GDRNet: A deep learning framework for dynamic sliding window determination and improved prediction accuracy in online IIoT systems

The application of machine learning (ML) on IoT data is pivotal for developing intelligent industrial systems. This involves using a sliding window to extract time slices for ML prediction models. Existing methods rely on empirical, statistical, and optimisation or heuristic-based approaches to determine optimal sliding window size. Recent deep learning methods adaptively adjust or internally restructure windows, but they do not explicitly predict window size as a supervised target, particularly in multivariate IIoT contexts. In this paper, we introduce GDRNet, a novel deep learning-based approach for dynamically determining the optimal sliding window size to achieve improved prediction accuracy in multivariate Industrial IoT (IIoT) systems. The generative component (G) of GDRNet employs an LSTM-based variational autoencoder model to generate a large labeled multivariate dataset. The deep regression (DR) component employs a multi-layer perceptron (MLP) that is trained offline and predicts common optimal sliding window online. Unlike transformer-based or self-supervised forecasting models that assume large fixed contexts, GDRNet frames sliding window determination as a regression problem, enabling explicit and efficient prediction of optimal window sizes. Through extensive experimentation with multiple machine learning models and real-world IIoT datasets, we demonstrate that GDRNet achieves competitive or superior prediction accuracy with substantially higher computational efficiency and scalability as compared to the traditional methods, making it suitable for online IIoT systems. We also establish the generalisability of GDRNet through applying it across three real world IIoT datasets of different domains and contexts.