IOT-INTEGRATED DEEP NEURAL PREDICTIVE MAINTENANCE SYSTEM WITH VIBRATION-SIGNAL DIAGNOSTICS IN SMART FACTORIES

Abstract

This quantitative study examined an IoT-Integrated Deep Neural Predictive Maintenance System with Vibration-Signal Diagnostics in Smart Factories through an empirical evaluation informed by a structured review of 92 peer-reviewed journal articles and conference papers on vibration-based diagnostics, deep learning architectures, and industrial IoT deployment practices. Vibration data were acquired from 24 rotating assets operating across two smart-factory production lines at a sampling rate of 25.6 kHz and were segmented into 2.0-second windows with 50% overlap, producing 1,152,000 diagnostic windows and 38,400 prognostic sequences composed of 30 windows per sequence. The experimental design compared signal representation types (raw 1D, STFT, wavelet), model architecture families (1D CNN, temporal CNN, LSTM/GRU), inference placement (edge versus cloud), and data-quality conditions (baseline, noise, and missingness). Descriptive and inferential analyses showed that time–frequency representations consistently outperformed raw time-domain inputs for fault diagnosis and remaining useful life estimation. Wavelet-based cloud configurations achieved a macro F1 of 0.934 and PR-AUC of 0.953, compared with macro F1 of 0.881 and PR-AUC of 0.904 for raw 1D edge configurations. Prognostic accuracy followed a similar pattern, with wavelet-based cloud pipelines producing RUL MAE of 6.8 hours and RMSE of 9.9 hours, compared with 8.6 hours MAE and 12.4 hours RMSE for raw 1D edge pipelines. Mixed-effects regression confirmed statistically significant effects of representation type on diagnostic performance (wavelet versus raw, β = 0.041 for macro F1, p < 0.001) and significant degradation under missingness conditions (β = −0.034, p < 0.001). System-level analysis showed that cloud placement increased median latency by 102.6 ms, P95 latency by 181.2 ms, and bandwidth usage by 8.63 Mbps (p < 0.001), while sustaining higher throughput by 78.4 windows/s. Overall, the findings demonstrated that predictive accuracy and prognostic reliability were primarily governed by representation and architecture choices, whereas IoT placement chiefly determined latency and bandwidth behavior under realistic smart-factory streaming conditions.