Kalmannet-Aided Target Tracking with 1-Bit Decisions in Wireless Sensor Networks

Abstract

In wireless sensor networks (WSNs), 1-bit quantization provides an energy-efficient and bandwidth-conserving solution, albeit at the expense of considerable information loss in state estimation tasks. This paper introduces a modified KalmanNet architecture tailored for scenarios where the process noise statistics are known, but the measurement noise characteristics are unknown due to binary quantization. In contrast to the original KalmanNet, which processes temporally differenced, continuous-valued observations, the proposed model operates directly on 1-bit sensor decisions, interpreting them as independent Bernoulli samples at each time step. To ensure scalability in dense sensor deployments, we incorporate architectural compression inspired by SqueezeNet, significantly reducing the number of trainable parameters without sacrificing model expressiveness. The proposed approach is benchmarked against the Extended Kalman Filter (EKF) using raw sensor measurements, and the Particle Filter (PF) with 1-bit decisions under both adaptive and non-adaptive thresholding schemes driven by approximate mutual information and Fisher information criteria, namely Mutual Information Upper Bound - Adaptive (MIUB-A), Mutual Information Upper Bound - Shared (MIUB-S) and the Fisher Information Matrix Adaptive (FIM-A) designs. Simulation results show that the non-adaptive KalmanNet not only outperforms PF without adaptive thresholding, but also closely approaches the estimation accuracy of a representative adaptive scheme (MIUB-S). These findings underscore the potential of data-driven filtering in quantized WSNs, paving the way for robust, scalable, and feedback-free tracking systems. © 2025 Elsevier B.V., All rights reserved.