diff --git a/mat翻译（测试音频）.py b/mat翻译（测试音频）.py new file mode 100644 index 0000000..4f6eba6 --- /dev/null +++ b/mat翻译（测试音频）.py @@ -0,0 +1,134 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- +""" +无标注测试音频特征提取：读取单个WAV，提取五维特征，保存为MAT和PKL（无标签） +""" + +from pathlib import Path +import numpy as np +import scipy.io.wavfile as wav +from scipy.io import savemat +from scipy.signal import hilbert +import librosa +import matplotlib.pyplot as plt +import os +import pickle # 用于保存PKL文件 + +plt.rcParams['font.sans-serif'] = ['SimHei'] # 中文显示 +plt.rcParams['axes.unicode_minus'] = False # 负号正常显示 + +# ---------- 参数设定（无需改，按原逻辑） ---------- +WAV_FILE = r"D:\SummerSchool\test2.wav" # 你的无标注测试音频路径 +WIN_SIZE = 1024 # 帧长 +OVERLAP = 512 # 帧移 +STEP = WIN_SIZE - OVERLAP # 帧步长 +THRESH = 0.01 # 降低阈值，确保能检测到片段（已调小） +SEG_LEN_SEC = 0.2 # 每段音频长度（秒） +# 输出文件路径（可自定义，默认保存在WAV文件同目录） +OUT_MAT = Path(WAV_FILE).parent / f"{Path(WAV_FILE).stem}_features.mat" +OUT_PKL = Path(WAV_FILE).parent / f"{Path(WAV_FILE).stem}_features.pkl" + +# ---------- 工具函数（完全保留原特征提取逻辑，确保和训练集一致） ---------- +def segment_signal(signal: np.ndarray, fs: int): + """按能量切分音频片段（原逻辑不变）""" + if signal.ndim > 1: # 双声道转单声道 + signal = signal[:, 0] + signal = signal / (np.max(np.abs(signal)) + 1e-12) # 归一化 + + # 分帧+计算帧能量 + frames = librosa.util.frame(signal, frame_length=WIN_SIZE, hop_length=STEP).T + energy = np.sum(frames ** 2, axis=1) + + # 筛选能量高于阈值的帧，切出有效片段 + idx = np.where(energy > THRESH)[0] + if idx.size == 0: + return [] + hit_mask = np.diff(np.concatenate(([0], idx))) > 5 # 新敲击起始帧 + hit_starts = idx[hit_mask] + + seg_len = int(round(SEG_LEN_SEC * fs)) + segments = [] + for h in hit_starts: + start = h * STEP + end = min(start + seg_len, len(signal)) + segments.append(signal[start:end]) + return segments + + +def extract_features(sig: np.ndarray, fs: int): + """提取五维特征（和训练集完全一致，保证特征匹配）""" + sig = sig.flatten() + if sig.size == 0: + return np.zeros(5) + + # 1. RMS（均方根） + rms = np.sqrt(np.mean(sig ** 2)) + # 2. 主频（频谱峰值对应的频率） + L = sig.size + f = np.fft.rfftfreq(L, d=1 / fs) + Y = np.abs(np.fft.rfft(sig)) + main_freq = f[np.argmax(Y)] + # 3. 频谱偏度 + P = Y + centroid = np.sum(f * P) / (np.sum(P) + 1e-12) + spread = np.sqrt(np.sum(((f - centroid) ** 2) * P) / (np.sum(P) + 1e-12)) + skewness = np.sum(((f - centroid) ** 3) * P) / ((np.sum(P) + 1e-12) * (spread ** 3 + 1e-12)) + # 4. MFCC第一维均值 + try: + mfccs = librosa.feature.mfcc(y=sig, sr=fs, n_mfcc=13) + mfcc_mean = float(np.mean(mfccs[0, :])) + except Exception: + mfcc_mean = 0.0 + # 5. 包络峰值（希尔伯特变换） + env_peak = np.max(np.abs(hilbert(sig))) + + return np.array([rms, main_freq, skewness, mfcc_mean, env_peak]) + +# ---------- 主程序（核心：去掉标签，只提特征+保存） ---------- +def main(): + # 1. 读取音频文件 + wav_path = Path(WAV_FILE) + if not (wav_path.exists() and wav_path.suffix == ".wav"): + print(f"❌ 错误：{WAV_FILE} 不存在或不是WAV文件！") + return + + # 用librosa读取（兼容性更好，避免格式问题） + y, fs = librosa.load(wav_path, sr=None, mono=True) + print(f"✅ 成功读取音频：{wav_path.name}，采样率：{fs} Hz") + + # 2. 切分有效片段 + segments = segment_signal(y, fs) + if len(segments) == 0: + print(f"⚠️ 未检测到有效音频片段，尝试再降低THRESH（当前{THRESH}）！") + return + print(f"✅ 检测到 {len(segments)} 个有效片段") + + # 3. 