Improving the Event-Based Classification Accuracy in Pit-Drilling Operations: An Application by Neural Networks and...

Area of Science:

• Forestry science
• Machine learning applications
• Signal processing

Background:

• Forestry relies on resource and process monitoring, traditionally using manual time studies.
• Modern methods use data collection platforms and machine learning to improve data management.
• Current activity recognition accuracy needs improvement for efficient forest operations.

Purpose of the Study:

• To evaluate digital signal processing and Artificial Neural Networks (ANNs) for enhanced event-based classification accuracy.
• To test these methods in a case study of mechanized pit-drilling operations.
• To identify optimal signal processing and ANN tuning strategies for forestry monitoring.

Main Methods:

• Applied median filtering to triaxial accelerometer data with varying window sizes (3, 5, 21 observations at 1 Hz).
• Tuned Artificial Neural Networks (ANNs) using regularization hyperparameters.
• Tested a specific combination: median filter (window size 3) and ANN with regularization parameter α = 0.01.

Main Results:

• The best strategy involved median filtering (window size 3) and ANN tuning (α = 0.01).
• This approach improved event-based classification accuracy by 1% to 8%, depending on the event.
• Effectiveness varied with event duration and process type.

Conclusions:

• Digital signal processing and ANNs can significantly improve activity recognition accuracy in forestry.
• Optimal parameter selection for signal processing and ANNs is crucial and process-dependent.
• Further research may be needed for specific monitoring applications and event types.