Fig. 1. (a) Setup of DAS system; (b) raw signals collected by the system; (c) spatiotemporal maps of various events after denoising and quantization. NLL, narrow linewidth laser; OC, optical coupler; AOM, acousto-optic modulator; DAQ, digital acquisition card; Cir, circulator; EDFA, erbium-doped fiber amplifier; BPD, balanced photodetector; PG, pulse generator; BPF, bandpass filter; LNA, low-noise amplifier; DM, demodulation module.

Fig. 2. Photonic neural network accelerator.

Fig. 3. Experimental setup of the TWM-PNNA. AWG, arbitrary waveform generator; MZM, Mach–Zehnder modulator; WSS, wavelength-selective switch; OSW, optical switch; WDM, wavelength-division multiplexer; PD, photodetector.

Fig. 4. Results of convolution operations in the TWM-PNNA. (a1)–(a10) 1D feature maps after optical convolution; (b1)–(b10) 10 different convolution kernels; (c1)–(c10) restored 2D feature maps.

Fig. 5. Classification accuracy under different image sizes. (a) Electrical neural networks and optical neural networks in binary images; (b) electrical neural networks in binary images and nonbinary images.

Fig. 6. Comparison of chirp effects. (a) Without chirp. (b) With chirp.

Fig. 7. Confusion matrix for the DAS classification task with image size of 36×36. (a) Electrical CNN confusion matrix; (b)–(e) photonic convolution kernel confusion matrices; (b) push–pull modulation with wavelength spacing of 2 nm; (c) single-drive modulation with wavelength spacing of 2 nm; (d) push–pull modulation with wavelength spacing of 0.4 nm; (e) single-drive modulation with wavelength spacing of 0.4 nm.

Fig. 8. Relationship between wavelength spacing and classification accuracy under different modulation schemes.

Fig. 9. Pruning scheme in the TWM-PNNA.

Fig. 10. Pruning results in TWM-PNNA.