Maximising Detection Performance Using High Performance Processing of Multi-Sensor Data

该项目将开发用于潜艇战争的最先进的信号处理算法。未来对潜艇的传感将涉及机器人潜艇和水面舰艇，以及部署在海底的自主传感器。为了避免泄露传感器位置的可能性和潜艇躲避探测的能力，使用水听器阵列被动监听信号比主动监听更可取。不幸的是，声能在水下的传播是复杂的。被称为先跟踪后检测的信号处理算法已被开发用于检测来自单个传感器的隐身目标。这些算法使用顺序数值贝叶斯推理算法来处理长时间尺度上的原始传感器数据，在这些时间尺度上，目标的轨迹导致传感器伪影，这些伪影不太可能是噪声的结果。这些算法对计算要求很高，但易于并行处理，并且已被证明在检测性能方面具有显著优势。这类融合算法的高级变体仔细考虑物理模型和统计模型的组合：物理模型可以准确地预测数据，但通常计算要求非常高，或者很难进行强有力的校准；统计模型可以捕捉没有物理建模的现象的影响，但根据定义，无法像物理模型那样准确地预测将被感知的内容。该项目的重点将是开发检测前跟踪算法，该算法具有多个传感器的模型，在配置信号处理以最大化检测性能的同时，考虑物理模型和统计模型的组合。

This project will develop state-of-the-art in signal processing algorithms to be used in submarine warfare. Future sensing of submarines will involve robotic submarines and surface-ships as well as autonomous sensors deployed on the seabed. To avoid the potential to give away the position of the sensors and for the ability of a submarine to avoid detection, passive listening for signals using an array of hydrophones is preferable to active sensing. Unfortunately, the underwater propagation of acoustic energy is complicated. Signal processing algorithms called Track-Before-Detect have been developed to detect stealthy targets from single sensors. These algorithms use sequential numerical Bayesian inference algorithms to process the raw sensor data over long timescales over which the trajectories of the targets cause sensor artefacts that are unlikely to be the result of noise. These algorithms are computationally demanding but amenable to parallel processing and have been demonstrated to provide significant advantages in terms of detection performance. Advanced variants of such fusion algorithms carefully consider a combination of physical and statistical models: physical models can accurately predict the data but are typically very computationally demanding or difficult to calibrate robustly; statistical models can capture the effect of phenomena that are not modelled physically, but cannot, by definition, predict what will be sensed as accurately as physical models. The focus of the project will be developing Track-Before-Detect algorithms with models for multiple sensors that consider a combination of physical and statistical models while configuring the signal processing to maximise detection performance.