import numpy as np
from scipy.signal import spectrogram

def extract_ae_features(signal, fs):
    """提取声发射信号基础特征"""
    amplitude = np.max(np.abs(signal))
    duration = len(signal)/fs
    energy = np.sum(signal**2)

    # 计算RA值（需预设阈值）
    threshold = 0.5 * amplitude
    rising_indices = np.where(signal > threshold)[0]
    if len(rising_indices) > 0:
        rise_time = (rising_indices[-1] - rising_indices[0])/fs
        ra_value = rise_time / amplitude
    else:
        ra_value = 0

    # 频谱分析
    f, t, Sxx = spectrogram(signal, fs, nperseg=256)
    af = np.sum(Sxx * f) / np.sum(Sxx)  # 平均频率
    pf = f[np.argmax(Sxx)]              # 峰值频率

    return {
        'amplitude': amplitude,
        'duration': duration,
        'energy': energy,
        'ra_value': ra_value,
        'af': af,
        'pf': pf
    }

import tensorflow as tf
from tensorflow.keras import layers

def build_dphnn(input_shape):
    """构建双通道混合神经网络"""
    # 时序通道
    time_input = layers.Input(shape=input_shape)
    lstm_out = layers.Bidirectional(layers.LSTM(64, return_sequences=True))(time_input)
    time_branch = layers.GlobalMaxPooling1D()(lstm_out)

    # 频域通道
    freq_input = layers.Input(shape=(input_shape[0]//4,))
    freq_branch = layers.Dense(64, activation='relu')(freq_input)

    # 注意力融合
    merged = layers.Concatenate()([time_branch, freq_branch])
    attention = layers.Dense(2, activation='softmax')(merged)
    weighted = layers.Dot(axes=1)([merged, attention])

    output = layers.Dense(3, activation='softmax')(weighted)  # 三类裂纹分类
    model = tf.keras.Model(inputs=[time_input, freq_input], outputs=output)
    model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
    return model