提取五维特征 + features = [extract_features(seg, fs) for seg in segments] + features_matrix = np.vstack(features).astype(np.float32) # 特征矩阵（N行5列，N=片段数） + print(f"✅ 提取特征完成，特征矩阵形状：{features_matrix.shape}（行=片段数，列=5维特征）") + + # 4. 保存为MAT文件（兼容MATLAB） + savemat(OUT_MAT, {"matrix": features_matrix}) # 只存特征矩阵，无label + print(f"✅ MAT文件已保存：{OUT_MAT}") + + # 5. 保存为PKL文件（兼容Python后续模型推断） + with open(OUT_PKL, "wb") as f: + pickle.dump({"matrix": features_matrix}, f) # 和训练集PKL结构一致（只少label） + print(f"✅ PKL文件已保存：{OUT_PKL}") + + # （可选）绘制特征可视化图 + plt.figure(figsize=(10, 6)) + feature_names = ["RMS", "主频(Hz)", "频谱偏度", "MFCC均值", "包络峰值"] + for i in range(5): + plt.subplot(2, 3, i+1) + plt.plot(range(1, len(features_matrix)+1), features_matrix[:, i], "-o", linewidth=1.5) + plt.xlabel("片段编号") + plt.ylabel("特征值") + plt.title(f"特征：{feature_names[i]}") + plt.grid(True) + plt.tight_layout() + plt.show() + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/交叉验证.py b/交叉验证.py new file mode 100644 index 0000000..47a0240 --- /dev/null +++ b/交叉验证.py @@ -0,0 +1,166 @@ +# -*- coding: utf-8 -*- +""" +交叉验证（最小改动版）——与原结构一致，仅在第6段训练后新增模型与scaler导出 +""" + +from pathlib import Path +import pickle +import numpy as np +import matplotlib.pyplot as plt + +# 新增/补全的 import +import joblib +from scipy.stats import randint, loguniform, norm +from sklearn.svm import SVC +from sklearn.model_selection import StratifiedKFold, RandomizedSearchCV, cross_val_score +from sklearn.preprocessing import StandardScaler +from sklearn.feature_selection import SelectKBest, mutual_info_classif +from sklearn.pipeline import Pipeline + +# ---------- 1. 数据路径 ---------- +BASE_DIR = Path(r'D:\SummerSchool\mat_cv\mat_cv') +TRAIN_PKL = BASE_DIR / 'cv10_train.pkl' +TEST_FILES = [BASE_DIR / 'cv10_test.pkl'] # 也可放多个测试集 pkl 文件 + +# ---------- 2. 工具 ---------- +def load_pkl_matrix(path: Path): + with open(path, 'rb') as f: + data = pickle.load(f) + return data['matrix'], data.get('label') + +# ---------- 3. 读取训练集 ---------- +X_train, y_train = load_pkl_matrix(TRAIN_PKL) +if y_train is None: + raise ValueError('训练集缺少 label 字段') +y_train = y_train.ravel() +# {0,1} → {-1,+1} +y_train_signed = np.where(y_train == 0, -1, 1) + +# ---------- 4. 标准化 ---------- +scaler = StandardScaler().fit(X_train) +X_train_std = scaler.transform(X_train) +n_features = X_train_std.shape[1] + +# ---------- 5. RandomizedSearchCV 搜索 ---------- +pipe = Pipeline([ + ('sel', SelectKBest(mutual_info_classif)), + ('svm', SVC(kernel='rbf', class_weight='balanced', probability=True)) +]) + +param_dist = { + 'sel__k': randint(1, n_features + 1), + 'svm__C': loguniform(1e-3, 1e3), + 'svm__gamma': loguniform(1e-6, 1e1) +} + +cv_inner = StratifiedKFold(n_splits=5, shuffle=True, random_state=42) +search = RandomizedSearchCV( + pipe, + param_distributions=param_dist, + n_iter=60, # 搜索次数 + scoring='roc_auc', + cv=cv_inner, + n_jobs=-1, + random_state=42, + verbose=1 +) +search.fit(X_train_std, y_train_signed) +best_params = search.best_params_ +print("\n▶ RandomizedSearch 最佳参数:", best_params) +print(f" 内层 5-折 AUC ≈ {search.best_score_:.4f}") + +# ---------- 6. 