Machine learning enhanced smart tactile specklegram fiber sensor using spatial mode downsampling.

Area of Science:

• Fiber Optic Sensors
• Machine Learning Applications
• Tactile Sensing Technology

Background:

• Multimode fiber (MMF) specklegram sensors traditionally require significant data processing.
• Existing methods often involve complex free-space optics and cameras, increasing system cost and complexity.
• There is a need for more computationally efficient and integrated tactile sensing solutions.

Purpose of the Study:

• To demonstrate a proof-of-concept for a tactile sensor utilizing a seven-core fiber (SCF) and machine learning (ML).
• To reduce data processing workload and computational time in MMF-based tactile sensing.
• To achieve high-resolution, multi-parameter sensing (force and position) in an all-fiber configuration.

Main Methods:

• Data condensation from MMF high-resolution images into seven power values from SCF cores.
• Application of Gaussian process regression (GPR) for demodulation and force distribution analysis.
• Development of a downsampling technique analogous to ML pooling layers for enhanced efficiency.

Main Results:

• Achieved a spatial resolution of 0.075 mm for normal force detection with 1 mm sampling.
• Demonstrated a 20x reduction in computation time compared to conventional MMF specklegram ML algorithms.
• Attained over 96% accuracy for simultaneous measurement of contact force and position.

Conclusions:

• The SCF-based sensor significantly streamlines data acquisition and storage while maintaining signal integrity.
• The proposed technique eliminates the need for external cameras and free-space coupling, paving the way for all-fiber sensors.
• This approach offers a high-speed, cost-effective, and multi-parameter tactile sensing solution.