# TSC-CNN
基于一维卷积神经网络（1D-CNN）的多元时间序列分类
## 项目背景
该项目为基于一维卷积神经网络的多元时间序列分类方法，实际问题被抽象为时间序列的分类问题，实际输入为4个传感器信号，分别对应16个类别，搭建1D-CNN然后训练网络对多元时间序列进行分类。
## 1D-CNN
无论是一维、二维还是三维，卷积神经网络（CNNs）都具有相同的特点和相同的处理方法。关键区别在于输入数据的维数以及特征检测器（或滤波器）如何在数据之间滑动，一维和二维CNN处理过程对比。
## 网络搭建
首先分析网络的输入输出，输入为包含4个时间序列的信号，长度为1050，即输入为（1050，4），而输出对应16种类别，输出为（，16）。实质上输入时先将单个训练数据的4个时间序列展平为（4200，1），传入网络后再reshape为（1050，4）。网络构建代码如下：
```python
model = Sequential()
model.add(Reshape((TIME_PERIODS, num_sensors), input_shape=(TIME_PERIODS*num_sensors,)))
model.add(Conv1D(16, 8,strides=2, activation='relu',input_shape=(TIME_PERIODS,num_sensors)))
model.add(Conv1D(16, 8,strides=2, activation='relu',padding="same"))
model.add(MaxPooling1D(2))
model.add(Conv1D(64, 4,strides=2, activation='relu',padding="same"))
model.add(Conv1D(64, 4,strides=2, activation='relu',padding="same"))
model.add(MaxPooling1D(2))
model.add(Conv1D(256, 4,strides=2, activation='relu',padding="same"))
model.add(Conv1D(256, 4,strides=2, activation='relu',padding="same"))
model.add(MaxPooling1D(2))
model.add(Conv1D(512, 2,strides=1, activation='relu',padding="same"))
model.add(Conv1D(512, 2,strides=1, activation='relu',padding="same"))
model.add(MaxPooling1D(2))
model.add(GlobalAveragePooling1D())
model.add(Dropout(0.3))
model.add(Dense(num_classes, activation='softmax'))
```
运行得到网络结构如下:
```
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
reshape_1 (Reshape) (None, 1050, 4) 0
_________________________________________________________________
conv1d_1 (Conv1D) (None, 522, 16) 528
_________________________________________________________________
conv1d_2 (Conv1D) (None, 261, 16) 2064
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 130, 16) 0
_________________________________________________________________
conv1d_3 (Conv1D) (None, 65, 64) 4160
_________________________________________________________________
conv1d_4 (Conv1D) (None, 33, 64) 16448
_________________________________________________________________
max_pooling1d_2 (MaxPooling1 (None, 16, 64) 0
_________________________________________________________________
conv1d_5 (Conv1D) (None, 8, 256) 65792
_________________________________________________________________
conv1d_6 (Conv1D) (None, 4, 256) 262400
_________________________________________________________________
max_pooling1d_3 (MaxPooling1 (None, 2, 256) 0
_________________________________________________________________
conv1d_7 (Conv1D) (None, 2, 512) 262656
_________________________________________________________________
conv1d_8 (Conv1D) (None, 2, 512) 524800
_________________________________________________________________
max_pooling1d_4 (MaxPooling1 (None, 1, 512) 0
_________________________________________________________________
global_average_pooling1d_1 ( (None, 512) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 512) 0
_________________________________________________________________
dense_1 (Dense) (None, 16) 8208
=================================================================
Total params: 1,147,056
Trainable params: 1,147,056
Non-trainable params: 0
_________________________________________________________________
None
```
