft measurement technology has high computational complexity and a certain time delay in measurement; and it cannot handle non-Gaussian data well. To solve them, this paper introduces an over-complete broad learning system(OBLS) based on feature fusion to deal with the problems of real-time measurement of BOD5 in sewage treatment industrial process. In view of the data characteristics, the feature extraction ability of the BLS is improved, the non-Gaussian characteristic of sewage data is captured by the method of Overcomplete Independent Component Analysis (OICA), and the OBLS is used to deal with the real-time soft measurement. Compared with state-of-the-art methods on the sewage standard test platform, the measurement accuracy of the proposed algorithm is found to be higher and the performance is more stable.
... Yu and Zhao (2019) put forward the broad convolutional neural network method for fault diagnosis in industrial processes. Chang et al. (2020) proposed multi-stage learning system for batch process fault detection. Chang and Lu (2021) put forward overcomplete broad learning network(OBLS) considering both the non-Gaussian and nonlinearity of the raw data. ...
In the real-world batch process, the minor faults caused by aging equipment and catalyst failure have subtle difference from normal data, making it difficult to monitor them timely with the traditional monitoring methods based on variance measurement and the deep learning methods based on distance measurement. In this research, a novel monitoring approach based on fuzzy broad learning system (FBLS) is employed to address the aforementioned issues. To fully extract the feature of raw data, the Takagi-Sugeno (TS) fuzzy system is first adopted to process the input data in order to identify minor faults effectively. Incremental learning algorithm is then employed to reconstruct network model quickly without retraining the entire network, which contributes to better accuracy and lower computation complexity and achieves online fault monitoring. After that, the classification of monitoring results is visualized to evaluate the fault type intuitively so as to take corresponding remedial actions quickly. Consequently, this algorithm is conducted into the penicillin fermentation platform and real industrial process. The results demonstrate that FBLS can better capture the fuzzified feature and quickly update monitoring model to accomplish the self-increase of fault database without retraining the entire process.
... With the rapid changes of market demands, batch process with small batches, multiple varieties and high value-added products has received extensive attention (Chang et al., 2020;Hui and Zhao, 2018b;Zhang and Zhao, 2019). Typical batch processes may include food processing (Simoglou et al., 2005), injection molding (Jiang et al., 2019), biomedicine production (Zhu et al., 2019), and so forth. ...
... The fault detection rate (FDR) and false alarm rate (FAR) are commonly used as evaluation indicators of fault monitoring field. The calculation formulas are as follows (Chang et al., 2020) FDR = 1 À No: of miss alarms total fault samples 3 100% ð23Þ ...
Batch process generally has varying dynamic characteristic that causes low fault detection rate and high false alarm rate, and it is necessary and urgent to monitor batch process. This paper proposes a global enhanced multiple neighborhoods preserving embedding based fault detection strategy for dynamic batch process. Firstly, the angle neighbor is defined and selected to compensate for the insufficient expression for the spatial similarity of samples only by using the distance neighbor, and the time neighbor is introduced to describe the time correlations between samples. These three types of neighbors can fully characterize the similarity of the samples in time and space. Secondly, considering the minimum reconstruction error and the order information of three types of neighbors, an enhanced objective function is constructed to prevent the loss of order information when neighborhood preserving embedding (NPE) calculates the reconstruction weights. Furthermore, the enhanced objective function and a global objective function are organically combined to extract both global and local features, to describe process dynamics and visualize process data in a low-dimensional space. Finally, a monitoring index based on support vector data description is constructed to eliminate adverse effects of non-Gaussian data for monitoring performance. The advantages of the proposed method over principal component analysis, neighborhood preserving embedding, dynamic principal component analysis and time NPE are demonstrated by a numerical example and the penicillin fermentation process simulation.
