Mechademy · Turbomechanica · 2026
Turbomechanica's DiagnosticOps solution is the only one of its kind to provide comprehensive lifecycle management of Diagnostic rules. This includes authoring, backtesting, versioning, state management, and retirement.
Gas and steam turbines, gas engines, centrifugal, reciprocating and screw compressors, gas and liquid expanders, pumps, motors, and gearboxes.
Capabilities
Section 01
This section traces a single Diagnostic rule, Stage 3 Throw 2 Valve Fault on a reciprocating compressor, through eight lifecycle stages. Each platform capability is introduced at its point of use.
Step 01 · Equipment types
An Equipment Type defines the driver-driven equipment model/configuration that the rule can be broadly applied to. A rule declares the Equipment Types it applies to, and the platform binds it to every asset matching any declared Equipment Type.
Step 02 · Cases
The highest confidence Case is listed first. All Cases evaluate on every cycle, and the highest priority match is reported.
Valve failure and valve leakage are the same physics at different severities. Each requires different detection logic, different Insight text, and different urgency. A single Case cannot represent both.
Each Case includes its own detection logic, Insights, Recommendations, and Priority. Priority is High, Medium, or Low, and is configurable per customer.
Step 03 · Qualifier config
An event is raised if the condition defined in the rule has been true for at least 90 minutes in a rolling span of 2 hours. Once raised, it stays open until the rule has been quiet for more than 12 hours in a row.
Adiabatic efficiency loss from a failing valve does not recover until the valve is repaired. Persistence is set to a high fraction of the evaluation window: 1.5 hours true within a 2 hour window. Transient excursions cannot satisfy it.
Instrument looseness and grounding faults appear and clear within seconds. Persistence is set to 1 minute within a 2 hour window. A longer requirement would suppress the detection entirely.
One rule grammar serves both failure modes despite time constants two orders of magnitude apart. Qualifier configuration is the mechanism.
Step 04 · Simulation
Draft events are visible only to the author. No event reaches an operator before promotion.
The backtest answers three questions about a candidate rule:
Adjust the qualifier, re-run, and compare. Deployment is authorized on evidence rather than on author confidence.
Step 05 · Collections
User created. The rule and its events are visible only to the author, so backtesting produces no external effect.
Local to the tenant and not traversed to other Mechademy customers. Holds bespoke customer and equipment specific rules, authored by Mechademy engineers or by the customer.
Authored and validated by Mechademy diagnostics engineers, then deployed across the entire monitored fleet.
Superseded rules are removed from service with their history retained.
Step 06 · On Demand
Any rule can be fired On Demand across historical data. A newly authored rule therefore functions as a search: every prior occurrence of the failure mode, on every asset matching the rule's Equipment Types, is returned in a single execution.
Step 07 · Insights and recommendations
The following issues have been detected on Stage 3 Throw 2 of this compressor:
The above findings are strongly indicative of valve damage.
We recommend inspecting and replacing/reconditioning the compressor throw 2 valves soon.
This Case is high confidence, so the Recommendation specifies a single action. A lower confidence Case returns a ranked list of actions, ordered by likelihood of explaining or resolving the issue.
The reasoning path is retained with the event. A reliability engineer reviewing the event months later can reconstruct the basis for the diagnosis.
Step 08 · Versions and artifacts
Both versions detect the same four confirmed failures. v3 produces six fewer false positives and one day less lead time. Both versions were replayed against the same history, so the reliability engineer can weigh the six false positives against the days of lead time.
The loop
Equipment Types determine which assets the rule applies to. One rule, one machine class, many assets.
Detection logic, Insights, Recommendations, and Priority, per Case, ordered by confidence.
Min occurrence, event trigger, and event reset, matched to the physics of the failure mode.
Private replay over historical data. Detections, lead times, and false positives counted before promotion.
Draft to Facility to Default. Customer authored rules are tenant isolated.
Execute the new rule across the fleet to identify occurrences already in progress.
Insights, Recommendations, and Chain of Thought, bound to the logic that evaluated true.
Candidate against deployed, on identical history. Artifacts retained. Old versions retired, not deleted.
