AI in manufacturing: quality inspection and condition-based maintenance

Two programs, one plant floor

Almost every manufacturer who asks us about AI is asking about one of two things. The first is inspection: parts are leaving the line that should not, and the people catching them are expensive, inconsistent, or both. The second is maintenance: equipment is stopping the line, and the calendar-based preventive maintenance program is not catching the failures that matter. Different sensors, different math. Both fail for the same three reasons: nobody budgeted for labeling, nobody priced the false alarms, and the ROI case used numbers the plant controller does not recognize.

For defense industrial base suppliers there is a third pressure. An escape that reaches a depot is not a warranty line item; it is a corrective action request, a source-inspection posture change, sometimes a stop-ship. It is also how a supplier loses a place on a qualified products list that took three years to earn.

What machine-vision inspection actually replaces

It does not replace inspectors first. It replaces a sampling plan. Lines running acceptance sampling use ANSI/ASQ Z1.4, the civilian successor to MIL-STD-105E after that standard was cancelled in 1995. A sampling plan bets that the lot resembles the sample; vision changes the bet by inspecting every part at line rate. "We deployed AI inspection" and "we moved from AQL sampling to 100 percent inspection on three defect classes" are different sentences to an auditor.

Before any model gets built, measure the humans. An attribute agreement analysis under the AIAG Measurement Systems Analysis manual, scored with Cohen's kappa, tells you how often two trained inspectors agree on borderline parts. On cosmetic defects the answer is regularly worse than plant leadership expects, which is what happens when the spec says "no objectionable scratches." Without that baseline, every later claim of improvement is unfalsifiable.

Then comes the physics, which decides more inspection projects than model architecture does. Telecentric lenses for dimensional gauging. Dark-field lighting for scratches, dome for specular curves, coaxial for flat reflective surfaces. Line-scan cameras for continuous web, area-scan for discrete parts. GigE Vision and GenICam so the stack stays standard. The engineers we assign to a vision build spend the first two weeks on optics, lighting, and fixturing repeatability.

Not every defect needs a learned model. Presence, hole count, edge break, and dimensional checks against a tolerance are classical machine vision: easier to validate, easier to defend in an audit, and they do not drift. Reserve learned models for texture, cosmetic judgment, weld quality, corrosion, and rare defects nobody can enumerate.

The labeling burden is the project

Here is the arithmetic that ends most inspection pilots. A defect occurring at 0.3 percent requires imaging 30,000 parts to collect 90 real examples. If the model needs pixel-level masks, each example costs a trained annotator minutes rather than seconds, and that annotator has to know the part. It surfaces in week five as a schedule problem that looks like a technical problem.

Four levers cut the burden, and they compose:

Train on normal parts only. Anomaly-detection methods model the distribution of good product and flag deviation from it, so they need no defect labels. Published work in this family, evaluated on the MVTec AD dataset of 5,354 images across 15 industrial categories, reports image-level AUROC above 99 percent for the stronger memory-bank approaches. Your parts are not that dataset, but the technique takes labeling off the critical path.

Pre-label, then correct. A weak first model proposes; a human accepts, adjusts, or rejects. Correcting a mask is far faster than drawing one, and the gain compounds as the model learns from its own corrections.

Label what the model is unsure about. Uncertainty sampling puts annotation hours where they move the decision boundary instead of on the thousandth clearly-good part. On skewed defect rates that is the difference between a viable labeling budget and an unviable one.

Generate defects you cannot collect. Compositing real defect crops onto real good parts, or rendering from CAD, fills classes that occur twice a year. The rule that keeps it honest: synthetic data may appear in training, never in evaluation.

Build a golden set first: frozen, scored by two inspectors with a third breaking ties, held out of training permanently. It is the only artifact that lets you claim the system got better and be believed.

Condition-based maintenance versus the schedule you already have

Time-based preventive maintenance assumes failure arrives with age. For much industrial equipment it does not. The 1978 Nowlan and Heap study for United Airlines and the Department of Defense, the work that produced Reliability-Centered Maintenance, found that only about 11 percent of the failure modes examined showed a wear-out age a schedule could anticipate. RCM was codified in SAE JA1011 and JA1012, and DoD made Condition-Based Maintenance Plus department policy through DoD Instruction 4151.22.

