Noise-based self-supervised anomaly detection in washing machines using a deep neural network with operational information

Abstract

To ensure the reliable use and maintenance of a washing machine, condition monitoring and detection of anomalous operations at an early stage are necessary. In this study, we propose a deep neural network architecture for the detection of anomalies in a washing machine based on the noise spectra generated during its operation. Although several self-supervised learning techniques have been developed for the efficient training of an anomaly detection model using only normal data, high diversity in operational noise depending on operating conditions such as laundry weight and imbalance makes anomaly detection in a washing machine difficult. To build a deep neural network model that understands the context of washing conditions and actions, we develop architectures that utilize two types of operational information provided by a washing machine: the spin speed of the drum and laundry weight information. The main self-supervision task of the proposed architecture is to predict the future noise spectrum from the past spectra. However, to utilize the operational information, we investigate two different architectures: one that uses the operational information as a conditioning input to the main task and another that uses it as a secondary objective for a multitask model. Through a cross-validation test with various cloth types and weights, we demonstrate that the proposed architectures can be generalized to various washing machine data and can robustly detect anomalies, even for cloth types unseen during the training stage.