APPLICATION OF ARTIFICIAL INTELLIGENCE FOR ANOMALY DETECTION IN ENERGY GENERATION SYSTEMS

In recent years, two major subjects have emerged in the scientific community: artificial intelligence and sustainability. These topics converge in developing data-driven techniques for detecting malfunctions or inefficiencies in energy generation systems. Extensive research exists on applications for large-scale generators, however, smaller generators, particularly microcogenerators, have received minor attention. This study focuses on the YANMAR’s best-selling micro-cogenerator model in Europe. However, findings from this investigation can be extended to a diverse range of sizes and models. Dealing with real data, this thesis explores various facets related to the challenge of anomaly detection in industrial components, thus giving rise to multiple contributions. In particular, it addresses anomaly detection in energy systems by tailoring a general deep learning technique, the autoencoder, to handle time series data. The study presents a general methodology to determine the optimal dataset size for training an autoencoder. An application of the algorithm to predict cogenerator faults is proposed after estimating the dataset size that offers the best compromise for the YANMAR micro-cogenerator under analysis. False positives are reduced through a frequency-based technique. Additionally, a failure root cause analysis is conducted to identify features associated with abnormalities. The proposed approach is validated by predicting various fault types several weeks prior, potentially preventing breakdowns and inefficiencies. Furthermore, the study applies a technique to quantify the algorithm’s detection confidence. This technique enables the development of condition-based maintenance strategies tailored to the fault’s uncertainty level. Finally, a method is proposed to retrain the model, ensuring its performance despite variations due to the system’s aging and seasonal operations which may lead to erroneous abnormality detections. By addressing these challenges, this research significantly advances autoencoder-based methodologies and anomaly detection in microcogenerators. Rooted in real-world industrial data, the developed procedures offer practical solutions to enhance the reliability and efficiency of energy systems. This work attempts to fill the gap between artificial intelligence techniques and their application in industrial contexts, providing valuable insights for future research and practical implementations.