AI Across Industries: Manufacturing, Logistics, and Beyond

UPI Study › Courses › Introduction to Artificial Intelligence › Free lessons › Lesson 9.4 Free with a free account

AI Across Industries: Manufacturing, Logistics, and Beyond

Introduction to Artificial Intelligence · Chapter 9: AI Applications, Ethics, and the FutureSection 9.4 · lesson 65 of 6915 min · free with a free account

← Transforming Healthcare Through Intelligent SystemsNext lesson: Emerging Directions in AI and Robotics →

Earn 3 college credits for Introduction to Artificial IntelligenceRecommended for credit by NCCRS · Computer Science 373 · self-paced · one payment, lifetime access
Enroll in Introduction to Artificial Intelligence · $250 →Download the syllabus (PDF)

In this lesson

How narrow AI models improve maintenance, inspection, logistics, finance and insurance.

  • Compare reactive, preventive and predictive maintenance, and value a prediction model in money terms
  • Explain why rare defects and frauds lead to low precision, and judge systems on precision and recall together
  • Apply forecasting, safety stock and routing heuristics to logistics decisions
  • Assess data readiness, integration and payback before adopting industrial AI

Free study notes

Get this lesson's notes by email

The key points and 4 practice questions with answers, in your inbox. One email, free.

One email. Unsubscribe in one click.

Transcript

Everything said in the video, section by section (1,977 words). Read along, or search it.

Opening Question

Is Predictive Maintenance Worth Buying

Priya runs maintenance at a plant with 200 machines. Last year, they broke down 200 times without warning, at about $20,000 a repair. That's 4 million dollars. A vendor now offers an AI system that predicts failures. Is it worth buying? And is her plant even ready for it?

The Industrial Loop

Sense, Predict, Act and Feed Back

Industrial AI is rarely one dramatic machine. Usually it's a narrow model attached to one decision: which machine to repair, which order to visit stops in, whether to hold a payment. Most follow the same loop: sense, predict, act, and feed the results back. The feedback step keeps a model trustworthy as conditions change. This lesson's numbers are made-up round figures that show the method, not industry statistics.

Maintenance Strategies

Reactive, Preventive and Predictive

Maintenance balances repairing too late, an expensive breakdown, against too early, wasted parts and downtime. Three strategies mark the field's progress. Reactive maintenance repairs after the machine breaks. No up-front effort, but costly downtime. Preventive maintenance follows a fixed schedule, say every three months. It's simple, but healthy parts get replaced, and some failures still happen between visits. Predictive maintenance repairs when the data says a failure is likely soon. It needs sensors, history and a model kept up to date.

Sensor Signals

Anomaly Detection and Failure Prediction

Sensors measure vibration, temperature, electric current, pressure or sound. A model looks for two kinds of signal. Anomaly detection flags readings that differ from the machine's normal behavior. It works even with few recorded failures. Prediction uses past failures as labeled examples, to estimate the chance of failure soon, or the remaining useful life. And filters from earlier lessons, like a Kalman filter or a hidden Markov model, track a hidden wear level from noisy readings.

Worked Example

Pricing a Failure Prediction Model

A prediction is only useful in money terms. Say Priya's model finds 80 percent of failures, and half of its alerts are right. The 40 failures it misses still cost $20,000 each: $800,000. The 160 it catches become planned repairs at $3,000: $480,000. It raises 320 alerts, so 160 are false alarms, inspected at $500 each: $80,000. The system itself costs $400,000 a year. Total: $1.76 million, a saving of $2.24 million. Catching more failures usually adds false alarms, so the best balance depends on these costs.

Visual Inspection

Cameras Judging Parts at Line Speed

On a production line, cameras photograph each part, and a vision model judges whether it's defective: a scratched panel, a missing component, a cracked weld. It inspects every item at line speed, with the same standard all shift. The difficulties are practical. Defects are rare, so there are few examples of the ones that matter most. Lighting, camera angle, supplier or product changes can quietly lower accuracy. And labels are subjective: two inspectors may disagree about a faint mark. But the least obvious problem is what rare defects do to the numbers.

Rare Defects

High Recall, Low Precision

Ten thousand parts pass the camera, and 1 percent are truly defective. The model catches 95 percent of defective parts, and wrongly flags 2 percent of good ones. That's 95 real catches, but 2 percent of 9,900 good parts is 198 false flags. So of 293 flagged parts, only 95 are defective. Recall is 95 percent, but precision is only 32 percent. About two of every three flagged parts are fine. Overall accuracy is about 98 percent, which hides the problem. So judge inspection on precision and recall together, and send flagged parts to a second check, not straight to scrap.

Quick Check

Cracked Bottles Among the Flags

Quick check. A bottling line checks 20,000 bottles, and half a percent are cracked. The camera catches 90 percent of cracked bottles, and wrongly flags 1 percent of good ones. Of the bottles it flags, roughly what share is really cracked? Pause, and work it out. About 30 percent. A hundred bottles are cracked, and the camera catches 90. But 1 percent of the 19,900 good bottles is 199 false flags. So 90 of 289 flags are real: 31 percent. Ninety percent is the trap. That's the recall, not the precision.

