Autoencoder-driven fault detection in drilling machines: a study of hybrid deep learning models

Detecting faults promptly and accurately in manufacturing is crucial for maintaining efficiency, productivity, and product quality. This research explores the potential of using autoencoders, including models such as Long Short-Term Memory (LSTM) networks, Bidirectional Long Short-Term Memory (BiLSTM), WaveNet, and hybrid models that combine 1D Convolutional Neural Networks (1D CNN) with BiLSTM layers, for real-time fault detection in drilling operations. The autoencoders were trained on data from normal operational conditions to learn standard patterns and then tested on datasets containing fault instances. Results show that autoencoders with BiLSTM and LSTM layers excelled at capturing temporal dependencies. Hybrid models, such as WaveNet-BiLSTM and CNN-BiLSTM, which integrate LSTM and convolutional layers, significantly enhanced fault detection performance, achieving perfect accuracy, precision, recall, and F1 scores with just 10% of normal training samples. The models’ effectiveness was also evaluated using shorter time windows, resulting in an accuracy of 98%. This advancement facilitates the integration of such models into intelligent manufacturing systems for proactive maintenance and operational optimization, potentially reducing downtime and maintenance costs while maintaining high product quality.