【基於深度學習的腦電圖識別】手把手教你使用 1D 卷積和 LSTM 混合模型做 EEG 信號識別

1. 數據集

1.1 數據集下載：

https://archive.ics.uci.edu/ml/datasets/Epileptic+Seizure+Recognition

打開後是這樣的：

點擊 Data Folder，就可以看到保存數據的csv文件，右鍵下載下來：

打開看一下：

1.2 數據集解釋：

表頭爲 X* 的是電信號數據，共有 11500 行，每行有 178 個數據，表示 1s 時間內截取的 178 個電信號；表頭爲 Y 的一列是該時間段數據的標籤，包括 5 個分類：

5-記錄大腦的EEG信號時病人睜開了眼睛；

4-記錄大腦的EEG信號時患者閉上了眼睛；

3-健康大腦區域的腦電圖活動；

2-腫瘤所在區域的腦電圖活動；

1-癲癇發作活動；

2. 讀取數據：

import pandas as pd

data = "data.csv"

df = pd.read_csv(data, header=0, index_col=0)
df1 = df.drop(["y"], axis=1)
lbls = df["y"].values - 1

這裏使用 pandas 庫讀取 data.csv，df1 保存電位數據，lbls 保存標籤；

import numpy as np

wave = np.zeros((11500, 178))

z = 0
for index, row in df1.iterrows():
    wave[z, :] = row
    z+=1

mean = wave.mean(axis=0)
wave -= mean
std = wave.std(axis=0)
wave /= std

def one_hot(y):
    lbl = np.zeros(5)
    lbl[y] = 1
    return lbl

target = []
for value in lbls:
    target.append(one_hot(value))
target = np.array(target)
wave = np.expand_dims(wave, axis=-1)

我們將數據保存在數組 wave 和 target 中，將點位數據標準化（減去均值後除以方差），並將標籤轉換成 one hot 的形式；

3. 搭建模型：

我們使用 keras 搭建一個模型，包括 1D 卷積層和幾個堆疊的 LSTM 層：

from keras.models import Sequential
from keras import layers

model = Sequential()
model.add(layers.Conv1D(64, 15, strides=2,input_shape=(178, 1), use_bias=False))
model.add(layers.ReLU())
model.add(layers.Conv1D(64, 3))
model.add(layers.Conv1D(64, 3, strides=2))
model.add(layers.ReLU())
model.add(layers.Conv1D(64, 3))
model.add(layers.Conv1D(64, 3, strides=2))  # [None, 54, 64]
model.add(layers.BatchNormalization())
model.add(layers.LSTM(64, dropout=0.5, return_sequences=True))
model.add(layers.LSTM(64, dropout=0.5, return_sequences=True))
model.add(layers.LSTM(32))
model.add(layers.Dense(5, activation="softmax"))
model.summary()

網絡結構如圖：

即該模型使用 1D 卷積進行特徵提取，使用 LSTM 進行時域建模，最後通過一個全連接層預測類別；

4. 訓練模型：

我們使用 Adam 優化器，並設置學習率衰減來進行訓練：

import matplotlib.pyplot as plt
import pandas as pd
from keras.models import Sequential
from keras import layers
from keras import regularizers
import os
import keras

import keras.backend as K

save_path = './keras_model.h5'

if os.path.isfile(save_path):
    model.load_weights(save_path)
    print('reloaded.')

adam = keras.optimizers.adam()

model.compile(optimizer=adam,
              loss="categorical_crossentropy", metrics=["acc"])
# 計算學習率
def lr_scheduler(epoch):
    # 每隔100個epoch，學習率減小爲原來的0.5
    if epoch % 100 == 0 and epoch != 0:
        lr = K.get_value(model.optimizer.lr)
        K.set_value(model.optimizer.lr, lr * 0.5)
        print("lr changed to {}".format(lr * 0.5))
    return K.get_value(model.optimizer.lr)

lrate = LearningRateScheduler(lr_scheduler)

history = model.fit(wave, target, epochs=400,
                    batch_size=128, validation_split=0.2,
                    verbose=1, callbacks=[lrate])

model.save_weights(save_path)

這樣就可以開始訓練啦：

訓練的模型參數保存在 sace_path 中；

5. 展示結果：

print(history.history.keys())
# summarize history for accuracy
plt.plot(history.history['acc'])
plt.plot(history.history['val_acc'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()
# summarize history for loss
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()

這時我們可以查看訓練結果（因爲時間有限，我只訓練了 100 個 epoch：

6. 完整代碼：

import matplotlib.pyplot as plt
import pandas as pd
from keras.models import Sequential
from keras import layers
from keras import regularizers
import os
import keras

import keras.backend as K

import numpy as np

from keras.callbacks import LearningRateScheduler

data = "data.csv"

df = pd.read_csv(data, header=0, index_col=0)
df1 = df.drop(["y"], axis=1)
lbls = df["y"].values - 1

wave = np.zeros((11500, 178))

z = 0
for index, row in df1.iterrows():
    wave[z, :] = row
    z+=1

mean = wave.mean(axis=0)
wave -= mean
std = wave.std(axis=0)
wave /= std

def one_hot(y):
    lbl = np.zeros(5)
    lbl[y] = 1
    return lbl

target = []
for value in lbls:
    target.append(one_hot(value))
target = np.array(target)
wave = np.expand_dims(wave, axis=-1)

model = Sequential()
model.add(layers.Conv1D(64, 15, strides=2,
                        input_shape=(178, 1), use_bias=False))
model.add(layers.ReLU())
model.add(layers.Conv1D(64, 3))
model.add(layers.Conv1D(64, 3, strides=2))
model.add(layers.BatchNormalization())
model.add(layers.Dropout(0.5))
model.add(layers.Conv1D(64, 3))
model.add(layers.Conv1D(64, 3, strides=2))
model.add(layers.BatchNormalization())
model.add(layers.LSTM(64, dropout=0.5, return_sequences=True))
model.add(layers.LSTM(64, dropout=0.5, return_sequences=True))
model.add(layers.LSTM(32))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(5, activation="softmax"))
model.summary()

save_path = './keras_model3.h5'

if os.path.isfile(save_path):
    model.load_weights(save_path)
    print('reloaded.')

adam = keras.optimizers.adam()

model.compile(optimizer=adam,
              loss="categorical_crossentropy", metrics=["acc"])
# 計算學習率
def lr_scheduler(epoch):
    # 每隔100個epoch，學習率減小爲原來的0.5
    if epoch % 100 == 0 and epoch != 0:
        lr = K.get_value(model.optimizer.lr)
        K.set_value(model.optimizer.lr, lr * 0.5)
        print("lr changed to {}".format(lr * 0.5))
    return K.get_value(model.optimizer.lr)

lrate = LearningRateScheduler(lr_scheduler)

history = model.fit(wave, target, epochs=400,
                    batch_size=128, validation_split=0.2,
                    verbose=2, callbacks=[lrate])

model.save_weights(save_path)

print(history.history.keys())
# summarize history for accuracy
plt.plot(history.history['acc'])
plt.plot(history.history['val_acc'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()
# summarize history for loss
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()