The rapid advancement of AI has led to its integration into various areas, especially with Large Language Models (LLMs) significantly enhancing capabilities in Artificial Intelligence Generated Content (AIGC). However, the complexity of AI systems has also exposed their vulnerabilities, necessitating robust methods for Failure Analysis (FA) and Fault Injection (FI) to ensure resilience and reliability. Despite the importance of these techniques, there lacks a comprehensive review of FA and FI methodologies in AI systems. This study fills this gap by presenting a detailed survey of existing FA and FI approaches across six layers of AI systems. We systematically analyze 142 studies to answer three research questions including (1) what are the prevalent failures in AI systems, (2) what types of faults can current FI tools simulate, (3) what gaps exist between the simulated faults and real-world failures. Our findings reveal a taxonomy of AI system failures, assess the capabilities of existing FI tools, and highlight discrepancies between real-world and simulated failures. Moreover, this survey contributes to the field by providing a framework for fault diagnosis, evaluating the state-of-the-art in FI, and identifying areas for improvement in FI techniques to enhance the resilience of AI systems.

Deep learning models have been successfully adopted in anomaly detection for multivariate time series data in various fields. These models are good at capturing complex time dependencies and extracting meaningful patterns from time series data. However, the trained models may become outdated due to unforeseen changes in real-world data, which can lead to a decrease in the quality of model service. Therefore, it is crucial to continuously monitor the performance of the model and analyze its behavior to ensure its reliability and availability. We propose an online data drift detection method that uses an unsupervised deep learning network, Variational Autoencoder (VAE), to monitor deep learning models in the field of multivariate time series anomaly detection. This method consists of three main steps namely data collection and statistical analysis, real-time drift detection, and drift interpretation. We collect raw time series data and model prediction data non-invasively from the model server. Then they are separated into windows for drift detection. Furthermore, the method can provide analysis and interpretation when drift is detected. Our evaluation experiments involve three real-world datasets from various industrial domains and four different structured anomaly detection models. We validate the effectiveness of drift detection in multivariate time series, and then test how the anomaly detection models perform during data drift detection. The highest improvement in F1 score is approximately 0.16. In addition, we provide an analysis of the interpretability of the model performance.