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 65b1e79aa7
10 files changed
+869 -119

No files matched your search

+3
View File
@@ -18,3 +18,6 @@ Thumbs.db
desktop.ini
.DS_Store
~$*
# 测试生成的临时数据库与数据
tests/_work/
+28 -12
View File
@@ -25,10 +25,11 @@ Python + Flask + SQLite 后端,React 19 + [Liquid Glass UI](https://tsdsj.gith
```
Local-sever/
├── app.py Flask 路由:认证与权限、增删改查、导入、统计、导出、日志、用户;托管前端页面
├── core.py 核心算法:建库入库、两级根因关联、时长与考核指标
├── core.py 核心算法:建库入库、时间窗标定、三级根因关联、告警关联量化、时长与考核指标
├── importer.py 数据导入:读取 Excel/CSV、自动识别文件类型、逐行校验(见下方“支持导入的文件”)
├── data_gen.py 第 18 题数据生成器(Data-Generated-Core-Py/data.py 的改进版,字段口径一致)
├── publish_online.py 同步到 ../Online-sever(1Panel 线上版)
├── bench.py 性能与时长准确率测试(对应开题报告预期目标)
├── tests/ 根因准确率扰动测试、考核指标口径测试(见下方“测试”)
├── data/ sites.csv / data_outage.csv / data_alarm.csv 种子数据,outage.db 自动生成
├── static/ 前端(全部可直接阅读和修改)
│ ├── index.html 页面入口,importmap 把 'vendor' / 'echarts' 指向 vendor/ 下的文件
@@ -42,8 +43,6 @@ Local-sever/
│ │ ├── components/ ui(卡片/指标/标签/下拉)、DataTable(表格+分页)、FormDialog、EChart、icons
│ │ └── pages/ index.js 页面注册表 + 10 个业务页面,一页一个文件
│ └── vendor/ 第三方库(React、液态玻璃组件库、htm、ECharts),压缩文件,不要手改
├── tools/vendor-build/ 只在升级第三方库时使用,见下文
└── _legacy_frontend/ 最早的原生 JS 前端,仅供对照,可删除
```
## 修改前端
@@ -68,7 +67,7 @@ html`<${Section} title="标题" span=${6}><${EChart} option=${opt}/></${Section}
### 升级第三方库(很少需要)
`static/vendor/` 由 `tools/vendor-build/` 一次性生成。只有升级 React、组件库或 ECharts,或需要新的 React API / 新的图表类型时才需要 Node.js 20.19+:
`static/vendor/` 由 `tools/vendor-build/`(不在本仓库中)一次性生成。只有升级 React、组件库或 ECharts,或需要新的 React API / 新的图表类型时才需要 Node.js 20.19+:
```bash
cd tools/vendor-build
@@ -112,12 +111,29 @@ npm run build # 重新生成 ../../static/vendor/
## 核心算法(core.py)
- **退服时长**:一律按“恢复时间 − 开始时间”重算;上报时长保存为 `reported_min`,两者相差超过 1 分钟即标记为“时长差错”。
- **退服率** = 总退服时长 ÷(站点数 × 统计时长);**网络可用率** = 1 − 退服率。
- **超长退服站点**:月均超长退服 ≥ 1 次(单次时长 ≥ 场景阈值)。**高频重复站点**:月均退服次数 ≥ 场景次数阈值。两个阈值都可在页面上配置。
- **根因判定分两级**:
1. 衍生关系:“小区退服”告警带有 `parent_alarm_id` 时,直接取父告警的类别。
2. 时间窗:回溯退服前未清除的同站告警,按优先级判定:割接施工(30 分钟)→ 电源故障(300 分钟)→ 机房环境(180 分钟)→ 传输中断(15 分钟)→ 天馈故障(30 分钟)→ 设备硬件(15 分钟)。
- **统计周期**:顶栏筛选了月份 / 日期范围时取筛选范围(同时有则取交集),否则取全部流水首日 0 点到末日 24 点。
- **退服率** = 总退服时长 ÷(站点数 × 统计周期);**网络可用率** = 1 − 退服率。站点数按场景、片区、站型、厂家、站名关键字筛选。看板、对标与导出用同一口径。
- **超长退服站点**:月均超长退服 ≥ 1 次(单次时长 ≥ 场景阈值)。**高频重复站点**:月均退服次数 ≥ 场景次数阈值。“月均”按统计周期的自然月数计算。**重复退服率** = 高频重复站点数 ÷ 站点数。
- **根因判定分三级**(每条结果带置信度 高 / 中 / 低):
0. 衍生告警归并:带 `parent_alarm_id` 的告警按父告警类别计(停电引起的传输告警归为电源故障)。
1. 衍生关系(高):与退服开始相差 ≤ 2 分钟的“小区退服”告警带父告警时,取父告警类别。
2. 时间窗(中):同站在退服时仍未清除、提前量在该类时间窗内的告警。
3. 疑似(低):时间窗放宽到 3 倍才命中,提示人工复核。
准确率按“上报原因大类 × 判定根因”交叉矩阵计算,样例数据约 98%。
多条命中时依次比较:级别 → 是否随业务恢复清除(±15 分钟)→ 规则优先级(割接施工 → 电源故障 → 机房环境 → 传输中断 → 天馈故障 → 设备硬件)→ 时间早;都没有命中判为“其他”。
- **时间窗标定**:每次全量分析前,用已闭环流水中同站同类告警的提前量 95% 分位数标定各类时间窗(不小于默认值 30/300/180/15/30/15 分钟,不大于 720 分钟;样本少于 30 条时沿用默认值)。
- **告警—退服关联量化**:对每类告警计算支持度(退服前 W 分钟内有该类告警的退服占比)、置信度(该类告警其后 W 分钟内发生退服的比例)、提升度(置信度 ÷ 随机窗口遇到退服的概率)。
- **断站分级**:重要 / 主要 / 次要,对应不同的处置响应时限。
- **影响评估**:受影响用户 = 小区数 × 单小区在线用户 × 忙闲时系数;业务损失 = 小区话务 × 退服时长。
- **影响评估**:受影响用户 = 小区数 × 单小区在线用户 × 忙闲时系数 × 覆盖补偿系数;受影响用户·小时 = 受影响用户 × 退服时长;业务损失 = 小区话务 × 忙闲时系数 × 覆盖补偿系数 × 退服时长。覆盖补偿:500 米内每有一个邻站下降 10%(室分 5%),最低 40%。
## 测试
测试都在 `tests/_work/` 下建临时数据库,不影响 `data/outage.db`。
```bash
python tests/test_rootcause.py # 根因判定准确率:基线 + 6 组扰动数据(去父告警、时钟偏移、告警缺失、干扰告警、提前量变宽、综合)
python tests/test_kpi.py # 统计周期、稀疏筛选、月均口径、看板与导出一致、影响评估、关联量化
python bench.py # 单批 1 万条导入耗时、全量分析耗时、时长计算准确率
```
根因测试的时间窗只用 1~6 月流水标定,在 7~9 月流水上计分(留出检验)。数据为合成数据,准确率反映的是算法在设定扰动下的表现,论文中应同时说明扰动设置。
+40 -18
View File
@@ -5,7 +5,7 @@
接口分组(与下方代码顺序一致):
页面与认证 / /login /logout /healthz
看板与统计 /api/meta /api/dashboard /api/stats /api/confusion /api/sites/rank
看板与统计 /api/meta /api/dashboard /api/stats /api/confusion /api/association /api/sites/rank
退服流水 /api/outages(查询、新增、修改、删除、批量) /api/analyze
数据导入 /api/import/preview /api/import/confirm /api/template
站点与场景 /api/sites /api/scenes
@@ -154,7 +154,8 @@ def dashboard():
started = time.time()
filters = current_filters()
df = core.frame(filters)
stats = core.site_stats(df)
minutes, months = core.period(filters)
stats = core.site_stats(df, months)
# 月度趋势:次数与时长(小时)
trend = df.groupby('month').agg(n=('id', 'size'), h=('duration_min', 'sum')).reset_index()
@@ -175,11 +176,12 @@ def dashboard():
scene_sites = core.site_count({**filters, 'scene': scene})
if scene_sites == 0:
continue
scene_kpi = core.kpis(df[df['scene'] == scene], scene_sites)
scene_kpi = core.kpis(df[df['scene'] == scene], scene_sites, minutes, months)
bench.append(dict(scene=scene, **scene_kpi))
return jsonify(
kpi=core.kpis(df, core.site_count(filters)),
kpi=core.kpis(df, core.site_count(filters), minutes, months),
period=dict(days=round(minutes / 1440, 1), months=months),
elapsed=round(time.time() - started, 3),
rank=to_records(stats.head(15), ['site_name', 'scene', 'total_min', 'count']),
trend=trend.to_dict('records'),
@@ -209,10 +211,10 @@ def stats():
grouped = df.groupby(dimension).agg(
count=('id', 'size'), total_min=('duration_min', 'sum'), avg_min=('duration_min', 'mean'),
max_min=('duration_min', 'max'), long_count=('is_long', 'sum'), users=('affected_users', 'sum'),
loss_gb=('loss_gb', 'sum'), sites=('site_name', 'nunique'),
user_h=('user_hours', 'sum'), loss_gb=('loss_gb', 'sum'), sites=('site_name', 'nunique'),
).reset_index()
grouped['long_ratio'] = (grouped['long_count'] / grouped['count'] * 100).round(2)
grouped = grouped.round({'total_min': 1, 'avg_min': 1, 'loss_gb': 1})
grouped = grouped.round({'total_min': 1, 'avg_min': 1, 'loss_gb': 1, 'user_h': 1})
sort = request.args.get('sort', 'total_min')
if sort not in grouped.columns:
@@ -231,14 +233,29 @@ def confusion():
df = df[df['cause_category'].isin(core.CAUSES)]
matrix = pd.crosstab(df['cause_category'], df['root_cause'])
matrix = matrix.reindex(index=core.CAUSES, columns=core.CAUSES, fill_value=0)
return jsonify(causes=core.CAUSES, matrix=matrix.values.tolist())
# 按置信度分组的准确率:用来说明“高 / 中 / 低”置信度标签是否可信
by_confidence = []
for level in ['高', '中', '低']:
part = df[df['confidence'] == level]
if len(part):
by_confidence.append(dict(level=level, n=len(part),
acc=round((part['cause_category'] == part['root_cause']).mean() * 100, 2)))
return jsonify(causes=core.CAUSES, matrix=matrix.values.tolist(), by_confidence=by_confidence)
@app.route('/api/association')
