Abstract
Utilizing optical fibers to detect and pinpoint vibrations, Distributed Optical Fiber Vibration Sensing (DVS) technology provides real-time monitoring and surveillance of wide-reaching areas. This field has been leveraging Convolutional Neural Networks (CNN). Recently, a study has accomplished end-to-end vibration event recognition, enabling utilization of CNN-based DVS algorithms as real-time embedded system for edge computing in practical application situations. Considering the power consumption of central processing unit (CPU) and graphics processing unit (GPU), and the inflexibility of application-specific integrated circuit (ASIC), field-Programmable gate array (FPGA) is the optimal computing platform for the system. This paper proposes to compress pre-trained network and adopt a novel hardware structure, to design a fully on-chip, pipelined inference accelerator for CNN-based DVS algorithm, without fine tuning or re-training. This design allows for real-time processing with low power consumption and system requirement. An examination has been executed on an existing DVS algorithm based on a 40-layer CNN model comprising 2.7 million parameters. It is completely implemented on-chip, pipelined, with no reduction in accuracy.
Introduction
Distributed Fiber Optic Vibration Sensing (DVS) turns standard telecommunication fiber into a massive, continuous sensor capable of detecting vibrations (like digging or hammering) over tens of kilometers. While powerful, DVS generates a massive volume of data that requires high-performance processing for real-time event recognition.
Processing this data usually requires Convolutional Neural Networks (CNNs). However, running deep CNNs at the "edge"—directly where the data is collected—is difficult. CPUs and GPUs are too power-hungry for remote deployments, while traditional FPGA implementations often rely on external DRAM, creating bottlenecks for speed and energy efficiency. This research presents a fully on-chip, pipelined FPGA accelerator that eliminates the need for external memory, enabling low-power, real-time "edge" monitoring.
Methodology: The Shift-Add Architecture
The core innovation is the move away from traditional Multiply-Accumulate (MAC) operations. MAC units are resource-intensive on hardware. By representing weights as sums of powers of two, the authors replace expensive multiplications with bit-wise shift operations and additions.
Shift Layer Implementation
In the proposed architecture, weights are mapped to a "Shift Layer." Instead of multiplying an input matrix by a weight matrix, the hardware performs bit-wise shifts. For example, a weight of $2^n$ corresponds to a left shift by $n$ bits. This significantly reduces the logic resources (LUTs) required compared to traditional multipliers, allowing the entire model to fit on-chip.
Fully On-Chip Pipelined Design
By compressing the 40-layer CNN (2.7 million parameters), the design fits entirely within the FPGA’s internal Block RAM (BRAM). This "fully on-chip" approach bypasses the "memory wall"—the performance bottleneck caused by moving data between a processor and external DRAM. The hardware is designed as a pipeline, allowing continuous processing of incoming sensing data.
Experimental Setup and Datasets
The system was evaluated using data from a phase-sensitive Optical Time-Domain Reflectometry ($\phi$-OTDR) system. The setup involved a 12.5-meter fiber sampled at 80 MSa/s, providing a spatial resolution of 5 meters. The dataset included three types of vibration events: Excavator, Hammer, and Air Pick. Input data was processed into matrices of size 256 × 11 for the CNN.
Results and Discussion
The implementation achieved remarkable results on a 40-layer CNN model:
- Accuracy: The compressed model achieved 99.20% accuracy, matching the original uncompressed model.
- Efficiency: The model size was reduced to 37.26% of its original size.
- Hardware Performance: Operating at 250 MHz, the system has a latency of ~397 ms per inference cycle and consumes only 10.023W.
- Resource Utilization: The design utilized 98.6% of available BRAM and 45.5% of LUTs on the target FPGA.
Conclusion
This work demonstrates that high-performance vibration recognition can be brought to the edge with high efficiency. By replacing multiplication with a Shift-Add architecture, the authors achieved real-time performance with zero loss in accuracy. Future work will focus on optimizing the buffering system to reduce latency further and expanding the model to recognize a broader range of environmental events.
Summary updated