Network intrusion detection technology is a vital approach in the current network security systems, mainly for identifying abnormal traffic behavior in large-scale cyberspace. Existing network intrusion detection methods are limited by feature redundancy and category imbalance, making it difficult to effectively identify complex and variable attacks with high false alarm rates. We introduce DCBANET, an intrusion detection approach leveraging a multi-scale deep neural network, reinforced by embedded feature selection and EQLv2. A denoising autoencoder is first applied to the preprocessed data for feature selection, producing a subset that effectively represents the original high-dimensional heterogeneous network data. Next, we adopt the Equalization Loss v2 (EQLv2), which adjusts the gradient flow of positive and negative samples in a dynamic manner to mitigate class imbalance. Meanwhile, a deep learning architecture is developed that fuses multiple channels, combining convolutional neural network, bi-directional LSTM, and multi-head attention mechanism to capture local spatial patterns, model temporal dependencies, and strengthen key feature representation. Experimental findings show that the proposed method exhibits better performance across all four datasets, NSL-KDD, UNSW-NB15, CIC-IDS-2017 and CICIDOS2019, with accuracy levels of 99.65%, 98.79%, 99.46%, and 99.62%, respectively. The detection ability against intrusions is effectively improved.