Choosing the right signal processing tools for mechanical systems

• Simon Braun one of the first to appreciate the special characteristics of mechanical signals. • Machine health monitoring (MHM) a very important application of mechanical signal processing. • Causal processing is not only unnecessary but detrimental for MHM. • Non-causal processing allows ideal filtration with zero phase shift. • Non-causal processing allows error-free differentiation, integration, Hilbert transformation. • Causal real-time processing a requirement for active control, including robotics. • Causal processing an advantage for octave-based filtration over several decades. Simon Braun was one of the first to recognise the special requirements for processing of signals from mechanical systems, which was why he launched the journal Mechanical Systems and Signal Processing. Mechanical engineers are typically not well trained in signal processing, so signal processing specialists are often recruited from other areas, e.g. electrical engineering, speech processing, acoustics, which often have different requirements. A very important application area of signal processing in mechanical engineering (and mechatronics) is robotics and active control. This requires causal processing in real-time, but that places restrictions on the results, since causal filters have poor characteristics and phase distortion. There are also problems with differentiation, integration, and Hilbert transformation when performed directly in the time domain. Machine Condition Monitoring is another very important area of signal processing for mechanical engineers. This paper shows that causal signal processing is not only not required for Machine Condition Monitoring, even for online monitoring of critical machines, but gives problems and distortions that can be avoided with non-causal signal processing. The paper illustrates the advantages gained using non-causal processing, mostly based on FFT (fast Fourier transform) analysis, for ideal filtration with zero phase shift, as well as error-free differentiation/integration and Hilbert transformation via the frequency domain. However, the circularity of the (non-causal) FFT algorithm gives “wraparound effects”, which must be mitigated. The paper has a short discussion of the situations, apart from active control, where causal processing is of advantage, such as octave-based filtration, and real-time zoom as a precursor to FFT analysis. Finally, the paper discusses the special requirements of machine health indicators obtained by signal processing, because unlike structural health monitoring, they are based more on changes in forcing functions, varying greatly between different machines and components, and not just on dynamic (i.e. modal) properties.