|
|
---
|
|
|
license: mit
|
|
|
tags:
|
|
|
- audio-classification
|
|
|
- wav2vec2
|
|
|
- sound-detection
|
|
|
- few-shot-learning
|
|
|
- pytorch
|
|
|
language:
|
|
|
- zh
|
|
|
datasets:
|
|
|
- custom
|
|
|
metrics:
|
|
|
- accuracy
|
|
|
- precision
|
|
|
- recall
|
|
|
- f1
|
|
|
library_name: transformers
|
|
|
pipeline_tag: audio-classification
|
|
|
---
|
|
|
|
|
|
# 🎯 热水器开关声音检测器 (Heater Switch Sound Detector)
|
|
|
|
|
|
基于Wav2Vec2的热水器开关声音实时检测模型。这是一个少样本学习项目，仅用6个音频样本就能达到100%的检测准确率。
|
|
|
|
|
|
## 模型描述
|
|
|
|
|
|
该模型使用Facebook的Wav2Vec2预训练模型作为特征提取器，在热水器开关声音数据上进行微调，实现对开关按下声音的精确识别。
|
|
|
|
|
|
### 模型架构
|
|
|
|
|
|
```
|
|
|
原始音频 [48000 samples]
|
|
|
↓ Wav2Vec2特征编码器 (7层1D卷积)
|
|
|
局部特征 [1199, 768]
|
|
|
↓ Wav2Vec2上下文网络 (12层Transformer)
|
|
|
上下文特征 [1199, 768]
|
|
|
↓ 全局平均池化
|
|
|
固定特征 [768]
|
|
|
↓ 分类头 (2层全连接)
|
|
|
分类结果 [2] (开关/背景)
|
|
|
```
|
|
|
|
|
|
## 训练数据
|
|
|
|
|
|
- **正样本**: 6个热水器开关声音 (3.2-5.2秒)
|
|
|
- **负样本**: 6个自动生成的背景噪音
|
|
|
- **总样本**: 12个 (训练集8个，测试集4个)
|
|
|
- **采样率**: 16kHz
|
|
|
- **格式**: 单声道WAV
|
|
|
|
|
|
### 数据特征分析
|
|
|
|
|
|
| 样本类型 | 时长范围 | RMS能量 | 频谱质心 | 过零率 |
|
|
|
|----------|----------|---------|----------|--------|
|
|
|
| 开关声音 | 3.2-5.2s | 0.0079-0.0115 | 1587-1992Hz | 0.0657-0.1215 |
|
|
|
| 背景噪音 | 2.0-4.0s | 0.005-0.02 | 500-1500Hz | 0.05-0.15 |
|
|
|
|
|
|
## 性能指标
|
|
|
|
|
|
| 指标 | 数值 |
|
|
|
|------|------|
|
|
|
| **准确率** | 100% |
|
|
|
| **精确率** | 100% |
|
|
|
| **召回率** | 100% |
|
|
|
| **F1分数** | 100% |
|
|
|
| **训练轮数** | 15 epochs |
|
|
|
| **模型大小** | 361MB |
|
|
|
| **推理延迟** | <100ms |
|
|
|
|
|
|
### 混淆矩阵
|
|
|
|
|
|
```
|
|
|
实际\预测 无开关 有开关
|
|
|
无开关 2 0
|
|
|
有开关 0 2
|
|
|
```
|
|
|
|
|
|
## 使用方法
|
|
|
|
|
|
### 安装依赖
|
|
|
|
|
|
```bash
|
|
|
pip install torch torchaudio transformers huggingface_hub
|
|
|
```
|
|
|
|
|
|
### 加载模型
|
|
|
|
|
|
```python
|
|
|
from huggingface_hub import hf_hub_download
|
|
|
import torch
|
|
|
import torchaudio
|
|
|
from transformers import Wav2Vec2Model
|
|
|
|
|
|
# 下载模型
|
|
|
model_path = hf_hub_download(
|
|
|
repo_id="lemonhall/heater-switch-detector",
|
|
|
filename="switch_detector_model.pth"
|
|
|
)
|
|
|
|
|
|
# 加载模型
|
|
|