@login_required
def association():
"""告警类别 → 退服 的关联强度(支持度、置信度、提升度)与当前使用的时间窗。"""
return jsonify(rows=core.association(current_filters()), windows=core.window_table())
@app.route('/api/sites/rank')
@login_required
def site_rank():
"""超长退服站点(kind=long)/ 高频重复站点(kind=repeat)清单,附处置建议。"""
stats = core.site_stats(core.frame(current_filters()))
filters = current_filters()
stats = core.site_stats(core.frame(filters), core.period(filters)[1])
kind = request.args.get('kind', 'long')
if kind == 'long':
stats = stats[stats['long']]
@@ -254,8 +271,8 @@ def site_rank():
# ================================================================ 退服流水管理
OUTAGE_LIST_COLUMNS = ['id', 'record_id', 'ne_id', 'cell_id', 'site_name', 'scene', 'site_type', 'region', 'vendor',
'start_time', 'end_time', 'duration_min', 'reported_min', 'dur_diff', 'cause_category',
'cause_detail', 'cause_code', 'ticket_id', 'root_cause', 'evidence', 'level', 'status',
'handler', 'remark', 'affected_users', 'loss_gb', 'advice']
'cause_detail', 'cause_code', 'ticket_id', 'root_cause', 'evidence', 'confidence', 'level',
'status', 'handler', 'remark', 'affected_users', 'user_hours', 'loss_gb', 'advice']
# 新增 / 管理员修改时写入的字段,顺序与 outage_values() 返回值一致
OUTAGE_EDIT_FIELDS = ['record_id', 'ne_id', 'cell_id', 'start_time', 'end_time', 'duration_min', 'cause_category',
@@ -380,12 +397,12 @@ def outage_batch():
@app.route('/api/analyze', methods=['POST'])
@admin_required
def reanalyze():
"""对全部退服流水重新做根因分析(修改阈值或导入告警后使用)。"""
"""先按历史流水重新标定时间窗,再对全部退服流水重新做根因分析(导入告警或流水后使用)。"""
started = time.time()
count = core.analyze_all()
accuracy = core.kpis(core.frame())['acc']
log(session['user'], '全量根因分析', f'{count} 条')
return jsonify(n=count, elapsed=round(time.time() - started, 3), acc=accuracy)
return jsonify(n=count, elapsed=round(time.time() - started, 3), acc=accuracy, windows=core.window_table())
# ================================================================ 数据导入
@@ -523,6 +540,7 @@ EXPORT_COLUMN_NAMES = {
'duration_min': '系统计算时长(分)', 'reported_min': '上报时长(分)', 'cause_category': '上报原因大类',
'cause_detail': '原因细项', 'cause_code': '原因码', 'ticket_id': '工单号', 'root_cause': '系统判定根因',
'evidence': '关联告警', 'level': '断站分级', 'status': '闭环状态', 'advice': '处置建议', 'record_id': '流水号',
'confidence': '判定置信度', 'user_h': '受影响用户·小时', 'user_hours': '受影响用户·小时',
}
@@ -546,20 +564,23 @@ def excel_response(sheets, filename):
@login_required
def export(kind):
"""三类报表:stats 退服统计表、long 超长退服站点清单、advice 根因处置建议清单。"""
df = core.frame(current_filters())
stats = core.site_stats(df)
filters = current_filters()
df = core.frame(filters)
minutes, months = core.period(filters)
stats = core.site_stats(df, months)
stamp = datetime.now().strftime('%Y%m%d%H%M')
if kind == 'stats':
k = core.kpis(df)
k = core.kpis(df, core.site_count(filters), minutes, months) # 与看板同一口径
kpi_rows = [
['退服次数', k['count']], ['累计退服时长(小时)', k['total_h']], ['平均退服时长(分钟)', k['avg_min']],
['退服率(%)', k['outage_rate']], ['网络可用率(%)', k['availability']], ['超长退服占比(%)', k['long_ratio']],
['重复退服率(%)', k['repeat_rate']], ['受影响用户数', k['users']], ['业务损失(GB)', k['loss_gb']],
['重复退服率(%)', k['repeat_rate']], ['受影响用户数', k['users']], ['受影响用户·小时', k['user_h']],
['业务损失(GB)', k['loss_gb']], ['统计站点数', k['sites']], ['统计周期(天)', round(minutes / 1440, 1)],
['根因判定准确率(%)', k['acc']], ['闭环率(%)', k['closed']], ['上报时长差错(条)', k['dur_fix']],
]
site_columns = ['ne_id', 'site_name', 'region', 'scene', 'site_type', 'count', 'total_min', 'avg_min',
'max_min', 'long_count', 'users', 'top_cause']
'max_min', 'long_count', 'users', 'user_h', 'top_cause']
sheets = {
'考核指标': pd.DataFrame(kpi_rows, columns=['指标', '数值']),
'站点退服统计': stats[site_columns].rename(columns=EXPORT_COLUMN_NAMES),
@@ -589,7 +610,8 @@ def export(kind):
for cause, level in zip(long_outages['root_cause'], long_outages['level'])]
long_outages = long_outages.sort_values('duration_min', ascending=False)
columns = ['record_id', 'ne_id', 'site_name', 'scene', 'start_time', 'end_time', 'duration_min',
'cause_category', 'cause_code', 'root_cause', 'evidence', 'level', 'status', 'ticket_id', 'advice']
'cause_category', 'cause_code', 'root_cause', 'evidence', 'confidence', 'level', 'status',
'ticket_id', 'advice']
sheets = {'根因处置建议': long_outages[columns].rename(columns=EXPORT_COLUMN_NAMES)}
title = '根因处置建议清单'
+71
View File
@@ -0,0 +1,71 @@
"""性能与时长计算准确率测试(对应开题报告的预期目标),走 Flask 接口,结果可直接写入论文测试章节。
- 单批 10000 条退服流水导入(预览校验 + 确认入库 + 新记录根因分析)≤ 15 秒
- 单次全量退服时长计算与根因分析 ≤ 8 秒
- 退服时长计算准确率 ≥ 98%(以生成器真值为准),并核对人工上报时长差错的检出
运行:python bench.py 数据库建在 tests/_work 下,不影响 data/outage.db
"""
import io
import os
import sys
import time
sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tests'))
import harness # noqa: E402 设置数据库路径
from harness import check, core, finish, generate # noqa: E402
if os.path.exists(os.environ['OUTAGE_DB']):
os.remove(os.environ['OUTAGE_DB'])
core.DB_PATH = os.environ['OUTAGE_DB']
started = time.time()
import app as webapp # noqa: E402 导入时 init_db:建库 + 样例数据入库 + 全量分析
print(f'首次建库、样例入库与全量分析: {time.time() - started:.2f} s')
client = webapp.app.test_client()
client.post('/login', json={'username': 'admin', 'password': 'admin123'})
def timed(label, request):
begin = time.time()
response = request()
elapsed = time.time() - begin
print(f'{label}: {elapsed:.2f} s(HTTP {response.status_code})')
return response, elapsed
# 10000 条新流水:同一站点库,取生成序列第 6001~16000 条,避免与样例数据重复
_, alarms_big, outages_big = generate('bench', rows=16000)
batch = outages_big.iloc[6000:16000]
upload = io.BytesIO(batch.to_csv(index=False).encode('utf-8-sig'))
response, t_preview = timed('导入预览 10000 条', lambda: client.post(
'/api/import/preview', data={'file': (upload, 'batch.csv')}, content_type='multipart/form-data'))
preview = response.get_json()
print(f' 合格 {preview["ok"]},问题 {preview["bad"]},上报时长差错 {preview["diff"]}')
response, t_confirm = timed('导入确认(入库 + 新记录根因分析)', lambda: client.post(
'/api/import/confirm', json={'token': preview['token']}))
check('单批 10000 条导入 ≤ 15 s', t_preview + t_confirm <= 15, f'({t_preview + t_confirm:.2f} s)')
upload = io.BytesIO(alarms_big.to_csv(index=False).encode('utf-8-sig'))
response, _ = timed(f'告警预览 {len(alarms_big)} 条', lambda: client.post(
'/api/import/preview', data={'file': (upload, 'alarm.csv')}, content_type='multipart/form-data'))
