Bruno Miguel Veloso

Abstract
Early identification of failures is a critical task in predictive maintenance, preventing potential problems before they manifest and resulting in substantial time and cost savings for industries. We propose an approach that predicts failures in the near future. First, a deep learning model combining long short-term memory and convolutional neural network architectures predicts signals for a future time horizon using real-time data. In the second step, an autoencoder based on convolutional neural networks detects anomalies in these predicted signals. Finally, a verification step ensures that a fault is considered reliable only if it is corroborated by anomalies in multiple signals simultaneously. We validate our approach using publicly available Air Production Unit (APU) data from Porto metro trains. Two significant conclusions emerge from our study. Firstly, experimental results confirm the effectiveness of our approach, demonstrating a high fault detection rate and a reduced number of false positives. Secondly, the adaptability of this proposal allows for the customization of configuration of different time horizons and relationship between the signals to meet specific detection requirements.

Abstract
Nearest neighbor search (NNS) is one of the main concerns in data stream applications since similarity queries can be used in multiple scenarios. Online NNS is usually performed on a sliding window by lazily scanning every element currently stored in the window. This paper proposes Sliding Window-based Incremental Nearest Neighbors (SWINN), a graph-based online search index algorithm for speeding up NNS in potentially never-ending and dynamic data stream tasks. Our proposal broadens the application of online NNS-based solutions, as even moderately large data buffers become impractical to handle when a naive NNS strategy is selected. SWINN enables efficient handling of large data buffers by using an incremental strategy to build and update a search graph supporting any distance metric. Vertices can be added and removed from the search graph. To keep the graph reliable for search queries, lightweight graph maintenance routines are run. According to experimental results, SWINN is significantly faster than performing a naive complete scan of the data buffer while keeping competitive search recall values. We also apply SWINN to online classification and regression tasks and show that our proposal is effective against popular online machine learning algorithms.

Abstract
Hyper-parameter tuning of machine learning models has become a crucial task in achieving optimal results in terms of performance. Several researchers have explored the optimisation task during the last decades to reach a state-of-the-art method. However, most of them focus on batch or offline learning, where data distributions do not change arbitrarily over time. On the other hand, dealing with data streams and online learning is a challenging problem. In fact, the higher the technology goes, the greater the importance of sophisticated techniques to process these data streams. Thus, improving hyper-parameter self-tuning during online learning of these machine learning models is crucial. To this end, in this paper, we present MESSPT, an evolutionary algorithm for self-hyper-parameter tuning for data streams. We apply Differential Evolution to dynamically-sized samples, requiring a single pass-over of data to train and evaluate models and choose the best configurations. We take care of the number of configurations to be evaluated, which necessarily has to be reduced, thus making this evolutionary approach a micro-evolutionary one. Furthermore, we control how our evolutionary algorithm deals with concept drift. Experiments on different learning tasks and over well-known datasets show that our proposed MESSPT outperforms the state-of-the-art on hyper-parameter tuning for data streams.

Abstract
Predictive maintenance methods play a crucial role in the early detection of failures and errors in machinery, preventing them from reaching critical stages. This paper presents a comprehensive study on a real-world dataset called MetroPT3, with data from a Metro do Porto train's air production unit (APU) system. The dataset comprises data collected from various analogue and digital sensors installed on the APU system, enabling the analysis of behavioural changes and deviations from normal patterns. We propose a data-driven predictive maintenance framework based on a Long Short-Term Memory Autoencoder (LSTM-AE) network. The LSTM-AE efficiently identifies abnormal data instances, leading to a reduction in false alarm rates. We also implement a Sparse Autoencoder (SAE) approach for comparative analysis. The experimental results demonstrate that the LSTM-AE outperforms the SAE regarding F1 Score, Recall, and Precision. Furthermore, to gain insights into the reasons for anomaly detection, we apply the Shap method to determine the importance of features in the predictive maintenance model. This approach enhances the interpretability of the model to support the decision-making process better.

Abstract