Predictive maintenance has become a key enabler for improving the reliability, availability, and lifecycle management of modern lift systems. Current solutions typically rely on connected sensor devices that transmit operational data to cloud-based platforms where advanced analytics and artificial intelligence models are executed. While this architecture enables powerful diagnostics and fleet-level insights, it introduces dependencies on connectivity, latency constraints, and additional data transfer requirements.

This paper presents the next development stage of a predictive maintenance system for lifts in which artificial intelligence models are executed directly on-site at the edge device installed within the lift system. The approach builds on an existing cloud-connected monitoring architecture in which substantial preprocessing and feature extraction already take place locally. However, until now the final inference of the deep learning models has been performed in the cloud. By introducing an enhanced hardware platform equipped with dedicated Neural Processing Units (NPUs), it becomes possible to execute Deep Neural Network (DNN) models locally while maintaining acceptable processing time and energy consumption.

Running complex DNN models on embedded systems presents a significant challenge due to limited computing resources. Although NPUs can accelerate neural network inference compared to CPUs, this can lead to trade-offs between speed and model accuracy. Therefore, a balanced approach is necessary. Initial results from the presented architecture are promising and enable predictive real-time diagnostics directly within the lift system. This reduces reliance on cloud connections, improves response times, and paves the way for more autonomous and scalable predictive maintenance solutions in the lift industry.