Self, Semi and Fully Supervised Training for Autoencoders using Ternary Classification
Keywords: self supervised learning, semi supervised learning, fully supervised learning, autoencoder, loss function, contrastive learning, anomaly detection
TL;DR: A new label is introduced and loss functions are modified to enable self, semi and fully supervised training of autoencoders using ternary classification.
Abstract: Autoencoders are usually trained in a self-supervised fashion. In the context of anomaly detection, research shows that they can also be trained in a fully supervised one, using binary class labels, namely HEALTHY and FAULTY. However, when working with real world data, such an approach might not be suitable. It is hard to binary classify data coming from equipment that has been in operation for a long time, is affected by wear and tear. In additional, its real current health status is unknown. Moreover, historical data is not usually labeled, and only maintenance interventions are recorded. To alleviate this problem, a third label is introduced, UNKNOWN, which enables the autoencoder to learn the structure of healthy and faulty data from the correspondingly labelled data points. This structure is used in reconstructing the UNKNOWN inputs. This can increase the performance of autoencoders in a wide range of anomaly detection cases, especially when the timeseries data used to train the autoencoder comes from machines that have been in operation for a long time. This is especially relevant in the case of industrial machinery. Different label-aware loss functions which can enable the training of an autoencoder, using the three aforementioned labels, in any combination of self, semi and fully supervised training are investigated in this work. The loss functions presented in this paper enable an autoencoder to achieve particularly good anomaly detection performance on a clutch-slip detection dataset acquired from a test bench which simulates the drivetrain of an electric Range Rover Evoque. The dataset is presented in the appendix.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5584
Loading