Although different infrastructures are used to create the blockchain structure in the studies, Hyperledger Fabric infrastructure is generally preferred. In addition, data integrity, data security, tamper resistance and distributed structure have been successfully carried out in many studies. In this study, a four-layer architecture is presented to increase the learning success of data processing systems in distributed locations, to achieve higher accuracy and to create a secure storage. The data from the nodes are incrementally collected to the central server and trained with the LSTM deep learning algorithm. The weights containing the learning results are transferred to the decision layer and all node devices are used. In addition, post-training data is safely stored in a blockchain-based storage system. Thus, learning capabilities have been transferred to all devices. In the tests and evaluations, each node showed high learning performance in new data, apart from its own data diversity. As a result, it is aimed to solve the problems experienced in cases where the data obtained from a location, the number of data and the diversity are insufficient, with the central architecture. In particular, it is foreseen that distributed systems will be processed securely and with higher accuracy, and that IoT systems will be used widely and effectively.