Control of Combustion Instabilities on a Rijke Tube by a Neural Network
Control of Combustion Instabilities on a Rijke Tube by a Neural Network
复制标题
通过神经网络控制 Rijke 管上的燃烧不稳定性
DOI：
--
复制
发表时间：
2000
期刊：
影响因子：
--
通讯作者：
A. Laverdant
中科院分区：
文献类型：
--
作者：
R. Blonbou;A. Laverdant
燃烧不稳定性仍然是动力装置发展中的一个主要问题。本文考虑了一种Rijke管，它在某些工作条件下，在145 dB/Hz以下的压力水平下出现不稳定性。为了控制非线性对象的不稳定性，提出了一种利用两个神经网络的非线性对象内模控制系统。第一个是内部模型，它近似于工厂的前向动态（学习过程）。第二种是自适应控制输入。证明了该方法对不稳定性噪声信号（麦克风或光电倍增管的OH辐射与热释放成正比）的逼近能力，并具有良好的精度。衰减的不稳定性，为固定或可变的operatingconditions，与压力水平衰减高达60 dB/Hz.IntroductionCombustion不稳定性是一个重要的困难，在航空或航空航天动力装置的概念遇到[1-4]。这些振荡机制具有灾难性的后果，如结构振动、增加的热传递与稳定状态、不期望的滚转扭矩，并且有时破坏动力装置。根据威廉姆斯的研究[2]，观察到了几种类型的振荡（系统不稳定性、固有不稳定性或腔室不稳定性）。在本文中，只考虑最后一种情况。例如，在冲压式喷气发动机中，进气口和/或燃烧室的一个或几个声学模态与不稳定燃烧放热耦合。在这里，涉及的燃烧器是一个Rijke管与丙烷-空气混合物在多孔板上燃烧。在剑桥大学[5，6]或巴黎中央高等学校[7]以前的研究中已经考虑过类似的设置。控制不稳定性（被动或主动）是克服这些不必要的振荡的自然方法。第一种方法-被动控制[4，8] -包括通过引入挡板或亥姆霍兹谐振器来改变腔室的几何形状，这些挡板或亥姆霍兹谐振器改变了本征频率并引入了额外的不稳定性阻尼。不幸的是，这种方法增加了动力装置的质量和尺寸。此外，其阻尼效率是有限的。第二种方式-主动控制-（参见[9]的良好评论），包括引入具有正确相移的次级声学信号以获得稳定状态。本工作的出发点是采用一种非线性滤波技术--神经网络。主动控制已成功地用于抑制实验室燃烧室中的不稳定性。控制回路由一个或多个传感器（检测与热释放成比例的光发射的麦克风或光电倍增管）[5]，相移装置和致动器（通常是电动气动驱动器）[9，10]组成。控制方案控制不稳定模式在足够远的频率。尽管取得了一些成功，但这种技术存在局限性。首先，手动设置控制器的参数。其次，当几种模式存在于同一水平时，控制失败。为了克服这一限制，Huynh等人。[10]，Billoud等人。[11]，通过优化控制器参数（Widrow [ 12]），使用了基于全局响应传感器信号能量最小化的自适应控制算法。滤波器是线性的（无限脉冲响应-IIR）。在250 kW湍流燃烧器中控制不同的不稳定模式。Neumeier和Zinn [13]使用一个观测器来识别不稳定模式并确定它们的相位和幅度。这些信息是稳定每个不稳定模式的控制器的输入。这种方法在重建低频到高频过渡的不稳定压力信号方面取得了一定的成功。然而，控制方案受到火焰对声扰动响应的线性模型的限制。最近，Yu、Wilson和Schadow [14]使用液体喷射（乙醇、庚烷和JP 10）来抑制突扩燃烧室的燃烧不稳定性。液滴夹带和火焰/大尺度涡的相互作用限制阻尼为15 dB/Hz。注入通带似乎被限制在几百赫兹。然而，这项研究已经给出了一个更好的理解的物理-化学过程中涉及的不稳定的转储combustion.Paper介绍了在RTO A VT研讨会上“主动控制技术为增强性能作战能力的军用飞机，陆地车辆和海上车辆在布伦瑞克，德国，8-11五月2000，并发表在RTO MP-051。
AbstractCombustion instabilities still constitute a major problem for powerplant development. In this paper, a Rijke tube which pre-sents, for some operating conditions, instabilities with pressure level up to 145 dB/Hz, is considered. In order to control instabili-ties, an Internal Model Control System for nonlinear plants, that uses two neural networks, has been developed. The first one isan Internal Model which approximates the plant forward dynamic (the learning process). The second one gives the adaptivecontrol input. The capacity of approximating the noisy signals of instabilities (microphone or photomultiplier for OH emissionproportional to heat release), with a good precision, is demonstrated. Attenuation of instabilities, for fixed or variable operatingconditions, with pressure level attenuation up to 60 dB/Hz, has been obtained.IntroductionCombustion instabilities represent an important difficulty encountered during the conception of aeronautical or aerospacepowerplant [1-4]. These oscillatory mechanisms have disastrous consequences like structure vibrations, increased heat transferversus stable regime, undesired roll torque and, sometimes, destruction of the powerplant. According to Williams [2], severaltypes of oscillations are observed (system -, intrinsic -or chamber instability). In this paper, only the last case is considered. Asan example, in a ramjet, one or several acoustic modes of an air intake