The number that governs the program is the P-F interval: the time between the point an incipient failure becomes detectable and the point the asset stops doing its job. If a bearing's P-F interval is three weeks and the vibration route runs monthly, the program misses failures while producing a full binder of data. Sampling rate follows P-F interval. Ask for it on every asset before agreeing to a sensor plan.

Name the standards so the program stays legible to an auditor and to the next engineer: ISO 17359 for general guidance, ISO 20816 for vibration evaluation on machine sets (it supersedes ISO 10816 and ISO 7919), ISO 4406 for fluid cleanliness coding, and ISO 13374 with the MIMOSA OSA-CBM reference architecture. Our specialists build against OSA-CBM layers because it separates sensing from diagnosis from recommendation, which is where accountability otherwise blurs.

Failure mode	Sensing that detects it	Signal you are looking for	Typical warning window
Rolling-element bearing spall	Accelerometer, envelope / demodulation analysis; airborne or structure-borne ultrasound	Bearing defect frequencies (BPFO, BPFI, BSF, FTF) rising above the noise floor	Weeks to months
Gear tooth wear or crack	Vibration spectrum; oil debris sensing and ferrography	Sidebands around gear mesh frequency; ferrous particle count and morphology	Weeks
Rotor bar or winding degradation	Motor current signature analysis; partial discharge on medium-voltage assets	Current sidebands at slip-pole frequency; PD pulse rate trend	Weeks to months
Misalignment and imbalance	Triaxial vibration with phase reference	1x and 2x running speed, radial versus axial ratio, ISO 20816 broadband trend	Months
Lubrication breakdown	Scheduled oil analysis; inline viscosity and moisture	ISO 4406 particle counts, additive depletion, water ingress	Days to weeks
Hydraulic or pneumatic leakage	Pressure and flow trending from the PLC; ultrasonic leak survey	Cycle-time drift, actuator dwell growth, compressor duty creep	Days

Warning windows are asset-specific. Establish yours from your own failure history before setting a sampling rate.

Sensor retrofit realities

The asset you want to instrument was commissioned in 2004 and has no spare I/O. Four things decide whether the retrofit produces usable data.

Mounting and sampling decide what you can see. A stud- or adhesive-mounted accelerometer keeps usable response into the high-kilohertz range where early bearing energy lives; a magnet mount rolls off well before that, and a handheld probe is worse. Spectral analyzers conventionally sample at about 2.56 times the maximum analysis frequency, and resolving gear-mesh sidebands needs enough spectral lines to separate them. F-max and line count are diagnostic decisions, not storage decisions.

The environment decides the hardware. Class I Division 2 areas, ATEX or IECEx zones, washdown lines needing NEMA 4X or IP67, and hot sections above a sensor's rated temperature narrow the catalog fast. Wireless nodes quote multi-year battery life at low duty cycles; that number collapses under continuous high-rate sampling, which is the first thing an enthusiastic pilot does.

The network decides whether IT and OT will sign. Data comes off through OPC UA (IEC 62541), Modbus TCP, or MQTT with Sparkplug B, crossing the Purdue-model boundaries of ISA-95 under the segmentation expectations of IEC 62443 and NIST SP 800-82 Rev. 3. Controls engineers will not let a model write setpoints to a PLC, and they are right: design the first system read-only, off a historian or a mirrored port. On federally funded lines, check the camera and network bill of materials against the covered-equipment restrictions in Section 889 of the FY2019 NDAA, implemented at FAR 52.204-25, before anything is purchased.

Installation decides the budget. Conduit, cable tray, an electrician, a drawing update, an outage to mount on a running asset, and historian tag licensing are the cost. On most retrofits we scope, the sensors are a minority of the installed number.

The false-alarm problem, priced

This is where good models produce bad programs. Detection quality interacts with base rate, and at low base rates the interaction is brutal.

Alarm load is a governed quantity in process industries. ANSI/ISA-18.2 puts roughly 150 alarms per day per operator in the "very likely acceptable" band and about 300 at the maximum a person can manage. Every CBM alert lands in that budget alongside the process alarms already there. Forty low-confidence alerts a day is not information; it is spent attention.

