Missing data commonly exists in industry processes, primarily due to sensor failures, signal transmission issues. Effective key performance indicator (KPI)-related fault detection for processes with missing data is essential for improving product quality and yield. This paper proposes a novel framework to monitor KPI-related faults, which consists of data imputation and fault detection. During the data imputation process, a novel location embedding generative adversarial imputation network (LEGAIN) is proposed. The LEGAIN transforms the distance between each pair of observation points into a position-aware weight matrix, which captures the location information to facilitate accurate data imputation. LEGAIN introduces a symmetric window mask strategy to enhance its ability in capturing local dynamic characteristics. During the fault detection process, the deep feature correlation analysis (DFCA) is proposed. The DFCA introduces a dual-branch variational autoencoder (VAE) to separately extract nonlinear features from process variables and KPIs. The spatial-temporal correlation between these two sets of features is analyzed and considered as a part of the loss function. Generalized singular value decomposition is applied to effectively distinguish KPI-related subspace and KPI-unrelated subspace. Corresponding theory analysis is provided. Experiment results demonstrate the superiority of the proposed method.

Quality-oriented fault detection facilitates intelligent inspection, operation optimization, and product quality improvement in the industrial Internet of Things. Data-driven latent variable-based approaches are among the mainstream methods in this field. However, traditional latent variable methods establish static relational models, which are less suited to the dynamic characteristics of industrial processes. To address this issue, dynamic latent variable methods, such as dynamic inner partial least squares (DiPLS), have been proposed. Despite this, the DiPLS method only enhances the descriptive capability of the partial least squares (PLS) method for dynamic characterization and does not overcome the inherent limitations of the PLS method in quality-oriented fault detection. Specifically, the residual subspace obtained by the PLS still contains quality-related information, which undoubtedly increases the false alarm rates and missed alarm rates in subsequent detection. Additionally, the DiPLS method does not address the problem of overfitting during the modeling process. To deal with the above issues, this article proposes a quality-oriented dynamic sparse latent variable method for efficient fault detection with respect to quality indicators. This method introduces a weight matrix to sparsity the coefficient matrix, enhancing the interpretability of the model while addressing the overfitting problem that may occur during the DiPLS modeling process. Furthermore, an alternating direction method of multipliers-based technology is studied to solve the corresponding optimization problem. To further refine the orthogonal decomposition of the process variable space, a joint temporal and spatial decomposition strategy is presented. This strategy accomplishes the orthogonal decomposition of the process variable space based on latent variable and time lag components, extracting the latent variables accordingly. Subsequently, the detection statistics and logic are derived using the Bayesian fusion. Finally, the effectiveness of the proposed method is verified through a numerical simulation and two industrial cases.

In modern Industrial Production Processes(IPP), the increasing complexity poses challenges for operation analysis and fault path recognition. This study proposes a dynamic modeling approach using Boolean Networks (BNs) for IPP operation analysis, offering greater dynamism and interpretability compared to existing approaches. Furthermore, different abnormal events lead to the same fault. To bridge this gap, we propose using BNs attractor cycle and semi tensor product (STP) bijective technique to convert fault paths into data sequences. This enables precise identification of fault paths and impacts even when different root causes lead to similar failures. Subsequently, we provide strategies to mitigate fault impacts and use BNs nodes fault propagation topology graphs to prove the interpretability. The results show this method is effective in handling BNs complexity and fault path length.

Water level prediction is of great importance in alleviating the increasing water scarcity and preventing frequent floods. However, current water level prediction models do not consider the spatial and temporal features in water levels at monitoring stations. This study proposes a ladder water level prediction model for the Yangtze River based on transfer learning and Transformer to obtain more accurate predictions of water levels under tidal interactions. Our model utilizes the attention mechanism of the Transformer and incorporates spatial features and correlations of water level variations at the tidal limit of rivers. In addition, transfer learning is employed to explore and analyze the temporal characteristics of the wet season and dry season. Numerical experiments conducted at monitoring stations in the lower reach of the Yangtze River in China validate the effectiveness of our model. In 24-h water level prediction, our model achieves an average reduction of 52.14% in mean absolute error (MAE), 52.99% in root mean square error (RMSE), and an average increase of 5.70% in the pass rate within ±0.3 m compared to existing studies.

The inference of Boolean networks is crucial for analyzing the topology and dynamics of gene regulatory networks. Many data-driven approaches using evolutionary algorithms have been proposed based on time-series data. However, the ability to infer both network topology and dynamics is restricted by their inflexible encoding schemes. To address this problem, we propose a novel Boolean network inference algorithm for inferring both network topology and dynamics simultaneously. The main idea is that, we use a marker-based genetic algorithm to encode both regulatory nodes and logical operators in a chromosome. By using the markers and introducing more logical operators, the proposed algorithm can infer more diverse candidate Boolean functions. The proposed algorithm is applied to five networks, including two artificial Boolean networks and three real-world gene regulatory networks. Compared with other algorithms, the experimental results demonstrate that our proposed algorithm infers more accurate topology and dynamics.