AI守護高齡者AI Guardian for Elders

AI永續新世代 · 普台科技競賽

中暑之前，先一步察覺。 Catch heatstroke before it starts.

結合居家感測器、智慧手錶與個人化 AI，日夜留意長者的身體與環境。一出現異狀，手錶先提醒；需要降溫時，自動開冷氣；風險升高，家人立刻收到通知。

Home sensors, a smartwatch and an AI that knows each person’s normal keep watch day and night. At the first sign of trouble the watch nudges them, the air conditioner turns on when cooling is needed, and family is alerted as soon as risk rises.

看系統怎麼判斷Watch it decide 系統架構See the architecture
  • 感測Sensing每個房間一個節點，加上一支手錶A node in every room, plus one watch
  • 判斷Deciding三級風險，應對逐級累加Three risk levels, with actions that add up
  • 試辦Pilot20 位長者 · 一個夏季20 elders over one summer
悶熱的夜晚 · 臥室A hot night · bedroom 示意情境Illustrative c-001 · pi-0007 · w-0003
時間Time02:10
臥室溫度Bedroom31.4°C
心率Heart rate76bpm睡眠基準 +18
風險Risk輕度冷氣 開啟

  1. 102:00心率達到睡眠基準 +15 bpm，開始計時。Heart rate reaches 15 bpm over the sleep baseline, so the clock starts.
  2. 202:10持續 10 分鐘，判定輕度風險，臥室冷氣自動開啟。It holds for 10 minutes: mild risk, and the bedroom AC turns on.
  3. 302:38連續 15 分鐘恢復正常，事件結束，家人不必半夜被叫醒。Normal for 15 minutes in a row: the episode closes and nobody is woken up.
示意情境：臥室節點與手錶的資料每分鐘合併一次。把游標移到圖上，或聚焦後用方向鍵查看任一分鐘。Illustrative scenario: bedroom-node and watch data are combined every minute. Hover over the chart, or focus it and use the arrow keys, to read any minute.
查看數據表View data table

為什麼需要Why it matters

熱傷害，常發生在沒人注意的時候Heat illness often strikes when no one is watching

長者對熱與口渴的感覺比較遲鈍，也常為了省電不開冷氣。家人無法隨時在旁，等到出現明顯症狀，往往已經需要送醫。我們把應對提前到身體剛出現變化的時候。

Older adults feel heat and thirst less sharply, and many skip the air conditioner to save on electricity. Family can’t be there every hour, and by the time symptoms are obvious, a hospital visit is often already needed. We move the response up to the first change in the body.

熟睡的夜晚Asleep on a hot night

房間慢慢變熱，人在睡夢中很難察覺。A bedroom heats up slowly, and a sleeping person rarely notices.

心率持續偏高時，自動打開臥室冷氣。When heart rate stays high, the bedroom AC turns on.

獨自在家Home alone

身邊沒有人，早期徵兆沒人發現。With no one around, the early signs go unnoticed.

風險升高時依序通知家人，並說明原因。As risk rises, family is alerted in order, with the reasons.

外出途中Out in the midday heat

中午走路辦事，離涼爽的地方很遠。A midday errand on foot, far from anywhere cool.

提醒休息補水，必要時導航到最近的避暑地點。A prompt to rest and drink, and if needed, directions to the nearest cooling spot.

運作方式How it works

監測、預測、引導Monitor, predict, guide

家中每個房間放一個小型感測節點，長者戴一支市售智慧手錶。所有資料送到同一台雲端伺服器，合併分析後，再把結果送回手錶、家人與冷氣。

Each room gets a small sensor node, and the elder wears an off-the-shelf smartwatch. Everything goes to one cloud server, which combines and analyzes it, then sends results back to the watch, the family and the air conditioner.

