修复 P0:根因判定、退服率口径、影响评估与告警关联量化

- 根因判定改为三级(衍生关系/时间窗/疑似)并带置信度;小区退服告警 ±2 分钟匹配;
  衍生告警按父告警归类;优先选随业务恢复清除的告警;时间窗按历史流水 95% 分位数标定
- 退服率按日历统计周期计算;站点数分母支持站名关键字;重复退服率除以全部站点;
  月均按统计周期自然月数;导出与看板同一口径
- 影响评估新增受影响用户·小时与 500 米邻站覆盖补偿
- 新增告警—退服关联量化(支持度/置信度/提升度)及 /api/association,根因页展示
- 新增 tests/(扰动数据集留出检验、指标口径检验)与 bench.py;旧库启动时自动迁移

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
treeandClaude Opus 5.5 committed 2026-11-18 14:43:34 +08:00
1 parent f8d1ccc526
commit e9616af34d
10 files changed
+869 -119

No files matched your search

+54
View File
@@ -0,0 +1,54 @@
"""根因判定准确率检验:基线 + 6 组扰动数据集,时间窗只用 1~6 月流水标定,在 7~9 月(留出集)上计分。
运行:python tests/test_rootcause.py
"""
from harness import (WIDE_LEAD, add_noise, check, core, drop_root, finish, generate, jitter_clear, load, no_parent,
sample_data, score, skew_cell_alarm)
SPLIT = '2026-07-01' # 之前的流水用于标定时间窗,之后的用于计分
s, a, o = sample_data()
sw, aw, ow = generate('wide', lead=WIDE_LEAD)
# (名称, 数据, 准确率下限);下限为 None 的只报告不判定——根因告警缺失时,缺失的那部分本来就无法从告警判出
SCENARIOS = [
('S0 基线(仓库样例数据)', (s, a, o), 90),
('S1 去掉父告警(只靠时间窗)', (s, no_parent(a), o), 90),
('S2 小区退服告警时刻偏 5-120 秒', (s, skew_cell_alarm(a), o), 90),
('S3 S1 + 20% 根因告警缺失', (s, drop_root(no_parent(a), o, .2), o), None),
('S4 S1 + 30% 退服前干扰告警', (s, add_noise(no_parent(a), o, .3), o), 90),
('S5 S1 + 告警提前量变宽', (sw, no_parent(aw), ow), 90),
('S6 综合(S5 + 时刻偏移 + 30% 父告警 + 清除时刻 ±10 分 + 15% 缺失 + 25% 干扰)',
# 先删根因告警再扰动清除时刻:drop_root 靠“清除时间 = 恢复时间”识别根因告警
(sw, add_noise(jitter_clear(skew_cell_alarm(drop_root(no_parent(aw, keep=.3), ow, .15)), ow), ow, .25), ow),
None),
]
summary = []
for name, (sites, alarms, outages), floor in SCENARIOS:
df = load('rootcause', sites, alarms, outages, calibrate_until=SPLIT)
test = df[df['start_time'] >= SPLIT]
accuracy, rows = score(test)
other = (test['root_cause'] == '其他').mean() * 100
print(f'\n== {name}\n 留出集 {len(test)} 条,准确率 {accuracy}%,判为“其他” {other:.1f}%')
print(' 类别 precision recall F1')
for cause, precision, recall, f1 in rows:
print(f' {cause:<8} {precision:>7} {recall:>7} {f1:>6}')
confidence = test.groupby('confidence').apply(lambda x: (x['cause_category'] == x['root_cause']).mean() * 100,
include_groups=False).round(1)
print(' 各置信度准确率:', confidence.to_dict(), ' 条数:', test['confidence'].value_counts().to_dict())
print(' 时间窗:', {w['category']: w['window_min'] for w in core.window_table()})
summary.append((name, accuracy))
if floor is not None:
check(f'{name.split()[0]} 准确率 ≥ {floor}%', accuracy >= floor, f'({accuracy}%)')
else:
# 缺失场景的理论上限:真根因告警仍在的比例;判定应接近它,且“高/中”置信度的判定仍应可靠
reliable = test[test['confidence'].isin(['高', '中'])]
rel_acc = (reliable['cause_category'] == reliable['root_cause']).mean() * 100
check(f'{name.split()[0]} 高/中置信度判定准确率 ≥ 90%', rel_acc >= 90, f'({rel_acc:.1f}%,'
f'占 {len(reliable) / len(test) * 100:.0f}%)')
print('\n汇总(留出集准确率):')
for name, accuracy in summary:
print(f' {accuracy:>6}% {name}')
finish()