Diagnostic tests
Departure from a first principles asset model such as polytropic head or efficiency. These are synthetic sensors created on the platform.
Departure from a trained multivariate machine learning model of the asset's normal behavior in its current operating context.
Standard and Critical, low and high IOW set points provide tiered limits, usable directly as terms inside a machinery rule.
Dynamically calculated standard and critical bands around operating data based on near term performance.
Statistical departure on a single signal. Low cost and fast, and effective on instrument faults and step changes.
Logic the preceding five test classes cannot express, authored directly by the engineer who understands the asset.
Machine run states
Available run states: Running, Starting, Stopping, Stopped. A Case may be gated to one or more.
High vibration is not a valid test on a stopped machine. Reliability engineers mute these alerts, and muting removes a real detection later.
Coast down surge is the inverse case: it is evaluable only while the machine is stopping. A rule gated to Running never evaluates it, so the fault is never detected.
Run state gating is written into the detection statement, and specifies that the diagnostic logic under test is valid only in a given machine state. Suppression applied after the fact cannot make that distinction.
The event
Between 02:20 and 09:30 CT on 02-Nov-24, the following were detected:
No significant change was observed in stage 1 suction pressure, discharge pressure, or stage 3 discharge pressure, which is strongly indicative of early valve damage on stage 3 throw 2.
There is a high likelihood of valve damage on one or more Stage 3 Throw 2 Valves.
Dispatch field technician to inspect all Throw 2 suction and discharge valves for leakage. Schedule valve replacement or repair before the next run cycle.
Section 02
Each case study runs across two slides: the Diagnostic rule and its logic, then the data and how the fault was caught. Five faults across three equipment classes, each one requiring something a limit on a single tag cannot express.
| Case | Equipment | Fault | What the rule needed |
|---|---|---|---|
| 01 | Reciprocating compressor | Cylinder valve failure | Detection on computed signals no instrument on the machine measures. The measured temperatures indict the wrong cylinder. |
| 02 | Centrifugal compressor | Stonewall | Two coefficients evaluated together, because the fault is defined by position on the curve. |
| 03 | Gas engine | Cylinder head failure | Detection of high exhaust port temperature, validated against multiple field and synthetic sensors. |
| 04 | Centrifugal compressor | Flowmeter drift with surge risk | A third test that must stay False, separating an instrument fault from a real loss of performance. |
| 05 | Gas engine | Engine overload | Using multiple blocks of tiered diagnostic logic to handle unpredictability of available instrumentation. |
Each rule and its Cases are as configured in the platform. Trends are redrawn from the platform's own plots so they can be annotated.
Case 01 · Reciprocating compressor · The rule
Run Status ON, lower efficiency deviation, lower drop in pressure ratio. Fires earlier.
Adiabatic efficiency drifts with cylinder discharge temperature and gas composition. A mass flow increase across the stage with a drop in stage pressure ratio confirms the event is real.
The run state gate removes shutdown and startup excursions. A 90 minute minimum occurrence inside a 2 hour window removes short load swings, which cannot hold both deviations true for that long.
Case 1 requires all three tests and returns a specific action at High priority. Case 2 requires fewer sensors with lower thresholds to trigger.
Case 01 · Reciprocating compressor · The detection
Forty-two days of unit history, replayed at speed. Five signals, one rule.
Case 02 · Centrifugal compressor · The rule
Approaching Stonewall. Run Status ON, head coefficient and flow coefficient at the end-of-curve but not in deep stonewall.
Deep Stonewall without power deviation, for applications with multiple stages.
Approaching Stonewall without power deviation, for applications with multiple stages.
No power deviation implies that the calculated head and measured flow are real and the issue is not related to drifting transmitters.
Centrifugal compressors are commonly started and stopped with the antisurge valve 100% open, which can put the machine into a low power stonewall condition not generally considered detrimental to long term health. The run state gate and a 2 hour minimum occurrence remove those excursions. Long term operation at stonewall is documented to result in failures.
Case 1 requires both coefficients critical, with power confirming the instruments. Case 2 trades depth for earliness at the standard set points. Cases 3 and 4 drop the power test for multi-stage trains where driver power is split across stages, and are marked lower confidence for that reason.
