# -*- coding: utf-8 -*- """ 螺栓拧紧数据预处理与特征工程脚本 处理流程（对应「工业大数据预处理与特征工程」技术方向）： 1. 数据清洗 —— 去除空行程段、3σ 异常值剔除 + 线性插值、缺失值处理 2. 特征提取 —— 从力矩-转角时序提取 12 维时域特征 3. 特征选择 —— Pearson 相关 + 共线性分析，降维至 10 维核心特征 4. 特征标准化 —— Z-score（仅基于训练集统计量，避免数据泄露） 5. 数据集划分 —— 7:2:1 分层随机抽样（按质量标签） 输入：data/raw/{benchmark,extended}/records.jsonl 输出： data/processed/train.csv / val.csv / test.csv —— 标准化后特征集 data/processed/feature_metadata.json —— 特征元信息 用法： python data/preprocess.py """ from __future__ import annotations import json from pathlib import Path import numpy as np import pandas as pd BASE_DIR = Path(__file__).resolve().parent RAW_DIR = BASE_DIR / "raw" PROCESSED_DIR = BASE_DIR / "processed" # 各规格名义最终拧紧角度（用于计算「转角偏差量」特征） NOMINAL_ANGLE = {"M12": 95.0, "M16": 135.0, "M20": 175.0, "M24": 220.0} # 12 维时域特征定义 FEATURE_NAMES = [ "torque_max", # 力矩最大值 (N·m) "torque_mean", # 力矩平均值 (N·m) "torque_std", # 力矩标准差 (N·m) "torque_peak_angle", # 力矩峰值出现角度 (deg) "snug_torque", # 贴合点扭矩 (N·m) "total_angle", # 拧紧总转角 (deg) "avg_rate", # 拧紧平均速率 (N·m/deg) "rising_slope", # 力矩上升斜率 (N·m/deg) "torque_fluctuation",# 力矩波动率 (无量纲) "angle_deviation", # 转角偏差量 (deg) "final_torque", # 最终扭矩值 (N·m) "hold_fluctuation", # 扭矩保持段波动率 (N·m) ] def load_raw() -> pd.DataFrame: """加载原始数据（基准集 + 扩展集合并）。""" frames = [] for sub in ("benchmark", "extended"): path = RAW_DIR / sub / "records.jsonl" with open(path, encoding="utf-8") as f: recs = [json.loads(line) for line in f if line.strip()] df = pd.DataFrame(recs) df["source"] = sub frames.append(df) return pd.concat(frames, ignore_index=True) def clean_curve(angle: np.ndarray, torque: np.ndarray) -> tuple[np.ndarray, np.ndarray]: """数据清洗：去除空行程段 + 3σ 异常值剔除并线性插值补全。""" angle = np.asarray(angle, dtype=float) torque = np.asarray(torque, dtype=float) # 1) 去除空行程段：截取从贴合点（力矩首次超过峰值 30%）到终止 if len(torque) == 0: return angle, torque thresh = 0.3 * np.max(torque) snug_idx = int(np.argmax(torque >= thresh)) angle, torque = angle[snug_idx:], torque[snug_idx:] if len(torque) < 3: return angle, torque # 2) 3σ 异常值剔除：基于一阶差分（跳变点）识别离群点，线性插值补全 diff = np.abs(np.diff(torque, prepend=torque[0])) mu, sigma = diff.mean(), diff.std() if sigma > 1e-9: outliers = diff > (mu + 3 * sigma) if outliers.any(): idx = np.where(outliers)[0] valid = ~outliers # 仅当两端存在有效点时进行插值，否则保留原值 torque_clean = torque.copy() for i in idx: if 0 < i < len(torque) - 1: torque_clean[i] = (torque[i - 1] + torque[i + 1]) / 2.0 torque = torque_clean return angle, torque def extract_features(angle: np.ndarray, torque: np.ndarray, spec: str) -> dict: """提取 12 维时域特征。""" n = len(torque) torque_max = float(np.max(torque)) peak_idx = int(np.argmax(torque)) thresh = 0.3 * torque_max snug_idx = int(np.argmax(torque >= thresh)) if torque_max > 0 else 0 grad = np.gradient(torque, angle) hold_start = int(n * 0.8) # 保持段起点（末 20% 区间） snug_torque = float(torque[snug_idx]) denom = float(angle[peak_idx] - angle[snug_idx]) if peak_idx > snug_idx else 1e-6 return { "torque_max": torque_max, "torque_mean": float(np.mean(torque)), "torque_std": float(np.std(torque)), "torque_peak_angle": float(angle[peak_idx]), "snug_torque": snug_torque, "total_angle": float(angle[-1] - angle[0]), "avg_rate": float((torque_max - snug_torque) / denom), "rising_slope": float(np.max(grad[: peak_idx + 1])), "torque_fluctuation": float(np.std(torque) / (np.mean(torque) + 1e-6)), "angle_deviation": float(angle[-1] - NOMINAL_ANGLE.get(spec, np.nan)), "final_torque": float(torque[-1]), "hold_fluctuation": float(np.std(torque[hold_start:])), } def select_features(feat_df: pd.DataFrame, y: pd.Series, threshold_corr: float = 0.1, threshold_collinear: float = 0.97) -> tuple[list[str], dict]: """特征选择：剔除弱相关特征，合并共线性特征（连通分量聚类，保留每组代表）。""" corr_with_label = {f: float(np.corrcoef(feat_df[f], y)[0, 1]) for f in FEATURE_NAMES} abs_corr = {f: abs(c) for f, c in corr_with_label.items()} # 弱相关剔除 kept = [f for f in FEATURE_NAMES if abs_corr[f] >= threshold_corr] dropped_weak = [f for f in FEATURE_NAMES if abs_corr[f] < threshold_corr] # 共线性合并：|corr| 高于阈值的特征对构成连通分量，每组保留与标签相关性最强的一个 parent = {f: f for f in kept} def find(x: str) -> str: while parent[x] != x: parent[x] = parent[parent[x]] x = parent[x] return x def union(a: str, b: str) -> None: ra, rb = find(a), find(b) if ra != rb: parent[rb] = ra collinear_pairs = [] for i in range(len(kept)): for j in range(i + 1, len(kept)): f_i, f_j = kept[i], kept[j] c = float(np.corrcoef(feat_df[f_i], feat_df[f_j])[0, 1]) if abs(c) > threshold_collinear: collinear_pairs.append((f_i, f_j, round(c, 3))) union(f_i, f_j) clusters: dict[str, list[str]] = {} for f in kept: clusters.setdefault(find(f), []).append(f) selected, dropped_collinear = [], [] for members in clusters.values(): best = max(members, key=lambda f: abs_corr[f]) selected.append(best) for m in members: if m != best: dropped_collinear.append({"kept": best, "dropped": m, "corr": next(c for a, b, c in collinear_pairs if {a, b} == {best, m})}) selected.sort() meta = { "pearson_with_label": {f: round(v, 4) for f, v in corr_with_label.items()}, "dropped_weak": dropped_weak, "dropped_collinear": dropped_collinear, "selected_features": selected, } return selected, meta def main() -> None: raw = load_raw() print(f"[OK] 加载原始数据 {len(raw)} 条") # 1) 清洗 + 2) 特征提取 feature_rows = [] for _, row in raw.iterrows(): angle, torque = clean_curve(row["angle"], row["torque"]) if len(torque) < 3: continue feats = extract_features(angle, torque, row["spec"]) feats.update({"id": row["id"], "spec": row["spec"], "source": row["source"], "label": row["label"], "label_name": row["label_name"]}) feature_rows.append(feats) feat_df = pd.DataFrame(feature_rows) print(f"[OK] 特征提取完成，有效样本 {len(feat_df)} 条") # 3) 特征选择 selected, meta = select_features(feat_df, feat_df["label"]) print(f"[OK] 特征选择：{len(FEATURE_NAMES)} -> {len(selected)} 维") print(f" 剔除弱相关: {meta['dropped_weak']}") print(f" 合并共线性: {meta['dropped_collinear']}") # 4) 数据集划分（7:2:1 分层） from sklearn.model_selection import train_test_split train, temp = train_test_split(feat_df, test_size=0.3, stratify=feat_df["label"], random_state=42) val, test = train_test_split(temp, test_size=1 / 3, stratify=temp["label"], random_state=42) # 5) 特征标准化：仅基于训练集统计量 mean = train[selected].mean() std = train[selected].std().replace(0, 1e-9) def standardize(df: pd.DataFrame) -> pd.DataFrame: out = df.copy() out[selected] = (out[selected] - mean) / std return out train_s, val_s, test_s = standardize(train), standardize(val), standardize(test) PROCESSED_DIR.mkdir(parents=True, exist_ok=True) cols = ["id", "spec", "source", "label", "label_name"] + selected train_s[cols].to_csv(PROCESSED_DIR / "train.csv", index=False, encoding="utf-8-sig") val_s[cols].to_csv(PROCESSED_DIR / "val.csv", index=False, encoding="utf-8-sig") test_s[cols].to_csv(PROCESSED_DIR / "test.csv", index=False, encoding="utf-8-sig") meta["n_samples"] = {"train": int(len(train)), "val": int(len(val)), "test": int(len(test))} meta["standardization"] = {"mean": mean.round(6).to_dict(), "std": std.round(6).to_dict()} with open(PROCESSED_DIR / "feature_metadata.json", "w", encoding="utf-8") as f: json.dump(meta, f, ensure_ascii=False, indent=2) print(f"[OK] 数据集划分 -> train {len(train)} / val {len(val)} / test {len(test)}") print(f"[OK] 输出目录 {PROCESSED_DIR}") if __name__ == "__main__": main()