The Weld That Passed Visual and Failed in Service

Why this matters

A weld that passed inspection and then failed is usually not a weld that got worse. It is a weld that was measured against the wrong question. An inspection scoped to defect presence asks whether there is anything wrong with the metal that is there; a conformance check asks whether the metal that the detail called for is there at all. Those are different questions, and only one of them catches a weld that was never finished. This case was reconstructed after the fact, from a fracture and a folder, and the useful part of the finding turned out to be a field the record did not have.

Stop before anyone reads anything. A cracked bracket carrying rotating equipment is a live hazard, and the sequence is fixed. Shut the machine down and lock and tag its disconnect for the stored rotational energy under 29 CFR 1910.147, wait for the rotor to coast to a full stop and confirm zero motion by sight before anyone approaches, and do not rely on a control-side stop. Independently support the equipment on cribbing or a rated stand before any person goes under it or beside it, because the remaining brackets are of unknown condition and one has already failed. Keep everyone out from under the other supports until each has been checked. Only then does anybody start measuring. If the crack has to be removed later, gouging and grinding release the base metal and any coating as respirable dust, an inhalation route needing capture at the tool or a respirator issued under a written program meeting 29 CFR 1910.134, and air carbon arc gouging also puts the work inside a hearing conservation program under 29 CFR 1910.95.

What was on the ticket

A belt-driven blower on a fabricated steel stand, eight identical welded brackets tying a 3/8 in gusset to a 1/2 in base plate at each support point. The contact footprint of each gusset is 5 in long by 3 in wide, and the detail on the shop drawing reads 1/4 in fillet, all around.

The unit ran for about 14 months. At a routine service call the technician found a crack in the base plate at bracket 4, running out from the end of the weld and across the plate, roughly 1 in long, open enough to see without magnification.

Reading the fracture before touching the paperwork

The fracture face told the story before the folder did, and it is worth learning to read because it is free.

• Flat and square to the plate surface, with no necking and no rolled lip. Overload in a ductile steel plate tears with visible plastic deformation. There was none. That rules out a single overload event, which is what everyone reaches for first.
• Curved bands running out from a single origin. Beach marks, each one a boundary where the crack rested or the load changed. They are the signature of a crack that grew a little at a time over many load applications.
• A rougher, coarser region at the far edge. The final fast fracture, where the remaining section could no longer carry the load. Its size tells you how much section was left at the end, and here it was a small fraction of the plate width.
• The origin sits at the end of the fillet weld, on the corner of the gusset footprint, not in the weld metal and not at a weld defect.

So: high-cycle fatigue, initiating at a weld termination. That conclusion is available on site with a light and a magnifier and it constrains everything that follows, because fatigue at a welded detail is governed by the stress range and by the geometry of the detail, not by the strength of the steel.

Why "it ran fine for a year" is not evidence. Print the cycle count. At a shaft speed near 1,750 rpm, an unbalanced rotor loads a support once per revolution, which is 105,000 cycles per running hour. On a duty of roughly 12 hours a day over 14 months at about 30 days a month, that is on the order of 500 million cycles. Both the duty hours and the once-per-revolution assumption are estimates and they are stated as such, but the order of magnitude is the point: a detail that survives a year has not been proven, it has been sampled several hundred million times and finally found.

Walking the record backwards

The folder held five things. Four of them answered.

The visual report is a true statement. Every word of it is accurate about the weld that exists. It reports on defects in deposited metal, and it does not report on how much metal was deposited or where.

Measured on the failed bracket: the fillet ran the two 5 in sides only, terminating square at each corner. Total deposited length 10 in. The detail's all-around requirement on a 5 in by 3 in footprint is a 16 in perimeter. Missing length 6 in, which is 37.5 percent of the required 16 in, and four unreturned weld ends where the detail called for none.

Where the two stories separate

A weld end is a discontinuity in the load path. Load transferring from the gusset into the base plate has to turn the corner, and at a squared-off termination it turns through a sharp geometric change with the highest local stress in the connection sitting right there. That is why the detail said all around: the return carries the weld around the corner so the termination lands somewhere the load is low, instead of at the point of maximum stress. It is a fatigue detail, not a cosmetic one, and its requirement is owned by the engineer of record and the shop drawing, not by the welder's judgement at the bench.

So the two candidate stories, and how the evidence separates them:

• The weld was defective. Ruled out. The origin is not in weld metal, there is no defect at the origin, and the visual report on the deposited weld was accurate.
• The detail was not built. Confirmed. 10 in of a required 16 in perimeter, four square terminations, and the fatigue origin at one of them.

