This paper proposes an attention-enhanced Long Short-Term Memory (LSTM) model for anomaly detection in multivariate time series data from industrial servers. The model incorporates an attention mechanism into the standard LSTM architecture to weight its hidden state outputs, thereby capturing intrinsic temporal features more effectively and improving prediction accuracy for multivariate performance metrics in industrial servers. In practical application, the process begins with an autocorrelation function (ACF) analysis to evaluate the temporal dependencies within the multivariate performance data and determine an appropriate sliding window length. The prediction model is then trained on server data under normal operating conditions, using historical multivariate features to forecast future time series values. Finally, prediction errors are computed and anomalies are identified through a three-level detection strategy combining global thresholding, local maxima analysis, and peak width filtering. This approach effectively distinguishes true anomalies while mitigating the impact of long-term error drift. The method is validated on the Server Machine Dataset (SMD), and experimental results demonstrate its outstanding performance in detecting both long-duration and transient anomalies. Additionally, the proposed method offers rapid detection response, providing an efficient and reliable solution for intelligent and secure operation of industrial servers.