Blockchain-based Predictive Maintenance Application with Deep Learning

Abstract

This thesis presents a comprehensive predictive maintenance system and application interface that integrates deep learning and blockchain technologies in order to enhance maintenance strategies in industrial systems. Traditional predictive maintenance systems have significant issues regarding data security and decentralization. This study aims to address these limitations by leveraging blockchain technology, with a specific focus on improving the reliability and verifiability of predictive maintenance processes. In this study, an LSTM-CNN hybrid model was developed to evaluate complex patterns in both time and features, thereby enabling high-accuracy fault prediction. The proposed model is designed to perform binary classification for fault prediction in industrial equipment. During the implementation phase of the study, an open-source dataset was used to train and test the developed model. The Randomized Search method was used in the hyperparameter optimization process to increase the prediction success of the proposed model. The hybrid model was trained with 5-fold cross-validation, and class weighting and threshold value optimization methods were applied to eliminate the class imbalance problem. In the threshold optimization phase, F1-score-based methods are applied to maximize recall at three predefined minimum precision levels (0.05, 0.2, and 0.85), while identifying the most balanced trade-off between precision and recall. In the proposed system, sensor data are stored in a database (SQLite3), and cryptographic proofs generated using zero-knowledge techniques are transmitted to the Ethereum network. The Poseidon hash function is used to ensure data integrity, and the Groth16 protocol is used for Zk-Snark proof generation. This approach enables secure verification of data validity without publicly disclosing sensor data and simultaneously addresses scalability concerns. The system architecture is designed to include manager, operator, and engineer nodes, and all smart contracts are implemented using Solidity. In addition, a graphical user interface is developed using the Tkinter library in Python. The experimental results demonstrate that the proposed LSTM–CNN hybrid model produces successful outcomes in terms of fault prediction performance. According to scenario where the decision threshold is optimized based on the F1-score, the model achieves an accuracy of 0.987, an AUC value of 0.979, and an F1-score of 0.794. In future studies, the proposed system is planned to be implemented on the Ethereum mainnet instead of a test network, with a comprehensive evaluation of on-chain operational costs. However, instead of Zk-Snark proofs, which have a centralized structure, the use of Zk-Stark proofs, which are transparent and do not violate the principle of decentralization, is planned.