自确认传感器理论及应用研究

【作者】 高羽；

【导师】 张建秋；

【作者基本信息】 复旦大学 ， 电路与系统， 2008， 博士

【摘要】 自确认传感器是一种在输出测量值的同时,给出测量值精度及自身工作状态的下一代新型传感器。它是结合了测量理论,故障检测技术及数字通信技术的智能测量单元。本文针对自确认传感器中其观测噪声方差未知,存在动态测量偏差以及状态方程未知情况下的状态估计等问题开展研究,试图通过采用一些新的数学方法,引入新的建模方法,为自确认传感器寻求新的实现途径和方法,无疑对自确认传感器的发展具有重要的理论意义和实际工程应用价值。众所周知,在传感器观测噪声方差未知情况下,卡尔曼滤波等基于模型的滤波算法是可能失效的。为了解决这一问题,我们通过分析不准确的观测噪声对卡尔曼滤波性能的影响,并根据小波分析可以实时分离信号和噪声的特性,提出了首先利用小波分解的方法实时估计出传感器观测噪声方差,再利用估计出的观测噪声方差进行状态估计的未知观测噪声的卡尔曼滤波算法。最后给出了提出算法在多传感器数据融合中的应用。理论分析和仿真结果均表明,本文提出的算法可以实时准确的估计出传感器观测噪声方差,从而有效地避免了由于观测噪声方差不准确导致的卡尔曼滤波失效。我们知道无论何种精度的实际传感器,在复杂的应用环境中,随着应用时间的延续,传感器或多或少都会普遍存在动态测量偏差。无疑传感器动态测量偏差的存在都将或多或少直接影响其测量结果及相应状态估计的准确度。为了解决这个问题,本文开展了单传感器动态测量偏差的实时估计研究。通过采用多项式预测模型分别对传感器动态测量偏差及系统状态进行建模,并引入一个与偏差相关的可控可测物理量,解决了扩展状态卡尔曼滤波算法中可观测性条件难以满足的问题。理论分析及仿真结果表明,本文方法可以同时准确地估计出系统状态和传感器动态测量偏差。在研究了单传感器动态测量偏差估计算法的基础上,本文进一步研究了多传感器动态测量偏差的实时估计问题。尽管多传感器多目标跟踪的具体应用可以满足可观测性条件,但对未知传感器动态测量偏差的建模仍然是一个难题。本文通过多项式预测模型建立了传感器偏差的伪测量模型,提出了多传感器时变偏差的实时估计算法。理论分析及仿真结果表明,本文算法与文献中现有的算法相比,传感器动态测量偏差估计结果更加准确,估计方差更小。准确的状态方程是基于模型的状态估计算法准确有效的重要前提,但实际应用中,未知状态的先验信息是难以准确获得,这样会导致建立的状态模型存在大的模型不确定性。针对机动目标跟踪的具体应用,本文从牛顿运动定律出发,利用多项式预测模型为机动目标建立了一个自确认的状态方程。分析表明该模型不需要已知目标的具体运动参数,就可以自确认地描述目标的运动状态,这意味着基于该模型的相应跟踪算法则可以有效地跟踪无法预知的目标机动。本文模型和算法与交互多模型算法的比较结果表明,在交互多模型算法理想应用情况下,本文方法与交互多模型算法性能相当,在交互多模型算法非理想应用情况下,即:存在未能预知的目标机动时,本文模型和算法的跟踪性能优于交互多模型算法。从而验证了我们自确认模型分析的有效性。更多还原

【Abstract】 Self-validating sensor (SEVA sensor) is a new type of sensor which can generate the indicator of measurement quality and measurement value status at the same time of outputting the measurements. It is an intelligent measurement system including the use of fault detection techniques, the application of digital technology and the use of uncertainty analysis. By applying new mathematic methods and new modeling methods, the problem of state estimation of SEVA sensor in the condition of unknown measurement noise, time-varying measurement bias and unknown state equation are researched in this paper, aiming at finding new implementation of SEVA. It is of great importance in theory and practical application to the development of SEVA sensor.It is well known that the model-based state estimation algorithm such as Kalman filter is prone to invalid in the condition of unknown measurement noise. To solve the problem, by analyzing the effects of the inaccurate measurement noise covariance on the filter performance, a Kalman filter with unknown measurement noise is proposed. The feature of a wavelet transform separating a noise signal into the signal and noise in real time is combined into Kalman filter. The measurement noise covariance is estimated by wavelet transform and the covariance estimated is then used in state estimation algorithms. The proposed method applying for multisensor data fusion is discussed at last. The analytical results and simulation prove that the method proposed can estimate the measurement covariance in real time and making the Kalman filter under the condition of unknown measurement noise covariance valid.In the complex practical application environment, every sensor will have time-varying measurement bias with the time passing by, no matter its nominal precision. It is obvious that the measurement output of sensor and the state estimated from the measurement will be affected by the time-varying measurement more or less. In order to solve the problem, the method of estimating single sensor time-varying measurement bias in real time is proposed. By introducing a controllable and measurable physical variant which is linear with the measurement bias of a sensor and modeling the system states the sensor observed and the measurement bias of the sensor respectively by polynomial prediction models, both the system states and the measurement bias are observable and can be estimate with Kalman filter. The analytical results and simulation prove that the measurement bias of the sensor can be estimated by a Kalman filter together with its system states.Based on the research of single sensor time-varying measurement bias estimation method, the real-time estimation method of multisensor time-varying measurement bias is researched. Although there is no problem of observability, the modeling of unknown time-varying bias is another difficulty. In this paper, by modeling the state equations of the dynamically varying sensor biases with polynomial prediction model, the estimation method of multisensor time-varying bias in real time is proposed. The analytical results and simulation prove that the method proposed has better performance and less estimation covariance comparing with other methods.State equation is very important to the model-based state estimation method. In practical application, to obtain the prior information of unknown state is very difficult and as a result, the state model based on it has large uncertainty. Based on a constant acceleration motion law represented by a polynomial, a novel dynamic model of maneuver target--polynomial prediction model is proposed in this paper. Any target motion represented by the polynomials can be self-validating modeled by the polynomial prediction model which does not require any prior knowledge of the target dynamics, so the optimal filtering algorithm corresponding to the new model can track any maneuvering motion of a target. The simulation results of the maneuvering target tracking verify that the proposed model and algorithm have similar tracking performance as the interacting multiple model (IMM) method when IMM method works in ideal condition, but when IMM method works in nonideal condition, such as unexpected maneuver, the proposed model and algorithm have better trackingperformance.更多还原