Four levers reduce the damage. Move the threshold and accept lower recall where the failure is not safety-related. Require persistence, so an alert needs several consecutive readings rather than one excursion. Require corroboration across modalities, since a vibration trend confirmed by an oil-debris rise is a different claim than either alone. And put a qualified analyst between the alert and the work order, because intrusive maintenance carries its own infant-mortality risk: opening a healthy gearbox can create the failure it was meant to prevent. Then report precision at the chosen operating point. A vendor showing only AUROC is showing the number that flatters the model.

What a defensible ROI calculation looks like

Defensible means it survives a plant controller, an internal audit, and, on federal work, a program office that distinguishes cost savings from cost avoidance.

Start from a measured baseline. Twelve months of scrap, rework, and sort dollars from the quality system; downtime minutes from the MES or the historian, by asset and cause code. Not from the maintenance manager's estimate.

Count all four quality-cost buckets. Internal failure (scrap, rework, sorting), external failure (returns, warranty, field support, corrective action), appraisal (inspection labor, gauging, lab time), and prevention. Counting only scrap understates the value; counting only avoided downtime overstates it.

Price downtime at contribution margin, and only when the line is constrained. An hour lost on a line running below demand costs labor and little else. An hour lost on the plant's constraint costs margin on every unit not made. Vendor decks price every hour as if the line were constrained, and that assumption is where the credibility goes.

Separate savings from avoidance. A dollar removed from a budget line and a dollar of loss that did not occur are both real, and they are not the same argument. Federal customers, and good controllers, will make you say which one you mean.

Put the whole cost in the denominator. Sensors and cameras, install labor, the outage window, OT network changes, licensing, historian tags, labeling hours, retraining hours, and the fraction of an engineer who owns the model after go-live. Model maintenance is a permanent line item, and treating it as a one-time build cost is the most common error here.

Discount it and show payback. Federal benefit-cost analysis uses the real discount rates published annually in Appendix C of OMB Circular A-94; commercial buyers use their own hurdle rate. Show payback period and net present value together.

Show the sensitivity. A single-point ROI is a sales artifact. Vary defect rate, false-alarm rate, and downtime cost by 30 percent in each direction and state whether the decision flips.

A worked illustration, round numbers. A line makes 250,000 parts a year at a 0.6 percent escape rate: 1,500 escapes. At $65 per escape caught downstream, exposure is $97,500 a year. A vision system catching 70 percent recovers $68,250; a 0.5 percent false-reject rate diverts about 1,240 good parts for re-inspection at $4 each, costing $4,960. Against $140,000 installed and $30,000 a year to run, first-year net is roughly $33,300, so payback lands past four years and the project should probably not be funded. Now change one input: if an escape reaching the customer costs $2,000 in containment, sort, and corrective action response, exposure is $3M and payback is measured in weeks. The cost of an escape, not the model's F1 score, decides these projects.

How to structure the pilot: one line, one defect class

The pilot exists to produce a decision, not a demo. The decision rule gets written before any data arrives.

Three rules make the difference. Pilot on the worst line: a system proven on the showcase cell teaches nothing about the other eleven. Keep the scope to one defect class or one asset class. And run shadow mode long enough to cover a shift change, a material lot change, and a tooling change, because those are the events that break vision models and nobody schedules them into a four-week pilot.

Name the production owner at kickoff. Models drift when incoming material changes, when a supplier switches a coating, when a fixture is bumped. Somebody holds the retraining trigger, versions the model, and keeps the decision log. Without that name, the system degrades until an operator stops trusting it, and the failure gets recorded as "the AI did not work."

What changes for defense industrial base suppliers

The images and the sensor data may be controlled. Inspection imagery of a controlled part, and the dimensional data derived from it, can be technical data under ITAR (22 CFR Parts 120 through 130) or subject to the EAR. That determines where training runs, which cloud region and accounts are permissible, and who may hold a login.

Covered defense information carries a control set. DFARS 252.204-7012 requires NIST SP 800-171 protection for covered defense information on contractor systems, with Rev. 3 published in May 2024, and CMMC assessment requirements took effect under 32 CFR Part 170 on December 16, 2024 with a phased rollout. A quality-data lake holding drawings, inspection results, and nonconformance records sits inside that boundary.