Demand Forecasting

Safety Stock and the Reorder Point

Now logistics. Almost every supply decision starts with a forecast, from past sales, seasons, promotions and weather. Forecasts are never exact, so businesses hold safety stock. A standard rule sets the reorder point: expected demand during the supplier's lead time, plus z times the daily forecast error, times the square root of the lead time. A z of 1.65 gives about a 95 percent chance of not running out. With 50 units a day, 4 days' lead time and an error of 10, that's 200 plus 33: reorder at 233. A better forecast, with an error of 6, needs 19.8 of safety stock. Over 5,000 products at $20, that's about $1.3 million less stock tied up.

Route Planning

Why Trying Every Route Is Impossible

Delivering to many stops is the traveling salesman problem. With a depot and k stops, there are k factorial possible orders. Check a million routes a second, and 6 stops take under a millisecond. Ten take 3.6 seconds, and 12 take 8 minutes. Fifteen stops take 15 days. Twenty take about 77,000 years, and 25 take about 490 billion years. And a real fleet has hundreds of stops, many vehicles, time windows and changing traffic. Trying every route is impossible.

Heuristics

Nearest Neighbor and Reversal Improvements

So planners use heuristics. The simplest is nearest neighbor: always drive to the closest unvisited stop. It's fast, but the last stops can end up far apart. An improvement step then reverses a section of the tour whenever that makes it shorter. In one six-stop example, visiting in list order gives a length of 32.2. Nearest neighbor gives 26.7. Four reversals bring it to 20.8, the best of all 720 orders. On big problems, they get close, but can't promise the best. Machine learning adds realistic travel times from traffic, weather and history. And when a road closes, the plan is recomputed in seconds.

Warehouses

Slotting, Pick Routing and Robot Fleets

Inside a warehouse, the same ideas work at a smaller scale. Slotting stores fast-selling items near the packing area, so pickers walk less. Pick routing orders a picking list to cut walking: a small traveling salesman problem, solved thousands of times a day. Order batching collects items for several customers in one trip. And fleet software assigns tasks to robots, and plans paths so they don't block each other, using multi-agent search.

Other Fields

Farming, Energy and Retail Examples

The same building blocks serve other fields. In agriculture, a camera on a sprayer tells crop from weed, and sprays only the weeds. A 50-hectare field normally gets 2 liters of herbicide a hectare: 100 liters. Treating just 12 percent of the area uses 12 liters, 88 percent less. In energy, grids keep reserve for forecast error. For a 1,000-megawatt peak with 3 percent error, a 95 percent reserve is 1.65 times 30: 49.5 megawatts. Cut the error to 2 percent, and the reserve falls to 33, freeing 16.5 megawatts. That matters more as weather-driven wind and solar grow. Retailers combine demand forecasts with recommendations, and image models power visual search: upload a photo, find similar products.

Fraud Detection

Choosing the Cheapest Threshold

In finance, a fraud model scores each payment, and a threshold decides which are held for review. Say 100,000 payments hold 100 frauds. A review costs $5, and a missed fraud $300. A strict threshold raises 60 alerts and misses 55 frauds: $16,800 in all. A loose one raises 1,500 alerts and misses 5: $9,000. The medium setting, 250 alerts and 20 misses, costs just $7,250. The cheapest point lies between the extremes. Fraud is rare too. If a model catches 90 percent of a thousand frauds, but flags half a percent of a million good payments, only about 15 percent of its alerts are fraud.

Insurance Pricing

Premiums From Expected Loss

Insurance pricing rests on expected loss: the chance of a claim times its typical cost. Say the average claim is $8,000, and expenses and profit take 25 percent of each premium. Then the premium is the expected loss divided by 0.75. At a 3 percent claim chance, the expected loss is $240, so the premium is $320. At 6 percent, it's $480 and $640. Better risk models price more accurately, and image models speed up claims. But insurance is closely regulated. Models must not use factors the law forbids, and customers may be owed an explanation.

Quick Check

Pricing a Home Insurance Premium

Quick check. A home insurer expects a 4 percent chance of a claim, and an average claim of $5,000. Expenses and profit take 20 percent of each premium. What premium should it charge? Pause, and work it out. Two hundred fifty dollars. The expected loss is 0.04 times 5,000: $200. If expenses take 20 percent of the premium, the loss is the other 80 percent, so divide by 0.8. Two forty is the trap: it adds 20 percent of the loss, not 20 percent of the premium.

Integration

Where Industrial AI Projects Fail

Most failures of industrial AI have little to do with the algorithm. A model delivers value only when its output reaches whoever acts on it. That means connecting to old machine controllers and planning systems, moving data in real time, and choosing what runs at the machine, the edge, or in the cloud. Alerts must reach the right person, in time, in a form they trust. Connected equipment widens the attack surface. And the model needs monitoring and retraining. In one illustrative $600,000 budget, data work takes 35 percent, integration 25, and building the model only 15.