一位長者的家One elder’s home c-001 · 02:10
臥室Bedroom最熱Hottest
31.4°C濕度RH 72%
pi-0007
  • 溫濕度Temp & humidity
  • 人體存在Presence
  • 冷氣運轉AC running
  • 紅外線遙控IR remote
長者在這裡 · 手錶 w-0003 · 76 bpmThe elder is here · watch w-0003 · 76 bpm
客廳Living room
29.6°C濕度RH 70%
pi-0008
  • 溫濕度Temp & humidity
  • 人體存在Presence
  • 冷氣運轉AC running
  • 紅外線遙控IR remote
陽台 · 戶外Balcony · outdoors
28.6°C濕度RH 80%
pi-0009
  • 戶外溫濕度Outdoor temp & humidity
凌晨 02:10，臥室 31.4 °C，比戶外的 28.6 °C 還熱。每個房間節點量自己的空間，手錶跟著人走；資料累積之後，系統能告訴家人哪個房間最熱。At 02:10 the bedroom is 31.4 °C, hotter than the 28.6 °C outside. Each room node measures its own space and the watch goes wherever the person goes. Over time, the data shows the family which room runs hottest.
  1. 1

    監測Monitor

    房間節點每 10 秒量一次溫度、濕度、有沒有人、冷氣有沒有開，陽台節點量戶外溫濕度。手錶量心率、活動狀態與 GPS，支援的錶款再加上皮膚溫度。

    Room nodes read temperature, humidity, presence and whether the AC is running every 10 seconds, and a balcony node reads outdoor conditions. The watch tracks heart rate, activity and GPS, plus skin temperature on models that support it.

    • 溫濕度Temp & humidity
    • 人體存在Presence
    • 冷氣運轉AC running
    • 心率Heart rate
    • 活動狀態Activity
    • GPS
  2. 2

    預測Predict

    伺服器每分鐘把各房間與手錶的資料對齊成一條時間軸，拿長者自己的基準值來比：先排除運動，再看數值能不能自己恢復，最後決定風險等級。

    Every minute, the server lines up room and watch data into one timeline and compares it with the person’s own baseline. It rules out exercise first, waits to see whether the numbers recover, then sets the risk level.

    • 個人基準值Personal baseline
    • 活動判斷Activity check
    • 恢復觀察Recovery watch
    • 熱島門檻Heat-island thresholds
  3. 3

    引導Guide

    依風險等級逐步行動，從手錶輕震、補水提醒，到自動開冷氣、通知家人、導航到最近的避暑地點。

    Actions step up with the risk level, from a gentle watch vibration and a reminder to drink water to switching on the AC, alerting family and giving directions to the nearest cooling spot.

    • 手錶震動Watch vibration
    • 自動開冷氣AC on
    • 通知家人Family alerts
    • 避暑導航Cooling-spot directions

風險分級 · 規則 v1Risk levels · rule v1

三級風險，應對逐級累加Three levels, and the actions add up

每升一級，都保留前一級的行動，再多加一層保護。數值連續 15 分鐘恢復正常，才會一級一級降回來。

Each level keeps everything the level below does and adds more. Levels step back down one at a time, and only after the numbers stay normal for 15 minutes.

輕度Mild

自我降溫Self-cooling

何時WHEN環境偏熱、心率持續高於個人基準，而且不是在運動。It’s hot, heart rate stays above the personal baseline, and exercise doesn’t explain it.

  • 手錶震動提醒The watch vibrates
  • 提示補水、休息A prompt to drink water and rest
  • 若正在睡覺，自動開冷氣If they’re asleep, the AC turns on
中度Moderate

環境降溫Environmental cooling

何時WHEN輕度處理 30 分鐘後仍未改善，或心率高出基準 25 bpm 以上。Mild risk hasn’t improved 30 minutes after its actions, or heart rate is 25 bpm or more above baseline.

輕度的所有行動，再加上：Everything at mild, plus:

  • 通知家人，並說明原因Family is alerted, with the reasons
  • 在室內：開冷氣Indoors: the AC turns on
  • 在戶外：導航到最近的避暑地點Outdoors: directions to the nearest cooling spot
高度High

第三方介入Third-party help

何時WHEN中度處理 20 分鐘後仍未改善，或長者對提醒沒有回應、也沒有動作。Moderate risk hasn’t improved 20 minutes after its actions, or the elder doesn’t respond to prompts and isn’t moving.