The quality system has to absorb the model. Under AS9100 or IATF 16949, an automated inspection step joins the control plan and the process FMEA. It needs measurement system analysis, a documented reaction plan, first article inspection alignment under AS9102, and traceability from a serial number to the model version that judged it.

There is capital available for this work. The Manufacturing USA institutes fund exactly these projects: MxD for digital manufacturing, the ARM Institute for robotics, America Makes for additive. Defense Production Act Title III and the Industrial Base Analysis and Sustainment program address industrial capability gaps, and DoD SBIR and STTR fund the development when a supplier partners with a technology firm.

Bottom line

Machine-vision inspection and condition-based maintenance are mature, well-standardized engineering programs with a learned component in the middle, and the learned component rarely decides the outcome. What decides it is whether the lighting is repeatable, whether the accelerometer can hear the fault, whether the labeling budget was real, whether the false-alarm rate fits the attention available on the floor, and whether the business case counted the escape correctly. Precision Federal builds these systems end to end: sensing and optics, the data path through OPC UA and the historian, the models, the validation record a quality auditor will accept, and the integration into the MES and CMMS. Our engineers, licensed professional engineers, and manufacturing domain specialists run this work for federal, state, and commercial customers, as a prime or under a prime.

Frequently asked questions

Two programs, one plant floor

Almost every manufacturer who asks us about AI is asking about one of two things, even when the request arrives worded more broadly. The first is inspection: parts are leaving the line that should not, and the people catching them are expensive, inconsistent, or both. The second is maintenance: equipment is stopping the line, and the calendar-based preventive maintenance program is not catching the failures that matter. Different sensors, different math. Both fail for the same three reasons: nobody budgeted for labeling, nobody priced the false alarms, and the ROI case used numbers the plant controller does not recognize.

For defense industrial base suppliers there is a third pressure. An escape that reaches a depot is not a warranty line item; it is a corrective action request, a source-inspection posture change, sometimes a stop-ship. Escapes are also how a supplier loses a place on a qualified products list that took three years to earn. That changes the arithmetic below.

What machine-vision inspection actually replaces

It does not replace inspectors first. It replaces a sampling plan. Lines running acceptance sampling use ANSI/ASQ Z1.4, the civilian successor to MIL-STD-105E after that standard was cancelled in 1995. A sampling plan is a statistical bet that the lot resembles the sample. Vision changes the bet by inspecting every part at line rate. "We deployed AI inspection" and "we moved from AQL sampling to 100 percent inspection on three defect classes" are very different sentences to an auditor.

Before any model gets built, measure the humans. An attribute agreement analysis under the AIAG Measurement Systems Analysis manual, scored with Cohen's kappa, tells you how often two trained inspectors agree on borderline parts. On cosmetic defects the answer is regularly worse than plant leadership expects. That is what happens when the written spec says "no objectionable scratches." Without that baseline, every later claim of improvement is unfalsifiable.

Then comes the physics, which decides more inspection projects than model architecture does. Telecentric lenses for dimensional gauging so perspective error stops moving the measurement. Dark-field lighting for scratches, dome for specular curved surfaces, coaxial for flat reflective ones. Line-scan cameras for continuous web or cylindrical parts, area-scan for discrete ones. GigE Vision and GenICam so the acquisition stack is standard rather than a vendor SDK nobody can maintain in year three. The engineers we assign to a vision build spend the first two weeks on optics, lighting, and fixturing repeatability.

Not every defect needs a learned model. Presence, hole count, edge break, and dimensional checks against a tolerance are classical machine vision and metrology: easier to validate, easier to explain in an audit, and they do not drift. Reserve learned models for texture, cosmetic judgment, weld quality, corrosion, and rare defects nobody can enumerate in advance.

The labeling burden is the project

Here is the arithmetic that ends most inspection pilots. A defect occurring at 0.3 percent requires imaging 30,000 parts to collect 90 real examples. If the model needs pixel-level masks, each example takes a trained annotator minutes rather than seconds, and the annotator has to know the part. This surfaces in week five as a schedule problem that looks like a technical problem.

