Table 1. Absolute error of PINNs and G-PINNs when network configuration points are differentTable 1. Absolute error of PINNs and G-PINNs when network configuration points are different 表1. 网络配置点不同时PINNs与G-PINNs训练的绝对误差值
4.2. 弱噪声下PINNs与G-PINNs对Burgers方程的求解结果实验对PINNs与G-PINNs添加5%服从标准正态分布的噪声,模拟实际应用中各种随机因素对数据产生的噪声干扰,如测量误差、环境干扰等,使实验结果更贴近真实情况 [25] 。在 的计算域内采用LHS选取100个边界点与10,000个内部点训练至收敛,在整个计算域内对PINNs与G-PINNs的预测结果和绝对误差进行比较分析。通过对比两种网络架构在实际弱噪声干扰情况下的表现,评估噪声引入后模型的稳健性及其在噪声环境下的性能 [26] 。为了更详细的了解结果,仍选取t = 0.5时刻PINNs和G-PINNs的预测结果和误差值进行单独的展示分析,得到的PINNs和G-PINNs的训练结果如 图5 和 图6 所示。Figure 5. Training results of PINNs with 5% noise--图5. 添加5%弱噪声后PINNs的训练结果图--结果显示,与无噪声训练相比,PINNs和G-PINNs模型在包含噪声的训练条件下最终预测解误差均呈现出一定程度的增加,该现象反映了噪声对模型预测精度的不利影响。但是G-PINNs模型在相同条件下的误差相对于PINNs模型在全域范围内以及t = 0.5时刻均显著降低了近20%。这一结果表明,在5%噪声率的情况下G-PINNs模型在处理噪声数据方面具有更强的鲁棒性,能够更有效地抑制噪声带来的不利影响,从而在噪声存在的条件下提供更为准确的预测结果。[-rId167-]
Figure 6. Training results of G-PINNs with 5% noise
图6. 添加5%弱噪声后G-PINNs的训练结果图
表2 列出了添加5%弱噪声后,PINNs与G-PINNs在不同数据配置条件下迭代3000、6000、9000次后绝对误差值。 表2 中的误差结果表明,在受到弱噪声干扰的情况下,G-PINNs模型寻找最优解的能力仍然优于PINNs模型。由 表2 数据可看出PINNs处理噪声数据时表现出较低的稳定性和准确度,而G-PINNs引入方程导数的残差增强了模型对物理过程的拟合精度,提高了模型在面对噪声数据时的鲁棒性。Table 2. Absolute error of PINNs and G-PINNs when network configuration points are different (with 5% noise)Table 2. Absolute error of PINNs and G-PINNs when network configuration points are different (with 5% noise) 表2. 网络配置点不同时PINNs与G-PINNs训练的绝对误差值(5%噪声)
3000 (Iters)
6000 (Iters)
9000 (Iters)
Error of PINNs
(Nu= 100, Nf= 5000)
1.869648*10−1
5.516061*10−2
4.288181*10−2
Error of G-PINNs
(Nu= 100, Nf= 5000)
5.468906*10−2
4.320308*10−2
3.192474*10−2
Error of PINNs
(Nu= 100, Nf= 10,000)
1.015703*10−1
4.789887*10−2
3.911240*10−2
Error of G-PINNs
Nu= 100, Nf= 10,000)
3.728751*10−2
2.218031*10−2
1.591892*10−2