timed('告警确认(入库 + 全量重新标定与关联)', lambda: client.post(
'/api/import/confirm', json={'token': response.get_json()['token']}))
total = core.query('SELECT COUNT(*) AS n FROM outages')[0]['n']
response, t_analyze = timed(f'全量根因分析 /api/analyze({total} 条,含时间窗标定)',
lambda: client.post('/api/analyze'))
core.data_changed()
response, t_dashboard = timed('看板首次请求(全量时长与指标重算)', lambda: client.get('/api/dashboard'))
check('全量时长计算与根因分析 ≤ 8 s', t_analyze + t_dashboard <= 8, f'({t_analyze + t_dashboard:.2f} s)')
df = core.full_frame()
true_minutes = (core.pd.to_datetime(df['end_time']) - df['start']).dt.total_seconds() / 60
accuracy = ((df['duration_min'] - true_minutes).abs() < 0.05).mean() * 100
check('退服时长计算准确率 ≥ 98%', accuracy >= 98, f'({accuracy:.2f}%)')
injected = (df['reported_min'] - true_minutes).abs() > 1
missed = int((injected & ~df['dur_diff']).sum())
false_alarm = int((~injected & df['dur_diff']).sum())
check('人工上报时长差错全部检出且无误报', missed == 0 and false_alarm == 0,
f'(注入 {int(injected.sum())} 条,标记 {int(df["dur_diff"].sum())} 条)')
finish()
+363 -75
View File
@@ -38,9 +38,10 @@ SCENE_DEFAULT = {
'工业园区': (180, 5, 120, 7),
}
# 时间窗规则:(告警类别, 判定根因, 告警最多早于退服多少分钟),按优先级从高到低排列。
# 时间窗规则:(告警类别, 判定根因, 默认时间窗:告警最多早于退服多少分钟),按优先级从高到低排列。
# 割接操作最先排除;停电有蓄电池后备、机房高温有过热保护延时,所以窗口放宽;
# 停电往往同时引起传输告警,因为电源优先级更高,这种情况归为电源故障。
# 实际使用的窗口由 calibrate_windows() 按历史流水标定(只会放宽,不会小于这里的默认值)。
RULES = [
('割接施工', '割接施工', 30),
('电源故障', '电源故障', 300),
@@ -49,6 +50,21 @@ RULES = [
('天馈故障', '天馈故障', 30),
('设备硬件', '设备硬件', 15),
]
DEFAULT_WINDOW = {category: window for category, _, window in RULES}
PRIORITY = {category: rank for rank, (category, _, _) in enumerate(RULES)}
MATCH_TOL_MIN = 2 # 第一级:“小区退服”告警与退服开始时间允许相差的分钟数(网管与工单系统时钟不完全同步)
CLEAR_TOL_MIN = 15 # 告警清除与业务恢复相差在此范围内,视为“随业务恢复清除”,作为根因的佐证
SUSPECT_FACTOR = 3 # 第三级:时间窗放宽到几倍仍命中的,判为“疑似”根因(置信度低)
WINDOW_MAX = 720 # 标定后的时间窗上限(分钟)
CALIB_QUANTILE = 0.95 # 标定取历史提前量的分位数
CALIB_MIN_SAMPLES = 30 # 某类样本少于此数时不标定,沿用默认窗口
# 覆盖补偿:退服站点周边 COVER_RADIUS_M 米内每有一个邻站,受影响用户和话务损失按 COVER_STEP 递减(邻站吸收部分话务),
# 最低保留 COVER_FLOOR。室分站主要覆盖室内,室外邻站补偿有限,递减幅度减半。
COVER_RADIUS_M = 500
COVER_STEP = 0.1
COVER_FLOOR = 0.4
ADVICE = {
'电源故障': ['核查市电及开关电源状态,必要时调度油机发电', '检测蓄电池组容量,老化电池纳入更换计划', '评估加装智能电表与远程动环监控'],
@@ -123,23 +139,38 @@ CREATE TABLE IF NOT EXISTS sites(ne_id TEXT PRIMARY KEY, site_name TEXT, distric
CREATE TABLE IF NOT EXISTS outages(id INTEGER PRIMARY KEY, record_id TEXT, ne_id TEXT, cell_id TEXT,
start_time TEXT, end_time TEXT, duration_min REAL, reported_min REAL, cause_category TEXT, cause_detail TEXT,
cause_code TEXT, ticket_id TEXT, root_cause TEXT, evidence TEXT, status TEXT DEFAULT '待处理',
handler TEXT, remark TEXT);
handler TEXT, remark TEXT, confidence TEXT);
CREATE INDEX IF NOT EXISTS ix_out_ne ON outages(ne_id, start_time);
CREATE TABLE IF NOT EXISTS alarms(alarm_id TEXT PRIMARY KEY, ne_id TEXT, alarm_code TEXT, alarm_name TEXT,
alarm_level TEXT, category TEXT, start_time TEXT, clear_time TEXT, duration_s INTEGER, cleared TEXT,
parent_alarm_id TEXT);
CREATE INDEX IF NOT EXISTS ix_alarm_ne ON alarms(ne_id, start_time);
CREATE TABLE IF NOT EXISTS logs(id INTEGER PRIMARY KEY, username TEXT, action TEXT, detail TEXT, ts TEXT);
CREATE TABLE IF NOT EXISTS windows(category TEXT PRIMARY KEY, window_min REAL, default_min REAL, samples INTEGER,
source TEXT, updated_at TEXT);
"""
def migrate():
"""旧版数据库补字段:outages.confidence(根因判定置信度)。补了字段返回 True,需要重新做一次全量分析。"""
with db() as conn:
columns = {row['name'] for row in conn.execute('PRAGMA table_info(outages)')}
if 'confidence' not in columns:
conn.execute('ALTER TABLE outages ADD COLUMN confidence TEXT')
return True
return False
def init_db():
"""启动时调用:建表;数据库文件是第一次创建时,再写入默认账号、场景阈值,并导入 data/ 下的样例数据。"""
is_first_run = not os.path.exists(DB_PATH)
os.makedirs(os.path.dirname(DB_PATH), exist_ok=True)
with db() as conn:
conn.executescript(SCHEMA)
upgraded = migrate()
if not is_first_run:
if upgraded:
analyze_all() # 旧库升级:按新算法重新标定时间窗并补上置信度
return
now = datetime.now().strftime(TIME_FMT)
@@ -218,88 +249,279 @@ def save_sites(df):
# ================================================================ 5. 根因关联分析
def analyze(outages):
def read_alarms(ne_ids=None, categories=None):
"""读取告警(可按基站、类别过滤),时间列解析为 datetime,开始时间列改名为 alarm_time。"""
conditions = []
args = []
if ne_ids is not None:
ne_ids = list(ne_ids)
if not ne_ids:
return pd.DataFrame(columns=['alarm_id', 'ne_id', 'alarm_name', 'category', 'alarm_time', 'clear_time',
'parent_alarm_id'])
# 基站很多时(导入全网工参后)不用 IN (?, ?, ...),避免超过 SQLite 的参数个数上限
conditions.append('ne_id IN (SELECT value FROM json_each(?))')
args.append(pd.Series(ne_ids).to_json(orient='values', force_ascii=False))
if categories is not None:
conditions.append('category IN (SELECT value FROM json_each(?))')
args.append(pd.Series(list(categories)).to_json(orient='values', force_ascii=False))
sql = 'SELECT alarm_id, ne_id, alarm_name, category, start_time, clear_time, parent_alarm_id FROM alarms'
if conditions:
sql += ' WHERE ' + ' AND '.join(conditions)
with db() as conn:
alarms = pd.read_sql(sql, conn, params=args)
alarms['start_time'] = pd.to_datetime(alarms['start_time'], errors='coerce', format='mixed')
alarms['clear_time'] = pd.to_datetime(alarms['clear_time'], errors='coerce', format='mixed')
alarms = alarms.dropna(subset=['start_time'])
return alarms.rename(columns={'start_time': 'alarm_time'})
def still_active(alarm_clear, outage_start):
"""告警在退服开始时仍未清除(允许 1 分钟误差)才可能是退服的原因。"""
return alarm_clear.isna() | (alarm_clear >= outage_start - pd.Timedelta(minutes=1))
def calibrate_windows(until=None):
"""
两级根因判定。outages 需含 id, ne_id, start_time 三列,返回 [(根因, 证据, id), ...]。
第一级 衍生关系:同站、同一时刻的“小区退服”告警如果带 parent_alarm_id,直接取父告警的类别
(网管已经给出因果链,最可靠);
第二级 时间窗: 否则回溯同站在退服前仍未清除的告警,按 RULES 的优先级取第一个命中的类别;
都没有命中判为“其他”。
用历史流水标定各类根因告警的时间窗(告警最多早于退服多少分钟),结果写入 windows 表并返回 {类别: 分钟}。
样本:已闭环、上报原因属于规则类别的退服(until 指定时只用该日期之前的,便于留出检验);
在同站、同类别、退服时仍未清除、提前量不超过 WINDOW_MAX 的告警中,取离退服最近的一条,记其提前量;
窗口:提前量的 95% 分位数,不小于默认窗口、不大于 WINDOW_MAX;样本少于 30 条时沿用默认窗口。
这样时间窗随本地网络的实际情况(蓄电池后备时长、动环告警延时等)自动调整,而不是写死在代码里。
"""
categories = list(DEFAULT_WINDOW)
sql = ("SELECT id, ne_id, start_time, cause_category FROM outages WHERE status = '已闭环' "
"AND cause_category IN (SELECT value FROM json_each(?))")
args = [pd.Series(categories).to_json(orient='values', force_ascii=False)]
if until:
sql += ' AND start_time < ?'
