Sensorless estimation of manually operated ball valve states using edge artificial intelligence in cyber-physical industrial systems

Abstract

  In industrial process installations, the improper operation 
or misconfiguration of safety-critical components, such as 
manually operated ball valves, can seriously compromise both 
process performance and plant safety. This work proposes a 
sensorless Edge AI method to estimate hand-operated ball 
valves states without the use of physical position sensors. Using 
multivariate time-series data collected from a PLC-based pilot 
plant, a benchmark evaluation is conducted comparing four 
Deep Learning (DL) and four classical Machine Learning  (ML) 
models for classification and regression tasks. The models 
are deployed on an embedded platform, enabling real-time 
inference at the edge with a minimum latency of 500ms. 
Results show Decision Tree (DT) and Random Forest (RF) 
achieve high regression accuracy (R2 >0.98, MAE < 0.5), while 
all eight model reach high classification accuracy. Additionally, 
the computational efficiency metric that combines model 
accuracy, latency, and size, confirming DT as the most efficient 
model (1.83/(ms.KB) for edge deployment. This work contributes 
a cost-effective and scalable monitoring strategy, particularly 
suitable for complex industrial environments where physical 
sensing and visual inspection are limited, offering a viable path 
toward early anomaly detection and intelligent supervision 
within cyber-physical systems.