Condition Monitoring and Fault Diagnosis of Fluid Machines in Process Industries

Condition Monitoring (CM) of fluid machines plays a critical role in maintaining efficient
productivity in many processing industries. Conventional vibration techniques generally
provide more localised information with the need for many sensors, associated data acquiring
and processing efforts, which are difficult for system deployment and are reluctantly accepted
by those industries, for example paper mills and food production lines making marginal profits.

To find adequate CM techniques for such industries this research investigates a new cost-
effective scheme of implementing CM, which combines the high diagnostic capability of using
Surface Vibration (SV) with the global detection capability of using the Instantaneous Angular
Speed (IAS) measurements and Airborne Sound (AS). To address specific techniques involved
in the scheme, this research is arranged in three consecutive Phases: Phase I is the technical
evaluation; Phase II is the field implementation practices and Phase III is the application of AS
through Convolution Neural Networks (CNN).

In Phase I, widely used reciprocating compressor is investigated numerically and
experimentally, which clarifies the performances of SV, IAS, AS, pressure and motor current
in a quantitative way for differentiating common faults such as leakages happening in valves
and intercoolers, faulty motor drives and mechanical transmission systems. It paves the
foundations for the field implementation in Phase II.

In Phase II, this novel scheme is realised on three sets of vacuum pumps in a paper mill. Based
on an analytic study of dynamic responses to common faults on these pumps, a field test was
conducted to verify the feasibility of the scheme and the preliminary study shows that airborne
sound can show the relative spectral components for each machine to a good degree of
accuracy.

Knowledge gained from the preceding phases of study is now applied to Phase III. New
techniques based on airborne signal differences through CNN have been demonstrated to give
a good indication of the sound propagation and location of noise sources under all operating
discharge pressure conditions at 100% validation accuracy, proving that the state of the art deep
leaning approaches can be used to deal with complicated acoustic data.