中度的所有行動，再加上：Everything at moderate, plus:

  • 傳送位置給家人Location is sent to family
  • 持續監測，不中斷Continuous monitoring
  • 依序聯絡緊急聯絡人，是否撥打 119 由家人決定Contacts are reached in order; the family decides whether to call 119

時間與門檻是規則 v1 的起始值，會與護理顧問一起調整。Times and thresholds are rule v1 starting values, to be tuned with our nursing advisor.

每個人都有自己的基準A baseline for each person

每個人的正常心率都不一樣。系統用過去兩週休息與睡眠時的資料，分別建立清醒與睡眠的基準；前 7 天先用保守的預設值。

Everyone’s normal heart rate is different. The system builds separate awake and asleep baselines from the last two weeks of rest and sleep, and uses conservative defaults for the first 7 days.

先排除運動Exercise is ruled out first

走路、運動時心率本來就會上升。活動結束 10 分鐘後，系統才重新判斷。

Heart rate naturally rises with walking and exercise. The system re-checks 10 minutes after the activity ends.

先觀察能不能恢復Wait for recovery

每一級行動之後都先觀察一段時間，沒有改善才升級，避免太多誤報讓家人疲乏。

After each level’s actions, the system waits to see whether things improve before escalating, so families aren’t worn down by false alarms.

依地區調整門檻Thresholds tuned by area

用 QGIS 分析都市熱島，替比較熱的地區調低環境門檻。

A QGIS urban heat-island map lowers the heat threshold in hotter neighbourhoods.

情境重播Scenario replay

看系統怎麼判斷Watch it decide

選一個情境，一步一步看規則怎麼運作。數值為示意，規則與門檻依規則 v1。

Pick a scenario and step through how the rules respond. The numbers are illustrative; the rules and thresholds follow rule v1.

    3 / 5

    每日 07:30 · 勞動部熱指數分級Daily at 07:30 · Ministry of Labor heat index

    每天早上，先知道今天有多熱Every morning, a heads-up on the heat

    每天 07:30，系統用長者所在地區的氣象署預報最高溫與濕度計算熱指數，依勞動部分級送出一則出門前的提醒。熱指數只用來提醒，不用來判斷是否中暑。

    At 07:30 each day, the system takes the weather service’s forecast high and humidity for the elder’s area, computes the heat index and sends one reminder before they head out, graded on the Ministry of Labor scale. The heat index is only a reminder and never decides heatstroke risk.

    熱指數HEAT INDEX 等級LEVEL 提醒內容REMINDER
    26.7–32.2 °C輕度Mild多補充水分，減少日曬。Drink more water and limit time in the sun.
    32.2–40.6 °C中度Moderate避免中午外出，出門請規劃陰涼路線。Avoid going out at midday, and plan a shaded route.
    40.6–54.4 °C高度High盡量不要外出，待在通風或有冷氣的室內。Better to stay in, in a ventilated or air-conditioned room.
    33.0 °C
    55%
    今日熱指數Today’s heat index
    37.8°C
    中度
    AI守護高齡者AI Guardian for Elders
    今日熱指數 37.8 °C · 中度
    避免中午外出，出門請規劃陰涼路線。

    計算方式：美國國家氣象局 Rothfusz 迴歸式（含低濕與高濕修正），換算為攝氏後依勞動部分級。Computed with the US National Weather Service Rothfusz regression (with its low- and high-humidity adjustments), converted to °C and graded on the Ministry of Labor scale.

    系統架構Architecture

    一台伺服器，照顧許多長者One server, many elders

    樹莓派只負責量測、檢查、暫存與傳送，所有分析都在伺服器完成。每筆資料都帶著長者、節點、感測器、手錶與時間的編號，新增長者或設備只要登記，不用改程式。

    Each Raspberry Pi only measures, checks, buffers and sends. All analysis happens on the server. Every reading carries IDs for the elder, node, sensor, watch and time, so adding a person or a device is a registration with no code changes.

