Enhancing Siamese Neural Networks Through Expert Knowledge for Predictive Maintenance

Abstract

The data provided by cyber-physical production systems (CPPSs) to monitor their condition via data-driven predictive maintenance is often high dimensional and only a few fault and failure (FaF) examples are available. These FaFs can usually be detected in a (small) localized subset of data streams, whereas the use of all data streams induces noise that could negatively affect the training and prediction performance. In addition, a CPPS often consists of multiple similar units that generate comparable data streams and show similar failure modes. However, existing approaches for learning a similarity measure generally do not consider these two aspects. For this reason, we propose two approaches for integrating expert knowledge about class or failure mode dependent attributes into siamese neural networks (SNN). Additionally, we present an attribute-wise encoding of time series based on 2D convolutions. This enables that learned knowledge in the form of filters is shared between similar data streams, which would not be possible with conventional 1D convolutions due to their spatial focus. We evaluate our approaches against state-of-the-art time series similarity measures such as dynamic time warping, NeuralWarp, as well as a feature-based representation approach. Our results show that the integration of expert knowledge is advantageous and combined with the novel SNN architecture it is possible to achieve the best performance compared to the other investigated methods.