我的神经网络模型准确率始终为 50%

Question

训练我的模型后，测试准确率始终为 50%。我下面的代码有什么问题？

0~4000个正常信号数据，4001~6000个异常信号数据进行二分类。 数据维度为(6000, 8000)

data = np.load('data.npy') 
label = []
for i in range(len(data)): ## labeling
    if i < 4000:
        label.append(1)
    else:
        label.append(0)

label = np.array(label)

## each 100 data was extracted for test
test_data =  np.concatenate((data[:100], data[4001:4101]), axis=0)  
test_label = np.concatenate((label[:100], label[4001:4101]), axis=0)
train_data = np.concatenate((data[100:4001], data[4101:]))
train_label = np.concatenate((label[100:4001], label[4101:]))

## data shuffleing
tmp = [[x,y]for x, y in zip(train_data, train_label)]
tmp1 = [[x,y]for x, y in zip(test_data, test_label)]
random.shuffle(tmp)
random.shuffle(tmp1) 
train_data = [n[0] for n in tmp]
train_label = [n[1] for n in tmp]
train_data = np.array(train_data)
train_label = np.array(train_label)
teet_data = [n[0] for n in tmp1]
test_label = [n[1] for n in tmp1]
test_data = np.array(test_data)
test_label = np.array(test_label)

## scaling
mean = train_data.mean(axis=0)
std = train_data.std(axis=0)

train_data -= mean
train_data /= std
test_data -= mean
test_data /= std

model = models.Sequential()
model.add(layers.Dense(128, activation='relu', input_shape=(8000,)))
model.add(layers.Dense(128, activation='relu'))
model.add(layers.Dense(128, activation='relu'))
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(32, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))

model.compile(optimizer='Adam',
             loss='binary_crossentropy',
             metrics=['acc'])

history = model.fit(train_data,
                    train_label,
                    epochs=60,
                    batch_size=128,
                    shuffle=True,
                    validation_split=0.2)

损失曲线

loss, acc = model.evaluate(test_data, test_label)

200/200 [==============================] - 0s 140us/步

print(acc)

0.5

Answer 1

您的模型很可能只为您的测试数据预测一个类别。

这可能是由您的特征缩放方法引起的。您应该根据从训练集中提取的统计数据对测试数据进行标准化。

Answer 2

对于这么多功能，您的模型太弱/太小。就在您的第一层，您通过将 8000 个特征转换为 8 个特征来破坏所有信息！使用更多的单元，远不止于此，让它学习一些东西，而不是破坏你的数据集。您的模型目前无法比随机预测更好。

Answer 3

这是我的信号数据。

import sounddevice as sd
import numpy as np
from math import pi

fs = 4000

n = np.arange(0, 2, 1/fs)

f = 13000 # x 
f1 = 1310 # x1
f2 = 175  # x 2
f3 = 45 # x3
'''
(8000,)
'''

x = np.sin(2*pi*f*n)
x1 = np.sin(2*pi*f1*n)
x2 = np.sin(2*pi*f2*n)
y = np.random.rand(len(x))
x3 = np.sin(2*pi*f3*n)
y = np.random.rand(len(x))

fault =  y*0.2 + (x1+x2 + x3) + 0.15
normal =  y*0.2 +(x1 + x2) +2

y = np.random.rand(len(x))
normal = normal
normal.shape
fault.shape

(8000,)

normal_data=[]
for i in range (4000):
    y = np.random.rand(len(x))
    normal = 2*y*(x1 + x2)
    normal_data.append(normal)

normal_data = np.array(normal_data)
normal_data.shape

(4000, 8000)

fault_data=[]
for i in range (2000):
    y = np.random.rand(len(x))
    fault = 2*y*(x1 + x2)
    fault_data.append(fault)

fault_data = np.array(fault_data)
fault_data.shape

(2000, 8000)

## Final signal data
data = np.concatenate((normal_data, fault_data))
data.shape

(6000, 8000)