Big Data platform architecture for predictive diagnostics of high-voltage electrical equipment in agro-industrial complexes based on vibroacoustic signal analysis

This article presents the architecture of a Big Data platform for predictive diagnostics of highvoltage insulators at substations of agro-industrial complexes. The problem is that existing vibration diagnostics methods are inapplicable to rural substations due to remoteness, unstable LTE-M and VSAT communication, as well as high-frequency vibroacoustic signals in the range of 20–250 kHz, which are two orders of magnitude higher than the ordinary spectrum of mechanical components. A five-layer architecture has been developed, including MEMS and piezoelectric sensors, NVIDIA Jetson AGX Orin edge nodes with wavelet packet decomposition and Zstandard compression (compression ratio of 12.7×), Apache Kafka brokers, an Apache Spark Structured Streaming distributed computing layer, and a hybrid CNN-LSTM model with an attention mechanism. An experiment at five substations in the Republic of Tatarstan over 12 months showed that 2.34 PB of raw data were collected and 184 TB were transmitted. The classification accuracy for six condition classes reached 98.4% (F1 = 0.977). The integrated health index HI provides a scale from 0.92 (healthy) to 0.14 (pre-breakdown). Implementation of the platform reduced unpredicted failures by 71% and unit maintenance costs by 34.2%. The reliable remaining useful life prediction horizon was 38 days. Without periodic retraining, model accuracy decreases by 5.7 percentage points per year, requiring quarterly recalibration.