Four levers cut the burden, and they compose:

Train on normal parts only. Anomaly-detection methods that model the distribution of good product and flag deviation from it need no defect labels. Published work in this family, evaluated on the MVTec AD dataset of 5,354 images across 15 industrial categories, reports image-level AUROC above 99 percent for the stronger memory-bank approaches. Your parts are not that dataset, but the technique takes defect labeling off the critical path.

Pre-label, then correct. A weak first model proposes; a human accepts, adjusts, or rejects. Correcting a proposed mask is far faster than drawing one, and the gain compounds as the model improves on its own corrections.

Label what the model is unsure about. Uncertainty sampling puts annotation hours where they change the decision boundary instead of on the thousandth clearly-good part. On skewed defect rates this is the difference between a viable labeling budget and an unviable one.

Generate defects you cannot collect. Compositing real defect crops onto real good parts, or rendering from CAD, fills classes that occur twice a year. The rule that keeps it honest: synthetic data may appear in training, never in evaluation. You validate on real defects or you have validated nothing.

Build a golden set first: frozen, scored by two inspectors with a third breaking ties, held out of training permanently, stored with its adjudication record. It is the only artifact that lets you claim the system got better and be believed.

Condition-based maintenance versus the schedule you already have

Time-based preventive maintenance assumes failure arrives with age. For much industrial equipment it does not. The 1978 Nowlan and Heap study for United Airlines and the Department of Defense, the work that produced Reliability-Centered Maintenance, found that only about 11 percent of the failure modes examined showed a wear-out age a schedule could anticipate. RCM practice was later codified in SAE JA1011 and JA1012, and DoD made Condition-Based Maintenance Plus department policy through DoD Instruction 4151.22. The question is which assets on your floor carry a condition signal worth instrumenting.

The number that governs the program is the P-F interval: the time between the point at which an incipient failure becomes detectable and the point at which the asset stops doing its job. If a bearing's P-F interval is three weeks and the vibration route runs monthly, the program misses failures while producing a full binder of data. Sampling rate follows P-F interval. Ask for it on every asset before agreeing to a sensor plan.

Name the standards so the program stays legible to an auditor and to the next engineer: ISO 17359 for general guidance, ISO 20816 for vibration evaluation on machine sets (it supersedes ISO 10816 and ISO 7919), ISO 4406 for fluid cleanliness coding, ISO 18436 for analyst certification, and ISO 13374 with the MIMOSA OSA-CBM reference architecture. Our specialists write CBM architectures against OSA-CBM layers because it separates sensing from diagnosis from recommendation, which is exactly where accountability otherwise blurs.

Failure mode	Sensing that detects it	Signal you are looking for	Typical warning window
Rolling-element bearing spall	Accelerometer, envelope / demodulation analysis; airborne or structure-borne ultrasound	Bearing defect frequencies (BPFO, BPFI, BSF, FTF) rising above the noise floor	Weeks to months
Gear tooth wear or crack	Vibration spectrum; oil debris sensing and ferrography	Sidebands around gear mesh frequency; ferrous particle count and morphology	Weeks
Rotor bar or winding degradation	Motor current signature analysis; partial discharge on medium-voltage assets	Current sidebands at slip-pole frequency; PD pulse rate trend	Weeks to months
Misalignment and imbalance	Triaxial vibration with phase reference	1x and 2x running speed, radial versus axial ratio, ISO 20816 broadband trend	Months
Lubrication breakdown	Scheduled oil analysis; inline viscosity and moisture	ISO 4406 particle counts, additive depletion, water ingress	Days to weeks
Hydraulic or pneumatic leakage	Pressure and flow trending from the PLC; ultrasonic leak survey	Cycle-time drift, actuator dwell growth, compressor duty creep	Days

Warning windows are asset-specific and vary with duty cycle, load, and lubrication practice. Establish yours from your own failure history before setting a sampling rate.

Sensor retrofit realities

The asset you want to instrument was commissioned in 2004 and has no spare I/O. Four things decide whether the retrofit produces usable data.

