基于 CNN-LSTM 的阴燃蔓延时序预判，实现从 “事后识别火情” 升级为 “提前预判高温扩张风险”。整套流程分为：视频时序样本构建、CNN-LSTM 模型训练、ONNX 模型导出、嵌入式部署四大环节。

• 数据来源：红外摄像头录制1m 30s 分段 AVI 火情视频；
• 样本构造：滑动窗口截取连续 8 帧红外图像为输入序列，提取未来帧高温特征作为回归标签，划分安全 / 阴燃扩张 / 明火三类风险标签；
• 标签提取：对每组序列，提取未来帧的高温像素面积、图像平均灰度、高温区域中心坐标作为回归标签；同时标注 0 = 无扩张风险、1 = 阴燃持续扩张、2 = 即将出现明火三分类标签；
• 数据划分：训练集 80%、验证集 20%，统一归一化至 [0,1]，适配模型输入。
• 数据集规格：共 8 个.npz时序文件，数据集硬盘占用 12GB；
• 训练硬件：笔记本 CPU，内存有限，无法一次性加载全部数据。

初始代码一次性读取全部 npz 文件，加载数组时 numpy 直接申请 4GB 以上内存，程序初始化阶段直接崩溃：

MemoryError: Unable to allocate 4.17 GiB for an array with shape (559153152,) and data type float64

摒弃全局一次性加载，改为单文件分批训练不再扫描全部文件构建全局索引，循环逐个读取 npz，训练完成立刻关闭文件释放内存，同一时间仅占用单个文件内存；

import os import torch import torch.nn as nn import torch.optim as optim import numpy as np from torch.utils.data import Dataset, DataLoader device = torch.device("cuda" if torch.cuda.is_available() else "cpu") SEQ_LENGTH = 8 H, W = 192, 256 WEIGHT_SAVE = "temp_weight.pth" class FireSingleFileDataset(Dataset): def __init__(self, file_path): print(f"Load single file: {os.path.basename(file_path)}") self.data = np.load(file_path, mmap_mode="r") self.seq_arr = self.data["seq"] self.label_arr = self.data["label"] self.total = len(self.seq_arr) print(f"Samples in this file: {self.total}") def __len__(self): return self.total def __getitem__(self, idx): seq = self.seq_arr[idx] label = self.label_arr[idx] seq = torch.from_numpy(seq).permute(0, 3, 1, 2).float() label = torch.from_numpy(label).float() return seq, label def close(self): self.data.close() class ConvBlock(nn.Module): def __init__(self, in_c, out_c): super().__init__() self.block = nn.Sequential( nn.Conv2d(in_c, out_c, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2) ) def forward(self, x): return self.block(x) class CNNLSTM(nn.Module): def __init__(self): super().__init__() self.cnn = nn.Sequential( ConvBlock(3, 16), ConvBlock(16, 32), ConvBlock(32, 64) ) self.lstm = nn.LSTM(input_size=64*24*32, hidden_size=128, batch_first=True) self.reg_head = nn.Linear(128, 4) self.cls_head = nn.Sequential(nn.Linear(128, 64), nn.ReLU(), nn.Linear(64, 3)) def forward(self, x): b, seq, c, h, w = x.shape feat_list = [] for t in range(seq): feat = self.cnn(x[:, t]) feat_list.append(feat.flatten(1)) seq_feat = torch.stack(feat_list, dim=1) lstm_out, _ = self.lstm(seq_feat) last_feat = lstm_out[:, -1, :] reg_out = self.reg_head(last_feat) cls_out = self.cls_head(last_feat) return reg_out, cls_out def train_one_file(model, file_path, epoch_num=1): dataset = FireSingleFileDataset(file_path) loader = DataLoader(dataset, batch_size=1, shuffle=True, num_workers=0) opt = optim.Adam(model.parameters(), lr=1e-4) loss_mse = nn.MSELoss() loss_ce = nn.CrossEntropyLoss() print(f"===== Start train file {os.path.basename(file_path)} =====") for e in range(epoch_num): total_loss = 0.0 for seq, label in loader: seq, label = seq.to(device), label.to(device) pred_reg, pred_cls = model(seq) loss1 = loss_mse(pred_reg, label) cls_label = torch.clamp(torch.round(label[:, 0]).long(), 0, 2) loss2 = loss_ce(pred_cls, cls_label) loss = loss1 + loss2 opt.zero_grad() loss.backward() opt.step() total_loss += loss.item() avg_loss = total_loss / len(loader) print(f"Epoch {e+1}/{epoch_num}, Loss: {avg_loss:.4f}") dataset.close() torch.save(model.state_dict(), WEIGHT_SAVE) print(f"Temporary weight saved to {WEIGHT_SAVE}") if __name__ == "__main__": print("===== Start Initializing CNN-LSTM Model =====") model = CNNLSTM().to(device) print(f"Model loaded on device: {device}") dataset_path = r"D:\图像处理\venv\train_seq" all_files = [os.path.join(dataset_path, f) for f in os.listdir(dataset_path) if f.endswith(".npz")] print(f"Total npz files: {len(all_files)}") if os.path.exists(WEIGHT_SAVE): print(f"Load previous weight from {WEIGHT_SAVE}") model.load_state_dict(torch.load(WEIGHT_SAVE, map_location=device)) for f in all_files: train_one_file(model, f, epoch_num=1) print("Release memory, prepare next file\n") print("===== All files training complete, export ONNX Model =====") model.eval() dummy = torch.randn(1, SEQ_LENGTH, 3, H, W).to(device) save_name = "fire_predict.onnx" full_save_path = os.path.abspath(save_name) torch.onnx.export(model, dummy, full_save_path, opset_version=12) print(f"ONNX model saved successfully! Path: {full_save_path}") print("Upload fire_predict.onnx to RK3568 /home via MobaXterm")

• 每次仅打开单个 npz，训练结束调用close()释放内存，杜绝多文件同时占用内存；
• 每个文件训练完成保存临时权重，下次运行可加载权重接续训练，无需从头开始；
• CNN 提取单帧红外热区空间特征，LSTM 学习时序扩张规律，同时输出热区回归预测值 + 火情风险三分类结果。