双向LSTM补全轨迹缺失位置：数值缺失标记方案问询

我尝试用双向LSTM补全物体轨迹中缺失的空间位置[x, y, z]。训练时用mask_value = -1掩盖轨迹中的部分已知位置，以未掩盖的完整轨迹作为目标，思路类似掩码语言建模（填空任务，但这里是填充缺失数值）。但当前模型既没关注被掩盖的值，也没学会在输入未掩盖时复制对应值。想知道针对数值数据，有没有能明确指定需要预测目标的「缺失标记」方案？

示例数据 #

df_input = pd.DataFrame({'X': {0: 0.7819923262212789, 1: 0.8459722182552086, 2: -1.0, 3: 0.8694600744527876, 4: 0.854969390572484, 5: 0.801757542867478, 6: 0.741300353222318, 7: 0.8235697352260694, 8: -1.0, 9: 0.7546104557615212, 10: 0.7499357344834469, 11: 0.7829158454961754}, 
                         'Y': {0: 0.8626673097793318, 1: 0.8375979741951041, 2: -1.0, 3: 0.7829012420113346, 4: 0.8166525985771135, 5: 0.8517183166525982, 6: 0.8004341010490774, 7: 0.8551790666827446, 8: -1.0, 9: 0.823200289400699, 10: 0.8178704931870248, 11: 0.8507657060171225}, 
                         'Z': {0: 0.5208070617906684, 1: 0.4317360235393022, 2: -1.0, 3: 0.5188734762505255, 4: 0.5005464480874317, 5: 0.5905842791088692, 6: 0.4405632618747371, 7: 0.405380411937789, 8: -1.0, 9: 0.5386717108028583, 10: 0.4439680538041194, 11: 0.4710382513661201}})

df_target = pd.DataFrame({'X': {0: 0.7819923262212789, 1: 0.8459722182552086, 2: 0.8341664048441855, 3: 0.8694600744527876, 4: 0.854969390572484, 5: 0.801757542867478, 6: 0.741300353222318, 7: 0.8235697352260694, 8: 0.8037283520417406, 9: 0.7546104557615212, 10: 0.7499357344834469, 11: 0.7829158454961754}, 
                          'Y': {0: 0.8626673097793318, 1: 0.8375979741951041, 2: 0.8172555167008316, 3: 0.7829012420113346, 4: 0.8166525985771135, 5: 0.8517183166525982, 6: 0.8004341010490774, 7: 0.8551790666827446, 8: 0.8476546484987337, 9: 0.823200289400699, 10: 0.8178704931870248, 11: 0.8507657060171225}, 
                          'Z': {0: 0.5208070617906684, 1: 0.4317360235393022, 2: 0.4711223203026479, 3: 0.5188734762505255, 4: 0.5005464480874317, 5: 0.5905842791088692, 6: 0.4405632618747371, 7: 0.405380411937789, 8: 0.4345102984447247, 9: 0.5386717108028583, 10: 0.4439680538041194, 11: 0.4710382513661201}})

原示例模型 #

import numpy as np
import pandas as pd
import tensorflow as tf
import tensorflow.keras as K

from keras.models import Sequential
from keras.layers import LSTM
from keras.layers import Bidirectional
from keras.layers import Dense
from keras.layers import TimeDistributed

# define model
n_timesteps = 12
batch_size = 1
mask_value = -1
ndim_per_time_input = 3 # only 3 readout per input (x, y, z)
ndim_per_time_output = 3 # only 3 readout per output (x, y, z)

# reformat input
X = df_input.values
Y = df_target.values
X = X.reshape(1, n_timesteps, ndim_per_time_input)
Y = Y.reshape(1, n_timesteps, ndim_per_time_output)

#### define LSTM ####
model = Sequential()

## masking layer
input_shape = (n_timesteps, ndim_per_time_input)

## input layer
model.add(Bidirectional(LSTM(20, 
                             return_sequences = True,
                             dropout = 0.2), input_shape=input_shape, merge_mode='ave'))

