基于深度学习的无线电调制识别系统

基于深度学习的无线电调制识别系统

本项目实现了一个基于深度学习的无线电调制识别系统，使用LSTM（长短期记忆网络）模型对不同类型的
无线电信号进行自动分类识别。该系统能够在不同信噪比(SNR)条件下，准确识别多种调制类型，如BPSK、
QPSK、QAM16等。无线电调制识别是认知无线电、频谱监测和信号情报等领域的关键技术。传统方法依赖
专家设计的特征提取，而深度学习方法可以自动学习信号特征，提高识别准确率和鲁棒性。

数据集：本项目使用RadioML2016.10a数据集，这是一个广泛用于无线电调制识别研究的标准数据集。
数据集特点：包含10种调制类型，20种信噪比条件（从-20dB到18dB）
信号特征：每个样本包含128个时间步长，每个时间步长有I/Q两个通道
数据格式：输入形状为[样本数, 128, 2]
数据分割：数据集被分为训练集、验证集和测试集
在本项目中，选择以下调制类型进行识别：
GFSK (2FSK)、PAM4 (2ASK)、BPSK、QPSK、QAM16、QAM64、CPFSK、8PSK
模型架构
本项目使用了两层LSTM网络结构，具体如下：
CopyInsert
输入层: [128, 2]（128个时间步长，每步2个特征：I/Q两个通道）
LSTM层1: 128个单元，return_sequences=True
LSTM层2: 128个单元
全连接层: 神经元数量等于调制类型数量
Softmax激活: 输出各调制类型的概率
LSTM模型的优势在于能够捕捉信号中的时序特征，这对于调制识别非常重要，因为不同调制方式的时域特征差异明显。

模型定义代码：

CopyInsert
def LSTMModel(weights=None, input_shape=[128, 2], classes=11):if weights is not None and not os.path.exists(weights):raise ValueError('Invalid weights path.')

input_layer = Input(shape=input_shape, name='input')# 替代 CuDNNLSTM
x = LSTM(128, return_sequences=True, activation='tanh', recurrent_activation='sigmoid')(input_layer)
x = LSTM(128, activation='tanh', recurrent_activation='sigmoid')(x)output_layer = Dense(classes, activation='softmax', name='softmax')(x)model = Model(inputs=input_layer, outputs=output_layer)if weights:model.load_weights(weights)return model

5. 数据预处理
数据预处理是提高模型性能的关键步骤：
数据加载：从pickle文件加载RadioML2016.10a数据集
数据标准化：对每个样本进行L2归一化，提高模型的泛化能力

def norm_pad_zeros(X_train, nsamples):for i in range(X_train.shape[0]):X_train[i,:,0] = X_train[i,:,0]/la.norm(X_train[i,:,0],2)return X_train

幅度-相位转换：将I/Q数据转换为幅度和相位表示

def to_amp_phase(X_train, X_val, X_test, nsamples):X_train_cmplx = X_train[:,0,:] + 1j* X_train[:,1,:]X_train_amp = np.abs(X_train_cmplx)X_train_ang = np.arctan2(X_train[:,1,:], X_train[:,0,:]) / np.pi# ...

数据筛选：从所有调制类型中筛选出目标调制类型

selected_mods = ['GFSK', 'PAM4', 'BPSK', 'QPSK', 'QAM16', 'QAM64', 'CPFSK', '8PSK']
train_selected = [i for i in range(len(Y_train)) if mods[np.argmax(Y_train[i])] in selected_mods]
X_train_selected = X_train[train_selected]

标签重编码：将标签转换为独热编码(one-hot)形式

Y_train_selected_new = np.zeros((len(train_selected), len(selected_mods)))
for i, idx in enumerate(train_selected):mod = mods[np.argmax(Y_train[idx])]Y_train_selected_new[i, selected_mods_dict[mod]] = 1

