Canonical deep learning-based remaining useful life prediction relies on supervised learning methods, which in turn requires large data sets of run-to-failure data to ensure model performance. In a considerable class of cases, run-to-failure data is difficult to collect in practice as it may be expensive and unsafe to operate assets until failure. As such, there is a need to leverage data that are not run-to-failure but may still contain some measurable, and thus learnable, degradation signal. In this paper, we propose utilizing self-supervised learning as a pretraining step to learn representations of data which will enable efficient training on the downstream task of remaining useful life prediction. The self-supervised learning task chosen is time series sequence ordering, a task that involves constructing tuples each consisting of n sequences sampled from the time series and reordered with some probability p. Subsequently, a classifier is trained on the resulting binary classification task; distinguishing between correctly ordered and shuffled tuples. The classifier’s weights are then transferred to the remaining useful life prediction model and fine-tuned using run-to-failure data. To conduct our experiments, we use a data set of simulated run-to-failure turbofan jet engines. We show that the proposed self-supervised learning scheme can retain performance when training on a fraction of the full data set. In addition, we show indications that self-supervised learning as a pretraining step can enhance the performance of the model even when training on the full run-to-failure data set. 

Robust and accurate prognostics models for estimation of remaining useful life (RUL) are becoming an increasingly important aspect of research in reliability and safety in modern electronic components and systems. In this work, a data driven approach to the prognostics problem is presented. In particular, machine learning models are trained to predict the RUL of wire-bonded silicon carbide (SiC) metal-oxide-semiconductor field-effect transistors (MOSFETs) subjected to power cycling until failure. During such power cycling, ON-state voltage and various temperature measurements are continuously collected. As the data set contains full run-to-failure trajectories, the issue of estimating RUL is naturally formulated in terms of supervised learning. Three neural network architectures were trained, evaluated, and compared on the RUL problem: a temporal convolutional neural network (TCN), a long short-term memory neural network (LSTM) and a convolutional gated recurrent neural network (Conv-GRU). While the results show that all networks perform well on held out testing data if the testing samples are of similar aging acceleration as the samples in the training data set, performance on out-of-distribution data is significantly lower. To this end, we discuss potential research directions to improve model performance in such scenarios.

Remaining useful life prediction models are a central aspect of developing modern and capable prognostics and health management systems. Recently, such models are increasingly data-driven and based on various machine learning techniques, in particular deep neural networks. Such models are notoriously “data hungry”, i.e., to get adequate performance of such models, a substantial amount of diverse training data is needed. However, in several domains in which one would like to deploy data-driven remaining useful life models, there is a lack of data or data are distributed among several actors. Often these actors, for various reasons, cannot share data among themselves. In this paper a method for collaborative training of remaining useful life models based on federated learning is presented. In this setting, actors do not need to share locally held secret data, only model updates. Model updates are aggregated by a central server, and subsequently sent back to each of the clients, until convergence. There are numerous strategies for aggregating clients’ model updates and in this paper two strategies will be explored: 1) federated averaging and 2) federated learning with personalization layers. Federated averaging is the common baseline federated learning strategy where the clients’ models are averaged by the central server to update the global model. Federated averaging has been shown to have a limited ability to deal with non-identically and independently distributed data. To mitigate this problem, federated learning with personalization layers, a strategy similar to federated averaging but where each client is allowed to append custom layers to their local model, is explored. The two federated learning strategies will be evaluated on two datasets: 1) run-to-failure trajectories from power cycling of silicon-carbide metal-oxide semiconductor field-effect transistors, and 2) C-MAPSS, a well-known simulated dataset of turbofan jet engines. Two neural network model architectures commonly used in remaining useful life prediction, long shortterm memory with multi-layer perceptron feature extractors, and convolutional gated recurrent unit, will be used for the evaluation. It is shown that similar or better performance is achieved when using federated learning compared to when the model is only trained on local data.