Figure 2. The general architecture of the study
Our next top layer is the edge layer, hierarchically. The purpose of this layer is to reduce the danger of network traffic that may occur on the server if the data from a lower layer is transmitted to the central server at once. Another purpose is to prevent data sending devices from being suspended while waiting for a response in return for the data they send. A simply modified version of the developed IoT_TH board is used to build this layer. This change can be described as deactivating the original sensors. In this layer, there are potentially different number of node devices under the responsibility of IoT_TH devices. Ɐ(THx) transmits the values in its environment to the Nx device it is connected to. What is expected from this layer in the whole organization is that the collecting devices in the layer correctly label and accumulate the media data coming from the node devices they are responsible for and periodically direct them to the upper layer, the decision layer. The obtained values are sent to the collection devices in the upper layer, the Edge Layer, in a format (↋1,↋2,…,↋t) prepared in accordance with the smart contract format prepared for the blockchain. Each collection device ↋t then stores the data sent to it by the Nn devices as shown in Equation 1. The data that is encrypted and stored in each ↋tdevice is sent to the Decision Layer, which is the upper layer, in fixed periods.
\({}_{t}=\sum_{n=1}^{N_{n}}\left(N_{\text{Temp}_{n}}\right)_{n}\bigcup\ \left(N_{\text{Hum}_{n}}\right)_{n}\)(1)
The detailing of the layers in the hierarchical structure will be continued with the last layer, not the next layer. We can explain the reason for this as the training part and the classification part in machine learning algorithms are inverse sequential in the architecture realized in this study.
In the last layer, Education and blockchain layer, there is one local server and one cloud server. Two main operations are performed in this layer. The first of these processes is to carry out the training process of the artificial intelligence model with data from different locations. Another process is to convert the incoming data into a blockchain structure and store it. In the first operation, a model is trained using the LSTM algorithm. The weight values ​​of the input variables formed after the training are carefully extracted from the model. these extracted weights are meaningful for the developed model and meaningless for the human mind. The weights obtained are transferred to the decision layer, which is a lower layer, in order to perform the classification process. In the second main transaction, the incoming data is safely stored in its raw form, under the guarantee of blockchain, in a re-readable form when necessary. It is configured for both local and cloud server storage in the blockchain layer, which is the last layer where data reaches. The necessity of a dual storage system can be decided depending on the importance of the data to be obtained and transferred as a requirement of the infrastructure developed in this study. In this layer, a single piece of operation that can be performed through a machine learning algorithm is fragmented. this distribution process has brought much more effective and successful results on heterogeneous data. We proved this with numerical data in section 3.3.
Virtual machines (J1, J2,…, Jk) are located in the Decision Layer. Jk collects the encrypted data sent to it from ↋t and sends it to the final blockchain storage layer as shown in Equation 2.
\(J_{k}=\sum_{t=1}^{{}_{t}}\left({}_{N_{x}}\right)_{t}\) (2)
The decision layer contains virtual machines (J1, J2,…, Jk). Jk accumulates the encrypted data sent to it from ↋t and sends it to the last layer as shown in Equation 2. The decision layer completes the work of the fragmented machine learning algorithm thanks to the trained model weights it receives from the training layer. Classification processes are carried out without the need for new training only thanks to the weights, and the results are sent to the lower layer, the edge layer, for implementation.