Scalable state change detection for stimulus-aware digital twins

Abstract

One of the fundamental functionalities of a digital twin of any production system is its ability to serve as a ‘twin’, i.e. effective synchronization between the physical state and the digital state. These states can include being idle, working, failure, setup, slowdown, etc. Application-based top-down approaches to developing digital twins generally consider pre-calculated states or known data conditions, limiting their scope. In contrast, real-life production presents a myriad of possible states. This stimulus-awareness serves as the foundation for realistically modeling production behavior in real time, which leads to improved production planning, simulation, execution, and management. Thus, a bottom-up scalable state detection method and a systematic four-phase learning framework using minimal data are proposed for realizing stimulus-aware digital twins at the beginning of equipment operation. The proposed solution uses an unsupervised autoencoder for two-stage pattern recognition and PCA-T2 for change detection learning using a single state, followed by new state detection and adaptation using a novel implementation of a custom loss function to maximize the distribution discrepancy between the latent representation of the two states. Validation is performed over three real-life datasets, including a newly created gearbox dataset that varies in type of equipment, signal, operating parameters, etc. 80 experiments are conducted across different scenarios and repeated 5 times each, along with statistical testing to benchmark performance with the traditional approach. To further benchmark the performance of the proposed model, Generative Adversarial Networks (GANs) have also been explored. An additional 10 experiments were conducted across various scenarios, with each experiment repeated five times to assess the model's performance. The proposed method detects a change in state with over 97.5% accuracy and an F1 score of more than 95.2%, irrespective of the dataset or the states used for modeling, narrowing the gap between localized success stories and the large-scale use of digital twins in production industries.