## output layer
model.add(TimeDistributed(Dense(ndim_per_time_output, activation = 'linear')))

## loss function
model.compile(loss='mean_squared_error', optimizer = 'adam') 

# train LSTM
for epoch in range(1000):
    model.fit(X, Y, epochs = 1, batch_size = batch_size, verbose=2)

1. 显式添加掩码标识特征 #

给输入数据新增一个维度，明确标记每个时间步的每个坐标是否需要预测：

• 标记值1表示该位置是被掩码的，需要模型预测
• 标记值0表示该位置是已知的，模型应直接复制输入值

这样模型能清晰识别需要处理的目标位置，避免混淆。

2. 使用掩码加权损失函数 #

不要用全局均方误差（MSE），而是只计算被掩码位置的损失，让模型专注于需要预测的部分，同时鼓励模型在非掩码位置复制输入值。

3. 正确集成Keras Masking层 #

原模型没有添加Masking层，导致LSTM会处理-1这个掩码值，干扰模型学习。添加Masking层后，LSTM会自动忽略掩码值对应的时间步输入。

4. 增加训练数据多样性 #

原模型仅用1条样本训练，无法泛化。需要生成大量带随机掩码的样本，让模型学习不同的轨迹补全模式。

import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Bidirectional, Dense, TimeDistributed, Masking

# 配置参数
n_timesteps = 12
batch_size = 1
mask_value = -1
ndim_coords = 3  # x, y, z

# 处理输入数据，添加掩码标识
df_input = pd.DataFrame({'X': {0: 0.7819923262212789, 1: 0.8459722182552086, 2: -1.0, 3: 0.8694600744527876, 4: 0.854969390572484, 5: 0.801757542867478, 6: 0.741300353222318, 7: 0.8235697352260694, 8: -1.0, 9: 0.7546104557615212, 10: 0.7499357344834469, 11: 0.7829158454961754}, 
                         'Y': {0: 0.8626673097793318, 1: 0.8375979741951041, 2: -1.0, 3: 0.7829012420113346, 4: 0.8166525985771135, 5: 0.8517183166525982, 6: 0.8004341010490774, 7: 0.8551790666827446, 8: -1.0, 9: 0.823200289400699, 10: 0.8178704931870248, 11: 0.8507657060171225}, 
                         'Z': {0: 0.5208070617906684, 1: 0.4317360235393022, 2: -1.0, 3: 0.5188734762505255, 4: 0.5005464480874317, 5: 0.5905842791088692, 6: 0.4405632618747371, 7: 0.405380411937789, 8: -1.0, 9: 0.5386717108028583, 10: 0.4439680538041194, 11: 0.4710382513661201}})

df_target = pd.DataFrame({'X': {0: 0.7819923262212789, 1: 0.8459722182552086, 2: 0.8341664048441855, 3: 0.8694600744527876, 4: 0.854969390572484, 5: 0.801757542867478, 6: 0.741300353222318, 7: 0.8235697352260694, 8: 0.8037283520417406, 9: 0.7546104557615212, 10: 0.7499357344834469, 11: 0.7829158454961754}, 
                          'Y': {0: 0.8626673097793318, 1: 0.8375979741951041, 2: 0.8172555167008316, 3: 0.7829012420113346, 4: 0.8166525985771135, 5: 0.8517183166525982, 6: 0.8004341010490774, 7: 0.8551790666827446, 8: 0.8476546484987337, 9: 0.823200289400699, 10: 0.8178704931870248, 11: 0.8507657060171225}, 
                          'Z': {0: 0.5208070617906684, 1: 0.4317360235393022, 2: 0.4711223203026479, 3: 0.5188734762505255, 4: 0.5005464480874317, 5: 0.5905842791088692, 6: 0.4405632618747371, 7: 0.405380411937789, 8: 0.4345102984447247, 9: 0.5386717108028583, 10: 0.4439680538041194, 11: 0.4710382513661201}})

# 生成掩码标识：1表示需要预测，0表示已知
mask_indicator = (df_input == mask_value).astype(int).values
# 合并坐标数据