基于深度学习的大规模光纤激光相干合成相位控制的异常检测

李国豪; 顾静良; 唐乾轲; 李正东; 颜宏; 王锋

doi:10.11884/HPLPB202537.250019

基于深度学习的大规模光纤激光相干合成相位控制的异常检测

doi: 10.11884/HPLPB202537.250019

李国豪^{1, 2,},
顾静良²,
唐乾轲²,
李正东²,
颜宏^{1, 2, 3, ,},
王锋^2, ,

1.
中国工程物理研究院研究生院，北京 100088
2.
中国工程物理研究院应用电子学研究所，四川绵阳 621900
3.
中国工程物理研究院高能激光科学与技术重点实验室，四川绵阳 621900

详细信息

作者简介:
李国豪，liguohao22@gscaep.ac.cn

通讯作者:
颜　宏，yanhong@caep.cn

王　锋，yuan13381338@163.com。

中图分类号: TP391.4
计量
- 文章访问数: 41
- HTML全文浏览量: 12
- PDF下载量: 11
- 被引次数: 0
出版历程
- 收稿日期: 2025-01-21
- 修回日期: 2025-07-05
- 录用日期: 2025-06-25
- 网络出版日期: 2025-09-24

Anomaly detection for phase control of large-scale fiber laser coherent combination based on deep learning

Li Guohao^{1, 2
,},
Gu Jingliang²,
Tang Qianke²,
Li Zhengdong²,
Yan Hong^{1, 2, 3
, ,},
Wang Feng^{2
, ,}

1.
Graduate School of China Academy of Engineering Physics, Beijing 100088, China
2.
Institute of Applied Electronics, CAEP, Mianyang 621900, China
3.
The key Laboratory of Science and Technology on High Energy Laser, CAEP, Mianyang 621900, China:

摘要

摘要: 光纤激光相干合成技术通过精确控制各路光纤激光的相位，实现高功率的激光输出。然而，系统运行中存在多种影响因素，如相位控制精度、光强稳定性、通信链路可靠性以及环境干扰等，这些因素可能导致系统性能下降。针对大规模光纤激光相干合成相位控制中的异常检测问题，提出一种基于深度学习的多探测器串行共孔径相干合成检测新方法。首先，采集十路光纤激光相干合成数据，分析系统控制过程及其合束原理，归类系统中可能出现的异常情况，并仿真得到数据集。其次，设计一种结合轻量化高效多头注意力机制（EMA）的EMA-Transformer网络模型。在对比实验中，本算法相较于ResNet50，在验证集上的精度提升了约50%，在测试集上的精度提升了约2.20%。在算法的实际应用中，搭建八束光纤激光相干合成实验装置，使用TensorRT部署算法进行测试。实验结果表明，本算法推理耗时达2.153 ms，达到了相位控制异常检测的实时性要求。
- 深度学习 /
- 相干合成 /
- 异常检测 /
- 多头注意力机制
Abstract: Background
Fiber laser coherent beam combining technology enables high-power laser output through precise phase control of multiple laser channels. However, factors such as phase control accuracy, optical intensity stability, communication link reliability, and environmental interference can degrade system performance.
Purpose
This study aims to address the challenge of anomaly detection in phase control for large-scale fiber laser coherent combining by proposing a novel deep learning-based detection method.
Methods
First, ten-channel fiber laser coherent combining data were collected, system control processes and beam combining principles were analyzed, and potential anomalies were categorized to generate a simulated dataset. Subsequently, an EMA-Transformer network model incorporating a lightweight Efficient Multi-head Attention (EMA) mechanism was designed. Comparative experiments were conducted to evaluate the model's performance. Finally, an eight-beam fiber laser coherent combining experimental setup was established, and the algorithm was deployed using TensorRT for real-time testing.
Results
The proposed algorithm demonstrated significant improvements, achieving approximately 50% higher accuracy on the validation set and a 2.20% enhancement on the test set compared to ResNet50. In practical testing, the algorithm achieved an inference time of 2.153 ms, meeting real-time requirements for phase control anomaly detection.
Conclusions
The EMA-Transformer model effectively addresses anomaly detection in fiber laser coherent combining systems, offering superior accuracy and real-time performance. This method provides a promising solution for enhancing the stability and reliability of high-power laser systems.
- deep learning /
- coherent combination /
- anomaly detection /
- multi-head attention mechanism

