This paper presents a novel approach for anomaly detection in industrial processes. The system solely relies on unlabeled data and employs a 1D-convolutional neural network-based deep autoencoder architecture. As a core novelty, we split the autoencoder latent space in discriminative and reconstructive latent features and introduce an auxiliary loss based on k-means clustering for the discriminatory latent variables. We employ a Top-K clustering objective for separating the latent space, selecting the most discriminative features from the latent space. We use the approach to the benchmark Tennessee Eastman data set to prove its applicability. We provide different ablation studies and analyze the method concerning various downstream tasks, including anomaly detection, binary and multi-class classification. The obtained results show the potential of the approach to improve downstream tasks compared to standard autoencoder architectures.

This paper proposes a novel deep learning architecture for estimating the remaining useful lifetime (RUL) of industrial components, which solely relies on the recently developed transformer architectures. The RUL estimation resorts to analysing degradation patterns within multivariate time series signals. Hence, we propose a novel shared temporal attention block that allows detecting RUL patterns with the progress of time. Furthermore, we develop a split-feature attention block that enables attending to features from different sensor channels. The proposed shared temporal attention layer in the encoder fulfils the goal of attending to temporal degradation patterns in the individual sensor signals before creating a shared correlation across the feature range. We develop two transformer architectures that are specifically designed to operate with multivariate time series data based on these novel attention blocks. We apply the architectures to the well known C-MAPSS benchmark dataset and provide various hyperparameter studies to analyse their impact on the performance. In addition, we provide a thorough comparison with recently presented state-of-the-art approaches and show that the proposed transformer architectures outperform the existing methods by a considerable margin.

This article presents a novel neural network training approach for faster convergence and better generalization abilities in deep reinforcement learning (RL). Particularly, we focus on the enhancement of training and evaluation performance in RL algorithms by systematically reducing gradient's variance and, thereby, providing a more targeted learning process. The proposed method, which we term gradient monitoring (GM), is a method to steer the learning in the weight parameters of a neural network based on the dynamic development and feedback from the training process itself. We propose different variants of the GM method that we prove to increase the underlying performance of the model. One of the proposed variants, momentum with GM (M-WGM), allows for a continuous adjustment of the quantum of backpropagated gradients in the network based on certain learning parameters. We further enhance the method with the adaptive M-WGM (AM-WGM) method, which allows for automatic adjustment between focused learning of certain weights versus more dispersed learning depending on the feedback from the rewards collected. As a by-product, it also allows for automatic derivation of the required deep network sizes during training as the method automatically freezes trained weights. The method is applied to two discrete (real-world multirobot coordination problems and Atari games) and one continuous control task (MuJoCo) using advantage actor-critic (A2C) and proximal policy optimization (PPO), respectively. The results obtained particularly underline the applicability and performance improvements of the methods in terms of generalization capability.