Triangular-modular convolutions as a neural network alternative for technical condition forecasting

Бесплатный доступ

The relevance of the study stems from the limitations of neural network methods (LSTM, CNN) in forecasting the technical condition of industrial equipment, including high computational complexity, low interpretability, and vulnerability to cyberattacks. The paper introduces a deterministic triangular-modular convolution method, which combines triangular number theory and modular arithmetic for processing sensor data (vibration, temperature, pressure). Deterministic data aggregation (DRA) was applied to reduce noise (SNR improvement by 15 dB) and compress data (10:1). Testing on 12-month IoT datasets (100 samples) revealed TriModConv s superiority over LSTM in accuracy (F1-score: 0,96 vs. 0,78) and processing speed (1,8× faster, 800 ms/sample). The probability of a successful cyberattack was below 10−9. The proposed method is highly effective for real-time equipment monitoring, offering determinism, interpretability, and security.

Еще

Deterministic algorithms, sensor fusion, predictive maintenance, modular convolution, blockchain, IoT diagnostics

Короткий адрес: https://sciup.org/148331829

IDR: 148331829   |   УДК: 2.9.5