This paper presents an innovative optimization method for training datasets, aimed at enhancing the accuracy of network anomaly traffic detection. To address the issue of data imbalance, a hybrid sampling strategy is employed. During the oversampling phase, Conditional Tabular Generative Adversarial Network (CTGAN) technology is leveraged to generate high-quality synthetic samples, effectively augmenting data for rare attack types. In the undersampling phase, random selection is used to reduce the number of majority class samples, thereby balancing the data distribution. To tackle the challenge of high feature dimensionality, the Principal Component Analysis (PCA) algorithm is integrated, resulting in the development of the PCA_LSTM model. PCA reduces the dimensionality of the feature space, mitigating redundancy and noise, which enhances the efficiency and accuracy of Long Short-Term Memory (LSTM) networks in processing time series data. Experimental results on the CIC-IDS-2017 dataset demonstrate that the proposed method achieves an anomaly detection accuracy of 98.83%, showcasing its effectiveness and practical value in addressing data imbalance and high feature dimensionality issues.