UnifAI's AI outperforms traditional water sensing
Published on by Nuno Silva, CSTO - Chief Scientist and Technology Officer at UnifAI Technology in Technology
Assessment of UnifAI Technology’s Artificial Intelligence Tools and Capabilities, Using Real Industry Data on Water Quality in Pipe Infrastructure
SUMMARY
Independent testing of UnifAI Technology’s AI performance against a leading water industry incumbent solution using 3,368,928 historic data points collected over a two-year period concluded that the UnifAI AI outperformed the incumbent state of art digital sensing solution for on-line water quality measurement.
BACKGROUND
UnifAI Technology has a Data Acquisition and Visualisation Engine (DAVE) and Advanced Neural Networks and AI capability (ANNA). Together these provide a horizontal capability with applicability across multiple sectors.
Aliaxis worked with Water-link to evaluate the effectiveness of UnifAI Technology’s Artificial Intelligence capability within the water sector.
THE DATA
First batch : Water-link prepared 141,663 data samples, each with 12 parameters, from a surface water at the production intake covering the period April 2019 to December 2019.
In addition, Water-link provided a list of alerts and alarms for the same period identified by the incumbent on-line sensing system, and determined the thresholds for the alarms. No information was provided to UnifAI as to the nature of the alerts/alarms or how they were calculated.
Purpose: training data for the AI.
Second batch : subsequently, Water-link provided a further 139,081 data samples, each with the same 12 parameters, from the same infrastructure covering the period January 2020 to August 2020, without any data on alerts/alerts for this period.
This was the testing data for the AI.
THE CHALLENGE
The challenge was a ‘blind test’ in which the raw data was provided to UnifAI without context. No domain knowledge was implemented or embedded into the test. The outputs from UnifAI’s AI after it had ‘learned’ how and when to identify alerts/alarms was compared with the outputs from the incumbent on-line sensing solution.
Training
The UnifAI Data Acquisition tool was used to ingest and clean the training data provided by Water-link.
UnifAI’s ANNA then conducted a series of training exercises using the 80/20 approach:
· 113,330 samples (80% of the data) were used to train the neural network to identify and understand the multiparameter and multi-dimensional correlations between the raw data and the alerts/alarms.
· 28,330 samples (20% of the data) were used to test/validate the training, and to select the most appropriate neural network for the challenge.
The result was a trained neural network ready for testing.
Testing
The testing data was then given to UnifAI, and this was also ingested and cleaned.
All 139,081 samples, covering January to August 2020, were run through the neural network that ANNA had trained. The outcome was a set of alerts and alarms provided by ANNA for the second time period.
These AI-generated alerts/alarms were then independently compared with the alerts and alarms from the incumbent on-line sensing system for the same time period.
RESULTS
· 99.97% of alert and alarm outcomes from the test data (139,081 samples) were the same.
· There were 12 exceptions where alerts/alarms were raised in one solution but not the other.
· The exceptions were further investigated:
9 of the 12 anomalies, the UnifAI AI was right and the incumbent system was wrong.
2 of the 12 anomalies, the UnifAI AI was wrong and the incumbent system was right.
1 of the 12 anomalies was a null event which we assume resulted from untrained data.
CONCLUSION
The conclusion of Aliaxis and Water-link is that the UnifAI Technology AI outperformed the incumbent state of art on-line digital sensing solution for the identification of alerts and alarms based on the parameters collected.
Phil Hughes, CEO UnifAI Technology said: “our mission is to use AI to improve the health and well-being of people, the environment and the infrastructure we use. The work done by Aliaxis and Water-link demonstrates that AI can outperform incumbent systems and the benefit of this is that we can now begin to drive down the cost of water quality measurement and management in a way that helps water utilities to improve the product they deliver to their consumers with a capability that more easily help to manage evolving risk based compliance requirements. What’s really exciting is that we believe we can begin to predict when events are likely to happen so we can help make a material difference to the compliance, quality and operations of water utilities.”
Taxonomy
- artificial intelligence
- artificial neural network