Plant detectors: boiler fouling, tower fan effort, condenser bypass

Three detectors added in 0.92 for plant faults that the earlier rules could not see on the labelled LBNL chiller and boiler plants. All three are provisional (see API stability); their thresholds are screening-grade, set from the size of change worth acting on rather than tuned on the validation data. Each one also names the sensor problem that would fake its symptom, and reports it as a sensor problem (info) instead of an equipment fault.

Detector	Kind	Watches	Target fault	Fakes it
boiler_efficiency_drift (#13)	drift, boiler family	gas input per unit of heat delivered, at matched load	fire- or water-side fouling	a hot-water temperature or flow sensor that moved
cooling_tower_fan_effort_drift (#14)	drift, tower family	tower fan speed at matched load and wet-bulb	fill fouling, plugged nozzles, lost airflow on a controlled tower	a tower leaving-water sensor that reads warm
condenser_bypass_leak (#15)	single period, built-in	condenser water entering the chillers vs the towers' leaving water, bypass commanded shut	a leaking or stuck-open tower-bypass valve	an offset between those two sensors

The drift detectors take an injected BaselineStore and freeze their reference from a baseline window, like the chiller and pump drift families; run them with drift.families in a config ({"class": "...", "family": "boiler"} or "tower") or through Registry.run_periods. camber.plantdrift rolls each family up to a locus: the equipment, a sensor, or steady.

Since 0.98 a family can also declare its reference instead of reading a frozen baseline: {"class": "CHW_PLANT", "family": "tower", "reference": {"equip": "PLANT__fault_free"}} scores every other unit against a named healthy one, and "reference": {"period": [start, end]} against a known-good window of the same unit. The reference is fitted in memory on every run and never stored; each finding records it in baseline_source with a caveat, and the reference unit itself declines as is_reference. See the CLI guide for the windows and the declines, and Tuning for choosing a reference.

Boiler combustion efficiency (boiler_efficiency_drift)

A fouled boiler burns more gas for the same heat, so its input/output ratio (1 / efficiency) rises. The rule computes

ratio = gas input rate / heat delivered,   heat = 500 x gpm x (supply - return) / 3412.14 kW

on the firing hours (gas above zero, and the boiler status when one is mapped), fits it against the heat delivered (part-load efficiency varies) and, when it moves by 2 F or more, the return-water temperature (a condensing boiler loses efficiency as its return warms), and scores the current period's median residual. Without a flow meter, pump speed x delta-T indexes the heat instead: the drift is judged as a relative change, so the ratio's units cancel and a gas rate in any consistent unit works. Warn at +5 % and 1.5 sigma, fault at +15 % and 3 sigma.

The corroboration. A biased supply or return sensor moves the computed heat and the ratio without a single extra cubic foot of gas. So a second frozen model fits the gas input against OAT (the building's heating signature): fouling raises the gas burned at matched weather, a sensor bias does not. A ratio rise the gas does not share (a gas rise under a third of the ratio rise) is reported as attribution="heat_metering" at info, naming the sensors. Without OAT the rule cannot tell the two apart and caps its severity at warn.

Inputs: gas_input_rate (new in 0.92 -- a rate, kW; difference a totalizer first), hw_supply_temp, hw_return_temp, and hw_flow or hw_pump_speed; oat and boiler_status optional. Brick Natural_Gas_Flow_Sensor maps to the gas role; a Gas_Meter used as a point type is accepted with a caveat, and Natural_Gas_Usage_Sensor (a cumulative amount) is not mapped.

Cooling-tower fan effort (cooling_tower_fan_effort_drift)

Most towers run their fans under a leaving-water controller. A controlled tower that fouls keeps its approach and works its fans harder, which is why cooling_tower_approach_drift scores the LBNL tower-fouling runs as healthy. This rule fits the fan speed against the condenser load -- the tower's range (cw_return_temp - cw_supply_temp, heat per unit of a usually constant condenser flow), else the chilled-water tons -- and the wet-bulb (measured, or from OAT + RH), and scores the current period's median fan-speed residual. Warn at +5 %-points, fault at +10. The high-fan share (hours at 90 % or more) is reported alongside.

The cross-check. A tower controller whose leaving-water sensor reads 2 F warm cools the real water 2 F further and spends the fan effort to do it. When the water entering the chillers (cond_entering_water_temp) is trended, the rule compares the two sensors' offset -- with the bypass shut, if its valve is mapped -- in the baseline and current periods; a shift of 1 F or more is reported as attribution="sensor_offset" at info. Without that point the finding says it could not rule a sensor bias out.

Limits. The range stands in for heat rejected only while the condenser flow is steady; a plant that changes its condenser pumping between baseline and current (a pump swap, a VFD condenser loop) moves the fan-effort model for reasons that are not the tower. A tower whose fans are already at full speed shows fouling in its approach, not its effort -- run both rules (the tower family does).

