Development of an artificial intelligence based occupational noise induced hearing loss early warning system for mine...

Occupational Noise Induced Hearing Loss (ONIHL) is one of the most prevalent conditions among mine workers globally. This reality is due to mine workers being exposed to noise produced by heavy machinery, rock drilling, blasting, and so on. This condition can be compounded by the fact that mine workers often work in confined workspaces for extended periods of time, where little to no attenuation of noise occurs. The objective of this research work is to present a preliminary study of the development of a hearing loss, early monitoring system for mine workers.

The system consists of a smart watch and smart hearing muff equipped with sound sensors which collect noise intensity levels and the frequency of exposure. The collected information is transferred to a database where machine learning algorithms namely the logistic regression, support vector machines, decision tree and Random Forest Classifier are used to classify and cluster it into levels of priority. Feedback is then sent from the database to a mine worker smart watch based on priority level. In cases where the priority level is extreme, indicating high levels of noise, the smart watch vibrates to alert the miner. The developed system was tested in a mock mine environment consisting of a 67 metres tunnel located in the basement of a building whose roof top represents the "surface" of a mine. The mock-mine shape, size of the tunnel, steel-support infrastructure, and ventilation system are analogous to deep hard-rock mine. The wireless channel propagation of the mock-mine is statistically characterized in 2.4-2.5 GHz frequency band. Actual underground mine material was used to build the mock mine to ensure it mimics a real mine as close as possible. The system was tested by 50 participants both male and female ranging from ages of 18 to 60 years.

Preliminary results of the system show decision tree had the highest accuracy compared to the other algorithms used. It has an average testing accuracy of 91.25% and average training accuracy of 99.79%. The system also showed a good response level in terms of detection of noise input levels of exposure, transmission of the information to the data base and communication of recommendations to the miner. The developed system is still undergoing further refinements and testing prior to being tested in an actual mine.