训练最终流水线 ---------- +final_model = search.best_estimator_ +final_model.fit(X_train_std, y_train_signed) + +# ---------- 6.5 新增：导出模型与标准化器（供 GUI 使用） ---------- +# 输出到 BASE_DIR 下，也可按需改路径 +model_out = BASE_DIR / 'svm_model.pkl' +scaler_out = BASE_DIR / 'scaler.pkl' +joblib.dump(final_model, model_out) +joblib.dump(scaler, scaler_out) +print(f"\n✅ 已导出模型与标尺：\n 模型: {model_out}\n 标尺: {scaler_out}\n has_predict_proba: {hasattr(final_model, 'predict_proba')}") + +# ---------- 7. 外层 5-折交叉验证 ---------- +cv_outer = StratifiedKFold(n_splits=5, shuffle=True, random_state=42) +cv_auc = cross_val_score(final_model, X_train_std, y_train_signed, + cv=cv_outer, scoring='roc_auc', n_jobs=-1) +cv_acc = cross_val_score(final_model, X_train_std, y_train_signed, + cv=cv_outer, scoring='accuracy', n_jobs=-1) + +print('\n========== 外层 5-折交叉验证 ==========') +print(f'AUC = {cv_auc.mean():.4f} ± {cv_auc.std():.4f}') +print(f'ACC = {cv_acc.mean():.4f} ± {cv_acc.std():.4f}') + +# ---------- 8. 推断 ---------- +THRESHOLD = 0.5 +Z = norm.ppf(0.975) +infer_results = [] +print('\n========== 推断结果 ==========') + +for pkl_path in TEST_FILES: + X_test, _ = load_pkl_matrix(pkl_path) + X_test_std = scaler.transform(X_test) + pred_signed = final_model.predict(X_test_std) + proba_pos = final_model.predict_proba(X_test_std)[:, 1] + pred_label = np.where(pred_signed == -1, 0, 1) + + mean_p = proba_pos.mean() + sem_p = proba_pos.std(ddof=1) / np.sqrt(len(proba_pos)) if len(proba_pos) > 1 else 0.0 + ci_low, ci_high = mean_p - Z * sem_p, mean_p + Z * sem_p + file_label = int(mean_p >= THRESHOLD) + + print(f'\n▶ 文件: {pkl_path.name} (样本 {len(pred_label)})') + for i, (lbl, prob) in enumerate(zip(pred_label, proba_pos), 1): + print(f' Sample {i:02d}: pred={lbl} prob(1)={prob:.4f}') + print(' ---- 文件级融合 ----') + print(f' mean_prob(1) = {mean_p:.4f} (95% CI {ci_low:.4f} ~ {ci_high:.4f})') + print(f' Final label = {file_label} (阈值 {THRESHOLD})') + + infer_results.append(dict( + file=pkl_path.name, + pred=pred_label.tolist(), + prob=proba_pos.tolist(), + mean_prob=float(mean_p), + ci_low=float(ci_low), + ci_high=float(ci_high), + final_label=int(file_label) + )) + +# 打印测试文件的原始标签（若有） +try: + print("TEST_FILES 标签：", load_pkl_matrix(TEST_FILES[0])[1]) +except Exception: + pass + +# ---------- 9. 保存 & 可视化 ---------- +out_pkl = BASE_DIR / 'infer_results.pkl' +with open(out_pkl, 'wb') as f: + pickle.dump(infer_results, f) +print(f'\n所有文件结果已保存到: {out_pkl}') + +plt.rcParams['font.sans-serif'] = ['SimHei'] +plt.rcParams['axes.unicode_minus'] = False + +labels = [r['file'] for r in infer_results] +means = [r['mean_prob'] for r in infer_results] +yerr = [(r['mean_prob'] - r['ci_low'], r['ci_high'] - r['mean_prob']) + for r in infer_results] + +fig, ax = plt.subplots(figsize=(6, 4)) +ax.bar(range(len(means)), means, + yerr=np.array(yerr).T, capsize=5, alpha=0.8) +ax.axhline(THRESHOLD, color='red', ls='--', label=f'阈值 {THRESHOLD}') +ax.set_xticks(range(len(labels))) +ax.set_xticklabels(labels, rotation=15) +ax.set_ylim(0, 1) +ax.set_ylabel('mean_prob(空心=0)') +ax.set_title('文件级空心概率 (±95% CI)') +ax.legend() +plt.tight_layout() + +desktop = Path.home() / 'Desktop' +save_path = desktop / 'infer_summary.png' +fig.savefig(save_path, dpi=300, bbox_inches='tight') +print(f'可视化图已保存至: {save_path}') +plt.show()