Predictive Maintenance in Smart Systems with Temporal Convolutional Networks (TCN) and Autoencoders: Algorithm Optimization, Intelligent Systems, Blockchain, Cryptography and Cybersecurity

Predictive maintenance in smart systems demands accurate anomaly detection and fault prediction amidst noisy, multivariate time-series data. We propose a novel TCN-DTDAE framework, integrating temporal convolutional networks (TCN) with a dynamic threshold denoising autoencoder (DTDAE), to enhance predictive maintenance performance. The TCN leverages dilated convolutions to capture long-term temporal dependencies in sensor data, such as vibration and temperature, producing robust feature maps. The DTDAE, a key innovation, employs adaptive thresholding based on a Gaussian mixture model (GMM) of reconstruction errors, effectively distinguishing normal operations from anomalous and fault states. Evaluated on a simulated industrial dataset with 10,000 samples (80% normal, 15% anomalous, 5% faults), our method achieves a 0.92 F1 score, 0.93 precision, and a 0.05 false positive rate (FPR), outperforming baselines like Long Short-Term Memory Autoencoder (LSTM-AE) (0.81 F1 score) and TCN-AE (0.87 F1 score). The model detects 470 out of 500 defects and has 89% pre-fault early warning accuracy with the risk of downtime minimized. TCN-DTDAE with an inference time of 10 ms/sample is real-time feasible. The approach encourages the credibility of intelligent systems by a broad solution that is industry-applicable.