Deep learning techniques have recently shown promise in the field of anomaly detection, providing a flexible and effective method of modelling systems in comparison to traditional statistical modelling and signal processing-based methods. However, there are a few issues that Neural Networks (NN)s face, such as generalisation ability, requiring large volumes of labelled data to train effectively, and understanding spatial context in data. This paper introduces a novel NN architecture to tackle these problems, which utilises a Long-Short-Term-Memory (LSTM) encoder and Capsule decoder in a multi-channel input Autoencoder architecture for use on multivariate time series data. Experimental results show that using Capsule decoders increases the resilience of the model to overfitting and improves training efficiency, which is shown by the improvement of Mean Squared Error (MSE) on unseen data from an average of 10.61 to 2.08 for single channel architectures, and 10.08 to 2.05 for multi-channel architectures. Additionally, results also show that the proposed model can learn multivariate data more consistently, and was not affected by outliers in the training data. The proposed architecture was also tested on an open-source benchmark, where it achieved state-of-the-art performance in outlier detection, and performs best overall with a total accuracy of 0.494 over the metrics tested.

Prognostics and Health Monitoring (PHM) of machinery is a research area with great relevance to industrial applications as it can serve as a foundation for safer, more cost-efficient operation and maintenance. The prediction of Remaining Useful Life (RUL) plays an important part in this field and has seen significant advances from the introduction of machine learning methods. However, these methods typically require model training with a large number of run-to-failure sequences, which are often not feasible to obtain due to the required time and cost investments. The present study addresses this issue by introducing a novel methodology, which first quantifies the deviation from the machine’s health and fault state and then calculates a machine Health Index (HI) prior to the prediction of RUL. In addition, the start of a degradation state is determined. Alternative implementations of the proposed methodology are compared utilising several methods, including Support Vector Regression (SVR), Long Short-Term Memory (LSTM) Neural Network (NN), Mahalanobis Distance (MD), and LSTM Autoencoder (AE) NN. The methodology is applied to the open turbofan degradation (C-MAPSS) and bearing vibration (FEMTO-ST PROGNOSTIA) datasets. When a reduced subset of training sequences is used, the prediction results demonstrate that the proposed methodology largely outperforms the baseline method without HI generation. For example, when comparing prediction errors of the C-MAPSS dataset at a reduction of the available number of training sequences to 5%, the proposed method shows an average prediction improvement by 6.5% - 19.2% relative to the baseline method. The presented approach is therefore suitable to improve model generalisation for cases with a limited number of training sequences. When the full training set is utilised, the most resource-saving variant of the proposed approach achieves an average training duration of 8.9% compared to the baseline method. Hence, an additional contribution of the presented data-efficient approach is the reduction of required computing resources, which has implications on training time, energy consumption, and environmental impact.

This paper introduces a procedure to compare the functional behaviour of individual units of electronic hardware of the same type. The primary use case for this method is to estimate the functional integrity of an unknown device unit based on the behaviour of a known and proven reference unit. This method is based on the so-called virtual sensor network (VSN) approach, where the output quantity of a physical sensor measurement is replicated by a virtual model output. In the present study, this approach is extended to model the functional behaviour of electronic hardware by a neural network (NN) with Long-Short-Term-Memory (LSTM) layers to encapsulate potential time-dependence of the signals. The proposed method is illustrated and validated on measurements from a remote-controlled drone, which is operated with two variants of controller hardware: a reference controller unit and a malfunctioning counterpart. It is demonstrated that the presented approach successfully identifies and describes the unexpected behaviour of the test device. In the presented case study, the model outputs a signal sample prediction in 0.14 ms and achieves a reconstruction accuracy of the validation data with a root mean square error (RMSE) below 0.04 relative to the data range. In addition, three self-protection features (multidimensional boundary-check, Mahalanobis distance, auxiliary autoencoder NN) are introduced to gauge the certainty of the VSN model output.

The generalisation of Neural Networks (NN) to multiple datasets is often overlooked in literature due to NNs typically being optimised for specific data sources. This becomes especially challenging in time-series tasks due to difficulties in fusing temporal data from multiple sources. However, in a commercial environment, generalisation can effectively utilise available data and computational power which is essential to Green AI, the sustainable development of AI models. This paper introduces "Dataset Fusion," a novel dataset composition algorithm for fusing periodic signals from multiple homogeneous datasets whilst retaining unique features for generalised anomaly detection. The proposed approach, tested on a case study of three-phase current data from two different homogeneous Induction Motor (IM) fault datasets on anomaly detection, outperforms conventional training approaches with an Average F1 score of 0.879 and effectively generalises across all datasets. Furthermore, when tested with varying percentages of the training data, results show that using only 6.25% of the training data, translating to a 93.7% reduction in computational power, results in only a 4.04% decrease in performance, demonstrating the advantages of the proposed approach in terms of both performance and computational efficiency. Moreover, the algorithm’s effectiveness under imperfect conditions highlights its potential for use in real-world applications.