Mounting decides bandwidth. A stud-mounted or adhesive-mounted accelerometer keeps usable response into the high-kilohertz range where early bearing energy lives. A magnet mount rolls off well before that; a handheld probe is worse. If sensors go on with magnets because the outage window was tight, the program underperforms for a reason that never appears in the model report.

Sampling rate decides what is visible. Spectral analyzers conventionally sample at about 2.56 times the maximum analysis frequency, and resolving sidebands around a gear-mesh peak needs enough spectral lines to separate them. F-max and line count are diagnostic decisions, not storage decisions.

The environment decides the hardware. Class I Division 2 areas, ATEX or IECEx zones, washdown lines needing NEMA 4X or IP67, and hot sections above a sensor's rated operating temperature narrow the catalog fast. Wireless nodes quote multi-year battery life at low duty cycles; that number collapses under continuous high-rate sampling, which is the first thing an enthusiastic pilot does.

The network decides whether IT and OT will sign. Data comes off through OPC UA (IEC 62541), Modbus TCP, or MQTT with Sparkplug B, crossing the Purdue-model boundaries described in ISA-95 under the segmentation expectations of IEC 62443 and NIST SP 800-82 Rev. 3. Controls engineers will not let a model write setpoints to a PLC, and they are right. Design the first system read-only, off a historian or a mirrored port. On federally funded lines, check the camera and network bill of materials against the covered-equipment restrictions in Section 889 of the FY2019 NDAA, implemented at FAR 52.204-25, before anything is purchased.

Budget honestly. The sensor is the cheap part. Conduit, cable tray, an electrician, a drawing update, an outage to mount on a running asset, and historian tag licensing are the cost. On most retrofits we scope, hardware is a minority of the installed number.

The false-alarm problem, priced

This is where good models produce bad programs. Detection quality interacts with base rate, and at low base rates the interaction is brutal.

Alarm load is a governed quantity in process industries, and the governance applies here too. ANSI/ISA-18.2 puts roughly 150 alarms per day per operator in the "very likely acceptable" band and about 300 at the maximum a person can manage; EEMUA 191 gives comparable guidance. Every CBM alert lands inside that budget alongside the process alarms already there. Forty low-confidence alerts a day is not information; it is spent attention.

Four levers reduce the damage. Move the operating threshold and accept lower recall where the failure is not safety-related. Require persistence, so an alert needs several consecutive readings rather than one excursion. Require corroboration across modalities, since a vibration trend confirmed by an oil-debris rise is a different claim than either alone. And put a qualified analyst between the alert and the work order, which is how mature programs run. Intrusive maintenance carries its own infant-mortality risk: opening a healthy gearbox can create the failure it was meant to prevent.

Report precision at the chosen operating point. AUROC is a model-selection statistic, not a business statistic, and a vendor showing only AUROC is showing the number that flatters the model rather than the one that predicts your cost.

What a defensible ROI calculation looks like

Defensible means it survives a plant controller, an internal audit, and, on federal work, a program office that distinguishes cost savings from cost avoidance. Seven requirements:

Start from a measured baseline. Twelve months of scrap, rework, and sort dollars from the quality system. Downtime minutes from the MES or the historian, by asset and by cause code. Not from memory, and not from the maintenance manager's estimate.

Count all four quality-cost buckets. Internal failure (scrap, rework, sorting), external failure (returns, warranty, field support, corrective action), appraisal (inspection labor, gauging, lab time), and prevention. Counting only scrap understates the value; counting only avoided downtime usually overstates it.

Price downtime at contribution margin, and only when the line is constrained. An hour lost on a line running below demand costs labor and little else. An hour lost on the plant's constraint costs margin on every unit not made. Vendor ROI decks price every hour as if the line were constrained, and that single assumption is where the credibility goes.

Separate savings from avoidance. A dollar removed from a budget line and a dollar of loss that did not occur are both real, and they are not the same argument. Federal customers, and good controllers, will make you say which one you mean.

Put the whole cost in the denominator. Sensors and cameras, install labor, the outage window, OT network changes, licensing, historian tags, labeling hours, retraining hours, and the fraction of an engineer who owns the model after go-live. Model maintenance is a permanent line item; treating it as a one-time build cost is the most common error in these business cases.

