Simultaneously anomaly detection and forecasting for predictive maintenance using a zero-cost differentiable architecture search-based network
Laio Oriel Seman, Luiza Scapinello Aquino, Stéfano Frizzo Stefenon, Kin‐Choong Yow, Viviana Cocco Mariani, Leandro dos Santos Coelho
Abstract
To prevent costly failures and unplanned downtime, predictive maintenance for industrial machinery requires accurate forecasting and early anomaly recognition. This paper introduces a novel zero-cost Differentiable Neural Architecture Search framework for Vibration Analysis (DNAS-VA) that simultaneously optimizes forecasting and anomaly detection in vibration signals. The proposed approach automatically discovers the most appropriate neural network architectures by exploring a search space combining time and frequency-domain operations, including Fourier and wavelet transforms, attention mechanisms, and temporal modeling components. A Forecasting-Integrated Variational Autoencoder (FI-VAE) enhances anomaly detection by combining reconstruction error, latent space analysis, and temporal pattern assessment. The methodology employs a hierarchical training protocol to optimize both architecture search and anomaly detection performance. Experiments in real triaxial vibration data from an industrial motor demonstrate the framework’s effectiveness. The discovered architecture achieves superior forecasting performance, with mean absolute errors of 0.118–0.156 across vibration axes, and robust anomaly detection, outperforming baseline methods like Isolation Forest. Main innovations include a multi-fidelity evaluation strategy using zero-cost metrics, such as Fisher Information, correlation equal 0.90, to efficiently identify high-performing architectures without full training cycles. Latent space analysis reveals interpretable clusters corresponding to operational states, with anomalies detected at cluster boundaries. The results show that the integrated framework significantly improves predictive maintenance by enabling accurate forecasting and reliable early fault detection while reducing computational costs. The proposed method achieves state-of-the-art performance in both tasks, offering a scalable solution for industrial condition monitoring.