SSD Failure Prediction Model Based on Multimodal Data Fusion

Active failure prediction for solid-state drives (SSDs) is crucial for enhancing the reliability of cloud storage systems. This paper proposes a novel multimodal fusion model for SSD failure prediction, addressing the limitations of existing models that rely solely on time-series data. Our approach introduces the transformation of time-series data from SMART logs into image representations and the conversion of basic drive information and key features into textual data. By integrating these image and text features with the original time-series data, a multimodal framework is constructed to enhance prediction performance. To tackle data imbalance, a Conditional Generative Adversarial Network (CGAN) is employed to augment the failure data. The proposed multimodal neural network employs attention-based fusion to integrate diverse modalities, enhancing the extraction of critical failure-related insights and optimizing the prediction process. Experimental results show that the model achieves higher recall, accuracy, and F1 score while effectively reducing the false positive rate, demonstrating its potential for reliable failure prediction in cloud storage environments.