HTML全文

图 1 大规模光纤激光共孔径相干合成系统光路

Figure 1. Large-scale fiber laser common-aperture coherent combining system optical path

下载: 全尺寸图片幻灯片

图 2 开环和闭环效果正常情况的电压平均值、电压最大值以及电压均方根值

Figure 2. Average, maximum, and RMS values for open and closed loop cases

下载: 全尺寸图片幻灯片

图 3 Transformer 的总体架构

Figure 3. General architecture of Transformer

下载: 全尺寸图片幻灯片

图 4 EMA 的总体架构

Figure 4. General architecture of EMA

下载: 全尺寸图片幻灯片

图 5 Transformer+EMA 的总体架构

Figure 5. General architecture of Transformer+EMA

下载: 全尺寸图片幻灯片

图 6 Transformer模型的训练损失曲线和精度曲线

Figure 6. Training loss curves and accuracy curves for transformer models

下载: 全尺寸图片幻灯片

图 7 Transformer+EMA模型的训练损失曲线和精度误差曲线

Figure 7. Training Loss Curves and Accuracy Curves for Transformer+EMA Models

下载: 全尺寸图片幻灯片

图 8 Xception和ResNet50的网络结构图

Figure 8. The network structure diagrams of Xception and ResNet50

下载: 全尺寸图片幻灯片

图 9 光纤激光相干合成系统信号处理和控制流程

Figure 9. Signal processing and control flow of fibre laser coherent synthesis system

下载: 全尺寸图片幻灯片

表 1 各个模型的性能指标统计

Table 1. Performance metrics statistics for each model

model	validation error	test error	train time/s
Xception	0.0185	0.0176	400
Xception+EMA	0.0150	0.0135	432
ResNet50	0.0107	0.0091	621
ResNet50+EMA	0.0076	0.0079	701
Transformer	0.0103	0.0122	624
Transformer+EMA	0.0061	0.0089	668

下载: 导出CSV

表 2 各个方法的性能指标统计

Table 2. Performance metrics statistics for each method

	accuracy /%	inference time/ms
Pytorch	98.52	4.314
ONNX	98.64	4.145
TensorRT fp16	97.53	2.153

下载: 导出CSV

参考文献(20)