Condenser-water tower-bypass leak (condenser_bypass_leak)

With the bypass valve commanded shut, all the condenser water goes through the towers, so the water entering the chillers equals the towers' leaving water. A leaking or stuck valve mixes warm return water back in. The rule takes the samples with the valve at or below 2 % while a chiller runs (its status, else its power) and scores the median of cond_entering_water_temp - cw_supply_temp: warn at 2 F (twice the combined accuracy of two plant sensors), fault at 5 F. With the condenser return temperature it estimates the mixing fraction the temperatures imply, (entering - tower leaving) / (return - tower leaving).

Leak or offset? The temperatures alone cannot tell a leak from a miscalibrated sensor, but their shape can: mixed-in return water raises the entering water in proportion to the condenser range, a sensor offset adds the same amount at every load. The rule regresses the difference on the range; a difference that does not grow with it (slope under 0.05 per F, intercept carrying at least half of it) is reported as attribution="sensor_offset" at info, as is entering water reading colder than the tower's (mixing can only warm it). Without the return temperature the rule says it cannot tell them apart.

Two roles are new in 0.92: cond_entering_water_temp (condenser water after the bypass mixing) and cw_bypass_valve (the bypass command or position, %). Map the entering temperature only where the plant trends both points; with one condenser-supply sensor it is cw_supply_temp. Brick: a Condenser_Water_Bypass_Valve's valve points and a Bypass_Command on it map to the valve role; a cooling tower's own Leaving_/Entering_Water_Temperature_Sensor map to cw_supply_temp / cw_return_temp; and a chiller's Entering_Condenser_Water_Temperature_Sensor becomes cond_entering_water_temp when the model also has that tower point.

Hot-water pump warm-weather lockout (hw_pump_summer_lockout)

PNNL's heating-plant re-tuning locks out the hot-water pumps, not only the boiler, in warm weather. boiler_summer_lockout judges boiler firing alone, so a pump left running all summer with the boiler off passes it: the pump's energy is wasted, and the warm loop feeds heat into every air handler and box whose heating valve leaks. hw_pump_summer_lockout (0.103, #132) is the pump's own check, built-in and single-period, on hot-water plants and pumps.

• When the pump runs. Its run status (pump_status), else its speed (hw_pump_speed above 5 %, the running cut of hw_pump_dp_reset), else the loop flow (hw_flow above 5 % of its own 95th percentile). run_source says which, with a caveat when it is not the status.
• The check. The share of the pump's occupied running hours (the generic weekday schedule) with the outdoor temperature above summer_lockout_oat_f (summer_run_pct): the boiler check's semantics and severity, warn at 5 % and fault at 20 %. Without OAT it is not evaluated.
• The lockout is shared. The parameter has the boiler rule's name and default (65 °F). A config that sets it on boiler_summer_lockout only passes the value on; the pump finding records it in param_basis ("inherited from boiler_summer_lockout in this config").
• What else it reports. max_oat_running_f, the warmest outdoor temperature at which the pump ran (the boiler finding reports the same for firing since 0.103), and boiler_off_pct: of the pump's hours above the lockout, the share with the boiler not firing (from its status or gas input). Near 100 % means the pump was left on alone; lower means the whole plant was enabled.
• Which loop. A pump_status point does not say which loop a pump serves. On an equipment with no hot-water point at all (no hw_pump_speed, hw_flow, hot-water temperature or boiler signal) the rule declines rather than judge what may be a chilled-water pump.
• Why a separate rule. Each finding then names one piece of equipment to lock out (a pump can be its own equipment, class pump, and can be locked out by another sequence than the boiler), and the boiler's finding and its existing answers do not change. The RCx report groups a boiler and a pump finding on the same plant into one issue (the hw_lockout cause chain), the boiler first.
• Not a fault everywhere. A plant that serves dehumidification reheat, domestic hot water or freeze protection needs the loop in warm weather: set the lockout to its sequence, or leave the rule out.

It is scored on a synthetic injected fault (camber.faultlab.SCENARIOS, TPR 1.0 and FPR 0.0), a gated synthetic benchmark key since 0.103; no open dataset shows the fault, so the plant-lockout exercise works it on the synthetic catalog entry synthetic-hw-plant-lockout.

Validation

On the labelled LBNL chiller and boiler plants (examples/lbnl_fdd/plant_detectors.py; measured, not gated) the three detectors catch 3/3, 2/3 and 5/5 of their target faults with no false alarm on any other run; see Validation for the intervals and what the sensor-bias runs did.

Through a config (0.98). On these datasets each fault is its own year-long run, so the drift detectors have no before-and-after on one unit. The lbnl-chiller and lbnl-boiler templates (and the plant-cooling-tower and plant-boiler workbook exercises) therefore declare PLANT__fault_free as the reference, and camber run gives the same verdict, run by run, as the example script (tests/test_drift_reference_098.py checks it on the real data when present): cooling_tower_fan_effort_drift catches tower fouling 065 and 080 (+17.8 and +10.9 fan %-points) and misses 095 (+3.3), and boiler_efficiency_drift catches all three boiler foulings (+53.7 %, +24.9 %, +5.1 %; 095 clears the 5 % warn floor by 0.1 point). Both are now the entries' declared targets for camber datasets score. The scorer counts the fault-free run, which declines as the reference, as a correct negative, so the false-positive counts read 0 of 21 and 0 of 14.