    家中At homec-001

    房間節點 · 樹莓派ROOM NODES · RASPBERRY PI
    臥室Bedroompi-0007
    客廳Living roompi-0008
    陽台（戶外）Balcony (outdoor)pi-0009

    量測 → 檢查 → 先存本機 → 傳送Measure → check → save locally → send

    穿戴裝置WEARABLE
    智慧手錶Smartwatchw-0003

    藍牙 → 長者手機 AppBluetooth → the elder’s phone app

    c-002、c-003 … 每位長者都有自己的一組設備與獨立通道。c-002, c-003 … every elder has their own devices and a separate lane.

    節點：MQTT（TLS 加密）
    手錶：App → HTTPS    Nodes: MQTT over TLS
    Watch: app → HTTPS

    雲端伺服器Cloud server每分鐘every minute

    1. 接收資料Receive節點與手錶分兩條路進來Nodes and the watch arrive by separate roads
    2. 辨識長者Identify the elder用設備憑證與登記表確認屬於誰Device credential plus registry decide whose data it is
    3. 儲存資料Store原始與處理後資料，依長者分開保存Raw and processed data, kept per elder
    4. 合併資料Combine對齊成每分鐘一列的時間軸Aligned into one row per minute
    5. 分析資料Analyze個人基準、活動判斷、恢復觀察、風險等級Baseline, activity check, recovery watch, risk level
    6. 產生結果Generate results每次等級變化都留下原因與紀錄Every level change is logged with its reasons

    外部資料 → 分析OUTSIDE DATA → ANALYZE氣象署觀測與預報 · 勞動部熱指數 · QGIS 都市熱島 · Google 地圖CWA weather observations and forecasts · Ministry of Labor heat index · QGIS urban heat islands · Google Maps

    即時推播
    與控制指令    Live push
    and commands

    結果送到Results go to

    • 網頁儀表板Web dashboard團隊看全部長者，家人只看自家長者The team sees every elder; families see only their own
    • 家人通知Family alertsApp 與 LINE 推播，簡訊備援App and LINE alerts, with SMS backup
    • 手錶震動Watch vibration提醒長者補水、休息Nudges the elder to drink and rest
    • 冷氣控制AC control指令送回房間節點的紅外線遙控器Commands go back to the room node’s IR remote
    • 避暑導航Cooling-spot directions人在戶外時，找最近的避暑地點The nearest cooling spot when outdoors
    房間節點與手錶走兩條不同的路，只在伺服器裡會合。伺服器依設備身分判斷資料屬於哪位長者，之後的每一步都只在那位長者的通道裡進行。Room nodes and the watch take separate roads and meet only inside the server. The server uses each device’s identity to decide whose data it is, and every later step stays inside that elder’s own lane.

    斷網也不漏資料Nothing lost when Wi‑Fi drops

    每筆資料先寫進樹莓派本機，最多保存 7 天。恢復連線後依原始時間補傳，過去的時間不會再觸發舊警報。

    Every reading is saved on the Pi first and kept for up to 7 days. After reconnecting, it’s replayed with its original timestamp, and past minutes never set off stale alerts.

    設備只能代表自己Each device speaks only for itself

    每台樹莓派與每支手錶都有自己的憑證。伺服器用登記資料確認它屬於哪位長者，對不上的資料直接拒收。

    Every Pi and watch has its own credential. The server checks the registry to confirm which elder it belongs to and rejects anything that doesn’t match.

    加一位長者，不用改程式A new elder needs no new code

    新增長者、房間或感測器都只是登記一筆資料；手錶壞了，重新配對一支就好。

    Adding an elder, a room or a sensor is one registry entry. If a watch breaks, pair a new one.

    hp.telemetry.v1每筆資料都帶著身分與時間Every reading carries its identity and time
    {
  "schema": "hp.telemetry.v1",
  "client_id": "c-001",
  "node_id": "pi-0007",
  "seq": 18342,
  "readings": [{
    "sensor_id": "pi-0007.s01",
    "metric": "temp_c",
    "ts": "2027-07-14T05:30:00Z",
    "value": 31.4,
    "quality": "ok"
  }]
}
    規劃中的技術Planned stack Raspberry Pi Zero 2 W SHT4x 溫濕度SHT4x temp & humidity 毫米波人體存在mmWave presence MQTT · Mosquitto Python · FastAPI PostgreSQL + TimescaleDB Redis React 儀表板React dashboard Docker Compose Grafana