[1]	McNaught S J, Komine H, Weiss S B, et al. 100 kW coherently combined slab MOPAs[C]//Proceedings of the 2009 Conference on Lasers and Electro-Optics and 2009 Conference on Quantum electronics and Laser Science Conference. 2009: 1-2.
[2]	Uberna R, Bratcher A, Tiemann B G. Power scaling of a fiber master oscillator power amplifier system using a coherent polarization beam combination[J]. Applied Optics, 2010, 49(35): 6762-6765. doi: 10.1364/AO.49.006762
[3]	刘泽金, 周朴, 马鹏飞, 等. 4路高功率窄线宽、线偏振光纤放大器相干偏振合成实现5kW级高亮度激光输出[J]. 中国激光, 2017, 44: 0415004 doi: 10.3788/CJL201744.0415004 Liu Zejin, Zhou Pu, Ma Pengfei, et al. Coherent polarization synthesis of four high-power narrow-linewidth, linearly polarized fiber amplifiers to achieve 5 kW high-brightness laser output[J]. Chinese Journal of Lasers, 2017, 44: 0415004 doi: 10.3788/CJL201744.0415004
[4]	张嘉怡, 任博, 李灿, 等. 八路超快光纤激光高效相干偏振合成[J]. 中国激光, 2024, 51: 1516001 doi: 10.3788/CJL240842 Zhang Jiayi, Ren Bo, Li Can, et al. Eight-channel ultrafast fiber lasers achieve efficient coherent polarization beam combination[J]. Chinese Journal of Lasers, 2024, 51: 1516001 doi: 10.3788/CJL240842
[5]	周宏冰, 张昊宇, 李敏, 等. 大规模激光相干合成主动相位控制技术研究进展[J]. 强激光与粒子束, 2024, 36: 061001 Zhou Hongbing, Zhang Haoyu, Li Min, et al. Progress in active phase control for large-scale coherent laser beam combining[J]. High Power Laser and Particle Beams, 2024, 36: 061001
[6]	常琦, 高志强, 邓宇, 等. 强噪声下光纤激光相干合成突破千路[J]. 中国激光, 2023, 50: 0616001 doi: 10.3788/CJL221540 Chang Qi, Gao Zhiqiang, Deng Yu, et al. Fiber laser coherent beam combination breaks through 1000 channels under strong noise[J]. Chinese Journal of Lasers, 2023, 50: 0616001 doi: 10.3788/CJL221540
[7]	Goodno G D, Shih C C, Rothenberg J E. Perturbative analysis of coherent combining efficiency with mismatched lasers[J]. Optics Express, 2010, 18(24): 25403-25414. doi: 10.1364/OE.18.025403
[8]	颜宏, 叶一东, 卢飞, 等. 基于偏振鉴相的相干合成技术[J]. 强激光与粒子束, 2013, 25(s1): 5-8 Yan Hong, Ye Yidong, Lu Fei, et al. Coherent beam combining based on polarization phase discrimination[J]. High Power Laser and Particle Beams, 2013, 25(s1): 5-8
[9]	李彬, 叶一东, 颜宏, 等. 激光场相干叠加光束合成技术[J]. 强激光与粒子束, 2014, 26: 101023 doi: 10.3788/HPLPB20142610.101023 Li Bin, Ye Yidong, Yan Hong, et al. Technology of filled-aperture coherent beam combining[J]. High Power Laser and Particle Beams, 2014, 26: 101023 doi: 10.3788/HPLPB20142610.101023
[10]	刘家盈, 李自强, 杨然, 等. 光纤激光相控阵相干合成技术研究进展[J]. 强激光与粒子束, 2023, 35: 041003 Liu Jiaying, Li Ziqiang, Yang Ran, et al. Research progress of coherent beam combining technique of phased fiber laser array[J]. High Power Laser and Particle Beams, 2023, 35: 041003
[11]	Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. 2017: 6000-6010.
[12]	Al-Hardanee O F, Demirel H. Hydropower station status prediction using RNN and LSTM algorithms for fault detection[J]. Energies, 2024, 17: 5599. doi: 10.3390/en17225599
[13]	Chen Haipeng, Yan Jie, Fu Wenxing. Two-layer fault diagnosis model of aircraft based on LSTM[J]. Aerospace Science and Technology, 2025, 158: 109756. doi: 10.1016/j.ast.2024.109756
[14]	Ouyang Daliang, He Su, Zhan Guozhong, et al. Efficient multi-scale attention module with cross-spatial learning[C]//Proceedings of the ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 2023: 1-5.
[15]	Chollet F. Xception: deep learning with depthwise separable convolutions[C]//Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017: 1800-1807.
[16]	He Kaiming, Zhang Xiangyu, Ren Shaoqing, et al. Deep residual learning for image recognition[C]//Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016: 770-778.
[17]	Yang Xu, Qu Shuo’ou, Wang Zhilin, et al. The study on ultrasound image classification using a dual-branch model based on Resnet50 guided by U-net segmentation results[J]. BMC Medical Imaging, 2024, 24: 314. doi: 10.1186/s12880-024-01486-z
[18]	张海瑞. 基于ONNX与PyTorch框架的神经网络格式转换研究[D]. 西安: 西安电子科技大学, 2023 Zhang Hairui. Research on the conversion of neural network formats using ONNX and PyTorch frameworks[D]. Xi’an: Xidian University, 2023
[19]	Shafique M A, Munir A, Kong J. Deep learning performance characterization on GPUs for various quantization frameworks[J]. AI, 2023, 4(4): 926-948. doi: 10.3390/ai4040047
[20]	Zhang Shicheng, Zhang Laixian, Qin Mingyu, et al. TensorRT acceleration based on deep learning photoelectric target detection[C]//Proceedings of the International Conference on Intelligent and Human-Computer Interaction Technology (IHCIT 2022). 2022.