Semisupervised Learning for Noise Suppression Using Deep Reinforcement Learning of Contrastive Features

In this letter, we present DeDRLSSL: a generic semisupervised noise suppression framework. The proposed model is based on a reinforcement learning system for learning contrastive features to refine the features utilized in consistency matching for semisupervised learning (SSL). The proposed method outperforms the state-of-the-art supervised models in terms of error compensation for Inertial Measurement Unit data from various evaluation metrics and improves the baselines for yaw estimation on average by 38% and 28% across the benchmarks for 30% and 50% of labeled data, respectively. Our approach can be adapted to any SSL approach to compensate for the problem of label scarcity.

I. Introduction

The presence of noise, regardless of the signal type, is ubiquitous in signal processing, therefore leveraging the error compensation methods before an inference is inevitable. Noise suppression methods are divided into conventional approaches and data-driven models. The conventional methods for signal denoising, such as filtering and wavelet transforms, are widely used in industry and academia. Methods such as the Wiener filter [1] work well in removing low-frequency noises like biases, however, they failed in high-frequency noise removal. In the same way, error reduction methods based on a linear model for error compensation like Kalman filter [2] are obviously not applicable for large and complex errors. In addition, the methods such as wavelet processing rely on delicate analysis to choose the basis for signal decomposition, which requires prior knowledge of the nature of signals. Data-driven models recently were leveraged to perform the noise removal task. The deep neural modules have recently seen success to some extent in noise removal for datasets with clean ground truth labels [3], [4]. The existing methods typically employ encoder–decoder architecture for signal denoising and the reconstruction error is used for training and performance evaluation of the deep models. In [5], an unsupervised method based on autoencoders is proposed to remove noise artifacts from IMU signals. The method results in a competent F1-score for human-activity recognition. In [4], an adversarial model is introduced to distinguish the distribution of clean and noisy data toward the target of aligning the distribution of the latent representation of clean and noisy signals. In [6], dilated convolutional layers are applied for denoising gyroscope data, and it was shown that if the model is trained with a proper loss function, resulting accuracy competes with the accuracy of highly accurate methods like visual-inertial systems. Similarly, in [7] a deep model is constructed by using the temporal convolutional network for noise compensation of IMU gyroscope data.