and / or chamber, are coupled with unsteady combustionheat release. Here, the burner involved is a Rijke tube with propane-air mixture burning above a porous plate. Similar setups havebeen considered in previous studies at Cambridge University [5,6] or Ecole Centrale de Paris [7].The control of instabilities (passive or active) is a natural way to overcome these unwanted oscillations. The first way -thepassive control [4,8] -consists in modifying the chamber geometry by the introduction of baffles or Helmholtz resonators whichchange eigenfrequencies and introduce additional damping of instabilities. Unfortunately, this approach increases mass and di-mensions of the powerplant. Moreover, its damping efficiency is limited. The second way -the active control -(see [9] for a goodreview), consists in the introduction of a secondary acoustic signal with a correct phase shift to obtain a stable regime. The origi-nal aspect of the present work is the use of a nonlinear filtering technique : the neural network.The active control has been used with success to suppress instabilities in laboratory combustors [5-7, 9]. The control loopconsists of one or more sensors (microphone or photomultiplier which detects light emission proportional to heat release) [5], aphase shift device and an actuator (generally an electro-pneumatic driver) [9, 10]. The control schemes control unstable modes atsufficiently distant frequencies. Despite of some success, this techniques presents limitations. First, the parameters of the con-troller are set manually. Secondly, when several modes are present with the same level, the control fails.In order to overcome this limitation, Huynh et al. [10], Billoud et al. [11], have used an adaptive control algorithm basedupon energy minimization of the global response sensor signal, through optimization of the controller parameters (Widrow [ 12]).The filter is linear (Infinite Impulse Response -IIR). Different unstable modes are controlled in a 250 kW turbulent burner.Neumeier and Zinn [13] uses an observer which identifies the unstable modes and determines their phases and amplitudes.These informations are inputs for a controller which stabilizes each unstable mode. This approach has obtained some success forthe reconstruction of an unstable pressure signal with transition from low to high frequency. However, the control scheme islimited by the linear modeling of flame response to acoustic perturbations.More recently, Yu , Wilson and Schadow [14] have used liquid injection (ethanol, heptane and JP 10) to damp combustioninstabilities of a dump combustor. Droplets entrainment and flame/large scale vortex interaction limit the damping to 15 dB/Hz. Itseems that the injection pass band is limited to several hundred of Hertz. However, this research has given a better understandingof physico -chemical processes involved in an unstable dump combustor.Paper presented at the RTO A VT Symposium on "Active Control Technology forEnhanced Performance Operational Capabilities of Military Aircraft, Land Vehicles and Sea Vehiclesheld in Braunschweig, Germany, 8-11 May 2000, and published in RTO MP-051.