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
|
|
checkpoint = torch.load(model_path, map_location=device)
|
|
|
|
|
|
# 重建模型架构
|
|
|
wav2vec2_model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-base")
|
|
|
classifier = torch.nn.Sequential(
|
|
|
torch.nn.Linear(768, 256),
|
|
|
torch.nn.ReLU(),
|
|
|
torch.nn.Dropout(0.3),
|
|
|
torch.nn.Linear(256, 2)
|
|
|
)
|
|
|
|
|
|
# 加载权重
|
|
|
classifier.load_state_dict(checkpoint['classifier_state_dict'])
|
|
|
classifier.eval()
|
|
|
```
|
|
|
|
|
|
### 音频预测
|
|
|
|
|
|
```python
|
|
|
def predict_audio(audio_path):
|
|
|
# 加载音频
|
|
|
waveform, sample_rate = torchaudio.load(audio_path)
|
|
|
|
|
|
# 重采样到16kHz
|
|
|
if sample_rate != 16000:
|
|
|
resampler = torchaudio.transforms.Resample(sample_rate, 16000)
|
|
|
waveform = resampler(waveform)
|
|
|
|
|
|
# 转为单声道
|
|
|
if waveform.shape[0] > 1:
|
|
|
waveform = waveform.mean(dim=0, keepdim=True)
|
|
|
|
|
|
# 特征提取
|
|
|
with torch.no_grad():
|
|
|
features = wav2vec2_model(waveform).last_hidden_state
|
|
|
pooled_features = features.mean(dim=1) # 全局平均池化
|
|
|
|
|
|
# 分类预测
|
|
|
logits = classifier(pooled_features)
|
|
|
probabilities = torch.softmax(logits, dim=-1)
|
|
|
prediction = torch.argmax(probabilities, dim=-1)
|
|
|
|
|
|
return {
|
|
|
'prediction': '开关按下' if prediction.item() == 1 else '背景声音',
|
|
|
'confidence': probabilities.max().item(),
|
|
|
'probabilities': {
|
|
|
'背景声音': probabilities[0][0].item(),
|
|
|
'开关按下': probabilities[0][1].item()
|
|
|
}
|
|
|
}
|
|
|
|
|
|
# 使用示例
|
|
|
result = predict_audio("test_audio.wav")
|
|
|
print(f"预测结果: {result['prediction']}")
|
|
|
print(f"置信度: {result['confidence']:.3f}")
|
|
|
```
|
|
|
|
|
|
### 实时检测
|
|
|
|
|
|
```python
|
|
|
import pyaudio
|
|
|
import numpy as np
|
|
|
|
|
|
def realtime_detection():
|
|
|
# 音频参数
|
|
|
SAMPLE_RATE = 16000
|
|
|
CHUNK_SIZE = 1024
|
|
|
DETECTION_WINDOW = 3.0 # 3秒检测窗口
|
|
|
|
|
|
# 初始化音频流
|
|
|
audio = pyaudio.PyAudio()
|
|
|
stream = audio.open(
|
|
|
format=pyaudio.paFloat32,
|
|
|
channels=1,
|
|
|
rate=SAMPLE_RATE,
|
|
|
input=True,
|
|
|
frames_per_buffer=CHUNK_SIZE
|
|
|
)
|
|
|
|
|
|
print("🎤 开始实时检测...")
