A novel nonparametric method based on manifold learning is proposed for industrial process monitoring. In conventional algorithms, to preserve the global and local structure information of data, heat kernels containing two auxiliary parameters are introduced to define the global and local weight matrices, respectively. However, it is difficult to identify and choose these two parameters empirically. The inadequate selection of parameters can lead to one-sided and inappropriate global and local feature extractions, resulting in an inadequate fault detection performance. To resolve the above problems, a nonparametric strategy is used in this study to generate two nonparametric weight matrices to replace the heat kernel-based weight matrices. Consequently, the proposed method requires no auxiliary parameters in defining the weight matrices, making it more practical. Moreover, it automatically determines a good trade-off between global and local feature extractions. A process monitoring model based on the proposed method was developed. The feasibility and effectiveness of the new nonparametric method are evaluated using a synthetic example and the Tennessee Eastman chemical process.

Graph Neural Networks (GNNs) have recently received a surge of popularity due to their superiority in modeling realistic complex systems into graphs. Better yet, many approaches have made extraordinary contributions to this topic under the context of dynamic graph. However, current dynamic graph models merely focused on the design of the model but didn’t lay enough emphasis on the dynamism of data, which weakened the expressiveness of the output. Hence, we propose the Adaptive Sampling Temporal Graph Network (ASTGN), a Continuous-Time Dynamic Graph (CTDG) algorithm which casts the sampling strategy as a contextual bandit problem. To capture the dynamic information of graphs, we use a change-detection mechanism to keep the sampling strategy up with the times. Besides, we propose an additional constraint to keep our sampling path causal. We experimentally demonstrate the effectiveness of our approaches with six benchmarks and show the superiority over state-of-the-art baselines.

To capture higher-order structural features, most GNN-based algorithms learn node representations incorporating k-hop neighbors' information. Due to the high time complexity of querying k-hop neighbors, most graph algorithms cannot be deployed in a giant dense temporal network to execute millisecond-level inference. This problem dramatically limits the potential of applying graph algorithms in certain areas, especially financial fraud detection. Therefore, we propose Asynchronous Propagation Attention Network, an asynchronous continuous time dynamic graph algorithm for real-time temporal graph embedding. Traditional graph models usually execute two serial operations: first graph querying and then model inference. Different from previous graph algorithms, we decouple model inference and graph computation to alleviate the damage of the heavy graph query operation to the speed of model inference. Extensive experiments demonstrate that the proposed method can achieve competitive performance while greatly improving the inference speed. The source code is published at a Github repository.

Nowadays, the application of dynamic graphs in the modeling of complex systems has made a great achievement, which has aroused people's attention to anomaly detection of dynamic graphs. As an unsupervised learning task, anomaly detection is target at identifying the abnormal data that is different from the majority. One-class support vector machine, one of the classic anomaly detection algorithms, has been widely applied to find the outliers for it's stability, robustness and convenience. However, traditional anomaly detection algorithms always lose their effectiveness when applied to dynamic graph anomaly detection task. In order to solve the above problem, we design one-class temporal graph attention neural network (OCTGAT) for anomaly detection on dynamic graph. OCTGAT aims to integrate the powerful representation capabilities off temporal graph neural networks and the classical one-class objective. Compared with the given benchmarks, OGTGAT achieves significant improvements in the experiments.

Nowadays, graph-structured data are increasingly used to model complex systems. Meanwhile, detecting anomalies from graph has become a vital research problem of pressing societal concerns. Anomaly detection is an unsupervised learning task of identifying rare data that differ from the majority. As one of the dominant anomaly detection algorithms, one-class support vector machine has been widely used to detect outliers. However, those traditional anomaly detection methods lost their effectiveness in graph data. Since traditional anomaly detection methods are stable, robust and easy to use, it is vitally important to generalize them to graph data. In this work, we propose one-class graph neural network (OCGNN), a one-class classification framework for graph anomaly detection. OCGNN is designed to combine the powerful representation ability of graph neural networks along with the classical one-class objective. Compared with other baselines, OCGNN achieves significant improvements in extensive experiments.

