الفهرس 
CHAPTER 1 : INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 113 |  |  |
| | | from | 113 | | |
Abstract
A CIoT hardware platform was implemented as shown in Figure 5.1 to meet the proposed solution. The platform has three main parts; a WI-FI module for IoT MQTT connectivity, a PIN header for sensors’ interfacing, and a cortex microcontroller for the firmware implementation of the MLP-NN Model and the IoT MQTT client. It is based on the STMicroelectronics (STM) Board. An STM32F401RE microcontroller (MCU) is the main MCU for this board.
The experimental results are discussed. Experiments were carried out to study and compare the performance of our CIoT multi-topic protocol versus the standard MQTT using different parameters of the real network which affect the performance of protocols. Practical experiments were handled with a real environment using our NTI network infrastructure. The obtained results showed our protocol had lower delay and traffic for multi-topics compared to MQTT. Therefore, the proposed protocol outperformed the MQTT protocol for multi-topic messaging. Moreover, our protocol was better than the batching of multiple messages for real-time system applications.
A smart cognitive IoT devices solution that integrates the AI and IoT system was introduced in the industry automotive field. It has an AI implementation of the MLP-NN with the double hidden layer and the single hidden layer. As a result, the double hidden layer was better in terms of testing and validation accuracy. After that, two solutions were proposed: the smart device solution, which uses the implementation of the MLP-NN model on ARM-Cortex as a limited microcontroller; and the other solution is the cloud-based MLP-NN model, which is implemented on the cloud AI server.
For the first proposed solution, the hardware implementation of the proposed MLP-NN is successfully integrated with the implementation of IoT MQTT and real sensor interfaces on an ARM-Cortex as an example of a limited microcontroller.
The cloud-based solution required more traffic and delay than the smart device solution. The delay difference was bigger from 15% and above in the lossy network. The difference in traffic bytes was 15 bytes at first with no losses and increased with the increasing of the network loss until it reached 152 bytes at 30% network loss. However, the cloud-based system was simple to re-train, as the smart devices required training on another entity and then the re-update was carried out for the MLP model. In addition, updates were easier to be performed on one device in the case of the cloud model rather than the smart device solution.