Discount it and show payback. Federal benefit-cost analysis uses the real discount rates published annually in Appendix C of OMB Circular A-94; commercial buyers use their own hurdle rate. Show payback period and net present value together, because a two-year payback on a five-year asset and on a fifteen-year asset are different decisions.

Show the sensitivity. A single-point ROI is a sales artifact. Vary defect rate, false-alarm rate, and downtime cost by 30 percent in each direction and state whether the decision flips.

A worked illustration with round numbers, to show the shape. A line makes 250,000 parts a year at a 0.6 percent escape rate: 1,500 escapes. At $65 per escape caught downstream, that is $97,500 of annual exposure. A vision system catching 70 percent of them recovers $68,250, while a 0.5 percent false-reject rate diverts about 1,240 good parts for re-inspection at $4 each, costing $4,960. Against $140,000 installed and $30,000 a year to run, first-year net is roughly $33,300 and payback lands past four years. That project should probably not be funded on those inputs. Change one number: if an escape reaching the customer costs $2,000 in containment, sort, and corrective action response, exposure is $3M and payback is measured in weeks. The cost of an escape, not the model's F1 score, is what decides these projects, and it is the number most business cases never establish.

How to structure the pilot: one line, one defect class

The pilot exists to produce a decision, not a demo. That means the decision rule is written before any data arrives.

Three rules make the difference. Pilot on the worst line, not the best one: a system that works where conditions are ugly will travel, and one proven on the plant's showcase cell teaches nothing about the other eleven. Keep the scope to one defect class or one asset class. And run shadow mode long enough to cover a shift change, a material lot change, and a tooling change, because those are the three events that break vision models and nobody schedules them into a four-week pilot.

Name the production owner at kickoff. Models drift when incoming material changes, when a supplier switches a coating, when a lamp ages, when a fixture is bumped. Somebody holds the retraining trigger, versions the model, and keeps the decision log. Without that name, the system quietly degrades until an operator stops trusting it, and the failure gets recorded as "the AI did not work."

What changes for defense industrial base suppliers

Four things, and they all show up in the contract before they show up in the engineering.

The images and the sensor data may be controlled. Inspection imagery of a controlled part, and the dimensional data derived from it, can be technical data under ITAR (22 CFR Parts 120 through 130) or subject to the EAR. That determines where training runs, which cloud region and which accounts are permissible, and who may hold a login. Deciding it after the GPUs are provisioned is expensive.

Covered defense information carries a control set. DFARS 252.204-7012 requires NIST SP 800-171 protection for covered defense information on contractor systems, with Rev. 3 published in May 2024, and CMMC assessment requirements took effect under 32 CFR Part 170 on December 16, 2024 with a phased rollout. A quality-data lake holding drawings, inspection results, and nonconformance records sits inside that boundary.

The quality system has to absorb the model. Under AS9100 or IATF 16949, an automated inspection step joins the control plan and the process FMEA. It needs measurement system analysis, a documented reaction plan, first article inspection alignment under AS9102, and traceability from a serial number to the model version that judged it. Building that record from the start costs little; retrofitting it after an audit finding does not.

There is capital available for this work. The Manufacturing USA institutes fund exactly these projects: MxD for digital manufacturing and manufacturing cyber, the ARM Institute for robotics and automation, America Makes for additive. Defense Production Act Title III and the Industrial Base Analysis and Sustainment program address industrial capability gaps, and DoD SBIR and STTR fund the development work when a supplier partners with a technology firm on the build.

Bottom line

Machine-vision inspection and condition-based maintenance are mature, well-standardized engineering programs with a learned component in the middle. The learned component rarely decides the outcome. What decides it is whether the lighting is repeatable, whether the accelerometer is mounted where it can hear the fault, whether the labeling budget was real, whether the false-alarm rate fits the attention available on the floor, and whether the business case counted the escape correctly. Precision Federal builds these systems end to end: sensing and optics, the data path through OPC UA and the historian, the models, the validation record a quality auditor will accept, and the integration into the MES and CMMS where the work order gets cut. Our engineers, licensed professional engineers, and manufacturing domain specialists run this work for federal, state, and commercial customers, as a prime or under a prime.

Frequently asked questions