A Hybrid 1D CNN-GOA Approach for Fault Detection in Feed water Pumps: Case Study of Al-Sabiya Steam Power Plant

Abstract

Feed water pump fault detection is important to ensure the reliability and efficiency of the steam power plant. Fault detection based on threshold monitoring and rule-based systems cannot handle complex operational anomalies. To enhance fault detection ac-curacy whilst retaining computational efficiency, this study presents a hybrid 1D Convolutional Neural Network (CNN) and Grasshopper Optimization Algorithm (GOA) approach. GOA optimizes hyperparameters to improve the model’s predictive performance in tandem with the 1D CNN to extract temporal features from time-series sensor data. In this methodology, data pre-processing, model training on Google Colab using TensorFlow, then deployment on an ESP32 microcontroller using TensorFlow Lite (TFLite). Accuracy, precision, recall, F1-score, confusion matrices, ROC curves, AUC score and inference latency are used to evaluate model’s performance from precision. The results show significant better classification accuracy with 99.5% using the pro-posed hybrid model, compared to traditional machine learning techniques. Addition-ally, 97% accuracy on inference time 9 milliseconds is obtained on ESP32 deployment, making it fit for real time industrial applications. Validation of the efficiency of deep learning and metaheuristic optimization combination in predictive maintenance is shown by the findings. This technique enables online, at the edge fault detection in power plants with reduced downtime and improved operational reliability.