args.append(until)
with db() as conn:
outages = pd.read_sql(sql, conn, params=args)
outages['start_time'] = pd.to_datetime(outages['start_time'])
alarms = read_alarms(outages['ne_id'].unique(), categories)
pairs = outages.merge(alarms, left_on=['ne_id', 'cause_category'], right_on=['ne_id', 'category'])
lead = (pairs['start_time'] - pairs['alarm_time']).dt.total_seconds() / 60
usable = still_active(pairs['clear_time'], pairs['start_time']) & (lead >= -5) & (lead <= WINDOW_MAX)
pairs = pairs[usable].assign(lead=lead[usable].clip(lower=0))
nearest = pairs.sort_values('lead').drop_duplicates('id')
now = datetime.now().strftime(TIME_FMT)
result = {}
rows = []
for category in categories:
leads = nearest.loc[nearest['cause_category'] == category, 'lead']
default = DEFAULT_WINDOW[category]
if len(leads) >= CALIB_MIN_SAMPLES:
window = float(min(max(leads.quantile(CALIB_QUANTILE), default), WINDOW_MAX))
source = '历史标定'
else:
window = float(default)
source = '默认(样本不足)'
result[category] = round(window, 1)
rows.append((category, round(window, 1), default, int(len(leads)), source, now))
with db() as conn:
conn.execute('DELETE FROM windows')
conn.executemany('INSERT INTO windows VALUES(?,?,?,?,?,?)', rows)
return result
def load_windows():
"""当前使用的时间窗 {类别: 分钟};未标定的类别用默认窗口。"""
windows = dict(DEFAULT_WINDOW)
for row in query('SELECT category, window_min FROM windows'):
if row['category'] in windows and row['window_min']:
windows[row['category']] = row['window_min']
return windows
def window_table():
"""时间窗明细(页面展示用)。"""
saved = {row['category']: row for row in query('SELECT * FROM windows')}
rows = []
for category, _, default in RULES:
row = saved.get(category) or {'window_min': default, 'samples': 0, 'source': '默认', 'updated_at': ''}
rows.append(dict(category=category, priority=PRIORITY[category] + 1, default_min=default,
window_min=row['window_min'], samples=row['samples'], source=row['source'],
updated_at=row['updated_at']))
return rows
def analyze(outages, windows=None):
"""
三级根因判定。outages 需含 id, ne_id, start_time, end_time 四列,返回 [(根因, 证据, 置信度, id), ...]。
预处理 衍生告警归并:告警带 parent_alarm_id 且父告警属于根因类别时,按父告警的类别和名称计
(如停电引起的传输告警归为电源故障);
第一级 衍生关系(置信度高):同站“小区退服”告警与退服开始相差不超过 MATCH_TOL_MIN 分钟、且带父告警,
直接取父告警类别;
第二级 时间窗(置信度中):同站在退服时仍未清除、提前量在该类时间窗内的告警;
第三级 疑似(置信度低):时间窗放宽到 SUSPECT_FACTOR 倍才命中的告警,提示人工复核;
多条命中时依次比较:级别 → 是否随业务恢复清除(清除与恢复相差 ≤ CLEAR_TOL_MIN 分钟)→ 规则优先级 → 时间早;
都没有命中判为“其他”。
"""
if outages.empty:
return []
windows = windows or load_windows()
# 只取涉及到的基站的告警,避免全表读取
ne_ids = tuple(outages['ne_id'].dropna().unique())
placeholders = ','.join('?' * len(ne_ids))
with db() as conn:
alarms = pd.read_sql(f'SELECT alarm_id, ne_id, alarm_name, category, start_time, clear_time, parent_alarm_id '
f'FROM alarms WHERE ne_id IN ({placeholders})', conn, params=ne_ids,
parse_dates=['start_time', 'clear_time'])
alarms = alarms.rename(columns={'start_time': 'alarm_time'})
outages = outages[['id', 'ne_id', 'start_time']].copy()
outages = outages[['id', 'ne_id', 'start_time', 'end_time']].copy()
outages['start_time'] = pd.to_datetime(outages['start_time'])
outages['end_time'] = pd.to_datetime(outages['end_time'])
alarms = read_alarms(outages['ne_id'].dropna().unique())
results = {} # 退服 id -> (根因, 证据)
# ---- 预处理:衍生告警按父告警计
alarms['root_cat'] = alarms['category']
alarms['root_name'] = alarms['alarm_name']
parents = alarms.drop_duplicates('alarm_id').set_index('alarm_id')
parent_id = alarms['parent_alarm_id'].fillna('')
parent_cat = parent_id.map(parents['category'])
linked = parent_cat.isin(DEFAULT_WINDOW)
alarms.loc[linked, 'root_cat'] = parent_cat[linked]
alarms.loc[linked, 'root_name'] = parent_id[linked].map(parents['alarm_name'])
# 证据里显示的告警名:衍生告警写成“传输告警(衍生自 市电停电)”
alarms['show_name'] = alarms['alarm_name']
alarms.loc[linked, 'show_name'] = alarms.loc[linked, 'alarm_name'] + '(衍生自 ' + alarms.loc[linked, 'root_name'] + ')'
# ---- 第一级:衍生关系。找到带父告警的“小区退服”告警,再查出父告警的类别
child_alarms = alarms[(alarms['category'] == '小区退服') & alarms['parent_alarm_id'].fillna('').ne('')]
child_alarms = child_alarms[['ne_id', 'alarm_time', 'parent_alarm_id']].rename(columns={'alarm_time': 'start_time'})
parent_alarms = alarms.drop(columns=['ne_id', 'parent_alarm_id'])
child_with_parent = child_alarms.merge(parent_alarms, left_on='parent_alarm_id', right_on='alarm_id')
matched = outages.merge(child_with_parent, on=['ne_id', 'start_time']).drop_duplicates('id')
results = {} # 退服 id -> (根因, 证据, 置信度)
# ---- 第一级:衍生关系
child = alarms[(alarms['category'] == '小区退服') & linked][['ne_id', 'alarm_time', 'root_cat', 'root_name']]
matched = outages.merge(child, on='ne_id')
gap = (matched['alarm_time'] - matched['start_time']).abs()
matched = matched[gap <= pd.Timedelta(minutes=MATCH_TOL_MIN)].assign(gap=gap)
matched = matched.sort_values(['id', 'gap']).drop_duplicates('id')
for row in matched.itertuples():
if row.category in CAUSES:
evidence = f'衍生关系:{row.alarm_name}@{row.alarm_time:%m-%d %H:%M}'
results[row.id] = (row.category, evidence)
results[row.id] = (row.root_cat, f'衍生关系:{row.root_name} → 小区退服@{row.alarm_time:%m-%d %H:%M}', '高')
# ---- 第二级:时间窗。把剩余退服与同站的根因类告警两两配对,逐条规则打标
# ---- 第二、三级:时间窗与疑似
remaining = outages[~outages['id'].isin(results)]
pairs = remaining.merge(alarms[alarms['category'] != '小区退服'], on='ne_id')
minutes_before = (pairs['start_time'] - pairs['alarm_time']).dt.total_seconds() / 60
# 告警在退服时仍未清除(允许 1 分钟误差)才可能是原因
still_active = pairs['clear_time'].isna() | (pairs['clear_time'] >= pairs['start_time'] - pd.Timedelta(minutes=1))
candidates = alarms[(alarms['category'] != '小区退服') & alarms['root_cat'].isin(DEFAULT_WINDOW)]
pairs = remaining.merge(candidates, on='ne_id')
lead = (pairs['start_time'] - pairs['alarm_time']).dt.total_seconds() / 60
window = pairs['root_cat'].map(windows).astype(float)
active = still_active(pairs['clear_time'], pairs['start_time']) & (lead >= -5)
pairs['tier'] = np.select([active & (lead <= window), active & (lead <= window * SUSPECT_FACTOR)], [0, 1], 9)
clear_gap = (pairs['clear_time'] - pairs['end_time']).abs()
pairs['sync'] = clear_gap <= pd.Timedelta(minutes=CLEAR_TOL_MIN)
pairs['unsync'] = ~pairs['sync']
pairs['prio'] = pairs['root_cat'].map(PRIORITY)
pairs['rank'] = 99 # 命中规则的序号,越小优先级越高;99 表示没有命中
for rank, (category, _, window) in enumerate(RULES):
hit = ((pairs['category'] == category) & (minutes_before <= window) & (minutes_before >= -5)
& still_active & (pairs['rank'] > rank))
pairs.loc[hit, 'rank'] = rank
# 每条退服取优先级最高、时间最早的一条告警作为根因
best = pairs[pairs['rank'] < 99].sort_values(['id', 'rank', 'alarm_time']).drop_duplicates('id')
best = pairs[pairs['tier'] < 9].sort_values(['id', 'tier', 'unsync', 'prio', 'alarm_time']).drop_duplicates('id')
for row in best.itertuples():
evidence = f'时间窗:{row.alarm_name}@{row.alarm_time:%m-%d %H:%M}'
results[row.id] = (RULES[row.rank][1], evidence)
sync_note = ',随业务恢复清除' if row.sync else ''
if row.tier == 0:
results[row.id] = (row.root_cat, f'时间窗:{row.show_name}@{row.alarm_time:%m-%d %H:%M}{sync_note}', '中')
else:
results[row.id] = (row.root_cat, f'疑似(超出时间窗):{row.show_name}@{row.alarm_time:%m-%d %H:%M}'
f'{sync_note},建议人工复核', '低')
output = []
for outage_id in outages['id']:
cause, evidence = results.get(outage_id, ('其他', '时间窗内无关联告警'))
