An important application which we propose to investigate in our research is the development of machine fault detection and diagnostic systems. Specifically, we shall illustrate our approach in fault detection and diagnostics for mechanical transmissions in aircraft (e.g. an SH-60 helicopter) using signals collected by acoustical sensors. Mechanical transmissions can incur a variety of faults, from chipped teeth in transmission gears to bearing faults, which affect the vibration structures of the rotating system. Most of the prior work in fault monitoring and detection for such applications (e.g. [. Motalvao 1990, Mekdeci 1993, Kuczewski 1992, lopez heli Polyak 1995.]) has been based on processing accelerometer signals which monitor system vibrations. Such accelerometers must be hard-mounted on the vibrating system, which can alter the physical integrity and behavior of a component. In addition, the signals collected are very localized in nature. In contrast, acoustical sensors can be mounted external to the machinery and can have a wide area of regard.
The proposed application involves detection of helicopter gearbox failure using acoustic signals. Faults in the gearbox result in vibrational changes of long duration, thereby affecting the resulting acoustic signals. These vibrational changes are immersed in external interference which exhibits significant numbers of transient events. The acoustical signals are collected by a set of directional and omni-directional condenser microphones, which transduce the sound pressure signals effected by mechanical vibrations emitted by the transmission. The set of signals collected will contain ``fault signatures'' as well as signals from other interfering sources. Using an array of sensors provides the opportunity to focus the processing on the ``fault signatures'', suppressing the interference from other sources.
There are several critical challenges which make this failure detection/classification problems difficult. One of these challenges is to identify a concise yet distinctive set of features on which to base the subsequent decision process. In our proposed research, we will process the set of acoustic signals using the auditory representation models which are the focus of our research to extract a set of critical temporal and spectral features for the subsequent detection/classification problem. Another critical challenge is the development of robust classifiers which can adapt to temporal changes in loading conditions and in environmental conditions which lead to nonstationary environments for the classification problem. These temporal variations are spatially inhomogeneous across the array of sensors, creating further complications for classifier design. The results of our research on detection of temporal sequences and nonstationary HMM modeling discussed in Thrust area II(B) provides a basis for development of nonstationary, nonlinear adaptive classifiers which can robustly recognize the fault signatures embedded in the signal array.
We propose to evaluate and enhance our algorithms using data obtained from fault tests conducted at the Naval Air Warfare Center. The Aircraft Division has an SH-60 Helicopter test cell which consists of an entire SH-60 drive train, including 2 turbine engines, main and intermediate gearboxes, mechanical linkages and loading mechanisms. This test cell provides a controlled testing environment for collecting realistic signals over a broad range of operating conditions. Acoustical data collected from three sensors is available for a variety of seeded fault operations (such as chipped pinion teeth) and normal operations [. Lopez heli1 1996 .]. Recent work with these data sets [.lopez heli2 1996, lopez heli3 1996.] indicates that fault detection based on acoustical signals can be as effective as those based on accelerometer signals. As part of our research, we propose to compare the performance of our fault diagnosis algorithms with that of a recently-developed system based on multiresolution wavelet feature extraction, followed by neural network fault recognition, that is a strategy similar to the cortical multiscale representation described in Thrust area I.