实际流体力学场景复杂,难以获取完备且充分的数据集,数据集的准确度及质量也较难达到高水平。综合考虑数据集引入噪声后G-PINNs的表现可以得出,G-PINNs投入更多的计算成本换取模型在不确定性噪声环境下的稳健性和卓越性能,其模型对数据噪声的容忍度高于PINNs。G-PINNs模型的优良表现可大幅减少对数据集的依赖,降低实际数据采集过程中的成本及难度,为复杂流体力学场景的应用提供强有力的理论支持,而且在真实应用环境中也具有显著的求解潜力。本实验中对数据集分别引入1%、3%和5%服从标准正态分布的噪声,模拟PINNs与G-PINNs在不同噪声情况下的逼近能力。计算域内选取100个边界点与10,000个内部点进行弱噪声下的实验,并记录网络输出的预测解的误差值,结果如 图7 所示。Figure 7. Error of PINNs and G-PINNs with different noises--图7. 不同噪声下PINNs与G-PINNs的误差--针对Burgers方程数值求解的过程,以跳出循环后最终的相对误差值作为评估PINNs与G-PINNs两种网络架构表现力的依据。由 图7 可观察出在无噪声情况下PINNs与G-PINNs最终的相对误差存在显著差异,表明两者在求解精度上具有较大的性能差距。当引入数据噪声后,PINNs和G-PINNs预测解的误差均随着噪声强度的增加而增加,但在弱噪声环境下G-PINNs的误差始终保持在较低水平且明显优于PINNs。G-PINNs通过引入导数的残差加强对梯度的约束,使得网络在训练过程中不仅关注函数值的准确性,还考虑到导数信息的准确性,可以有效减少数值解在导数上的震荡,从而提高整体解的平滑度和精确度。传统PINNs缺乏对梯度的强约束从而导致求解精度的局限性,而G-PINNs则能够更好地捕捉解的物理特性,为数值求解偏微分方程提供了一种更为有效和鲁棒的方法。然而对于神经网络来说,在算力和算法都基本已经确定的前提下,数据的数量与质量是决定深度学习任务最终效果的关键因素。数据集噪声过多直接降低了数据的质量,对网络训练不利,最终的效果也随之不理想。当数据集中的噪声增大到一定程度时,两种网络结构的训练效果均变得不理想,最终误差都显著增加,性能差距也逐渐减小。因此,在数据集中噪声过多的情况下,网络能力的对比不再具有明显的实质性意义。5. 结论References
Alwehebi, F., Hobiny, A. and Maturi, D. (2023) Variational Iteration Method for Solving Time Fractional Burgers Equation Using Maple. Applied Mathematics, 14, 336-348. >https://doi.org/10.4236/am.2023.145021
Ohkitani, K. and Dowker, M. (2012) Numerical Study on Comparison of Navier-Stokes and Burgers Equations. Physics of Fluids, 24, Article ID: 055113. >https://doi.org/10.1063/1.4719787
Sirignano, J. and Spiliopoulos, K. (2018) DGM: A Deep Learning Algorithm for Solving Partial Differential Equations. Journal of Computational Physics, 375, 1339-1364. >https://doi.org/10.1016/j.jcp.2018.08.029
Xu, H., Zhang, D. and Zeng, J. (2021) Deep-Learning of Parametric Partial Differential Equations from Sparse and Noisy Data. Physics of Fluids, 33, Article ID: 037132. >https://doi.org/10.1063/5.0042868
韩忠华. Kriging模型及代理优化算法研究进展[J]. 航空学报, 2016, 37(11): 3197-3225.
孙斌, 姚海涛, 刘婷. 基于高斯过程回归的短期风速预测[J]. 中国电机工程学报, 2012, 32(29): 104-109.
Roska, T., Chua, L.O., Wolf, D., Kozek, T., Tetzlaff, R. and Puffer, F. (1995) Simulating Nonlinear Waves and Partial Differential Equations via CNN. I. Basic Techniques. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 42, 807-815. >https://doi.org/10.1109/81.473590
Long, Z., Lu, Y., Ma, X. and Dong, B. (2018) PDE-Net: Learning PDEs from Data. International Conference on Machine Learning (PMLR), Stockholm, 10-15 July 2018, 3208-3216.
