目录
1 概述
1.1 完备集合经验模态分解原理
1.2 鲸鱼优化
1.3 LSTM
2 运行结果
3 参考文献
4 Python代码实现
早期的 EMD 方法具有较强的自适应性,能够有效地分解时间序列;但是,算法在运算过程中
容易出现模态混叠现象。EEMD 分解方法的思想是:在原始信号中加入白噪声[16],使极值点分布更均衡;最终分量在EMD 的基础上进行集成平均而得。但是,这种方法具有计算量大且重构时残留噪音大的缺陷。CEEMDAN 是 EEMD 的改进算法。该算法通过添加有限次数的自适应白噪声,解决了集合平均次数限制下的重构误差较大的问题。
座头鲸有特殊的捕猎方法,这种觅食行为被称为泡泡网觅食法;标准 WOA 模拟了座头鲸特有的搜索方法和围捕机制,主要包括:围捕猎物、气泡网捕食、搜索猎物三个重要阶段。WOA 中每个座头鲸的位置代表一个潜在解,通过在解空间中不断更新鲸鱼的位置,最终获得全局最优解。
长短时记忆( long-short term memory,LSTM) 神经网络是 Hochreiter 等提出的一种改进后的循环式神经网络,可有效解决循环式神经网络存在的梯度爆炸和阶段性梯度消失的问题。在传统
循环式神经网络基础上,在隐含层增设记忆模块,可使信息较长时间地储存和遗传,其结构如图 1
所示。
......
Epoch 87/100
19/19 [==============================] - 0s 5ms/step - loss: 1.2908e-04 - accuracy: 5.3677e-04 - val_loss: 9.1420e-06 - val_accuracy: 0.0000e+00
Epoch 88/100
19/19 [==============================] - 0s 5ms/step - loss: 1.3659e-04 - accuracy: 5.3677e-04 - val_loss: 2.2255e-06 - val_accuracy: 0.0000e+00
Epoch 89/100
19/19 [==============================] - 0s 5ms/step - loss: 1.1987e-04 - accuracy: 5.3677e-04 - val_loss: 3.4974e-05 - val_accuracy: 0.0000e+00
Epoch 90/100
19/19 [==============================] - 0s 5ms/step - loss: 1.2746e-04 - accuracy: 5.3677e-04 - val_loss: 9.6258e-05 - val_accuracy: 0.0000e+00
Epoch 91/100
19/19 [==============================] - 0s 5ms/step - loss: 1.2758e-04 - accuracy: 5.3677e-04 - val_loss: 9.1996e-05 - val_accuracy: 0.0000e+00
Epoch 92/100
19/19 [==============================] - 0s 5ms/step - loss: 1.5623e-04 - accuracy: 5.3677e-04 - val_loss: 1.8761e-05 - val_accuracy: 0.0000e+00
Epoch 93/100
19/19 [==============================] - 0s 6ms/step - loss: 1.4421e-04 - accuracy: 5.3677e-04 - val_loss: 3.0035e-06 - val_accuracy: 0.0000e+00
Epoch 94/100
19/19 [==============================] - 0s 5ms/step - loss: 1.4949e-04 - accuracy: 5.3677e-04 - val_loss: 2.6891e-04 - val_accuracy: 0.0000e+00
Epoch 95/100
19/19 [==============================] - 0s 5ms/step - loss: 1.2961e-04 - accuracy: 5.3677e-04 - val_loss: 2.1525e-05 - val_accuracy: 0.0000e+00
Epoch 96/100
19/19 [==============================] - 0s 5ms/step - loss: 1.2142e-04 - accuracy: 5.3677e-04 - val_loss: 3.6751e-05 - val_accuracy: 0.0000e+00
Epoch 97/100
19/19 [==============================] - 0s 5ms/step - loss: 1.3616e-04 - accuracy: 5.3677e-04 - val_loss: 8.5641e-07 - val_accuracy: 0.0000e+00
Epoch 98/100
19/19 [==============================] - 0s 6ms/step - loss: 1.2854e-04 - accuracy: 5.3677e-04 - val_loss: 1.4613e-04 - val_accuracy: 0.0000e+00
Epoch 99/100
19/19 [==============================] - 0s 5ms/step - loss: 1.4222e-04 - accuracy: 5.3677e-04 - val_loss: 1.1871e-04 - val_accuracy: 0.0000e+00
Epoch 100/100
19/19 [==============================] - 0s 6ms/step - loss: 1.7137e-04 - accuracy: 5.3677e-04 - val_loss: 2.4105e-06 - val_accuracy: 0.0000e+00
65/65 [==============================] - 0s 1ms/step
16/16 [==============================] - 0s 1ms/step
部分理论来源于网络,如有侵权请联系删除。
[1]金子皓,向玲,李林春,胡爱军.基于完备集合经验模态分解的SE-BiGRU超短期风速预测[J].电力科学与工程,2023,39(01):9-16.
[2]蒋富康,陆金桂,刘明昊,丰宇.基于CEEMDAN和CNN-LSTM的滚动轴承故障诊断[J].电子测量技术,2023,46(05):72-77.DOI:10.19651/j.cnki.emt.2210775.