Lazy Estimation in Networked Systems

由电池驱动的、广泛分布的设备提供的传感器数据量正在稳步增长。由于传感器数据通常被输入信息处理单元，因此值得考虑的是如何利用信息处理本身来减少通信和能源需求。为此，本项目侧重于信息处理技术，这些技术可以将传输机制传递的隐式信息整合在一起。尽管传感器节点决定不发送其数据，但接收方仍然可以利用数据的缺失来更新其状态估计。例如，可以将传感器读数与阈值进行比较，以决定是否进行传输。接收方可以将该决策规则转换为有关数据的信息，尽管没有发生传输。发送方和接收方可以协商这些决策规则，以便在发送端使通信成本最小化，在接收端使可检索信息最大化。由于基于阈值的策略对于观察时变系统的限制太大，因此将研究基于模型和数据驱动的策略。本项目主要研究触发传输的随机决策规则。与确定性触发器相比，随机机制可以保持隐式信息的高斯性，简化了接收端的估计器设计。例如，当没有传输事件被触发时，卡尔曼滤波器只需要很小的调整来包含隐含信息。该项目的目标是推动随机触发原理向前发展，建立一个全面的延迟估计框架。首先，研究了惰性状态估计的一般性质和智能触发决策的设计，以提高惰性状态估计的有效性和鲁棒性。这些包括基于模型和数据驱动的触发机制，非周期性和异步传输和处理时间，以及对不可靠通信链路的研究。研究结果为多传感器系统和高维状态表示的大规模延迟估计提供了基础。例如，多个系统协同监控一个动态系统，并融合交换的传感器数据和估计。这种分布式数据融合问题导致需要自适应触发机制的依赖触发决策。特别是，该项目考虑在目标跟踪中的应用，以评估派生的概念。惰性估计在处理神经形态传感器数据和实现国家保密方法方面显示出巨大的潜力。这两个方向都是延迟估计的应用前景。

The amount of sensor data provided by battery-driven, widely distributed devices is steadily increasing. Since sensor data are typically fed into information processing units, it is worth considering how information processing itself can be exploited to reduce communication and energy demands. For this purpose, this project focuses on information-processing techniques that can incorporate implicit information conveyed by the transmission mechanism. Although a sensor node decides not to send its data, the receiver can still leverage the absence of data to update its state estimates. For instance, sensor readings can be compared against a threshold to decide for a transmission. The receiver can translate this decision rule into information about the data although no transmission took place. Sender and receiver can negotiate such decision rules in order to minimize communication costs, on the transmitting end, and to maximize the retrievable information, on the receiving end. Since threshold-based strategies are far too restrictive for time-varying systems being observed, model-based and data-driven policies will be investigated. This project primarily investigates stochastic decision rules to trigger transmissions. In contrast to deterministic triggers, stochastic mechanisms can preserve the Gaussianity of the implicit information simplifying the estimator design at the receiver. For instance, a Kalman filter only requires minor adaptions to incorporate implicit information when no transmission event is triggered. The goal of this project is to push the principles of stochastic triggering forward to establish a comprehensive framework of lazy estimation. First, the investigations are concerned with general properties and the design of intelligent trigger decisions to improve the effectiveness and robustness of lazy state estimation. These include model-based and data-driven trigger mechanisms, aperiodic and asynchronous transmission and processing times, as well as the study of unreliable communication links. The results provide the foundations for large-scale lazy estimation with respect to both multisensor systems and high-dimensional state representations. For instance, multiple systems collaboratively monitor a dynamic system and fuse exchanged sensor data and estimates. Such distributed data fusion problems lead to dependent trigger decisions that require self-adapting trigger mechanisms. In particular, the project considers applications in object tracking to evaluate the derived concepts. Lazy estimation shows great potential in the processing of neuromorphic sensor data and in implementing state secrecy methods. Both directions are studied as prospective fields of application of lazy estimation.