    隱私與安全Privacy & safety

    先同意，才收集Consent comes first

    心率與位置都是敏感的個人資料。系統比照《個人資料保護法》的敏感資料標準處理：每個設備只能代表自己，每個家庭只看得到自己的長者。

    Heart rate and location are sensitive personal data, and the system handles them to the standard for sensitive data under Taiwan’s Personal Data Protection Act. Each device can speak only for itself, and each family sees only its own elder.

    同意後才啟用Nothing before consent

    長者與家人簽署同意書後才開始收集。生理資料、位置與冷氣控制分開同意，任何一項都能撤回。

    Collection starts only after the elder and family sign consent. Physiology, location and AC control are agreed to separately, and any of them can be withdrawn.

    冷氣連動隨時可關AC control can be switched off

    自動開冷氣可能和長者的省電習慣衝突，所以這項功能隨時可以關閉。

    Automatic AC can clash with an elder’s habit of saving power, so it can be switched off at any time.

    位置平時保持模糊Location stays coarse

    平時只保留大略位置；只有在中度以上的風險事件中，才顯示精確位置。

    Location is kept coarse day to day and shown precisely only during a moderate-or-higher risk event.

    只看得到自家長者Families see only their own

    每個家庭帳號只能看見自家長者的資料，每一次查看與操作都會留下紀錄。

    Each family account sees only its own elder’s data, and every view and command is logged.

    由人做最後決定People make the call

    AI 負責提醒與通知；要不要撥打 119，由家人決定。

    The AI informs and alerts. Whether to call 119 is the family’s decision.

    輔助照護，不做診斷Support, not diagnosis

    本系統不是醫療器材，也不做診斷。手錶量到的皮膚溫度不等於核心體溫。

    The system is not a medical device and does not diagnose. Skin temperature from a watch is not core body temperature.

    開發進度Roadmap

    16 週，從規格到試辦Sixteen weeks from spec to pilot

    九個階段、五道檢核關卡。第 12 週的 G3 是競賽展示目標：一位真實長者的資料從頭到尾跑完，並在儀表板上觸發即時警示。

    Nine phases and five checkpoint gates. Gate G3 in week 12 is the competition demo target: one real elder’s data flows end to end and triggers a live alert on the dashboard.

    1. 0決策與資料規格Decisions & data contract目前Now第 1 週Week 1
    2. 1伺服器基礎Server foundationG1第 2–3 週Weeks 2–3
    3. 2樹莓派感測節點Raspberry Pi node第 3–6 週Weeks 3–6
    4. 3智慧手錶串接Smartwatch integration第 3–6 週Weeks 3–6
    5. 4資料合併與同步Fusion & syncG2第 7–8 週Weeks 7–8
    6. 5風險引擎與外部資料Risk engine & context第 8–10 週Weeks 8–10
    7. 6儀表板、通知與冷氣Dashboard, alerts & ACG3第 9–12 週Weeks 9–12
    8. 7多位長者與系統強化Multi-client & hardeningG4第 12–14 週Weeks 12–14
    9. 8試辦準備Pilot readinessG5第 14–16 週Weeks 14–16

    第 16 週之後：夏季試辦，20 位長者。階段 2 與 3 同時進行，因為樹莓派與手錶是兩條獨立的資料路線。After week 16: the summer pilot with 20 elders. Phases 2 and 3 run side by side because the Pi and the watch are independent data roads.

    1. G1第 3 週Week 3

      模擬資料能正確標記長者、節點、感測器與時間；標錯長者的資料會被拒收。

      Simulated data is stored with the right elder, node, sensor and time, and data labelled with the wrong elder is rejected.

    2. G2第 8 週Week 8

      一個真實測試家庭跑完一整天，合併後的時間軸與手寫活動紀錄相符。

      One real test home runs a full day, and its combined timeline matches a hand-written activity log.

