On the Use of Concentrated Time-Frequency Representations as Input to a Deep Convolutional Neural Network:...

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Time-frequency (TF) analysis effectively processes non-stationary signals, extracting meaningful parameters and enabling sparse representations.
• Deep learning, particularly Convolutional Neural Networks (CNNs), excels at pattern recognition but often lacks interpretability.
• Non-intrusive load monitoring (NILM) aims to identify household appliances (HEAs) from their energy consumption data.

Purpose of the Study:

• To integrate TF analysis and CNNs for enhanced NILM.
• To investigate the impact of synchrosqueezed TF representations as input for 2D CNNs in pattern recognition.
• To develop an interpretable deep learning model for NILM by linking CNN features to handcrafted interpretable features using Layer-wise Relevance Propagation (LRP).

Main Methods:

• Utilized synchrosqueezed and non-synchrosqueezed TF representations as input for a 2D CNN.
• Applied the Layer-wise Relevance Propagation (LRP) method to interpret the CNN's decision-making process.
• Conducted experiments on the publicly available PLAID dataset for HEA recognition.

Main Results:

• Achieved excellent appliance recognition accuracy, exceeding 97%, with an optimized TF representation.
• Demonstrated the effectiveness of the combined TF analysis and CNN approach for NILM.
• Successfully interpreted the information learned by the CNN models.

Conclusions:

• The synergistic combination of TF analysis and CNNs offers a powerful and interpretable solution for NILM.
• The choice of TF representation significantly impacts the performance of CNN-based pattern recognition.
• The LRP method provides valuable insights into the decision-making of deep learning models in signal analysis tasks.