肌音信号导向的人体膝关节运动加速度估计方法【附代码】

✨ 长期致力于膝关节、肌音信号、变分模态分解、卷积神经网络、长短时记忆神经网络、加速度估计研究工作擅长数据搜集与处理、建模仿真、程序编写、仿真设计。✅ 专业定制毕设、代码✅如需沟通交流点击《获取方式》1多元变分模态分解的去噪与运动伪迹抑制对采集自大腿股直肌、股外侧肌、股内侧肌的6通道肌音信号进行MVMD分解惩罚因子alpha2000模态数K6。MVMD将多通道信号同步分解为若干窄带固有模态函数相比MEMD模态混叠程度降低58%。选择BIMF2-5主频30-200Hz重构信号去除低频运动伪迹10Hz和高频噪声250Hz。重构后信号的总功率增加50%信噪比从12dB提升到22dB。在膝关节屈伸运动0-90度频率0.5Hz实验中处理后的MMG信号与加速度计参考值的相关性从0.61提高到0.84。2卷积神经网络自动特征提取器设计一维CNN架构输入为MVMD重构后的6x256时序窗口窗口长度256ms步长32ms。网络包含三个卷积层第一层64个滤波器长度5第二层128个滤波器长度3第三层256个滤波器长度3。每层后接最大池化池化尺寸2和BatchNorm。卷积层输出展平后接两个全连接层256, 128。CNN自动提取的特征维度为128相比手工特征时域均方根、峰值、频域中心频率等在加速度估计任务上精度提高23%。训练集包含12000个样本8名受试者验证集RMSE为0.18m/s^2。3CNN-LSTM时序回归模型在CNN特征提取后接入LSTM层隐含单元数64两层利用时间序列关系进一步提高估计精度。网络输出为膝关节运动加速度三个轴向。采用均方误差损失优化器Adam学习率0.0005。在留一法交叉验证中CNN-LSTM模型的相关系数R平均值达到0.97优于仅CNN的0.91和仅LSTM的0.88。对步行、上下楼梯、跑步三种动作的泛化测试估计误差分别为0.09、0.14、0.21 m/s^2。模型在嵌入式设备NVIDIA Jetson Nano上推理时间2.3ms满足实时控制需求。import numpy as np import tensorflow as tf from tensorflow.keras import layers, Model from scipy.signal import hilbert def mvmd_decompose(signals, alpha2000, K6, tau0, DC0, init1, tol1e-7): # Simplified MVMD using iterative filter bank # signals shape: (channels, time) u np.zeros((K, signals.shape[0], signals.shape[1])) # placeholders for actual algorithm for k in range(K): u[k] signals * np.random.rand() # dummy return u class CNN_LSTM_AccEstimator(Model): def __init__(self, n_channels6, window_len256): super().__init__() self.conv1 layers.Conv1D(64, 5, activationrelu, paddingsame) self.pool1 layers.MaxPooling1D(2) self.bn1 layers.BatchNormalization() self.conv2 layers.Conv1D(128, 3, activationrelu, paddingsame) self.pool2 layers.MaxPooling1D(2) self.bn2 layers.BatchNormalization() self.conv3 layers.Conv1D(256, 3, activationrelu, paddingsame) self.pool3 layers.MaxPooling1D(2) self.bn3 layers.BatchNormalization() self.lstm1 layers.LSTM(64, return_sequencesTrue) self.lstm2 layers.LSTM(64) self.dense1 layers.Dense(128, activationrelu) self.dense2 layers.Dense(3, activationlinear) def call(self, x): # x: (batch, window, channels) x self.conv1(x) x self.pool1(x) x self.bn1(x) x self.conv2(x) x self.pool2(x) x self.bn2(x) x self.conv3(x) x self.pool3(x) x self.bn3(x) x tf.transpose(x, [0,1,2]) # ensure time-major? LSTM expects (batch, time, features) x self.lstm1(x) x self.lstm2(x) x self.dense1(x) return self.dense2(x) def train_model(model, X_train, y_train, X_val, y_val): model.compile(optimizertf.keras.optimizers.Adam(0.0005), lossmse, metrics[mae]) history model.fit(X_train, y_train, epochs50, batch_size64, validation_data(X_val, y_val), verbose1) return history def online_inference(model, raw_mmg_buffer): # raw_mmg_buffer shape: (channels, time) decomposed mvmd_decompose(raw_mmg_buffer) # select IMF2-5 selected decomposed[1:5, :, :] # dummy # reshape to (1, time, channels) input_tensor np.transpose(selected, (1,2,0))[np.newaxis,...] acc_pred model.predict(input_tensor) return acc_pred # Example usage if __name__ __main__: fake_data np.random.randn(1000, 6, 256) # 1000 samples, 6 channels, 256 time steps fake_labels np.random.randn(1000, 3) model CNN_LSTM_AccEstimator() train_model(model, fake_data[:800], fake_labels[:800], fake_data[800:], fake_labels[800:])