Abstract: A lightweight and transparent approach for predicting maintenance of pump systems at the edge is outlined via statistical modeling of time series data, an alternative to other supervised learning techniques that typically require a large labeled data sets (usually tens of thousands) and substantial computational resources. The methodology consists of using Exponential Moving Average, Trend Strength Detection, and Z-score-based anomaly scores to assess the health of pump systems on-line, rather than relying on off-line processing techniques primarily based on deep-learning methodologies. Multi-sensor data including vibration, sound and temperature will be processed in real-time at the edge using a Raspberry Pi computing platform. Exponential Moving Average reduces data noise while retaining and displaying slow degradation trends; Trend Strength Analysis can identify gradual trends in machine health; and the Z-scores serve to quantify deviations from an expected or healthy condition. By combining these outputs into a single Health Index, an easy-to-understand measurement of machine health can be generated. Experimental validation performed under both healthy and simulated degraded conditions demonstrated successful identification of mechanical imbalance, acoustic disturbance and thermal overload. The model will not require the use of supervised learning, therefore providing lower latency, transparency and applicability to lower resourced industrial settings along with offering a cost effective, scalable and less computationally intensive alternative to existing predictive maintenance methodologies.

Keywords: Predictive Maintenance, Edge Computing, Statistical Anomaly Detection, Exponential Moving Average (EMA), Trend Analysis, Industrial Pump Monitoring

Downloads: | DOI: 10.17148/IJIREEICE.2026.14383