A multichannel autoencoder deep learning approach is developed to address the present intrusion detection systems’ detection accuracy and false alarm rate. First, two separate autoencoders are trained with average traffic and assault traffic. The original samples and the two additional feature vectors comprise a multichannel feature vector. Next, a one-dimensional convolution neural network (CNN) learns probable relationships across channels to better discriminate between ordinary and attack traffic. Unaided multichannel characteristic learning and supervised cross-channel characteristic dependency are used to develop an effective intrusion detection model. The scope of this research is that the method described in this study may significantly minimize false positives while also improving the detection accuracy of unknown attacks, which is the focus of this paper. This research was done in order to improve intrusion detection prediction performance. The autoencoder can successfully reduce the number of features while also allowing for easy integration with different neural networks; it can reduce the time it takes to train a model while also improving its detection accuracy. An evolutionary algorithm is utilized to discover the ideal topology set of the CNN model to maximize the hyperparameters and improve the network’s capacity to recognize interchannel dependencies. This paper is based on the multichannel autoencoder’s effectiveness; the fourth experiment is a comparative analysis, which proves the benefits of the approach in this article by correlating it to the findings of various different intrusion detection methods. This technique outperforms previous intrusion detection algorithms in several datasets and has superior forecast accuracy.