Configurable Multi-Layer Perceptron-Based Soft Sensors on Embedded Field Programmable Gate Arrays: Targeting Diverse...

Area of Science:

• Embedded Systems Engineering
• Machine Learning Hardware Acceleration
• Digital Signal Processing

Background:

• Developing efficient soft sensors for FPGAs is challenging due to varying hardware constraints.
• Existing workflows often lack the adaptability needed for diverse deployment objectives.
• Optimizing for precision, latency, and power consumption requires tailored hardware solutions.

Purpose of the Study:

• To present a comprehensive workflow for developing and deploying adaptable Multi-Layer Perceptron (MLP)-based soft sensors on embedded FPGAs.
• To introduce a novel, open-source toolchain, ElasticAI.Creator, to facilitate the entire development and deployment process.
• To demonstrate the workflow's effectiveness in achieving a balance between precision, inference latency, and energy efficiency for different deployment scenarios.

Main Methods:

• Developed a flexible workflow supporting configurable MLP architectures (layer/neuron counts) and quantization bitwidths.
• Utilized the open-source ElasticAI.Creator toolchain for quantization-aware training, integer-only inference, and automated VHDL accelerator generation.
• Performed case studies on fluid flow estimation using two distinct FPGA platforms: AMD Spartan-7 XC7S15 and Lattice iCE40UP5K.

Main Results:

• Achieved high precision (MSE: 56.56, MAPE: 1.61%) with low latency (23.87 μs) for a precision-focused MLP on the XC7S15.
• Demonstrated low power (2.06 mW) and energy efficiency (0.172 μJ/inference) for a compact MLP on the iCE40UP5K.
• Validated the workflow's capability to tailor FPGA accelerators to specific performance requirements.

Conclusions:

• The presented workflow enables the development of optimized MLP-based soft sensors for FPGAs.
• ElasticAI.Creator facilitates adaptable and efficient deployment across diverse hardware platforms.
• The study successfully balances key performance metrics like precision, latency, and energy consumption for embedded ML applications.