Long, Z., Lu, Y. and Dong, B. (2019) PDE-Net 2.0: Learning PDEs from Data with a numeric-Symbolic Hybrid Deep Network. Journal of Computational Physics, 399, Article ID: 108925. >https://doi.org/10.1016/j.jcp.2019.108925
Regazzoni, F., Dede, L. and Quarteroni, A. (2019) Machine Learning for Fast and Reliable Solution of Time-Dependent Differential Equations. Journal of Computational Physics, 397, Article ID: 108852. >https://doi.org/10.1016/j.jcp.2019.07.050
Schaeffer, H. and McCalla, S.G. (2017) Sparse Model Selection via Integral Terms. Physical Review E, 96, Article ID: 023302. >https://doi.org/10.1103/PhysRevE.96.023302
陈庆强, 王文剑, 姜高霞. 基于数据分布的标签噪声过滤[J]. 清华大学学报(自然科学版), 2019, 59(4): 262-269.
陈倩, 杨旻, 魏鹏飞. 标签带噪声数据的重加权半监督分类方法[J]. 烟台大学学报(自然科学与工程版), 2019, 32(3): 205-209.
Raissi, M., Perdikaris, P. and Karniadakis, G.E. (2019) Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward and Inverse Problems Involving Nonlinear Partial Differential Equations. Journal of Computational Physics, 378, 686-707. >https://doi.org/10.1016/j.jcp.2018.10.045
Bai, G., Koley, U., Mishra, S. and Molinaro, R. (2021) Physics Informed Neural Networks (PINNs) for Approximating Nonlinear Dispersive PDEs. >https://doi.org/10.4208/jcm.2101-m2020-0342
唐明健, 唐和生. 基于物理信息的深度学习求解矩形薄板力学正反问题[J]. 计算力学学报, 2022, 39(1): 120-128.
Cai, S., Mao, Z., Wang, Z., Yin, M. and Karniadakis, G.E. (2021) Physics-Informed Neural Networks (PINNs) for Fluid Mechanics: A Review. Acta Mechanica Sinica, 37, 1727-1738. >https://doi.org/10.1007/s10409-021-01148-1
Rasht‐Behesht, M., Huber, C., Shukla, K. and Karniadakis, G.E. (2022) Physics‐Informed Neural Networks (PINNs) for Wave Propagation and Full Waveform Inversions. Journal of Geophysical Research: Solid Earth, 127, e2021JB023120. >https://doi.org/10.1029/2021JB023120
Burgers, J.M. (1948) A Mathematical Model Illustrating the Theory of Turbulence. Advances in Applied Mechanics, 1, 171-199. >https://doi.org/10.1016/S0065-2156(08)70100-5
Hopf, E. (1950) The Partial Differential Equation ut + uux = μxx. Communications on Pure and Applied mathematics, 3, 201-230. >https://doi.org/10.1002/cpa.3160030302
Meng, X., Yang, L., Mao, Z., Águila Ferrandis, J. and Karniadakis, G.E. (2022) Learning Functional Priors and Posteriors from Data and Physics. Journal of Computational Physics, 457, Article ID: 111073. >https://doi.org/10.1016/j.jcp.2022.111073
Vadyala, S.R., Betgeri, S.N. and Betgeri, N.P. (2022) Physics-Informed Neural Network Method for Solving One-Dimensional Advection Equation Using PyTorch. Array, 13, Article ID: 100110. >https://doi.org/10.1016/j.array.2021.100110
刘金国, 许开来. 自动微分及其在物理模拟中的应用[J]. 物理学报, 2021, 70(14): 97-107.
张钰, 刘建伟, 左信. 多任务学习[J]. 计算机学报, 2020, 43(7): 1340-1378.
Knowles, I. and Renka, R.J. (2014) Methods for Numerical Differentiation of Noisy Data. Electronic Journal of Differential Equations, 21, 235-246.
蒲削锟. 标签噪声下的故障诊断算法研究[D]: [硕士学位论文]. 杭州: 浙江大学, 2023: 27-39.