Feature extraction plays a key role in the data-driven process monitoring. Recently, manifold learning approaches have shown good effectiveness of preserving manifold structure features of process data. However, conventional algorithms only focus on low-order statistics while ignore high-order statistics, which makes them unable to extract non-Gaussian features of data effectively. Moreover, process data may not strictly follow Gaussian distribution in the complex modern industrial processes. In this paper, to address the above issues, a new manifold learning method named statistics local and nonlocal embedding (SLNLE) is proposed for non-Gaussian process monitoring. Firstly, both low-order and high-order process statistics are conducted by statistics pattern analysis (SPA) algorithm under given time window. Then, local and nonlocal embedding (LNLE) is adopted to preserve manifold structure information of them. Compared with locality preserving projections (LPP) algorithm only retaining local structure features, LNLE implements global and local feature extraction simultaneously, which can preserve non-Gaussian features more effectively. Lastly, a SLNLE-based monitoring model is developed and kernel density estimation (KDE) is applied to obtain more accurate control limits for better process monitoring performance. The feasibility and superiority of the proposed approach are tested on the Tennessee Eastman (TE) process.

Recently, deep generative models have become increasingly popular in unsupervised anomaly detection. However, deep generative models aim at recovering the data distribution rather than detecting anomalies. Moreover, deep generative models have the risk of overfitting training samples, which has disastrous effects on anomaly detection performance. To solve the above two problems, we propose a self-adversarial variational autoencoder (adVAE) with a Gaussian anomaly prior assumption. We assume that both the anomalous and the normal prior distribution are Gaussian and have overlaps in the latent space. Therefore, a Gaussian transformer net T is trained to synthesize anomalous but near-normal latent variables. Keeping the original training objective of a variational autoencoder, a generator G tries to distinguish between the normal latent variables encoded by E and the anomalous latent variables synthesized by T, and the encoder E is trained to discriminate whether the output of G is real. These new objectives we added not only give both G and E the ability to discriminate, but also become an additional regularization mechanism to prevent overfitting. Compared with other competitive methods, the proposed model achieves significant improvements in extensive experiments. The employed datasets and our model are available in a Github repository.

With the rapid development of the world economy and the continuous improvement of people’s living standards, users put forward higher requirements for the quality and reliability of the power system. As a representative of deep learning, variational autoencoders play an important role in processing high-dimensional big data. In this paper, a novel variational autoencoder based distribution network abnormal monitoring and positioning method is proposed. Through an indepth study of the variational autoencoder, we utilize the real-time dynamic distribution network status information provided by the wide-area measurement system, combined with the abilities of feature extracting and data reconstruction. Finally, the IEEE-33 simulation model was built in Matlab software and the experimental results were used to verify the correctness of this method. The results show that this method can quickly and accurately achieve the fault detection and location for the distribution network with strong robustness.

A new nonlinear process monitoring algorithm called ensemble local kernel principal component analysis (ELKPCA) is proposed. Conventionally, the performance of kernel-based model depends on the width parameter selected empirically in Gaussian kernel function, which means a single parameter corresponds to a single model. Once a poor width parameter is decided, a single kernel-based model may be only effective for part faults. As a typical kernel-based method, the local kernel principal component analysis (LKPCA), considering both global and local structure information of the original data, faces the problem of width parameter selection as well. Since a single model is one-sided, an available way is to combine different single models and take advantage of them. The ensemble kernel principal component analysis (EKPCA) uses single KPCA models as its sub-models and ensemble learning approach is used to combine them. Due to inherit drawbacks from KPCA models, EKPCA only preserves global structure information of data, but ignores important local structure information. In this paper, to solve the above issues, both LKPCA and EKPCA is unified in the proposed framework. First, single LKPCA models are chosen instead and combined by using ensemble learning strategy. Then two monitoring statistics are turned into fault probabilities through Bayesian inference approach and weighted combination strategy, which makes the monitoring behavior easier and more clear. The result shows that ELKPCA model can not only take advantage of sub-LKPCA models effectively, allowing it selecting width parameter more easily and stably, but also retain both global and local structure information from input data by introducing the local structure analysis in the EKPCA model. Case studies on synthetic example and Tennessee Eastman process demonstrate the proposed method outperforms LKPCA and EKPCA and enhances monitoring performance significantly. (C) 2019 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.