The increasing complexity and scalability of cloud IT infrastructures have made timely anomaly detection critical, especially in cold start scenarios with limited training data. This paper proposes an innovative approach for early detection of abnormal node load using a classification problem framework. The methodology focuses on optimizing input parameters and increasing metric reading frequency. A typical web application infrastructure was used to configure the approach, utilizing a native dataset of 20 hardware metrics. Feature vector optimization was achieved through feature engineering and correlation analysis. An LSTM model was employed for classification, with hyperparameters optimized using Bayesian methods. Experimental results demonstrate that the proposed approach outperforms existing solutions in accuracy and recall under cold start load balancing conditions. The framework offers a simple architecture with high scalability and adaptability across various business contexts, providing a robust solution for detecting IT infrastructure anomalies in dynamic computing environments.