transforming retail petroleum

Russell Dupuy 15 May 2018 By Russell Dupuy

Transforming maintenance in Retail Petroleum Part 2

This article leads on from prior articles regarding digital transformation through Machine Learning, Artificial Intelligence and Deep Learning techniques within a Retail Petroleum environment.

This article looks at what Deep Learning can do for Retail Petroleum in the context of Fuel System Asset Maintenance. Further articles will explore Logistics, Retail Pricing Compliance, OH&S, Leak Detection and driving Performance improvements.

To quickly recap, I’ve drawn upon simple definitions supported by an equally simple analogy [at the risk of over simplification]:

Artificial Intelligence in its simplest form is human logic embedded into a “Device, Machine or what is more commonly called Things”. This can range from basic logic routines through to complex algorithms.

Example: A garden watering system often features a timer, moisture, wind and rain sensors and a solenoid that replaces the human task of watering the garden.

The water system will turn the solenoid on/off as programmed to coincide with simple conditions such as, the soil is moist and it’s windy or raining. But it does not retain a history of weather conditions to reference.

Machine learning by contrast [using the same watering system example] is when the watering system has the ability to go beyond simple conditional on/off control logic based on sensor inputs and feedback. The Water system may have the ability to store historical data, develop patterns from the various inputs to make reason-based decisions on what it might do next based on change.

Example: It may decide that it’s been dry for a few days now, the soil is not moist, and it’s also been very windy, that will further dry out plants above the soil line. It might decide to wait until the wind calms down and water a little longer. It then might go further and decide, the wind is not calming, therefore plants are further dehydrating, and whilst against simple logic continue to water until the soil is moist. It then might reason, it has attempted to moisten the soil for an extended period to no avail, abort to save further watering until the wind calms.

Finally, Deep Learning in its simplest form is borne from the application of artificial intelligence melded to high fidelity – high frequency – high volume machine’s data that may be relational or disparate, that sets a population of facts for deeper enquiry.

Example: Using the Watering system example once again, then imagine 1,000+ watering systems across any country, the significant volumes of fact data, and what insights that may yield descriptively and inferentially. Further imagine this applied to Agriculture and the improvements in crop yield, productivity, and logistics and not to mention overall locale based weather monitoring. This really only touches the surface of what is possible and has actually been used in Agribusiness and Pig farming for a while now.

To the point: What will it do for maintenance in Retail Petroleum?

Replacing the Watering system with a Fuel system, consider a mid-sized retail network of 500 petrol sites, each featuring five underground storage tanks fitted with the following equipment: an Automatic Tank Gauge [ATG] with five level, five pressure and five discriminating sump sensors, five intelligent submersible turbine pumps [STP], that push fuel to four Dispenser island [5 hose].

The site may well feature stage 1 and 2 vapour recovery, a forecourt wastewater containment system, that is then blended with refrigeration, various vending machines, food warming and other reactive electrical loads; compressors, lighting, general power etc. The list of equipment or ‘Machines/Things’ and their intelligence is rapidly growing.

Sizing up the sensory inputs at a site level looks as follows:

  • 4 x five hose dispensers = 40 hoses x 500 sites = 20,000 hoses, to maintain uptime, flow, pressure and optimization
  • 5 x ATG probes = 3,000 tank compartments to monitor inventory, delivery, leak/losses, water, temperature and associated alarms
  • 5 x Sump sensors = 3,000 containment sumps to monitor hydrocarbon/water ingress
  • 5 x STP’s = 3,000 turbines to monitor rest/demand pressure, voltage/current and power factor plus alarms
  • 6 x reactive loads [Fridge, vending, power, lighting etc] to monitor power factor and status
  • 1 x Electronic Price sign – maintaining price compliance and know when something is not working
  • CO2 sensors located in Cool rooms or Fridges to detect oxygen depletion
  • Emergency stops [E-Stop]

Across a network of 500 sites, that is approximately 33,000 sensors – who’s watching over them?

Transforming fuel maintenance


The volume of data is typically 200MB per site per month to transport, sort, store, manipulate to test-measure before anything really useful can be done with it.

Another key point for consideration is how often to transport sensor data to the cloud, which is a serious consideration when sizing cloud database and network traffic. We have seen networks of this size transporting 3.5 million messages a day that translates into approximately 4 gigabytes to store any given day!

Sensors individually are limited, at a site level they offer a measurable return on investment – at a network level, they present significant data for harvesting!

Big data of itself is not always helpful, more often than not overwhelming in volume and underwhelming in terms of insights. At this level, the human can’t process, let alone look for relational, conditional behaviour or scenario patterns, not to mention the cerebral gymnastics to apply mathematical and statistical computations required to draw correlations and/or inferences.

Back to the point: What will it do for maintenance in Retail Petroleum?

Example 1: A common goal of many petroleum retailers is to maintain hose/nozzle uptime, which equates to consumers quickly filling their vehicles.

It becomes obvious to monitor hose flow rates, which is not new or that insightful of itself.

However, when flow is correctly monitored for every hose in real-time [not historically meaning every few hours or day-old data] compared to all other hose flow rates, then modelled against piping pressure, and tested against a few other useful data parameters, as seen in the data capture below – real insights can be mined.

Dispenser Diagnostics

This example illustrates hose/nozzle pattern analysis. The power is within the background computations and deep learning techniques that further model all hoses/nozzles across an entire network in terms of:

  • Is any given hose within its normal behaviour pattern, that is dynamically changing
  • Every hose is modelled against all other hoses on the same grade
  • For submersible turbines, the pressure is modelled as is transaction data to in-tank movements and vapour emissions

The outcome is network-wide:

  • Maximised hose delivery rates meaning improved uptime
  • Predictive alerts to impending maintenance events that impact uptime and delivery rates
  • Highly accurate prescriptive maintenance diagnostics highlighting route cause
  • Hose/nozzle meter over/under dispensing quantification in real-time
  • Site optimisation for grade changes and future designs

To unlock insights that deliver significant business performance improvement, applying artificial intelligence and deep learning techniques produce profound outcomes that drive asset management improvements.

To learn more about how this translates into value at the bottom-line, get in touch with us via