載入中…
📈 查看趨勢圖與技術指標📈 View trends and technical indicators
🔬 如何判定（完整方法與限制）🔬 How determinations are made (method & limitations)
1. 同時段季節基線1. Hour-of-week baseline
網路流量有很強的日／週週期 —— 凌晨三點本來就只有下午的一半，直接對原始值設門檻毫無意義。基線取「同一星期幾＋同一小時」其他週的中位數，而且是 leave-one-out：異常點不會把自己的基線拉下去。
Traffic has strong daily/weekly cycles — 3 a.m. is naturally half of mid-afternoon, so thresholding raw values is meaningless. The baseline is the median of the same weekday-and-hour slot in other weeks, computed leave-one-out so an anomaly cannot drag down its own baseline.
2. 穩健 z 分數2. Robust z-score
殘差除以 MAD×1.4826，不用標準差 —— 一次大掉點就會把 σ 撐大到之後什麼都偵測不到。另設尺度下限（基線的 0.5%）：序列極平穩時 MAD 會是 0，沒有下限就會放棄偵測；反之 MAD 過小時，1% 的抖動也會被算成異常。
Residuals are divided by MAD×1.4826 rather than the standard deviation — one large drop would inflate σ enough to mask everything after it. A floor (0.5% of baseline) is applied: on a very flat series MAD is 0 and detection would give up entirely, while a tiny MAD would turn 1% jitter into an "anomaly".
3. 持續性雙門檻3. Dual sustained threshold
要求連續 ≥2 小時且相對基線偏離 ≥10% 才成案。單一時段的尖刺多半是量測雜訊；海纜中斷的特徵是持續性的位階下移，因此嚴重度由「幅度 × 時數」決定（≥30% 且 ≥3 小時 → 嚴重）。
An event requires ≥2 consecutive hours and ≥10% deviation from baseline. Single-bucket spikes are usually measurement noise; a cable cut looks like a sustained level shift, so severity is graded by magnitude × duration (≥30% for ≥3 h → critical).
4. 船隻關聯4. Vessel correlation
取異常前 12 小時內、距海纜 5 公里內、航速 ≤5 節的船。海纜旁最密集的低速目標其實是漁網信標，因此沿用威脅評分引擎的排除規則；港內位置也一併排除 —— 海纜登陸點就在港邊，合法靠泊必然「0 節、離海纜很近」。候選依船型威脅權重排序，商船與公務船優先於漁船。
Vessels within 12 h before onset, 5 km of a cable and ≤5 kn. The densest slow targets near cables are actually fishing-net beacons, so the threat engine's exclusion rules are reused; in-port positions are dropped too — cable landings sit next to ports, so a lawfully berthed ship is always "0 kn, very close to a cable". Candidates are ranked by vessel-type threat weight, commercial and government vessels above fishing boats.
5. 縣市粒度從哪來5. Where county-level data comes from
縣市色塊有兩個來源，意義完全不同，不可混為一談。Cloudflare Radar 自 2025 年 9 月起提供第一級行政區（ADM1）資料，但實測（geoId 逐一測試）發現：Radar 的台灣 ADM1 只有 4 個分區，不是 22 個縣市 —— 臺北市、高雄市、福建省（金門＋馬祖）、臺灣省（其餘 18 縣市），用的是 GeoNames 的舊省制分區。因此地圖上最多只有 4 種顏色，每個縣市顯示的是它所屬分區的數值，彈出視窗會明白標示「非本縣市單獨量測」。分區內的指標依序嘗試「IQI 頻寬 → IQI 延遲 → NetFlows 流量指數」，第一個取得到的就是該分區顯示的指標 —— 頻寬與延遲是連線品質量測，流量指數則只是「這個縣市送出多少流量」，不能當成網速。彈出視窗另外附上 Speed Test 實測中位數（使用者實跑測速的下載／上傳），它與 IQI 品質指數是兩種不同的量測，數值量級也不同，因此分開標示。IODA 給的是可達性（BGP／主動探測／darknet），不需要憑證，是縣市粒度的保底來源。每一格都會標明實際用了哪一個指標與來源。
County fills come from two sources with different meanings. Cloudflare Radar has offered first-level administrative division (ADM1) data since September 2025, but a live geoId probe showed that Radar splits Taiwan into only 4 regions, not 22 counties — Taipei, Kaohsiung (Takao), Fukien (Kinmen + Matsu) and Taiwan Province (the other 18 counties), following the old GeoNames provincial scheme. The map therefore carries at most four colours: each county shows the value of the region it belongs to, and the popup says explicitly that it is not a measurement of that county alone. Within a region the pipeline tries IQI bandwidth → IQI latency → NetFlows traffic index and shows the first one that returns data — bandwidth and latency measure connection quality, while the traffic index only says how much traffic the county sends and is not a speed. Popups additionally carry the Speed Test median (download/upload from user-run tests), listed separately because it is a different measurement from the IQI index and lands on a different order of magnitude. IODA provides reachability (BGP / active probing / darknet telescope) without credentials and is the fallback that guarantees county-level coverage. Every popup states which metric and source that county actually used.
6. 解讀限制6. Interpretation limits
流量變化可能來自維護、路由調整、量測缺口或一般故障；AIS 也可能缺漏、關閉或遭偽造。船隻在時間與空間上接近海纜只構成調查線索，不足以判定故障，更不能證明肇因或行為者。
Traffic changes can reflect maintenance, routing, measurement gaps, or ordinary failures; AIS can be missing, disabled, or spoofed. Proximity in time and space is an investigative lead only, not proof of an outage, cause, or actor.
本頁重點
What this page shows
本頁用網路流量與延遲辨識可能的持續性連線異常，並比對海纜附近船隻。訊號與船位僅供交叉查核，代表關聯而非故障或破壞的因果證明。
This page flags sustained connectivity anomalies from traffic and latency, then cross-checks vessels near cables. Signals and vessel positions support triage only: correlation is not proof of an outage or its cause.