output.append((cause, evidence, int(outage_id)))
cause, evidence, confidence = results.get(outage_id, ('其他', '时间窗内无关联告警', '低'))
output.append((cause, evidence, confidence, int(outage_id)))
return output
def analyze_all(ids=None):
"""对指定 id(不传则全部)的退服流水重新做根因分析并写回数据库,返回分析条数。"""
sql = 'SELECT id, ne_id, start_time FROM outages'
if ids:
sql += ' WHERE id IN (' + ','.join(str(i) for i in ids) + ')'
def analyze_all(ids=None, recalibrate=True):
"""
对指定 id 的退服流水(ids=None 表示全部)重新做根因分析并写回数据库,返回分析条数。
全量分析前先按历史流水重新标定时间窗(recalibrate=False 时沿用已有窗口)。
"""
if ids is not None and len(ids) == 0:
return 0
if ids is None and recalibrate:
calibrate_windows()
sql = 'SELECT id, ne_id, start_time, end_time FROM outages'
if ids is not None:
sql += ' WHERE id IN (' + ','.join(str(int(i)) for i in ids) + ')'
with db() as conn:
outages = pd.read_sql(sql, conn)
results = analyze(outages)
with db() as conn:
conn.executemany('UPDATE outages SET root_cause=?, evidence=? WHERE id=?', results)
conn.executemany('UPDATE outages SET root_cause=?, evidence=?, confidence=? WHERE id=?', results)
data_changed()
return len(results)
def association(filters=None):
"""
告警类别与退服的关联强度,规则为“同站出现 X 类告警 → 其后 W 分钟内发生退服”(W 为该类时间窗):
告警数 n_A、命中数 n_AB(其后 W 分钟内同站发生退服的告警数)
支持度 support = 开始前 W 分钟内同站有 X 类告警的退服数 ÷ 退服总数,即 P(X 类告警 | 退服)
置信度 confidence = n_AB ÷ n_A,即 P(退服 | X 类告警)
提升度 lift = 置信度 ÷ P0,P0 = 退服数 × W ÷(站点数 × 统计时长),即随机一个 W 分钟窗口内遇到退服的概率
提升度 > 1 说明该类告警出现后退服概率高于随机水平,数值越大关联越强。
"""
filters = filters or {}
df = frame(filters)
lo, hi = period_bounds(filters)
minutes = max((hi - lo).total_seconds() / 60, 1440)
n_sites = max(site_count(filters), 1)
windows = load_windows()
site_where, site_args = site_conditions(filters)
ne_ids = [row['ne_id'] for row in query('SELECT ne_id FROM sites' + site_where, site_args)]
alarms = read_alarms(ne_ids, list(DEFAULT_WINDOW))
alarms = alarms[(alarms['alarm_time'] >= lo) & (alarms['alarm_time'] < hi)]
outages = df[['id', 'ne_id', 'start']].dropna().sort_values('start')
rows = []
for category, _, _ in RULES:
window = windows[category]
tolerance = pd.Timedelta(minutes=window)
cat_alarms = alarms.loc[alarms['category'] == category, ['ne_id', 'alarm_time']].sort_values('alarm_time')
hits = 0
if len(cat_alarms) and len(outages):
forward = pd.merge_asof(cat_alarms, outages, left_on='alarm_time', right_on='start', by='ne_id',
direction='forward', tolerance=tolerance)
hits = int(forward['start'].notna().sum())
covered = 0
if len(cat_alarms) and len(outages):
backward = pd.merge_asof(outages, cat_alarms, left_on='start', right_on='alarm_time', by='ne_id',
direction='backward', tolerance=tolerance)
covered = int(backward['alarm_time'].notna().sum())
p0 = min(len(outages) * window / (n_sites * minutes), 1) if len(outages) else 0
confidence = hits / len(cat_alarms) if len(cat_alarms) else 0
rows.append(dict(
category=category, window_min=window, alarms=len(cat_alarms), hits=hits,
support=round(covered / len(outages) * 100, 2) if len(outages) else 0,
confidence=round(confidence * 100, 2),
lift=round(confidence / p0, 2) if p0 else 0,
))
return rows
# ================================================================ 6. 指标核算
def site_cover():
"""
各站覆盖补偿系数 {ne_id: 系数}:COVER_RADIUS_M 米内的邻站数 n,系数 = max(COVER_FLOOR, 1 − COVER_STEP × n),
室分站递减幅度减半。距离按经纬度近似换算(城市范围内误差可忽略);没有经纬度的站系数为 1。
"""
sites = pd.DataFrame(query('SELECT ne_id, site_type, lng, lat FROM sites'))
if sites.empty:
return {}
sites = sites.dropna(subset=['lng', 'lat'])
if sites.empty:
return {}
lat0 = np.radians(sites['lat'].astype(float).mean())
x = sites['lng'].to_numpy(float) * 111320 * np.cos(lat0)
y = sites['lat'].to_numpy(float) * 110540
neighbors = np.zeros(len(sites), dtype=int)
for begin in range(0, len(sites), 1000): # 分块计算,站点很多时不会一次占用大量内存
dx = x[begin:begin + 1000, None] - x[None, :]
dy = y[begin:begin + 1000, None] - y[None, :]
neighbors[begin:begin + 1000] = ((dx * dx + dy * dy) <= COVER_RADIUS_M ** 2).sum(axis=1) - 1
step = np.where(sites['site_type'].eq('室分'), COVER_STEP / 2, COVER_STEP)
factor = np.clip(1 - step * neighbors, COVER_FLOOR, 1)
return dict(zip(sites['ne_id'], factor.round(3)))
def full_frame():
"""
全部退服流水 + 站点信息 + 场景阈值,并计算派生字段:
is_long 是否超长、dur_diff 上报时长差错、affected_users 受影响用户、loss_gb 业务损失、level 断站分级。
is_long 是否超长、dur_diff 上报时长差错、cover_factor 覆盖补偿系数、affected_users 受影响用户、
user_hours 受影响用户·小时、loss_gb 业务损失、level 断站分级。
结果缓存到数据变化为止(见 data_changed)。
"""
global _frame_cache
@@ -319,11 +541,18 @@ def full_frame():
df['is_long'] = df['duration_min'] >= long_min
df['dur_diff'] = (df['reported_min'] - df['duration_min']).abs() > 1 # 人工上报时长与系统计算相差超过 1 分钟
# 影响评估:受影响用户 = 小区数 × 单小区在线用户 × 忙闲时系数;业务损失 = 小区数 × 单小区话务 × 系数 × 时长
# 影响评估:
# 受影响用户 = 小区数 × 单小区在线用户 × 忙闲时系数 × 覆盖补偿系数(同一时刻掉线的用户数)
# 受影响用户·小时 = 受影响用户 × 退服时长(小时),反映影响的持续程度
# 业务损失 = 小区数 × 单小区每小时话务 × 忙闲时系数 × 覆盖补偿系数 × 退服时长(小时)
busy_factor = np.where(df['start'].dt.hour.between(9, 22), 1.0, 0.35) # 9-22 点为忙时
cells = df['cells'].fillna(3).astype(float)
df['affected_users'] = (cells * df['users_per_cell'].fillna(200) * busy_factor).round().astype(int)
df['loss_gb'] = (cells * df['gb_per_cell_h'].fillna(10) * busy_factor * df['duration_min'] / 60).round(2)
df['cover_factor'] = df['ne_id'].map(site_cover()).fillna(1.0).astype(float)
hours = df['duration_min'] / 60
users = cells * df['users_per_cell'].fillna(200) * busy_factor * df['cover_factor']
df['affected_users'] = users.round().astype(int)
df['user_hours'] = (users * hours).round(1)
df['loss_gb'] = (cells * df['gb_per_cell_h'].fillna(10) * busy_factor * df['cover_factor'] * hours).round(2)
# 断站分级(按整列一次计算,比逐行判断快几十倍):
# 重要:超长,且是关键场景或时长超过阈值 2 倍
@@ -362,8 +591,49 @@ def frame(filters=None):
return df.copy()
def site_count(filters=None):
"""满足站点类筛选条件(场景、片区、站型、厂家)的站点数,用作退服率的分母。"""
def period_bounds(filters=None):
"""
统计周期 [起, 止)。有月份 / 日期范围筛选时取筛选范围(同时有则取交集),
否则取全部退服流水首日 0 点到末日 24 点。
"""
filters = filters or {}
df = full_frame()
if df.empty:
today = pd.Timestamp.now().normalize()
lo, hi = today, today + pd.Timedelta(days=1)
else:
lo = df['start'].min().normalize()
hi = df['start'].max().normalize() + pd.Timedelta(days=1)
lowers = []
uppers = []
if filters.get('month'):
month_start = pd.Timestamp(filters['month'] + '-01')
lowers.append(month_start)
uppers.append(month_start + pd.offsets.MonthBegin(1))
if filters.get('date_from'):
lowers.append(pd.Timestamp(filters['date_from']))
if filters.get('date_to'):
uppers.append(pd.Timestamp(filters['date_to']) + pd.Timedelta(days=1))
if lowers:
lo = max(lowers)
if uppers:
hi = min(uppers)
if hi <= lo:
hi = lo + pd.Timedelta(days=1)
return lo, hi
def period(filters=None):
"""统计周期的 (分钟数, 自然月数):分钟数是退服率的时间分母,月数是“月均”的分母。"""
lo, hi = period_bounds(filters)
minutes = (hi - lo).total_seconds() / 60
months = len(pd.period_range(lo, hi - pd.Timedelta(seconds=1), freq='M'))
return minutes, max(months, 1)
def site_conditions(filters=None):
"""站点类筛选(场景、片区、站型、厂家、站名/站号关键字)对应的 WHERE 子句与参数。"""
filters = filters or {}
column_of = {'scene': 'scene', 'region': 'district', 'site_type': 'site_type', 'vendor': 'vendor'}
conditions = []
@@ -372,30 +642,42 @@ def site_count(filters=None):
if filters.get(key):
conditions.append(f'{column} = ?')
