BMTA Newsletter BMTA Newsletter - Spring 2020 | Page 11
bmta.co.uk
Therefore, advanced predictive models can be used to perform predictions based on the available data as well as extracting
significant information for the end-user which will aid their decision-making processes and ultimately reduce operating costs.
To fully understand the potential and benefits which may be realised by incorporating machine learning algorithms and predictive
models in an industry such as flow measurement, several case studies have been conducted within TÜV SÜD National Engineering
Laboratory using historical data generated from flow meters.
Case Study: Error prediction using a machine learning model
A research project investigating the effect of improper installation on the performance of ultrasonic flow meters (USMs) was
conducted in 2011. It was observed that scenarios such as vertical and horizontal misalignment triggered drifts in the meter’s digital
diagnostic variables. However, it raised the challenge of distinguishing between different installation related errors based on drifts
seen in variables. Furthermore, due to the vast amount of data involved, it is difficult to manually digest the information provided
without using advanced modelling techniques. As a result, to unlock the potential within this historical data, a machine learning
model was constructed in analysing and predicting the cause of drifts in different USMs based entirely on the data produced from
flow meters. Over the years, machine learning models have become increasingly popular in the world of data modelling, due to its
ability to “learn” complex patterns, relationship and trends in a large dataset and applying the “learnt” knowledge in predicting
another data set. Certain machine learning models can even handle incomplete observations and “break-down” interrelationships
between variables to prevent over-optimistic predictions being made and improve the generalisation of models.
In this case study, a machine learning model was constructed based on historical USM data. The following predictions were made
using the constructed model to distinguish between four different types of operating conditions based entirely on the relationship
and trends observed in the USM’s digital output data. The four types of operating conditions are given below:
USM set up in an ideal condition
USM set up in a vertical misaligned manner
USM set up in a horizontal misaligned manner
USM set up with step change error induced by altered line build.
Note that vertical and horizontal misalignments can happen in practice if smaller bolts were used instead of the appropriate size
bolts. A summary of the prediction results is given in the following figure, where green bar denotes the correct prediction of error
responsible for the data drifts and the rest denote the wrongly predicted cause of error.
The number attached to each bar represents the probability of that class of error being predicted by the machine learning model.
For example, for Data 1, the machine learning model had predicted correctly, with probability 0.99, that the condition which was
represented by the data indicates a normal operating USM under an ideal set up. In other words, the model is 99% certain that the
data indicates a healthy performing USM and thus no further maintenance required. Whereas, for Data 2, the model is 91% certain
that this USM had the meter installed incorrectly in a vertical misaligned manner and thus rectifying action would be required.
Therefore, one of the main potentials from using predictive models such as machine learning algorithms in flow measurement is
that we can predict with high accuracy the exact cause of measurement error.
Furthermore, through the use of advanced predictive modelling techniques in flow measurement, we are capable of distinguishing
between errors. This shapes the foundation of future research in understanding the importance of diagnostic data as well as
identifying indicative variables under different operation conditions.
Predictive models are agnostic and are not limited to a particular industry. For end-users who wish to incorporate advanced
modelling techniques in their research interest, it is advised that simple baseline models should be constructed using existing
historical data as a proof of concept on the capability, before extending models to apply on a much larger data set and spending
considerable time and money in gathering new data.