Machine Learning Developer in Perth, Western Australia, Australia

Developed an automatic processing system that outperforms the analysts for routine processing in terms of processing quality and speed.

Our processing scheme mixes novel approaches and a more traditional method relying on energy variation-based event detection and on the picking of arrival times. Our platform directly employs the raw events detected by the seismic system. The event detection is set to be highly sensitive. When an event is detected, a block of continuous data is extracted and the event is located using a grid-based event location method. It classifies the event into ten categories, picks the arrival times, locates the events, and calculates the source parameters. The Oyu Tolgoi mine is currently using this system to process data in a production setting.

The system is deployed on Kubernetes clusters. Automatically processing seismic data requires the adoption of a holistic approach that extends to the design of the seismic array, the choice of sensors and the type and shape of the input data.

Using convolution neural networks and mixed networks with other features available, an automatic seismic identifier was created that outperforms other models in the market.

The project is to auto-tag production-level music records to run under AWS to enable auto-tagging of music as they are used. The auto-tagging included mood/fee, genre, Instruments, vocals, and Type. The project converted the audio track into a mel spectrogram and did image classification. For the beats or beats per minute (BPM), I used wave analysis and some property software written to estimate BPM. The project achieved over 80% macro precision and 60% macro recall.

Electricity demand prediction using current population categories and weather conditions helps companies plan their usage of electricity to reduce costs.

Western Power had a hack-a-gig competition to help them use their smart meter collected data, weather conditions, and population categories to determine their electricity demand. This helps clients to achieve minimum costs and helps Western Power plan its electricity network.

Mine was the winning solution and was written in Python using Keras for neural network and a mixed input model neural network, which consisted of the following inputs:

• Embedding of categorical fields like household categories, days to public holidays, and days after a public holiday (up to 5 days).
• Tabular numerical data for weather conditions.
• LSTM of the previous usage pattern of electricity.

These inputs were then concatenated into a fully connected deep neural network with one output variable to estimate the current usage of electricity per household. The data is then aggregated to reduce error by localities. My solution gave the best result in the competition, and the best reward was getting contracted by Western Power to do the project.

BHP mines have a large number of haul trucks, light vehicles, dozers, excavators, and others. There is a need to detect when drivers exceed the speed limit, hit the brakes harshly, and take turns at high speed. These events lead to fast wear of tires and required coaching drivers to stop doing that. Light vehicles should not approach haul trucks less than 50m unless they have positive communication since the haul truck driver may not be able to see the light vehicle. Detecting these events helps increase safety in mines.

I wrote a C# application to collect GPS location data from various sources streamed through Kafka and then calculate the speed and orientation using the Kalman filter to reduce errors in the GPS measurements, and I stored the data in an Elasticsearch spatiotemporal database. I then used this data to determine events using the Elasticsearch database. I was the lead developer in the project.