args.append(filters[key])
sql = 'SELECT COUNT(*) AS n FROM sites'
if conditions:
sql += ' WHERE ' + ' AND '.join(conditions)
return query(sql, args)[0]['n']
if filters.get('site'):
conditions.append('(instr(site_name, ?) > 0 OR instr(ne_id, ?) > 0)') # 与 frame() 一样区分大小写
args += [filters['site'], filters['site']]
if not conditions:
return '', []
return ' WHERE ' + ' AND '.join(conditions), args
def kpis(df, n_sites=None):
def site_count(filters=None):
"""满足站点类筛选条件的站点数,用作退服率、重复退服率的分母。"""
where, args = site_conditions(filters)
return query('SELECT COUNT(*) AS n FROM sites' + where, args)[0]['n']
def kpis(df, n_sites=None, minutes=None, months=None):
"""
运维考核指标。
运维考核指标。n_sites / minutes / months 为站点数、统计时长(分钟)、统计月数,不传时按全量计算。
退服率 = 总退服时长 ÷(站点数 × 统计时长);网络可用率 = 1 − 退服率
重复退服率 = 高频重复站点数 ÷ 站点数
根因判定准确率 = 上报原因与系统判定一致的比例(只统计上报原因已填写的记录)
"""
if not n_sites:
if n_sites is None:
n_sites = site_count()
if minutes is None or months is None:
default_minutes, default_months = period()
minutes = minutes or default_minutes
months = months or default_months
if df.empty:
return dict(count=0, total_h=0, avg_min=0, outage_rate=0, availability=100, long_ratio=0, repeat_rate=0,
users=0, loss_gb=0, sites=n_sites, acc=0, dur_fix=0, closed=0)
users=0, user_h=0, loss_gb=0, sites=n_sites, acc=0, dur_fix=0, closed=0)
total_minutes = df['duration_min'].sum()
period_minutes = max((df['start'].max() - df['start'].min()).total_seconds() / 60, 1)
outage_rate = total_minutes / (n_sites * period_minutes) * 100
denominator = max(n_sites, 1) * max(minutes, 1)
outage_rate = total_minutes / denominator * 100
stats = site_stats(df)
stats = site_stats(df, months)
known = df[df['cause_category'].isin(CAUSES) & df['root_cause'].notna()]
if len(known):
@@ -410,8 +692,9 @@ def kpis(df, n_sites=None):
outage_rate=round(outage_rate, 4),
availability=round(100 - outage_rate, 4),
long_ratio=round(df['is_long'].mean() * 100, 2),
repeat_rate=round(stats['repeat'].mean() * 100, 2),
repeat_rate=round(stats['repeat'].sum() / max(n_sites, 1) * 100, 2),
users=int(df['affected_users'].sum()),
user_h=int(round(float(df['user_hours'].sum()))),
loss_gb=round(df['loss_gb'].sum(), 1),
sites=n_sites,
acc=accuracy,
@@ -420,19 +703,23 @@ def kpis(df, n_sites=None):
)
def site_stats(df):
def site_stats(df, months=None):
"""
站点维度统计,按累计退服时长从大到小排序。
站点维度统计,按累计退服时长从大到小排序。months 为统计周期的自然月数,不传时按全量周期。
超长退服站点 long:月均超长退服 ≥ 1 次
高频重复站点 repeat:月均退服次数 ≥ 场景次数阈值
"""
if df.empty:
return pd.DataFrame(columns=['site_name', 'count', 'total_min', 'long', 'repeat'])
return pd.DataFrame(columns=['ne_id', 'site_name', 'region', 'scene', 'site_type', 'count', 'total_min',
'avg_min', 'max_min', 'long_count', 'users', 'user_h', 'long_min', 'freq_count',
'top_cause', 'long', 'repeat'])
if months is None:
months = period()[1]
stats = df.groupby(['ne_id', 'site_name', 'region', 'scene', 'site_type'], dropna=False).agg(
count=('id', 'size'), total_min=('duration_min', 'sum'), avg_min=('duration_min', 'mean'),
max_min=('duration_min', 'max'), long_count=('is_long', 'sum'), users=('affected_users', 'sum'),
long_min=('long_min', 'first'), freq_count=('freq_count', 'first'),
user_h=('user_hours', 'sum'), long_min=('long_min', 'first'), freq_count=('freq_count', 'first'),
).reset_index()
# 主要根因:每站出现次数最多的根因;次数相同时取名称排序靠前的(与 pandas 的 mode 结果一致)
@@ -441,11 +728,12 @@ def site_stats(df):
top_cause = cause_counts.drop_duplicates('ne_id').set_index('ne_id')['root_cause']
stats['top_cause'] = stats['ne_id'].map(top_cause).fillna('')
months = max(df['month'].nunique(), 1)
months = max(months, 1)
stats['long'] = stats['long_count'] / months >= 1
stats['repeat'] = stats['count'] / months >= stats['freq_count'].fillna(4)
stats['total_min'] = stats['total_min'].round(1)
stats['avg_min'] = stats['avg_min'].round(1)
stats['user_h'] = stats['user_h'].round(1)
return stats.sort_values('total_min', ascending=False)
+16 -7
View File
@@ -9,7 +9,8 @@ const BENCH_COLUMNS = [
{key: 'scene', title: '场景'}, {key: 'sites', title: '站点数'}, {key: 'count', title: '退服次数'},
{key: 'total_h', title: '累计时长(h)'}, {key: 'avg_min', title: '平均时长(min)'}, {key: 'outage_rate', title: '退服率(%)'},
{key: 'availability', title: '可用率(%)'}, {key: 'long_ratio', title: '超长占比(%)'}, {key: 'repeat_rate', title: '重复退服率(%)'},
{key: 'users', title: '受影响用户', render: num}, {key: 'loss_gb', title: '业务损失(GB)', render: num},
{key: 'users', title: '受影响用户', render: num}, {key: 'user_h', title: '用户·小时', render: num},
{key: 'loss_gb', title: '业务损失(GB)', render: num},
];
/** 场景柱状图配置:最大的场景用强调色标出 */
@@ -24,14 +25,15 @@ export default function Impact() {
const {filters} = useApp();
const {data} = useFetch('/api/dashboard', filters);
if (!data) return html`<${Loading}/>`;
const {kpi: k, bench} = data;
const {kpi: k, bench, period} = data;
return html`
<div className="grid">
<${Section} title="影响评估">
<${Section} title="影响评估" hint=${`统计周期 ${period.days} 天(${period.months} 个月)· ${num(k.sites)} 个站`}>
<div className="kpis">
<${Kpi} title="受影响用户" value=${num(k.users)} unit="人次" sub="小区数 × 单小区在线用户 × 忙闲时系数"/>
<${Kpi} title="业务损失" value=${num(k.loss_gb)} unit="GB" sub="小区话务 × 退服时长"/>
<${Kpi} title="受影响用户" value=${num(k.users)} unit="人次" sub="小区数 × 单小区在线用户 × 忙闲时 × 覆盖补偿"/>
<${Kpi} title="受影响用户·小时" value=${num(k.user_h)} unit="人·h" sub="受影响用户 × 退服时长"/>
<${Kpi} title="业务损失" value=${num(k.loss_gb)} unit="GB" sub="小区话务 × 忙闲时 × 覆盖补偿 × 退服时长"/>
<${Kpi} title="退服率" value=${fixed(k.outage_rate)} unit="%"/>
<${Kpi} title="网络可用率" value=${fixed(k.availability)} unit="%"/>
<${Kpi} title="平均退服时长" value=${k.avg_min} unit="min"/>
@@ -42,8 +44,15 @@ export default function Impact() {
<${Section} span=${6} title="分场景退服率" hint="单位:%">
<${EChart} option=${sceneBar(bench, bench.map(x => x.outage_rate))}/>
</${Section}>
<${Section} span=${6} title="分场景受影响用户" hint="单位:万人次">
<${EChart} option=${sceneBar(bench, bench.map(x => +(x.users / 1e4).toFixed(1)))}/>
<${Section} span=${6} title="分场景受影响用户·小时" hint="单位:万人·h">
<${EChart} option=${sceneBar(bench, bench.map(x => +(x.user_h / 1e4).toFixed(1)))}/>
</${Section}>
<${Section} title="口径说明">
<div className="note">
退服率 = 总退服时长 ÷(站点数 × 统计周期);统计周期取顶栏的月份 / 日期范围,未筛选时为全部流水首日至末日。
重复退服率 = 高频重复站点数 ÷ 站点数。“月均”按统计周期的自然月数计算。
覆盖补偿:站点 500 米内每有一个邻站,受影响用户与话务损失下降 10%(室分站 5%),最低保留 40%。
</div>
</${Section}>
<${Section} title="通信场景运维考核指标横向对标">
<${DataTable} rowKey="scene" columns=${BENCH_COLUMNS} rows=${bench}/>
+34 -7
View File
@@ -7,11 +7,23 @@ import {Cause, LevelTag, Loading, Section, Select, StatusTag} from '../component
const PAGE_SIZE = 15;
const WINDOW_COLUMNS = [
{key: 'priority', title: '优先级'}, {key: 'category', title: '告警类别', render: v => html`<${Cause} value=${v}/>`},
{key: 'default_min', title: '默认'}, {key: 'window_min', title: '当前'}, {key: 'samples', title: '标定样本', render: num},
{key: 'source', title: '来源'},
];
const ASSOC_COLUMNS = [
