All research
Research

Machine learning for hearing loss risk in mine workers

Machine learning for noise induced hearing loss in mine workers

Two 2019 IFAC papers: a recurrent network estimating hearing threshold shift in mine workers, and a classifier based noise policy advising system.

Status
Published
Period
2019
Role
Co-author (third of four authors on both papers)
Co-authors
MCI Madahana, JED Ekoru, OTC Nyandoro
Topics

The question

Noise exposure in mines causes permanent hearing damage. The two papers asked whether machine learning can estimate a worker's hearing threshold shift, and whether it can support decisions about which tasks to assign new employees given their hearing baseline.

How it was done

The first paper used a recurrent neural network to estimate the hearing threshold shift of mining employees and compared optimisation methods for training it. The second clustered mine workers with K-means and compared logistic regression, support vector machines, decision trees and random forests for classifying new employees, with task recommendations based on each worker's baseline and predicted future threshold shift.

What was found

  • The recurrent network predicted threshold shift with an accuracy of 95%, and the adaptive subgradient method (Adagrad) was preferred among the optimisers for its fast convergence.
  • In the policy advising system the decision tree had the highest accuracy, 91.25% average on testing and 99.79% on training, while logistic regression generalised best on the test set.

Limitations

The first paper notes performance could improve with more inputs; the second lists a usable interface for mine administrators as future work.

What it might mean

The first paper suggests the results could be used to build an early intervention and monitoring system for mines. That is a stated possibility, not something the papers built.

Publications

Sources for this page