Resource-Efficient Pet Dog Sound Events Classification Using LSTM-FCN Based on Time-Series Data

Area of Science:

• Computer Science
• Animal Behavior
• Internet of Things

Background:

• Internet of Things (IoT) devices are increasingly used for pet dog management, including monitoring behavior through sound classification.
• Classifying dog vocalizations (barking, growling, etc.) is crucial for understanding their emotions and needs when left alone.
• Traditional sound sensors generate large data volumes and consume significant power, posing challenges for resource-constrained IoT devices.

Purpose of the Study:

• To develop a method for classifying pet dog sound events that enhances resource efficiency without sacrificing accuracy.
• To address the limitations of using only sound intensity data for classification.

Main Methods:

• Utilized a resource-efficient noise sensor to acquire only sound intensity data.
• Applied a deep learning model, long short-term memory-fully convolutional network (LSTM-FCN), for time-series analysis.
• Employed bicubic interpolation to compensate for information loss from intensity-only data.

Main Results:

• The proposed method, using noise sensors and LSTM-FCN, achieved a tenfold improvement in energy efficiency.
• Accuracy was maintained without significant degradation compared to conventional methods using sound sensors and feature extraction techniques like MFCC.
• The Shapelet-based noise sensor approach demonstrated effectiveness in analyzing time-series data.

Conclusions:

• Classifying pet dog vocalizations using intensity data with LSTM-FCN and bicubic interpolation is feasible and energy-efficient.
• This approach offers a viable solution for resource-constrained IoT applications monitoring pet behavior.
• Significant energy savings can be achieved without compromising the accuracy of canine sound event classification.