Abstract
Deep learning models are widely used in multivariate time series forecasting, yet, they have high computational costs. One way to reduce this cost is by reducing data dimensionality, which involves removing unimportant or low importance information with the proper method. This work presents a study on an explainability feature selection framework composed of four methods (IMV-LSTM Tensor, LIME-LSTM, Average SHAP-LSTM, and Instance SHAP-LSTM) aimed at using the LSTM black-box model complexity to its favour, with the end goal of improving the error metrics and reducing the computational cost on a forecast task. To test the framework, three datasets with a total of 101 multivariate time series were used, with the explainability methods outperforming the baseline methods in most of the data, be it in error metrics or computation time for the LSTM model training.

Abstract
Outlier detection is a widely used technique for identifying anomalous or exceptional events across various contexts. It has proven to be valuable in applications like fault detection, fraud detection, and real-time monitoring systems. Detecting outliers in real time is crucial in several industries, such as financial fraud detection and quality control in manufacturing processes. In the context of big data, the amount of data generated is enormous, and traditional batch mode methods are not practical since the entire dataset is not available. The limited computational resources further compound this issue. Boxplot is a widely used batch mode algorithm for outlier detection that involves several derivations. However, the lack of an incremental closed form for statistical calculations during boxplot construction poses considerable challenges for its application within the realm of big data. We propose an incremental/online version of the boxplot algorithm to address these challenges. Our proposed algorithm is based on an approximation approach that involves numerical integration of the histogram and calculation of the cumulative distribution function. This approach is independent of the dataset's distribution, making it effective for all types of distributions, whether skewed or not. To assess the efficacy of the proposed algorithm, we conducted tests using simulated datasets featuring varying degrees of skewness. Additionally, we applied the algorithm to a real-world dataset concerning software fault detection, which posed a considerable challenge. The experimental results underscored the robust performance of our proposed algorithm, highlighting its efficacy comparable to batch mode methods that access the entire dataset. Our online boxplot method, leveraging dataset distribution to define whiskers, consistently achieved exceptional outlier detection results. Notably, our algorithm demonstrated computational efficiency, maintaining constant memory usage with minimal hyperparameter tuning.

Abstract
With technological advancements, much data is being captured by sensors, smartphones, wearable devices, and so forth. These vast datasets are stored in data centres and utilized to forge data-driven models for the condition monitoring of infrastructures and systems through future data mining tasks. However, these datasets often surpass the processing capabilities of traditional information systems and methodologies due to their significant size. Additionally, not all samples within these datasets contribute valuable information during the model training phase, leading to inefficiencies. The processing and training of Machine Learning algorithms become time-consuming, and storing all the data demands excessive space, contributing to the Big Data challenge. In this paper, we propose two novel techniques to reduce large time-series datasets into more compact versions without undermining the predictive performance of the resulting models. These methods also aim to decrease the time required for training the models and the storage space needed for the condensed datasets. We evaluated our techniques on five public datasets, employing three Machine Learning algorithms: Holt-Winters, SARIMA, and LSTM. The outcomes indicate that for most of the datasets examined, our techniques maintain, and in several instances enhance, the forecasting accuracy of the models. Moreover, we significantly reduced the time required to train the Machine Learning algorithms employed.

Abstract
Traditional recurrent neural networks (RNNs) are essential for processing time-series data. However, they function as read-only models, lacking the ability to directly modify the data they learn from. In this study, we introduce the corrector long short-term memory (cLSTM), a Read & Write LSTM architecture that not only learns from the data but also dynamically adjusts it when necessary. The cLSTM model leverages two key components: (a) predicting LSTM’s cell states using Seasonal Autoregressive Integrated Moving Average (SARIMA) and (b) refining the training data based on discrepancies between actual and forecasted cell states. Our empirical validation demonstrates that cLSTM surpasses read-only LSTM models in forecasting accuracy across the Numenta Anomaly Benchmark (NAB) and M4 Competition datasets. Additionally, cLSTM exhibits superior performance in anomaly detection compared to hierarchical temporal memory (HTM) models. © The Author(s) 2024.

Abstract
In manufacturing chatter is an unwanted phenomenon that can lead to product quality reduction and tool wear. Real time chatter detection is key to preventing these issues and improving overall machining efficiency. In this paper we propose an unsupervised chatter detection method using autoencoders (AE) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering algorithm that uses internal signals of Computer Numerical Control (CNC) machines. The proposed method starts by using an AE to extract features from raw internal signals collected from CNC machines. This step reduces the dimensionality of the data and captures the underlying patterns of chatter. Then the extracted features are fed into DBSCAN clustering algorithm which is a density based algorithm that groups similar data points and identifies outliers. We tested the proposed method with real world data collected from various CNC machines. The results show that our unsupervised chatter detection method has high accuracy, precision and recall, can detect chatter and distinguish it from normal machining. Also the method is robust to noise and can adapt to dynamic machining conditions. In summary our work presents an unsupervised chatter detection method using AE and DBSCAN clustering that uses internal signals of CNC machines. This method is a reliable and efficient solution for real time chatter detection so manufacturers can improve product quality, optimize machining process and reduce tool wear during machining.

Abstract
Spatio-temporal prediction tasks play a crucial role in facilitating informed decision-making through anticipatory insights. By accurately predicting future outcomes, the ability to strategize, preemptively address risks, and minimize their potential impact is enhanced. The precision in forecasting spatial and temporal patterns holds significant potential for optimizing resource allocation, land utilization, and infrastructure development. While existing review and survey papers predominantly focus on specific forecasting domains such as intelligent transportation, urban planning, pandemics, disease prediction, climate and weather forecasting, environmental data prediction, and agricultural yield projection, limited attention has been devoted to comprehensive surveys encompassing multiple objects concurrently. This paper addresses this gap by comprehensively analyzing techniques employed in traffic, pandemics, disease forecasting, climate and weather prediction, agricultural yield estimation, and environmental data prediction. Furthermore, it elucidates challenges inherent in spatio-temporal forecasting and outlines potential avenues for future research exploration.