Combining photonic neural networks with distributed acoustic sensing for infrastructure monitoring

Distributed acoustic sensing (DAS) systems represent cutting-edge technology in infrastructure monitoring, capable of detecting minute vibrations along fiber optic cables spanning tens of kilometers. These systems have proven invaluable for applications ranging from earthquake detection and oil exploration to railway monitoring and submarine cable surveillance.

However, the massive amounts of data generated by these systems create a significant bottleneck in processing speed, limiting their effectiveness for real-time applications where immediate responses are crucial.

Machine learning techniques, particularly neural networks, have emerged as a promising solution for processing DAS data more efficiently. While the processing capabilities of traditional electronic computing using CPUs and GPUs have massively improved over the past decades, they still face fundamental limitations in speed and energy efficiency. In contrast, photonic neural networks, which use light instead of electricity for computations, offer a revolutionary alternative, potentially achieving much higher processing speeds at a fraction of the power.

Unfortunately, integrating these optical computing systems with DAS technologies has presented significant technical challenges, particularly in handling the complex data structures and ensuring accurate signal processing.

Against this backdrop, researchers from Nanjing University, China, led by Ningmu Zou, have been working on an innovative approach to overcome these major obstacles. Their report, published in Advanced Photonics, explores the application of their newly developed Time-Wavelength Multiplexed Photonic Neural Network Accelerator (TWM-PNNA) to process data from DAS systems.

In Dr. Zou’s words, “This groundbreaking work represents the first successful integration of photonic neural networks with DAS systems that can handle real-time data processing.”

The researchers developed a system architecture that transforms traditional electronic neural network operations into optical processes. Their approach uses multiple tunable lasers emitting light at different wavelengths to represent the neural network’s convolution kernels—the mathematical filters that extract features from input data.

To make this work, they first had to convert two-dimensional data from the DAS systems into one-dimensional vectors that could be encoded onto optical signals using the well-established Mach-Zehnder modulator. The team employed a wavelength-selective switch to assign specific weights to different wavelength channels, effectively implementing the convolution operations using light signals rather than electronic calculations.

The researchers also focused on two major technical challenges: mitigating the effects of modulation chirp (frequency variations) on optical convolutions and developing reliable methods for achieving optical full-connection operations. Through detailed experiments, they found that the ratio of wavelength shift caused by modulation chirp to the wavelength spacing between adjacent laser channels is a key metric for assessing performance impact.

More specifically, when this ratio exceeds 0.1, recognition accuracy is significantly affected. By implementing a technique known as push-pull modulation or by reducing this ratio, the researchers could greatly mitigate the impact of chirp and achieve a classification accuracy above 90%, approaching the 98.3% realized by conventional electrical systems.

Additionally, the researchers discovered that the system maintained its classification accuracy above 90% as long as at least 60% of the full connection parameters were retained after pruning. This finding opens the door to further reducing model size and computational burden without sacrificing performance, making these optical systems less expensive and simpler to produce.

The proposed TWM-PNNA system demonstrated impressive computational capabilities, performing 1.6 trillion operations per second (TOPS) with an energy efficiency of 0.87 TOPS per watt. Theoretically, the system could reach speeds of 81 TOPS with an energy efficiency of 21.02 TOPS per watt, outperforming comparable electrical GPUs by orders of magnitude.

Overall, TWM-PNNA provides a novel computational framework for DAS systems, paving the way for the all-optical fusion of DAS with high-speed computational systems. This research represents a significant step toward next-generation infrastructure monitoring technology, capable of processing massive amounts of sensor data in real-time. With any luck, unlocking the true power of DAS could transform applications in critical infrastructure protection, seismic monitoring, and transportation safety.