Non-orthogonal multiple access (NOMA) with a grant-free access has received a lot of attention due to its support to massive machine-type communication (mMTC) devices. The devices in grant-free systems are allowed to transmit information without undergoing an authentication process. Therefore, in such systems base station needs to distinguish between active and non-active devices, called the active user detection process. This process is challenging as the active device needs to be detected from received signals that are superimposed. Furthermore, the identification of the Internet of Things (IoT) devices from these signals also poses a great challenge, that could help to allocate resources in future generation communication systems. Motivated from the aforementioned facts, this paper proposes a device detection and identification (DDI) architecture for joint active user detection and IoT device identification from the received superimposed signals. The architecture extracts Fourier patterns as the representative feature vector, which results in improved detection and identification process. Experimental results show that the architecture not only outperforms the conventional schemes and deep neural network-based approaches in terms of success probability for the active user detection task but also yields lower computational complexity. The evaluation of DDI architecture for IoT device identification problems has also been performed and compared with various shallow learning methods to prove its efficacy.