Machine learning, with its advances in deep learning has shown great potential in analyzing time series. In many scenarios, however, additional information that can potentially improve the predictions is available. This is crucial for data that arise from e. g., sensor networks that contain information about sensor locations. Then, such spatial information can be exploited by modeling it via graph structures, along with the sequential (time series) information. Recent advances in adapting deep learning to graphs have shown potential in various tasks. However, these methods have not been adapted for time series tasks to a great extent. Most attempts have essentially consolidated around time series forecasting with small sequence lengths. Generally, these architectures are not well suited for regression or classification tasks where the value to be predicted is not strictly depending on the most recent values, but rather on the whole length of the time series. We propose TISER-GCN, a novel graph neural network architecture for processing, in particular, these long time series in a multivariate regression task. Our proposed model is tested on two seismic datasets containing earthquake waveforms, where the goal is to predict maximum intensity measurements of ground shaking at each seismic station. Our findings demonstrate promising results of our approach—with an average MSE reduction of 16.3%—compared to the best performing baselines. In addition, our approach matches the baseline scores by needing only half the input size. The results are discussed in depth with an additional ablation study.

Graph neural networks (GNNs) haven proven to be an indispensable approach in modeling complex data, in particular spatial temporal data, e.g., relating to sensor data given as time series with according spatial information. Although GNNs provide powerful modeling capabilities on such kind of data, they require adequate input data in terms of both signal and the underlying graph structures. However, typically the according graphs are not automatically available or even predefined, such that typically an ad hoc graph representation needs to be constructed. However, often the construction of the underlying graph structure is given insufficient attention. Therefore, this paper performs an in-depth analysis of several methods for constructing graphs from a set of sensors attributed with spatial information, i.e., geographical coordinates, or using their respective attached signal data. We apply a diverse set of standard methods for estimating groups and similarities between graph nodes as location-based as well as signal-driven approaches on multiple benchmark datasets for evaluation and assessment. Here, for both areas, we specifically include distance-based, clustering-based, as well as correlation-based approaches for estimating the relationships between nodes for subsequent graph construction. In addition, we consider two different GNN approaches, i.e., regression and forecasting in order to enable a broader experimental assessment. Typically, no predefined graph is given, such that (ad hoc) graph creation is necessary. Here, our results indicate the criticality of factoring in the crucial step of graph construction into GNN-based research on spatial temporal data. Overall, in our experimentation no single approach for graph construction emerged as a clear winner. However, in our analysis we are able to provide specific indications based on the obtained results, for a specific class of methods. Collectively, the findings highlight the need for researchers to carefully consider graph construction when employing GNNs in the analysis of spatial temporal data.

With the developments in improved computation power and the vast amount of (automatic) data collection, industry has become more data-driven. These data-driven approaches for monitoring processes and machinery require different modeling methods focusing on automated learning and deployment. In this context, deep learning provides possibilities for industrial diagnostics to achieve improved performance and efficiency. These deep learning applications can be used to automatically extract features during training, eliminating time-consuming feature engineering and prior understanding of sophisticated (signal) processing techniques. This paper extends on previous work, introducing one-dimensional (1D) CNN architectures that utilize an adaptive wide-kernel layer to improve classification of multivariate signals, e.g., time series classification in fault detection and condition monitoring context. We used multiple prominent benchmark datasets for rolling bearing fault detection to determine the performance of the proposed wide-kernel CNN architectures in different settings. For example, distinctive experimental conditions were tested with deviating amounts of training data. We shed light on the performance of these models compared to traditional machine learning applications and explain different approaches to handle multivariate signals with deep learning. Our proposed models show promising results for classifying different fault conditions of rolling bearing elements and their respective machine condition, while using a fairly straightforward 1D CNN architecture with minimal data preprocessing. Thus, using a 1D CNN with an adaptive wide-kernel layer seems well-suited for fault detection and condition monitoring. In addition, this paper clearly indicates the high potential performance of deep learning compared to traditional machine learning, particularly in complex multivariate and multi-class classification tasks.

Industrial fault detection has become more data-driven due to advancements in automated data analysis using deep learning. Such methods make it possible to extract useful features, e. g., from time series data retrieved from sensors, which is typically of complex nature. This allows for effective fault detection and prognostics that boost the efficiency and productivity of industrial equipment. This work explores the influence of a variety of architectural hyperparameters on the performance of one-dimensional convolutional neural networks (CNN). Using a multi-method approach, this paper focuses specifically on wide-kernel CNN models for industrial fault detection, that have proven to perform well for tasks such as classifying vibration signals retrieved from sensors. By varying hyperparameters such as the kernel size, stride and number of filters, an extensive hyperparameter space search was conducted; to identify optimal settings, we collected a total of 12,960 different combinations on three datasets into a model hyperparameter dataset, with their respective performance on the underlying fault detection task. Afterwards, this dataset was explored with follow-up analysis including statistical, feature, pattern and hyperparameter impact analysis. We find that although performance varies substantially depending on hyperparameter choices, there is no single simple strategy to optimise performance across the three datasets. However, an optimal setting in terms of performance can be found in the number of filters used in the later layers of the architecture for all datasets. Furthermore, hyperparameter importance differs across and within the datasets, and we found nonlinear relationships between hyperparameter settings and performance. Our analysis highlights key considerations when applying a wide-kernel CNN architecture to new data within the field of industrial fault detection. This supports practitioners who wish to apply and train state-of-the-art convolutional learning methods to apply to similar fault detection settings, e. g., vibration data arising from new combinations of sensors and/or machinery in the context of bearing faults.

Machine learning, with its advances in deep learning has shown great potential in analyzing time series. In many scenarios, however, additional information that can potentially improve the predictions is available. This is crucial for data that arise from e.g., sensor networks that contain information about sensor locations. Then, such spatial information can be exploited by modeling it via graph structures, along with the sequential (time series) information. Recent advances in adapting deep learning to graphs have shown potential in various tasks. However, these methods have not been adapted for time series tasks to a great extent. Most attempts have essentially consolidated around time series forecasting with small sequence lengths. Generally, these architectures are not well suited for regression or classification tasks where the value to be predicted is not strictly depending on the most recent values, but rather on the whole length of the time series. We propose TISER-GCN, a novel graph neural network architecture for processing, in particular, these long time series in a multivariate regression task. Our proposed model is tested on two seismic datasets containing earthquake waveforms, where the goal is to predict maximum intensity measurements of ground shaking at each seismic station. Our findings demonstrate promising results of our approach -- with an average MSE reduction of 16.3% - compared to the best performing baselines. In addition, our approach matches the baseline scores by needing only half the input size. The results are discussed in depth with an additional ablation study.