fix(inference): equipment type

README.md

@@ -753,7 +753,7 @@ inference/
 
 ### Цикл `python -m inference.runner`
 
-Старт: прогрев окон из `events` за 72 ч, pickle ядра не используется. Такт: новые строки `events` пачкой → один прогноз на лог (`channel`, `CHANNEL_EVENT`, `horizon_until = ts + 30 ч`) → запись пачки одной транзакцией. TTL удаляет истёкшие `catboost-aft-*` без решения диспетчера. Обрыв БД — reconnect с backoff.
+Старт: прогрев окон из `events` за 72 ч, pickle ядра не используется. Такт: новые строки `events` пачкой → один прогноз на лог (`channel`, `EQUIPMENT_FAILURE` у канала без класса инцидента, иначе `CHANNEL_EVENT`; `horizon_until = ts + 30 ч`) → запись пачки одной транзакцией. TTL удаляет истёкшие `catboost-aft-*` без решения диспетчера. Обрыв БД — reconnect с backoff.
 
 Пропуски и сбои считаются в строке `tick` раз в минуту: `skipped_stale`, `skipped_ooo`, `skipped_unknown`, `failed`, `failed_write_prep`.
 

docs/incident-v4-500.md

@@ -2,8 +2,10 @@
 
 The second inference container uses `data/incident4/model_500.cbm` and
 `PREDICTION_KIND=INCIDENT`. It writes FIRE, FLOOD, GAS or INTRUSION with a
-30-hour horizon. The equipment container retains `CHANNEL_EVENT` and its
-own cursor and model version.
+30-hour horizon. The equipment container keeps its own cursor and model version. It writes
+`EQUIPMENT_FAILURE` only for channels without an incident class: its target is
+any alarm, which on a classed channel mixes faults with incidents, so there it
+keeps `CHANNEL_EVENT` (hidden by Backend as LEGACY).
 
 `hours_since_prev_target` is the time since the previous target alarm on the
 same channel, excluding the current event. `target_seed.json` contains the

inference/incident.py

@@ -52,6 +52,15 @@
 """
 
 
+def classed_channels(conn: psycopg.Connection) -> frozenset[int]:
+    """Каналы с классом инцидента — по реестру, а не по фичам модели."""
+    rows = conn.execute(
+        "SELECT channel_id FROM dim_channels_current WHERE sensor_type = ANY(%s)",
+        (list(SENSOR_CLASSES),),
+    ).fetchall()
+    return frozenset(int(row[0]) for row in rows)
+
+
 def incident_class(sensor_type: str | None) -> str | None:
     return SENSOR_CLASSES.get(sensor_type)
 

inference/runner.py

@@ -24,7 +24,9 @@
 from inference.config import connect as pg_connect
 from inference.cursor import EventCursor
 from inference.health import HEARTBEAT_FILE, READY_FILE
-from inference.incident import TargetHistory, incident_class, is_target_alarm
+from inference.incident import (
+    TargetHistory, classed_channels, incident_class, is_target_alarm,
+)
 from inference.predictions import (
     PredictionRow,
     delete_expired,
@@ -177,6 +179,7 @@ def __init__(
         self.totals = TickCounters()
         self.last_tick = TickCounters()
         self.targets: TargetHistory | None = None
+        self.classed: frozenset[int] = frozenset()
         self._tick_shap_ms = 0.0
         self._logged_at = float("-inf")
         self._traceback_at = float("-inf")
@@ -214,6 +217,7 @@ def start(self, conn: psycopg.Connection) -> None:
                     low, lag_seconds=self.options.reorder_lag_seconds
                 )
             rows = self._warmup(conn, service, cursor)
+            self.classed = classed_channels(conn)
             if self.settings.prediction_kind == "INCIDENT":
                 self.targets = TargetHistory.load(
                     self.path(TARGETS_FILE), cursor.low, cursor.seen
@@ -365,7 +369,11 @@ def _process(
         sensor_type = service.metadata[event.channel_id].get("тип_датчика")
         target_alarm = False
         previous_hours = None
-        incident = "CHANNEL_EVENT"
+        # Модель поломок прогнозирует любую тревогу: на канале без класса это
+        # отказ оборудования, на канале с классом — смесь с инцидентом, её
+        # не выдаём за отказ. Тип датчика берём из реестра: в фичах его нет.
+        incident = ("CHANNEL_EVENT" if event.channel_id in self.classed
+                    else "EQUIPMENT_FAILURE")
         if self.targets is not None:
             incident = incident_class(sensor_type)
             if incident is None:
@@ -485,6 +493,7 @@ def _reload_metadata(self, conn: psycopg.Connection, reason: str) -> None:
         service = self.service
         assert service is not None
         metadata = fetch_metadata(conn, service.compiled.static_features)
+        self.classed = classed_channels(conn)
 
         # Rolling-state не трогаем: меняется только справочная часть.
         with service.lock:

tests/test_runner.py

@@ -30,11 +30,12 @@ def __call__(self) -> float:
         return self.now
 
 
-def add_channel(conn: psycopg.Connection, channel_id: int) -> None:
+def add_channel(conn: psycopg.Connection, channel_id: int,
+                sensor_type: str = "КД АВ") -> None:
     conn.execute(
         "INSERT INTO dim_channels VALUES "
-        "(%s, %s, 'Охранная подсистема', 'КД АВ', '15-11.1.131.2.', 'КД АВ', 20)",
-        (channel_id, SNAPSHOT),
+        "(%s, %s, 'Охранная подсистема', %s, '15-11.1.131.2.', 'КД АВ', 20)",
+        (channel_id, SNAPSHOT, sensor_type),
     )
 
 
@@ -240,6 +241,23 @@ def test_unknown_channel_picked_up_after_reload(db, make_runner) -> None:
     assert [row[1] for row in predictions(db)] == [str(NEW_CHANNEL)]
 
 
+def test_classless_channel_writes_equipment_failure(db, make_runner) -> None:
+    runner, conn = started(make_runner, db, threshold=0.0)
+    now = datetime.now(timezone.utc)
+
+    with db.admin() as admin:
+        add_channel(admin, NEW_CHANNEL, "ИБП")
+    runner.monotonic.now += 700
+    runner._reload_metadata(conn, "test")
+    with db.admin() as admin:
+        add_event(admin, NEW_CHANNEL, now - timedelta(minutes=1), "Неисправен")
+        add_event(admin, CHANNEL, now - timedelta(minutes=1), "Неисправен")
+    runner.tick(conn)
+
+    types = {row[1]: row[2] for row in predictions(db)}
+    assert types == {str(NEW_CHANNEL): "EQUIPMENT_FAILURE", str(CHANNEL): "CHANNEL_EVENT"}
+
+
 def test_ttl_keeps_decisions_and_foreign_rows(db, make_runner) -> None:
     runner, conn = started(make_runner, db)
     past = datetime.now(timezone.utc) - timedelta(minutes=1)