Optimisation of vibration monitoring nodes in wireless sensor networks

This PhD research focuses on developing a wireless vibration condition monitoring
(CM) node which allows an optimal implementation of advanced signal processing
algorithms. Obviously, such a node should meet additional yet practical requirements
including high robustness and low investments in achieving predictive maintenance.
There are a number of wireless protocols which can be utilised to establish a wireless
sensor network (WSN). Protocols like WiFi HaLow, Bluetooth low energy (BLE),
ZigBee and Thread are more suitable for long-term non-critical CM battery powered
nodes as they provide inherent merits like low cost, self-organising network, and low
power consumption. WirelessHART and ISA100.11a provide more reliable and robust
performance but their solutions are usually more expensive, thus they are more suitable
for strict industrial control applications.
Distributed computation can utilise the limited bandwidth of wireless network and
battery life of sensor nodes more wisely. Hence it is becoming increasingly popular in
wireless CM with the fast development of electronics and wireless technologies in
recent years. Therefore, distributed computation is the primary focus of this research in
order to develop an advanced sensor node for realising wireless networks which allow
high-performance CM at minimal network traffic and economic cost.
On this basis, a ZigBee-based vibration monitoring node is designed for the evaluation
of embedding signal processing algorithms. A state-of-the-art Cortex-M4F processor is
employed as the core processor on the wireless sensor node, which has been optimised
for implementing complex signal processing algorithms at low power consumption.
Meanwhile, an envelope analysis is focused on as the main intelligent technique
embedded on the node due to the envelope analysis being the most effective and general
method to characterise impulsive and modulating signatures. Such signatures can
commonly be found on faulty signals generated by key machinery components, such as
bearings, gears, turbines, and valves.
Through a preliminary optimisation in implementing envelope analysis based on fast
Fourier transform (FFT), an envelope spectrum of 2048 points is successfully achieved
on a processor with a memory usage of 32 kB. Experimental results show that the simulated bearing faults can be clearly identified from the calculated envelope
spectrum. Meanwhile, the data throughput requirement is reduced by more than 95% in
comparison with the raw data transmission. To optimise the performance of the
vibration monitoring node, three main techniques have been developed and validated:
1) A new data processing scheme is developed by combining three subsequent
processing techniques: down-sampling, data frame overlapping and cascading. On
this basis, a frequency resolution of 0.61 Hz in the envelope spectrum is achieved on
the same processor.
2) The optimal band-pass filter for envelope analysis is selected by a scheme, in which
the complicated fast kurtogram is implemented on the host computer for selecting
optimal band-pass filter and real-time envelope analysis on the wireless sensor for
extracting bearing fault features. Moreover, a frequency band of 16 kHz is analysed,
which allows features to be extracted in a wide frequency band, covering a wide
category of industrial applications.
3) Two new analysis methods: short-time RMS and spectral correlation algorithms are
proposed for bearing fault diagnosis. They can significantly reduce the CPU usage,
being over two times less and consequently much lower power consumption