|
|
|
|
|
|
buffer = []
|
|
|
window_size = int(DETECTION_WINDOW * SAMPLE_RATE)
|
|
|
|
|
|
try:
|
|
|
while True:
|
|
|
# 读取音频数据
|
|
|
data = stream.read(CHUNK_SIZE)
|
|
|
audio_chunk = np.frombuffer(data, dtype=np.float32)
|
|
|
buffer.extend(audio_chunk)
|
|
|
|
|
|
# 保持窗口大小
|
|
|
if len(buffer) > window_size:
|
|
|
buffer = buffer[-window_size:]
|
|
|
|
|
|
# 检测
|
|
|
if len(buffer) == window_size:
|
|
|
waveform = torch.FloatTensor(buffer).unsqueeze(0)
|
|
|
|
|
|
with torch.no_grad():
|
|
|
features = wav2vec2_model(waveform).last_hidden_state
|
|
|
pooled_features = features.mean(dim=1)
|
|
|
logits = classifier(pooled_features)
|
|
|
probabilities = torch.softmax(logits, dim=-1)
|
|
|
|
|
|
switch_prob = probabilities[0][1].item()
|
|
|
|
|
|
if switch_prob > 0.93: # 高置信度阈值
|
|
|
print(f"🔥 检测到开关按下! 置信度: {switch_prob:.3f}")
|
|
|
|
|
|
except KeyboardInterrupt:
|
|
|
print("\n⏹️ 检测停止")
|
|
|
finally:
|
|
|
stream.stop_stream()
|
|
|
stream.close()
|
|
|
audio.terminate()
|
|
|
|
|
|
# 运行实时检测
|
|
|
realtime_detection()
|
|
|
```
|
|
|
|
|
|
## 技术特点
|
|
|
|
|
|
### 🚀 优势
|
|
|
|
|
|
- **少样本学习**: 仅需6个样本即可达到完美分类
|
|
|
- **端到端训练**: 从原始音频波形直接学习特征
|
|
|
- **预训练优势**: 利用Wav2Vec2的大规模预训练知识
|
|
|
- **实时检测**: 支持麦克风实时音频流处理
|
|
|
- **高精度**: 测试集100%准确率
|
|
|
|
|
|
### 🎯 应用场景
|
|
|
|
|
|
- **智能家居**: 自动检测热水器使用状态
|
|
|
- **设备监控**: 远程监控设备操作
|
|
|
- **节能管理**: 记录设备使用时间和频率
|
|
|
- **安全监控**: 异常使用模式检测
|
|
|
|
|
|
### ⚙️ 技术细节
|
|
|
|
|
|
- **基础模型**: facebook/wav2vec2-base
|
|
|
- **训练策略**: 冻结预训练参数，只训练分类头
|
|
|
- **优化器**: AdamW (lr=1e-4)
|
|
|
- **损失函数**: CrossEntropyLoss
|
|
|
- **数据增强**: 自动生成负样本
|
|
|
|
|
|
## 限制和改进
|
|
|
|
|
|
### 当前限制
|
|
|
|
|
|
- 训练数据较少，可能对新环境泛化能力有限
|
|
|
- 只能检测特定类型的开关声音
|
|
|
- 需要相对安静的环境以减少误报
|
|
|
|
|
|
### 未来改进
|
|
|
|
|
|
- [ ] 收集更多样化的训练数据
|
|
|
- [ ] 支持多类别检测（开/关/故障）
|
|
|
- [ ] 添加噪音鲁棒性训练
|
|
|
- [ ] 模型压缩和量化
|
|
|
- [ ] 支持更多设备类型
|
|
|
|
|
|
## 引用
|
|
|
|
|
|
如果您使用了这个模型，请引用：
|
|
|
|
|
|
```bibtex
|
|
|
@misc{heater-switch-detector-2024,
|
|
|
title={基于Wav2Vec2的热水器开关声音检测器},
|
|
|
author={lemonhall},
|
|
|
year={2024},
|
|
|
howpublished={\url{https://huggingface.co/lemonhall/heater-switch-detector}}
|
|
|
}
|
|
|
```
|
|
|
|
|
|
## 许可证
|
|
|
|
|
|
MIT License
|
|
|
|
|
|
## 联系方式
|
|
|
|
|
|
如有问题或建议，请通过以下方式联系：
|
|
|
- GitHub: [项目地址](https://github.com/lemonhall/heater_click)
|
|
|
- Email: [email protected]
|
|
|
|
|
|
---
|
|
|
|
|
|
*该模型仅用于研究和教育目的。在生产环境中使用前，请进行充分的测试和验证。* |
