Every day, we have thousands of operators climbing the wind turbine tower for maintenance. Now, why do we not have lifts inside to hoist the climbers up? Since most onshore wind turbines are shorter than 90m (296 feet approx.), it is difficult to have an elevator inside them. Instead, they usually have a ladder with some assistance mechanism that helps hoist the climbers up with relative ease. That’s where Moshman Research helps out in building their first machine learning powered climb assist subsystem.
Most of these climb assist subsystems have an assistance level that is set by the climbers. Instead of following this traditional method, Moshman Research is revolutionizing this technology. With the help of a vast array of sensors, the power of predictive modeling, and Data Science will help predict the climber’s weight.
We use a variation of the LSTM models to predict the climber’s weight for our problem space. These models are commonly used in Gmail’s autocomplete to predict the next word in a sentence. Gmail uses the English language’s grammar to guide the machine learning process. We took a similar approach to predict the user’s weight. Instead of using grammar to guide us, we used data from sensors for our deep learning models.
Evolution of machine learning models
During the initial stages, we used thorough data analysis and testing to figure out the correlation between features. This information was then used to find a suitable a machine learning model. With many iterations including Logistic Regression, ensemble tree models and Gradient Boosting Regressors (GBR), we were able to fit a GBR with the little data that we had.
With improved data acquisition techniques, we fit a model that works well with a large dataset — LSTMs. Long short-term memory models, or LSTMs are essential for leading autocomplete algorithms. This altered neural network is able to learn from previous layers as well as carry useful information forward from those layers.