Tram Air Conditioning Fault Prediction Using Machine Learning

In fault detection and diagnosis (FDD) research, most focus has been on accurate supervised learning techniques, but they heavily depend on well-organized datasets, which can be tough due to a lack of real-world faulty samples. To tackle this, our study introduces an unsupervised learning approach that uses new methods to create synthetic faulty samples and optimized Principal Component Analysis (PCA) for efficient computations in large datasets. The main aim is to blend unsupervised learning using Isolation forest, specifically clustering techniques, into fault detection systems to handle challenges from unclassifiable or imbalanced data, a problem for traditional supervised methods. The methodology combines innovative PCA algorithms with Rapid PCA, K means, LSTM, and clustering methods to build a strong fault detection framework. The study assesses the impacts of data imbalance, examines the benefits of unsupervised clustering, and evaluates algorithm adaptability in finding faults without labeled data. Challenges include accurately clustering with limited labeled data and identifying subtle fault behaviours in normal operational data. Split into sections like introduction, literature review, methodology, experiments, and conclusion, this research aims to offer a detailed look at using unsupervised learning (i.e., Isolation Forest) to improve fault detection, especially in situations with complex data challenges.