• Initial data mining showed some loggers making multiple recordings resulting in excess battery/ data use. More than 10% of loggers had made at least one recording per month and 5% had made 3 or more in a month.
• Firmware updated to wait for more than one day of ‘no leak’ before new recording would be made.

• Initial data mining showed that of 20000 units deployed, 85% of units have very low level-spread values that indicate not likely to be leaks.
• In these cases the determination can be stopped after the first minute of the 5 minute determination to make significant power saving of nearly 2Ah (~20%) over 5 years.
• firmware updated to check result of determination after 1 minute and not continue with the determination as the result will be negative.

This feature can be remotely configured so can be deployed inactive then subsequently activated.

• Total of 500 leak, 500 non leak sound files were segregated and analysed using MATLAB.
• Approaches taken - General audio characteristics, Frequency analysis, Wavelet transformations, Audio fingerprinting. These approaches were unsuccessful in creating a robust classifier.
• Hence a Machine Learning model becomes necessary to solve the problem of determining leaks from non leaks.

• PRV noise is biggest single source of false alarms.
• PRV noise is highest during the time when we perform measurements – this is because in the night the flow is lowest and PRV valves have to provide highest resistance.
• Night line improvement – average sensor noise reduces by 4% after system deployed