6. 训练过程
损失函数：分类交叉熵（categorical_crossentropy）
优化器：Adam优化器
批量大小：400
训练轮数：最多100轮，配合早停策略
回调函数：
ModelCheckpoint：保存最佳模型
ReduceLROnPlateau：在验证损失停滞时降低学习率
EarlyStopping：在验证损失长时间不改善时提前停止训练
训练代码：

model = LSTMModel(weights=None, input_shape=[128, 2], classes=len(selected_mods))
model.compile(loss='categorical_crossentropy', metrics=['acc'], optimizer='adam')filepath = 'weights/weights.h5'
history = model.fit(X_train_selected,Y_train_selected_new,batch_size=batch_size,epochs=nb_epoch,verbose=2,validation_data=(X_val_selected, Y_val_selected_new),callbacks=[ModelCheckpoint(filepath, monitor='val_loss', save_best_only=True),ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=5, min_lr=1e-6),EarlyStopping(monitor='val_loss', patience=50)])

7. 性能评估
模型评估采用多种方法：
混淆矩阵：直观展示各调制类型的识别准确率和错误类型

confnorm, _, _ = mltools.calculate_confusion_matrix(Y_test_selected_new, test_Y_hat, display_classes)
mltools.plot_confusion_matrix(confnorm, labels=display_classes, save_filename='picture/lstm_total_confusion.png')

不同信噪比下的性能：分析模型在不同信噪比条件下的表现

for i, snr in enumerate(snrs):indices = [j for j, s in enumerate(test_SNRs_selected) if s == snr]test_X_i = X_test_selected[indices]test_Y_i = Y_test_selected_new[indices]test_Y_i_hat = model.predict(test_X_i)confnorm_i, cor, ncor = mltools.calculate_confusion_matrix(test_Y_i, test_Y_i_hat, display_classes)acc[snr] = cor / (cor + ncor)

# 构建数据集
X = []
lbl = []
train_idx = []
val_idx = []
np.random.seed(2016)
a = 0# 遍历所有调制类型和信噪比
for mod in mods:for snr in snrs:X.append(Xd[(mod,snr)])for i in range(Xd[(mod,snr)].shape[0]):lbl.append((mod,snr))# 划分训练集和验证集train_idx += list(np.random.choice(range(a*1000,(a+1)*1000), size=600, replace=False))val_idx += list(np.random.choice(list(set(range(a*1000,(a+1)*1000))-set(train_idx)), size=200, replace=False))a += 1# 堆叠数据
X = np.vstack(X)
n_examples = X.shape[0]# 划分测试集
test_idx = list(set(range(0,n_examples))-set(train_idx)-set(val_idx))
np.random.shuffle(train_idx)
np.random.shuffle(val_idx)
np.random.shuffle(test_idx)# 提取数据子集
X_train = X[train_idx]
X_val = X[val_idx]
X_test = X[test_idx]# 转换为独热编码
def to_onehot(yy):yy1 = np.zeros([len(yy), len(mods)])yy1[np.arange(len(yy)), yy] = 1return yy1# 生成标签
Y_train = to_onehot(list(map(lambda x: mods.index(lbl[x][0]), train_idx)))
Y_val = to_onehot(list(map(lambda x: mods.index(lbl[x][0]), val_idx)))
Y_test = to_onehot(list(map(lambda x: mods.index(lbl[x][0]), test_idx)))# 转换为幅度-相位表示
X_train, X_val, X_test = to_amp_phase(X_train, X_val, X_test, 128)# 截断到最大长度
X_train = X_train[:,:maxlen,:]
X_val = X_val[:,:maxlen,:]
X_test = X_test[:,:maxlen,:]# 标准化
X_train = norm_pad_zeros(X_train, maxlen)
X_val = norm_pad_zeros(X_val, maxlen)
X_test = norm_pad_zeros(X_test, maxlen)return (mods, snrs, lbl), (X_train, Y_train), (X_val, Y_val), (X_test, Y_test), (train_idx, val_idx, test_idx)

selected_mods = ['GFSK', 'PAM4', 'BPSK', 'QPSK', 'QAM16', 'QAM64', 'CPFSK', '8PSK']
selected_mods_dict = {mod: i for i, mod in enumerate(selected_mods)}
display_classes = selected_mods

筛选与重新编码训练集

train_selected = [i for i in range(len(Y_train)) if mods[np.argmax(Y_train[i])] in selected_mods]
X_train_selected = X_train[train_selected]
Y_train_selected_new = np.zeros((len(train_selected), len(selected_mods)))
for i, idx in enumerate(train_selected):mod = mods[np.argmax(Y_train[idx])]Y_train_selected_new[i, selected_mods_dict[mod]] = 1