Case 02 · Centrifugal compressor · The detection
Shows 24 hours of operation with about 14 hours in deep stonewall. Both axes are computed, so speed changes, and suction volumetric flow are already normalized out.
Gas demand downstream of the compressor reduced and suction pressure at the station increased. The antisurge valve opened to accommodate the reduced head requirement. The operating point moved to the right until it went into deep stonewall, with isentropic head coefficient and efficiency falling off a cliff.
On variable speed applications the flow and differential pressure at the end of the curve can vary significantly, so a single value cannot be relied on.
Case 03 · Gas engine · The rule
Univariate anomaly detection catches quick changes, and the False flag on engine speed ensures the temperature rise is not simply a speed transient.
The adaptive high on turbocharger inlet temperature confirms cylinder 6 is running hot enough to raise the temperature entering the turbo. A positive deviation against the ML predicted exhaust port temperature confirms the value is high for the operating conditions.
The adaptive low on ignition timing tests whether retarded timing is a candidate cause of the high exhaust port temperature, so the event includes a first line of investigation.
Exhaust valve leak or seat recession on cylinder 6. this is the Case that fired
Cylinder 6 gas admission valve sticking partially open or leaking.
Low magnitude of increase in cylinder 6 exhaust port temperature, same potential causes.
Case 03 · Gas engine · The detection
Twelve days of history from this machine, replayed at speed.
Blind backtest. Historical data from this machine was evaluated using Turbomechanica diagnostic rules with no knowledge of the outcome. The measure of value is how early the issue was detected against when the site detected it and addressed it.
Case 04 · Centrifugal compressor · The rule
For multi-stage trains where actual power is not available because driver power is split across two or more process stages. The Case rests on the suction volumetric flow deviation alone, and because curve shape influences that calculation the deviation margins are wider.
Estimated suction volumetric flow comes off the as-tested curve using calculated polytropic or isentropic head. Actual comes from the raw dP at the flowmeter. A positive deviation means the dP is producing more flow than the head says the machine can be passing.
Power deviation compares actual driver output power against gas power estimated thermodynamically. A dP that has drifted high inflates the estimate, so the deviation turns negative. The same drift moves both signals, in opposite directions.
This third test estimates gas power from calculated head and the curve-derived suction volumetric flow. It still matches the driver, so it evaluates False. A real loss of performance would have moved it too, so a False on this test indicates an instrument fault.
Case 04 · Centrifugal compressor · The detection
Each point is one sample of stage 1 polytropic head against suction volumetric flow measured at the flowmeter. Point color is the suction volumetric flow estimated from the as-tested curve, so the two flow values appear on the same plot: one as position, one as color.
The surge limit line is where the machine surges. The surge control line is where the antisurge controller opens the recycle valve. The controller places the operating point from the flowmeter dP, so a transmitter reading high holds the machine against a position it is not at.
Measured flow bottoms out at about 4,110 ft³/min, so no point crosses the surge control line. Point color is the head-based estimate, which uses no flowmeter dP and runs 200 ft³/min lower across the band. At the control line it reaches 4,000 ft³/min, the bottom of the color scale.
Case 05 · Gas engine · The rule
The engine cannot hold speed against demand. Throttle pins wide open, load factor climbs past 100%, and speed still droops against setpoint. Case 1 adds the misfire signature, one cylinder cold at its own exhaust port while the others make up the load.
Overload shows up on whichever instrument that site happens to have: a load factor above 100%, a univariate anomaly on load factor, or the throttle pinned against its critical high. Any one of them satisfies the same physical condition, so the rule does not need all three to fire.
Engine Overload with no misfire. this is the Case that fired
Engine Overload. Fuel flow sensors not available.
Engine Overload. Fuel flow sensors and throttle position not available.
Engine Overload catch all. Only speed sensors available.
Case 05 · Gas engine · The detection
Speed drops 150 to 190 rpm below its 1350 rpm setpoint in 16 separate spans, about 86 hours of the window. Through the same spans the throttle is pinned wide open at 102% and load factor sits above 100% for 24 hours.