Novel Multiattribute Space-Based LSTM for Industrial Soft Sensor Applications

The rapid development of industrial technology has led to the expansion of production scales. As a result, industrial data present hard-to-handle characteristics such as high dimensionality, time sequence coupling, and strong nonlinearity. Data-driven modeling techniques can reduce the difficulty in developing industrial soft sensors, and have been widely utilized in industrial production processes. The long short-term memory (LSTM) network is an effective tool for time series forecasting, but its performance is limited when dealing with high-dimensional data. To address this issue, this article proposes a novel approach called multiattribute space-based LSTM (MAS-LSTM). In MAS-LSTM, the K-shape clustering algorithm is utilized to divide the input attribute space of samples into different categories. Each attribute category is then handled by an individual LSTM network, allowing each LSTM network to contain additional attribute category information and process smaller dimension input data. Next, the processing outcomes of multiple LSTMs are combined into a fully connected layer to obtain the prediction results. Finally, a soft sensor based on MAS-LSTM has been established. To test its performance, real-world industrial data from the debutanizer column and the production process of purified terephthalic acid are used in the case study. Simulation results show that the prediction accuracy of MAS-LSTM is greatly improved compared to other models.