创建模型

model = culstm.LSTMModel(weights=None, input_shape=[128, 2], classes=len(selected_mods))
model.compile(loss='categorical_crossentropy', metrics=['acc'], optimizer='adam')

训练模型

filepath = 'weights/weights.h5'
history = model.fit(X_train_selected,Y_train_selected_new,batch_size=batch_size,epochs=nb_epoch,verbose=2,validation_data=(X_val_selected, Y_val_selected_new),callbacks=[ModelCheckpoint(filepath, monitor='val_loss', save_best_only=True),ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=5, min_lr=1e-6),EarlyStopping(monitor='val_loss', patience=50)])

模型评估

score = model.evaluate(X_test_selected, Y_test_selected_new, verbose=1, batch_size=batch_size)
print("测试集性能:", score)

预测与绘图函数

def predict(model):# 加载最佳模型权重model.load_weights(filepath)# 预测测试集test_Y_hat = model.predict(X_test_selected, batch_size=batch_size)# 计算总体混淆矩阵confnorm, _, _ = mltools.calculate_confusion_matrix(Y_test_selected_new, test_Y_hat, display_classes)mltools.plot_confusion_matrix(confnorm,labels=display_classes,save_filename='picture/lstm_total_confusion.png')# 计算每个信噪比下的性能acc = {}acc_mod_snr = np.zeros((len(selected_mods), len(snrs)))test_SNRs_selected = [lbl[test_idx[i]][1] for i in test_selected]for i, snr in enumerate(snrs):# 筛选特定信噪比的样本indices = [j for j, s in enumerate(test_SNRs_selected) if s == snr]test_X_i = X_test_selected[indices]test_Y_i = Y_test_selected_new[indices]# 预测test_Y_i_hat = model.predict(test_X_i)# 计算混淆矩阵和准确率confnorm_i, cor, ncor = mltools.calculate_confusion_matrix(test_Y_i, test_Y_i_hat, display_classes)acc[snr] = cor / (cor + ncor)# 保存准确率with open('acc111.csv', 'a', newline='') as f:csv.writer(f).writerow([acc[snr]])# 绘制混淆矩阵mltools.plot_confusion_matrix(confnorm_i,labels=display_classes,title="Confusion Matrix SNR={}".format(snr),save_filename="picture/Confusion(SNR={})(ACC={:.2f}).png".format(snr, 100.0 * acc[snr]))# 计算每种调制类型在当前信噪比下的准确率acc_mod_snr[:, i] = np.round(np.diag(confnorm_i) / np.sum(confnorm_i, axis=1), 3)# 绘制所有调制方式准确率曲线plt.figure(figsize=(12, 8))for i in range(len(selected_mods)):plt.plot(snrs, acc_mod_snr[i], marker='o', label=display_classes[i])for x, y in zip(snrs, acc_mod_snr[i]):plt.text(x, y, '{:.2f}'.format(y), fontsize=8, ha='center', va='bottom')plt.xlabel("SNR (dB)")plt.ylabel("Accuracy")plt.title("Per-Modulation Classification Accuracy vs SNR (All Mods)")plt.legend(loc='best')plt.grid(True)plt.tight_layout()plt.savefig("picture/all_mods_acc.png", dpi=300)plt.close()# 保存结果数据with open('predictresult/acc_for_mod_on_lstm.dat', 'wb') as f:pickle.dump(acc_mod_snr, f)with open('predictresult/lstm.dat', 'wb') as f:pickle.dump(acc, f)# 绘制总体准确率曲线plt.plot(snrs, [acc[snr] for snr in snrs])plt.xlabel("SNR")plt.ylabel("Overall Accuracy")plt.title("Overall Classification Accuracy on RadioML2016.10a")plt.grid()plt.tight_layout()plt.savefig('picture/each_acc.png')

主要贡献：

使用深度学习方法自动提取信号特征，避免了传统方法中复杂的特征工程在不同信噪比条件下对多种调制类型进行识别，并分析了各调制类型的识别难度提供了完整的数据处理、模型训练和评估流程，便于后续研究和应用通过本项目，我们可以看到深度学习在信号处理领域的巨大潜力，它不仅简化了传统的特征工程过程，还能在复杂环境下取得更好的性能。随着深度学习技术的不断发展，我们可以期待更多创新应用在无线通信领域涌现。
https://pan.baidu.com/s/16FN0BR0LUkfpcxZizn43xw