Paper #4

Review and Utilisation of AI in Signal Processing

Johan Note, Maaruf Ali and Lekë Pepkolaj

Abstract: Numerous challenges persist within Signal Processing (SP) in communication systems, i.e. maintaining system stability, complexity of multi-dimensional SP and filtering. Many resources are required to cope with these growing problems. The application of AI offers promising solutions. AI is already being applied to solve engineering, medical and scientific problems, including utilisation in SP and Digital Signal Processing (DSP). The research methodology began with common supervised AI algorithms in SP: Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNNs). Regular unsupervised AI learning algorithms investigated included: Sparse Auto-encoder, Deep Auto-encoder, Convolutional Auto-encoder and the De-noising Auto-encoder. Their benefits were ascertained for their utilisation in an intelligent, scalable and secure communication system. The paper also provides over 100 references. The conclusion is that using specifically identified AI algorithms will make communication systems more scalable, thus mitigating overloading, instability and premature breaking point failure. AI use will also eventually make communication systems less expensive, more agile and far more intelligent - essential for Cognitive Radio and SDR (Software Defined Radio). The communication channels will become more reliable with AI SP and aid in its increased robustness in noise, quality and error recovery, compared with using traditional SP. These benefits will cause a shift toward using new methods in SP such as the Advantage Actor Critic and Trust Region Policy Gradient Algorithms. Two different datasets comprising the MNIST (Modified National Institute of Standards and Technology) database consisting of 70,000 28�28 pixels greyscale images of handwritten integers and the CIFAR-10 (Canadian Institute for Advanced Research) database containing 60,000 32�32 pixels small colour images were utilised. These were then processed using the aforementioned algorithms to test their accuracy. Because of 6G radio channel data scarcity for training purposes, image datasets were used to check the feasibility of the effectiveness of AI in SP in general radio communications, due to their two-dimensional nature. The result showed that all seven AI models performed well: 80% - 99.77% (accuracies). The CCN-Autoencoder gave the best result overall using both datasets: 94.5%-99.77% (accuracies). Future work will use real radio communication channel data and other AI algorithms.

Keywords: AI; Artificial Neural Networks; Convolutional Neural Networks; Denoising Auto-encoder; Signal Processing; Sparse Auto-encoder.