    3. G3第 12 週Week 12
      競賽展示目標Competition demo target

      現場加熱測試臥室：儀表板從輕度升到中度，家人手機收到警示，冷氣自動開啟。

      Live demo: warming the test bedroom moves the dashboard from mild to moderate, the family’s phone gets the alert and the AC turns on.

    4. G4第 14 週Week 14

      20 位模擬長者加上 2 個真實家庭連續運作 7 天，不同長者的資料完全沒有互相外洩。

      20 simulated elders plus 2 real homes run for 7 days, with no data leaking between elders.

    5. G5第 16 週Week 16

      通過試辦準備檢核，團隊簽署開始試辦。

      The pilot readiness checklist passes and the team signs off on going ahead.

    試辦與評估Pilot

    20 位長者，一個夏天Twenty elders, one summer

    試辦用三項指標檢驗成效。我們也會清楚說明限制：一個夏天裡真正的熱傷害事件可能不多，所以每一條規則都同時用情境重播測試驗證。

    Three measures test whether it works. We are also upfront about the limits: one summer may bring few real heat events, so every rule is also checked with scripted scenario replays.

    風險判斷準確度Accuracy of risk calls

    每一次中度以上的事件都由護理顧問審查是否合理，並比對家人回報、系統卻漏掉的事件。

    Our nursing advisor reviews every moderate-or-higher event, and we check for events families report that the system missed.

    衡量方式MEASURED AS中度以上警示的精確率與召回率Precision and recall of moderate-or-higher alerts

    家人即時收到Family alerted in time

    從中度風險觸發，到家人按下確認的時間。

    The time from a moderate-risk trigger to a family member acknowledging it.

    衡量方式MEASURED AS中位數與第 90 百分位數Median and 90th percentile

    長者願意使用Elders keep using it

    每天配戴手錶的時數，以及夏季結束時的使用問卷。

    Daily watch wear time and an end-of-summer survey.

    衡量方式MEASURED AS每日配戴時數與問卷分數（1–5 分）Daily wear time and survey score (1–5)

    AI 怎麼一步步變聰明How the AI grows

    1. 1試辦期間使用Used in the pilot

      個人化基準規則Rules on personal baselines

      不需要訓練資料，每一個判斷都說得出原因。

      Needs no training data, and every decision states its reasons.

    2. 2試辦期間在背景試跑Runs in the background during the pilot

      個人異常偵測Personal anomaly model

      學習在某個室溫、活動與時段下，這個人的心率通常是多少，並標出差距過大的時刻。

      Learns what this person’s heart rate usually is for a given room temperature, activity and time of day, and flags large gaps.

    3. 3累積足夠事件之後Once enough events are labelled

      監督式風險模型Supervised risk model

      需要大量標註過的熱傷害事件，等資料足夠才會啟用。

      Needs many labelled heat episodes, so it waits until the data exists.

    未來展望What’s next

    從一個家，到一座城市From one home to a whole city

    同一套架構，只要加入新的資料與新的服務對象，就能擴大照顧範圍，不必重新打造。

    The same architecture grows by adding new data and new groups of people, with no rebuild.

    社區避暑地圖Community cooling map

    整合里辦公處、圖書館、活動中心等避暑地點，並找出風險最高的里。

    Map cooling spots such as neighbourhood offices, libraries and activity centres, and find the neighbourhoods at highest risk.

    照顧更多族群More people at risk

    戶外工作者、慢性病患者，以及家中沒有冷氣的低收入家庭。

    Outdoor workers, people with chronic illness, and low-income homes without air conditioning.

    街區熱地圖Street-level heat map

    匿名彙整各家的戶外感測資料，回饋成街區尺度的熱地圖。

    Anonymized outdoor readings from many homes become a street-by-street heat map.

    城市氣候調適City climate adaptation

    成為城市氣候調適系統的一部分，並與社福單位連結。

    Become part of the city’s climate-adaptation system, linked with social-welfare services.

    SDG 3健康與福祉Good health and well-being SDG 11永續城市與社區Sustainable cities and communities SDG 13氣候行動Climate action