Since the increase in internet attacks brings much damage, it is essential to take care of the security of network activities. networks must use different security systems, such as intrusion detection systems, to deal with attacks. This research proposes a reliable approach for intrusion detection systems based on anomaly networks. The network traffic data sets are large and unbalanced, affecting intrusion detection systems' performance. The imbalance has caused the minority class to be incorrectly identified by conventional data mining algorithms. By ignoring the example of this class, we tried to increase the overall accuracy, while the correct example of the minority class protocols is also essential. In the proposed method, network penetration detection based on the combination of multi-dimensional features and homogeneous cumulative set learning was proposed, which has three stages: the first stage, based on the characteristics of the data, several original datasets of raw data or datasets criteria are extracted. Then, the original feature datasets are combined to form multiple comprehensive feature datasets. Finally, the same basic algorithm is used to train different comprehensive feature datasets for the multi-dimensional subspace of features. An initial classifier is trained, and the predicted probabilities of all the basic classifiers are entered into a meta-module. In this research, an AdaBoost meta-algorithm has been used for unbalanced data according to a suitable design. Also, various single CNN models and multi-CNN fusion models have been proposed, implemented, and trained. This evaluation is done with the NSL-KDD dataset to solve some of the inherent problems of the KDD'99 dataset. Simulations were performed to evaluate the performance of the proposed model on the mentioned data sets. This proposed method's accuracy and detection rate obtained better results than other methods.