The failure was a conformance failure that a defect inspection cannot see, by design.

Confirming it on the other seven

The other seven brackets were measured against the same detail, with the machine still locked out and the stand still independently supported.

• Five of the seven had the same shortfall: two sides welded, no returns, 10 in of 16 in.
• Two of the seven were welded all around, 16 in, with returns present.
• Across all eight brackets, six were nonconforming and two were correct.

That distribution is the confirmation. A single bad bracket is an accident; six out of eight is a practice, and it points at a fit-up and instruction gap rather than at a bad day. It also predicts where the next crack goes: the two conforming brackets are not the ones to watch.

Two of the six nonconforming brackets, including the cracked one, carried the same welder stamp. The other four carried two different stamps. So the practice was not one person's, which changes the corrective action from retraining an individual to fixing what the shop's instructions and its inspection scope actually require.

What the acceptance stamp was actually worth

It was worth its scope, which is the general rule and the whole finding. "No undercut, no surface porosity, accept" bought a defensible statement about defects in the metal present on the day. It bought nothing about weld size, weld length, weld returns, or joint type, because none of those were fields on the form. The related card on what a weld record has to show later owns the field list that closes this gap; the point this case adds is that the missing field was not a hard one, and nobody noticed it was missing until a fracture asked.

The sibling card on visual inspection makes the same call from the other direction: a visual result recorded without its stage and its criterion cannot be interpreted afterwards. Here the stage was final only and the criterion was defect presence, and both of those facts were recoverable from the form. That is the small mercy in this case. The form was honest about its own limits, and nobody read them.

How the repair was scoped

The repair decision is not "weld the missing 6 in." Three things had to be settled first, and only the last one is a shop decision.

The cracked plate is not a welding question. Whether that plate is repairable by excavation and re-welding or has to be replaced belongs to the engineer of record, and the crack extent has to be established before that call rather than after. Magnetic particle examination on the ground surface is the practical way to find the real ends, since crack tips run further than the open portion you can see.

The five uncracked nonconforming brackets are a different case. They have taken the same cycle count without initiating. Completing the returns adds weld to a detail that has already been loaded, and welding onto a fatigue-loaded connection deposits fresh residual stress exactly where the load turns. That is a question for the engineer too, and the framing to hand them is: as-built length, required length, service history in cycles.

Re-examination after any repair uses a method that can find the flaw type in question. Fatigue cracks are surface-initiating and planar, so magnetic particle on ferromagnetic plate reaches them, in two magnetisation directions about 90 degrees apart, and the sibling card on method capability owns why one direction is not enough.

Check the reconstruction against the rules the siblings state

• Counts re-derived from the printed list, last mention first. The closing distribution says six nonconforming and two correct across eight brackets: 1 cracked plus 5 of the 7 measured is 6, and 7 minus 5 is 2 conforming. 6 plus 2 is 8.
• Length arithmetic printed. A 5 in by 3 in footprint has a 16 in perimeter; 10 in was deposited; 16 minus 10 is 6 in missing, and 6 over 16 is 37.5 percent, stated as 37.5 percent.
• The cycle figure carries its basis and stays an estimate. 1,750 rpm times 60 is 105,000 cycles per running hour; times 12 hours a day is 1.26 million a day; times 30 days times 14 months is about 529 million, stated as on the order of 500 million. The once-per-revolution assumption and the duty hours are both named as estimates, and no conclusion in the case depends on the exact value.
• The initiation site is named precisely, so it does not contradict a sibling. The undercut and overlap card has fatigue cracks in welded structures overwhelmingly starting at the weld toe. This origin is a weld end termination, which is a different geometric feature at the same connection, and that distinction is stated rather than glossed.
• The visual report is quoted with its stage and criterion. Stage: final only. Criterion: defect presence, specifically undercut and surface porosity. Nothing in the finding accuses that report of being wrong.

References

• AWS D1.1, structural welding code for steel, in the edition invoked by your project specification or the building code your authority having jurisdiction has adopted, for weld size, length and end return requirements as detailed by the engineer of record
• 29 CFR 1910.147 for lockout and stored rotational energy before work on driven equipment, 29 CFR 1910.134 for respiratory protection during removal, and 29 CFR 1910.95 for the hearing conservation program covering air carbon arc gouging
• See related: Why a Visual Inspection Finds Most of What Matters; What a Weld Record Has to Show Later; How Undercut and Overlap Each Form