{key: 'category', title: '告警类别', render: v => html`<${Cause} value=${v}/>`}, {key: 'window_min', title: '窗口(分)'},
{key: 'alarms', title: '告警数', render: num}, {key: 'hits', title: '其后退服', render: num},
{key: 'support', title: '支持度(%)'}, {key: 'confidence', title: '置信度(%)'}, {key: 'lift', title: '提升度'},
];
export default function RootCause() {
const {filters, isAdmin} = useApp();
const toast = useToast();
const {data: confusion} = useFetch('/api/confusion', filters);
const {data: stats} = useFetch('/api/stats', {...filters, dim: 'root_cause'});
const {data: assoc, reload: reloadAssoc} = useFetch('/api/association', filters);
const [lastRun, setLastRun] = useState(null);
const [running, setRunning] = useState(false);
@@ -20,6 +32,7 @@ export default function RootCause() {
try {
const r = await api('/api/analyze', {method: 'POST'});
setLastRun(r);
reloadAssoc();
toast({message: `全量分析完成:${num(r.n)} 条,耗时 ${r.elapsed}s,准确率 ${r.acc}%`, tone: 'success'});
} catch (e) {
toast({message: e.message, tone: 'error'});
@@ -27,7 +40,7 @@ export default function RootCause() {
setRunning(false);
};
if (!confusion || !stats) return html`<${Loading}/>`;
if (!confusion || !stats || !assoc) return html`<${Loading}/>`;
// 交叉矩阵:matrix[i][j] = 上报原因 i、系统判定 j 的条数;对角线 i === j 为判定一致
const {causes, matrix} = confusion;
@@ -75,19 +88,32 @@ export default function RootCause() {
<div className="grid">
<${Section} span=${5} title="根因判定规则" extra=${runButton}>
<div className="note">
<b>第一级 · 告警衍生关系</b>:“小区退服”告警若带 parent_alarm_id,直接取父告警类别作为根因。<br/>
<b>第二级 · 时间窗关联</b>:回溯同站在退服前仍未清除的告警,按优先级取首个命中:<br/>
① 割接施工 30 分钟 ② 电源故障 300 分钟(考虑蓄电池后备)③ 机房环境 180 分钟(考虑过热保护延时)<br/>
④ 传输中断 15 分钟 ⑤ 天馈故障 30 分钟 ⑥ 设备硬件 15 分钟<br/>
停电衍生的传输告警因优先级归为电源故障;均未命中判为“其他”,派单现场核查。
<b>预处理 · 衍生告警归并</b>:带父告警的告警按父告警类别计(停电引起的传输告警归为电源故障)。<br/>
<b>第一级 · 衍生关系(置信度高)</b>:与退服开始相差 ≤ 2 分钟的“小区退服”告警带父告警时,取父告警类别。<br/>
<b>第二级 · 时间窗(置信度中)</b>:同站在退服时仍未清除、提前量在该类时间窗内的告警。<br/>
<b>第三级 · 疑似(置信度低)</b>:时间窗放宽到 3 倍才命中,提示人工复核。<br/>
多条命中时依次比较:级别 → 是否随业务恢复清除(±15 分钟)→ 规则优先级 → 时间早。
时间窗在每次全量分析时按已闭环流水的提前量 95% 分位数重新标定(不小于默认值);均未命中判为“其他”。
</div>
<div className="chips">
<div className="chip"><span>参与校验</span>${num(total)} 条</div>
<div className="chip" data-tone="green"><span>判定准确率</span>${total ? (hit / total * 100).toFixed(2) : 0}%</div>
${confusion.by_confidence.map(c => html`
<div className="chip" key=${c.level}><span>置信度${c.level}</span>${num(c.n)} 条 / ${c.acc}%</div>`)}
<div className="chip"><span>上次分析</span>${lastRun ? `${num(lastRun.n)} 条 / ${lastRun.elapsed}s` : '—'}</div>
</div>
</${Section}>
<${Section} span=${7} title="上报原因 × 系统判定根因 交叉矩阵" hint="对角线为判定一致"><${EChart} option=${matrixChart}/></${Section}>
<${Section} span=${5} title="时间窗标定" hint="单位:分钟">
<${DataTable} rowKey="category" columns=${WINDOW_COLUMNS} rows=${assoc.windows}/>
</${Section}>
<${Section} span=${7} title="告警 → 退服 关联强度" hint="提升度 > 1 表示该类告警出现后退服概率高于随机">
<${DataTable} rowKey="category" columns=${ASSOC_COLUMNS} rows=${assoc.rows}/>
<div className="note">
支持度 = 开始前 W 分钟内有该类告警的退服 ÷ 退服总数;置信度 = 其后 W 分钟内发生退服的告警 ÷ 该类告警总数;
提升度 = 置信度 ÷ 随机一个 W 分钟窗口内遇到退服的概率(W 为该类时间窗)。
</div>
</${Section}>
<${Section} title="根因维度统计" hint="柱高为平均退服时长(分钟)"><${EChart} height=${260} option=${causeChart}/></${Section}>
<${AdviceList}/>
</div>`;
@@ -108,7 +134,8 @@ function AdviceList() {
const columns = [
{key: 'ne_id', title: '基站编号'}, {key: 'site_name', title: '基站名称'}, {key: 'scene', title: '场景'},
{key: 'start_time', title: '退服开始'}, {key: 'duration_min', title: '时长(分)'},
{key: 'root_cause', title: '根因', render: v => html`<${Cause} value=${v}/>`}, {key: 'evidence', title: '关联告警证据'},
{key: 'root_cause', title: '根因', render: v => html`<${Cause} value=${v}/>`}, {key: 'confidence', title: '置信度'},
{key: 'evidence', title: '关联告警证据'},
{key: 'level', title: '分级', render: v => html`<${LevelTag} value=${v}/>`},
{key: 'status', title: '状态', render: v => html`<${StatusTag} value=${v}/>`},
{key: 'advice', title: '处置建议', wrap: true},
+173
View File
@@ -0,0 +1,173 @@
"""测试公共工具:在 tests/_work 下的临时数据库里加载数据集并调用 core 的分析逻辑,不影响 data/outage.db。"""
import os
import random
import sys
PROJECT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.insert(0, PROJECT)
WORK = os.path.join(PROJECT, 'tests', '_work')
os.makedirs(WORK, exist_ok=True)
os.environ['OUTAGE_DB'] = os.path.join(WORK, 'app.db') # 必须在 import core 之前设置
import pandas as pd # noqa: E402
from werkzeug.security import generate_password_hash # noqa: E402
import core # noqa: E402
import data_gen # noqa: E402
import importer # noqa: E402
FAILED = []
def check(name, ok, detail=''):
"""打印一条检查结果;失败的记入 FAILED,脚本结束时以非 0 退出码返回。"""
print(f" [{'通过' if ok else '失败'}] {name} {detail}")
if not ok:
FAILED.append(name)
def finish():
print('\n全部检查通过' if not FAILED else f'\n{len(FAILED)} 项未通过:' + '、'.join(FAILED))
sys.exit(1 if FAILED else 0)
def use_db(name):
"""切换到 tests/_work/<name>.db(删除重建)。"""
path = os.path.join(WORK, f'{name}.db')
if os.path.exists(path):
os.remove(path)
core.DB_PATH = path
core.data_changed()
with core.db() as conn:
conn.executescript(core.SCHEMA)
for scene, values in core.SCENE_DEFAULT.items():
conn.execute('INSERT INTO scenes VALUES(?,?,?,?,?)', (scene, *values))
conn.execute("INSERT INTO users(username, password, role, created_at) VALUES('admin', ?, 'admin', '')",
(generate_password_hash('admin123'),)) # 测试库的接口测试账号
def load(name, sites, alarms, outages, calibrate_until=None):
"""建库并写入 sites / alarms / outages(全文本 DataFrame),做根因分析,返回 full_frame()。
calibrate_until 指定时,时间窗只用该日期之前的流水标定(留出检验,避免用测试数据标定)。"""
use_db(name)
with core.db() as conn:
sites.to_sql('sites', conn, if_exists='append', index=False)
core.save_alarms(alarms.astype(object).where(alarms.notna(), None))
good, _ = importer.validate_outages(outages)
core.save_outages(good)
if calibrate_until:
core.calibrate_windows(until=calibrate_until)
core.analyze_all(recalibrate=False)
else:
core.analyze_all()
core.data_changed()
return core.full_frame()
def read_dir(directory):
def read(filename):
return pd.read_csv(os.path.join(directory, filename), encoding='utf-8-sig', dtype=str)
return read('sites.csv'), read('data_alarm.csv'), read('data_outage.csv')
def sample_data():
return read_dir(os.path.join(PROJECT, 'data'))
def generate(name, lead=None, rows=6000, seed=2026):
"""用 data_gen 生成一份数据到 tests/_work/gen_<name>;lead 可覆盖告警提前量分布。"""
out = os.path.join(WORK, 'gen_' + name)
saved = dict(data_gen.LEAD)
if lead is not None:
data_gen.LEAD.clear()
data_gen.LEAD.update(lead)
try:
data_gen.generate(n=rows, seed=seed, out=out)
finally:
data_gen.LEAD.clear()
data_gen.LEAD.update(saved)
return read_dir(out)
def score(df):
"""以生成器真值(cause_category)计算总体准确率与各类 precision / recall。"""
d = df[df['cause_category'].isin(core.CAUSES)]
if d.empty:
return 0.0, []
accuracy = (d['cause_category'] == d['root_cause']).mean() * 100
rows = []
for cause in core.CAUSES:
tp = ((d['cause_category'] == cause) & (d['root_cause'] == cause)).sum()
