An Interpretable Sensor Selection Strategy for Array Optimization and Performance Enhancement.

Area of Science:

• Sensor technology
• Artificial intelligence
• Data science

Background:

• Increasing sensor integration in electronic nose (E-nose) systems improves gas detection but introduces challenges like cross-sensitivity and redundancy.
• Array optimization is critical for enhancing multisensor system performance and overcoming these limitations.

Purpose of the Study:

• To propose an interpretable sensor selection strategy, SHMI-Select, for optimizing multisensor arrays.
• To reduce hardware costs, computational complexity, and information redundancy in E-nose systems.
• To ensure system adaptability and stability for various gas detection tasks.

Main Methods:

• Developed SHMI-Select, a method combining Shapley values and mutual information for sensor selection.
• Implemented an interpretable primary sensor selection based on interpretability analysis.
• Utilized mutual information for secondary sensor identification and an incremental approach for optimal combination.
• Validated the method on human breath, wine quality, and environmental gas datasets.

Main Results:

• SHMI-Select significantly reduced sensor redundancy across all datasets.
• Achieved substantial performance gains: 62.5% fewer sensors with 10% accuracy increase on breath data.
• Demonstrated 83.3% sensor reduction with 18% accuracy improvement for wine classification.
• Showcased 62.5% sensor reduction with a 2% R² increase for environmental gas detection.

Conclusions:

• The SHMI-Select method effectively optimizes E-nose sensor arrays, reducing complexity and cost.
• Offers significant improvements in accuracy and performance compared to existing algorithms.
• Presents strong application value and economic benefits for the industrialization of E-nose systems.