As key equipment in modern industry, it is important to diagnose and predict the health status of bearings. Data-driven methods for remaining useful life (RUL) prognostics have achieved excellent performance in recent years compared to traditional methods based on physical models. In this paper, we propose a novel data-driven method for predicting the remaining useful life of bearings based on a deep graph convolutional neural network with spatiotemporal domain convolution. This network uses the average sliding root mean square (ASRMS) as the health factor to identify the healthy and degraded states, and then uses correlation coefficient analysis on the hybrid features of the degraded data to construct a spatial graph according to the strength of the correlation between the obtained features. In the time domain, we introduce historical data as the input to the temporal convolution. After the data are processed by the spatial map and the temporal dimension, we perform the prediction of the remaining useful life. The experimental results show the accuracy of the method.

Hot blast Stove (HBS) provides hot air for the blast furnace, and temperature prediction for its exhaust gas is of vital importance to control the process. In this paper, a novel Deep Memory Echo State Network (DMESN) is proposed for temperature prediction. Firstly, data preprocessing is performed, including outlier rejecting, missing data handling and lag time calculating, so that better dynamic characteristics of data set can be obtained. Then, in order to improve the prediction accuracy, an improved hidden layer structure is proposed, which consists of two parts: echo state formation and hidden state formation. Echo state formation is used in network starting phase. Aiming at the problem that information cannot be effectively retained, three gates are designed in the cell to obtain a stable echo state which will be used to initialize the hidden state. In this way, the hidden state can operate continuously and stably. Finally, comparing experiments were carried out to demonstrate the efficiency of the proposed method. The simulation results show that DMESN achieves excellent performance in terms of accuracy, stability and learning speed.

Data-driven soft sensor modeling is a significant research area that aims to extract reliable and comprehensive features from massive data in production processes, thereby realizing the practical value of data for guiding actual production. However, existing research primarily focuses on extracting features in a single dimension, overlooking the multidimensional dependencies among equipment in the production process. To address this issue, this article proposes a spatiotemporal graph attention network (ST-GAT). The network constructs a spatial directed graph model to capture the connectivity among multiple equipment while considering the temporal variations of equipment data and their impact on the target variable. ST-GAT performs spatiotemporal feature extraction on both the equipment themselves and the equipment connected to them, aiming to represent the data of each production equipment more comprehensively. Furthermore, to effectively utilize equipment features relevant to the target prediction, an adaptive weighting module (AW) is introduced to evaluate the importance of each equipment feature for the target variable, thereby improving the utilization of important features. In addition, to overcome the limitations of traditional single activation functions, a variable activation function (VAF) is proposed, which adaptively selects activation functions to enhance the network’s autonomous learning capability. Finally, experiments are conducted on a real dataset from the suspended magnetization roasting (SMR) production process. Through multiple modeling experiments and comprehensive performance evaluations, ST-GAT achieves the best prediction results on the SMR process dataset, demonstrating the crucial role of modeling spatiotemporal dependencies among equipment and further validating the effectiveness of ST-GAT.