precision = tp / max((d['root_cause'] == cause).sum(), 1) * 100
recall = tp / max((d['cause_category'] == cause).sum(), 1) * 100
f1 = 2 * precision * recall / (precision + recall) if precision + recall else 0
rows.append((cause, round(precision, 1), round(recall, 1), round(f1, 1)))
return round(accuracy, 2), rows
# ---------------------------------------------------------------- 扰动工具(固定随机种子,结果可复现)
rng = random.Random(7)
def root_alarm_mask(alarms, outages):
"""每条退服的“真根因告警”:同站、清除时间 = 退服恢复时间、类别 = 上报原因、本身不是衍生告警。"""
key = set(zip(outages['ne_id'], outages['end_time'], outages['cause_category']))
parent = alarms['parent_alarm_id'].fillna('')
return pd.Series([(n, c, cat) in key and p == ''
for n, c, cat, p in zip(alarms['ne_id'], alarms['clear_time'], alarms['category'], parent)],
index=alarms.index)
def no_parent(alarms, keep=0.0):
"""去掉父告警关系(keep 为保留比例)。"""
alarms = alarms.copy()
alarms['parent_alarm_id'] = [p if rng.random() < keep else '' for p in alarms['parent_alarm_id'].fillna('')]
return alarms
def drop_root(alarms, outages, fraction):
"""随机删掉一部分真根因告警(网管漏报 / 动环未接入)。"""
mask = root_alarm_mask(alarms, outages)
drop = [i for i in alarms.index[mask] if rng.random() < fraction]
return alarms.drop(drop)
def skew_cell_alarm(alarms, max_seconds=120):
"""“小区退服”告警时刻比工单开始时间晚 5~max_seconds 秒(两套系统时钟不同步)。"""
alarms = alarms.copy()
is_cell = alarms['category'] == '小区退服'
times = pd.to_datetime(alarms.loc[is_cell, 'start_time'])
times = times + pd.to_timedelta([rng.randint(5, max_seconds) for _ in range(len(times))], unit='s')
alarms.loc[is_cell, 'start_time'] = times.dt.strftime(core.TIME_FMT)
return alarms
def jitter_clear(alarms, outages, max_minutes=10):
"""真根因告警的清除时刻与业务恢复不再完全相同,随机偏移 ±max_minutes 分钟。"""
alarms = alarms.copy()
mask = root_alarm_mask(alarms, outages)
times = pd.to_datetime(alarms.loc[mask, 'clear_time'])
times = times + pd.to_timedelta([rng.randint(-max_minutes * 60, max_minutes * 60) for _ in range(len(times))],
unit='s')
alarms.loc[mask, 'clear_time'] = times.dt.strftime(core.TIME_FMT)
return alarms
def add_noise(alarms, outages, fraction):
"""一部分退服开始前 0~12 分钟出现一条无关的短时告警(与 data_gen 的干扰告警同分布)。"""
rows = []
for ne_id, start in zip(outages['ne_id'], pd.to_datetime(outages['start_time'])):
if rng.random() < fraction:
category = rng.choice(['传输中断', '设备硬件', '天馈故障', '机房环境'])
begin = start - pd.Timedelta(minutes=rng.randint(0, 12))
end = begin + pd.Timedelta(minutes=rng.randint(2, 30))
rows.append(dict(alarm_id=f'NZ{len(rows):07d}', ne_id=ne_id, alarm_code='', alarm_name='干扰' + category,
alarm_level='次要', category=category, start_time=begin.strftime(core.TIME_FMT),
clear_time=end.strftime(core.TIME_FMT), duration_s='', cleared='是',
parent_alarm_id=''))
return pd.concat([alarms, pd.DataFrame(rows)], ignore_index=True)
# 比 data_gen 默认更宽的告警提前量(传输/设备也可能提前半小时,停电后备可达 7 小时)
WIDE_LEAD = {'电源故障': (10, 420), '机房环境': (10, 240), '割接施工': (0, 45), '天馈故障': (0, 45),
'传输中断': (0, 30), '设备硬件': (0, 30)}
+87
View File
@@ -0,0 +1,87 @@
"""考核指标口径、影响评估与关联量化检验(仓库样例数据)。
运行:python tests/test_kpi.py
"""
import calendar
from harness import check, core, finish, load, sample_data # 先导入 harness:它把数据库指向 tests/_work
import app as webapp # noqa: E402
s, a, o = sample_data()
full = load('kpi', s, a, o)
N = core.site_count()
DAYS = 273 # 样例数据 2026-01-01 ~ 2026-09-30
def calendar_rate(df, n_sites, days):
return round(df['duration_min'].sum() / (n_sites * days * 1440) * 100, 4)
print('== 1. 统计周期')
minutes, months = core.period()
check('全量周期 = 273 天、9 个月', (round(minutes / 1440), months) == (DAYS, 9), f'({minutes / 1440} 天, {months} 月)')
check('全量退服率按日历周期', core.kpis(full, N)['outage_rate'] == calendar_rate(full, N, DAYS))
for m in ['2026-02', '2026-07']:
f = {'month': m}
days = calendar.monthrange(*map(int, m.split('-')))[1]
k = core.kpis(core.frame(f), N, *core.period(f))
check(f'{m} 月退服率按 {days} 天', k['outage_rate'] == calendar_rate(core.frame(f), N, days), f'({k["outage_rate"]}%)')
print('\n== 2. 稀疏筛选')
counts = full.groupby('ne_id').size().sort_values()
ne = counts.index[0]
f = {'site': ne}
df = core.frame(f)
k = core.kpis(df, core.site_count(f), *core.period(f))
check('单站筛选分母为 1 个站', k['sites'] == 1, f'({k["sites"]})')
check('单站退服率按 273 天', k['outage_rate'] == calendar_rate(df, 1, DAYS), f'({k["outage_rate"]}%)')
day = df['start_time'].iloc[0][:10]
f = {'site': ne, 'date_from': day, 'date_to': day}
df = core.frame(f)
k = core.kpis(df, core.site_count(f), *core.period(f))
check('单站单日退服率按 1 站 × 1 天', k['outage_rate'] == calendar_rate(df, 1, 1), f'({k["outage_rate"]}%)')
print('\n== 3. 月均口径:单独查看一个站与全量查看判定一致')
whole = core.site_stats(full, months).set_index('ne_id')
flips = 0
for ne in whole.index:
f = {'site': ne}
one = core.site_stats(core.frame(f), core.period(f)[1]).set_index('ne_id').loc[ne]
flips += (bool(one['long']) != bool(whole.loc[ne, 'long'])) + (bool(one['repeat']) != bool(whole.loc[ne, 'repeat']))
check('超长 / 高频判定翻转数为 0', flips == 0, f'({flips})')
check('重复退服率分母为全部站点', core.kpis(full, N)['repeat_rate'] == round(whole['repeat'].sum() / N * 100, 2))
print('\n== 4. 导出与看板同一口径')
client = webapp.app.test_client()
client.post('/login', json={'username': 'admin', 'password': 'admin123'})
for scene in core.SCENES:
dash = client.get('/api/dashboard', query_string={'scene': scene}).get_json()['kpi']['outage_rate']
f = {'scene': scene}
exported = core.kpis(core.frame(f), core.site_count(f), *core.period(f))['outage_rate']
check(f'{scene} 看板 = 导出', dash == exported, f'({dash}% / {exported}%)')
response = client.get('/api/export/stats', query_string={'scene': '商圈'})
check('导出接口可用', response.status_code == 200)
print('\n== 5. 影响评估')
corr = full['user_hours'].corr(full['duration_min'])
check('受影响用户·小时与时长正相关(修复前受影响用户与时长 r = 0.03)', corr > 0.5, f'(r = {corr:.3f})')
expected = full['affected_users'] * full['duration_min'] / 60
relative = ((full['user_hours'] - expected).abs() / expected.clip(lower=1)).max()
check('用户·小时 = 受影响用户 × 时长', relative < 0.02, f'(最大相对误差 {relative:.4f},来自用户数取整)')
check('覆盖补偿系数在 [0.4, 1]', full['cover_factor'].between(core.COVER_FLOOR, 1).all(),
f'(均值 {full["cover_factor"].mean():.2f})')
isolated = core.site_cover()
check('有邻站的站补偿系数 < 1', min(isolated.values()) < 1)
print(f' 用户·小时合计 {core.kpis(full, N)["user_h"]:,.0f},覆盖补偿系数分布 '
f'{full.drop_duplicates("ne_id")["cover_factor"].value_counts().sort_index().to_dict()}')
print('\n== 6. 告警—退服关联量化')
rows = core.association()
for r in rows:
print(f' {r["category"]:<6} 窗口 {r["window_min"]:>6} 分 告警 {r["alarms"]:>5} 支持度 {r["support"]:>6}% '
f'置信度 {r["confidence"]:>6}% 提升度 {r["lift"]:>6}')
check('6 类告警均有关联指标', len(rows) == 6 and all(r['alarms'] > 0 for r in rows))
check('各类提升度 > 1(告警出现后退服概率高于随机)', all(r['lift'] > 1 for r in rows))
check('关联接口可用', client.get('/api/association').status_code == 200)
finish()
+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()