ML engine diagnosing engine failure from engine sounds developed using Wekinator, SuperCollider, Next.js, React.js and OSC.
- Frontend
- CSV Parser
- Graphing Tools
- Test Data
- Training Data
- References
- Models
- Analysis
- SuperCollider Buffer
- License
The Machine Learning Automated Sound System Information Retrieval is designed to detect whether a steam engine is experiencing a failure. It does so based on the sounds that the engine is emitting. The system labels the sound emitted from the machine as one of four failure types, and notifies an operator if they need to shutdown the machine to prevent further failures. Five prototypes were developed. This report concerns itself with the precision and accuracy of the prototype implemented in the case when no pre-processing took place on audio used to train and test the system. The other four prototypes evaluated the system precision and accuracy when the the cutoff frequency of band-pass filters pre-processing audio entering the classification system was changed. Test infrastructure was developed to evaluate whether the prototype met the performance benchmark of correctly identifying 95% of features in a given test set. Meeting the performance benchmark of a 95% correct detection rate would prove the null hypothesis (H_0) that: A system meeting the performance benchmark would produce feature detection results statistically insignificant from the true feature characteristics. It then follows that the alternate hypothesis (HA) is the case in which there is a statistically significant different between the predicted features generated by the system and the true feature data. This means that the system prototype is under-performing and does not meet the design requirement of a 95% correct detection rate