Abstract:
The reliance on conventional industrial maintenance strategies like reactive and
preventive maintenance leads to unexpected production halts, unnecessary costs,
and reduced equipment efficiency. This proposed work addresses these inefficiencies
by developing a scalable predictive maintenance (PdM) system for industrial
machines, specifically lathe machines. It leverages real-time data from low-cost
sensors and advanced deep learning models to monitor parameters such as vibration,
temperature, and current using an ESP32 microcontroller. The primary contribution
is the implementation of a fault-tolerant architecture utilizing Message
Queuing Telemetry Transport (MQTT), Apache Kafka for high-volume message
buffering, and Apache Spark for real-time data pre-processing. Another identified
gap is the development of an effective AI-based model for predicting machine
condition. To address this, a Patch Time-Series Transformer (PatchTST) model
is employed rather than a traditional Long-Short-Term-Memory (LSTM) model,
which processes 30 days of historical readings (720 past hourly timesteps as the
lookback which is the context used by the model) to forecast the next 5 days (120
future hourly timesteps as the horizon which is the forecast prediction time period
by the model). Experimental results demonstrate high predictive accuracy,
with the ‘current’ feature achieving 91.89%, ‘vibration’ at 81.46%, ‘body temperature’
at 70.51%, and ‘shaft temperature’ at 86.49%. The goal of implementing
this framework is to potentially reduce equipment downtime by approximately
25% based on forecasting-led proactive interventions. This integrated approach
provides a robust, cost-effective solution for real-time industrial condition monitoring.
