Design of a distributed big data platform for streaming vibroacoustic diagnostics and predictive remaining-useful-life forecasting of the serverand-power infrastructure of a digital agricultural complex information network based on Apache Kafka and a hybrid CNN-LSTM model
Year: 2026
Type: article
Abstract: The paper presents the results of designing and experimentally verifying a distributed big data platform for streaming processing of vibroacoustic and telemetry data from the server-andpower infrastructure of an edge data center that serves the information network of a digital agricultural complex, together with predictive forecasting of the remaining useful life of its critical units. A five-layer architecture is proposed: data sources (accelerometers, acoustic-emission sensors, a non-contact laser vibrometer, current and temperature sensors), edge preprocessing on Raspberry Pi 4B, the Apache Kafka 3.6 event bus, processing on Apache Spark 3.5 with HDFS, and an analytical layer with the CNN-LSTM-Attention model. A pilot deployment was carried out on a modular edge data center of a digital agricultural complex in the Republic of Tatarstan with a 3.6 TB archive over 18 months. The architecture delivered a P99 latency of 372 ms under a peak load of 1.18 million events per second with losses below 0.02%. The CNN-LSTM-Attention model achieved R² = 0.931, RMSE = 44.7 h, outperforming LSTM by 14.2% in RMSE. Edge aggregation of vibroacoustic streams reduces network traffic to the core by 72%.
Source: Agrarian science
Institution: Kazan State Power Engineering University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7203490232