Payback

Testing Several Benefit Levels

Does it pay off? Say a project costs $600,000 up front, $120,000 a year to run, and brings $450,000 a year in benefits. The net gain is $330,000 a year, so payback takes 600 over 330: about 1.8 years. But benefits often fall short, and the running cost doesn't shrink. At 80 percent of the benefit, payback is 2.5 years. At 60 percent, 4 years. At 40 percent, it's 10 years, more than five times longer. So test the numbers at several benefit levels before committing.

Data Readiness

Questions Before a Model Can Learn

Barrier two is data readiness. A model can only learn from data that exists, is trustworthy, and matches the decision. Ask about availability: is it recorded and kept? Labels: are failures recorded, with dates? Quality: are values right and complete? Coverage: does it include the rare events that matter? Timeliness: can it reach the model in time? Consistency across sites? And governance: is it lawful to use? Picture a pump whose logs just say repaired, with sensors never connected. There's nothing to learn from yet. The first step is recording failures and storing sensor data, not a bigger model.

Go or No-Go

Checks Before Starting a Project

Other barriers: operators act only on alerts they understand, rules in some fields demand testing and oversight, and a tidy pilot in one plant is far easier than many messy sites. So before starting, ask: is there a specific decision, and someone ready to act? Does the payback survive a big shortfall? Are outcomes already recorded in the data? Is there a path into daily work? And a plan to monitor and retrain? If the answers are weak, start with the data and the integration, not the model.

Back to Priya

Logs First, Then the Model

Back to Priya. On her numbers, a model that finds 80 percent of failures could save over 2 million dollars a year, and even a big shortfall leaves a profit. But she checks her logs first. If they only say repaired, the vendor's model has nothing to learn from. Her first project is recording failures and connecting the sensors.

Key Takeaways

Industrial AI in Summary

To sum up. Industrial AI attaches narrow models to single decisions, in a loop of sensing, predicting, acting and feedback. Rare events make precision low even for good models, so judge precision and recall together, in money. Logistics runs on forecasts and search heuristics. The same building blocks serve farming, energy, retail, finance and insurance. Data work and integration often cost more than the model, and a benefit shortfall can sink a project.

Think About It

Rejecting Cracked Biscuits

Here's something to think about. A food factory wants a camera to reject cracked biscuits, and defects are half a percent of output. Would you ask for its recall, its precision, or both? What would you do with the biscuits it flags, and how would you notice if it slowly got worse?

About Introduction to Artificial Intelligence at UPI Study

Introduction to Artificial Intelligence (Computer Science 373) is a self-paced online course of 69 video lessons in 9 chapters, from intelligent agents and search to game strategy, neural networks, logic and reasoning, machine learning, language, robotics and AI ethics. It carries a NCCRS credit recommendation for 3 semester credits.

Every lesson transcript is free on pages like this one, and the videos are free with a free UPI Study account. Enrolling adds a practice quiz after each lesson, the graded work and final exam that earn the credit, lifetime access, and an official transcript on completion. UPI Study has helped over 130,000 students with college credits. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Enroll now → Course details Download the syllabus All Computer Science courses Every course with PRO →

Compare: AI vs Current Trends in CS and IT

Why take it here

Why students take Introduction to Artificial Intelligence with UPI Study

How machines search, reason and learnIntroduction to Artificial Intelligence covers how AI systems are built: intelligent agents, search and heuristics, game strategy, logic, neural networks, machine learning, language and perception, and the ethics of AI. It is a common upper-level course in computer science degrees.
Credit at a fraction of the priceThe same 3 credits can cost $750 or more at a US college. Here it is one payment, self-paced, with lifetime access.
Finish your degree soonerTransfer credits reduce the courses you still owe, so you can graduate earlier or lighten a heavy semester.
Strengthen an applicationA completed college-level course on an official transcript shows admissions and scholarship committees you can do the work.
Study on your own scheduleNo start dates, no deadlines. Watch the lessons, practice, and take the final when you are ready.
130,000+students helped with college creditsACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Where the credit goes

Universities that work with NCCRS

Introduction to Artificial Intelligence carries a NCCRS credit recommendation. These universities work with NCCRS; some are UPI Study's signed partners. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.

Thomas Edison State UniversityThomas Edison State UniversitySigned partnerUniversity of the PeopleUniversity of the PeopleSigned partnerCharter Oak State CollegeCharter Oak State CollegeSigned partnerStrayer UniversityStrayer UniversitySigned partnerColumbia Southern UniversityColumbia Southern UniversitySigned partnerIndiana TechIndiana TechSigned partnerPenn StatePenn StateWorks with NCCRSSouthern New Hampshire UniversitySouthern New Hampshire UniversityWorks with NCCRSSUNYSUNYWorks with NCCRS +many more cooperating universitiesFind my college →

Find my college → Not listed? Ask Athena in the corner of this page, or book a 15-minute call.

FAQ

Questions students ask us

About AI Across Industries: Manufacturing, Logistics, and Beyond

What is the difference between preventive and predictive maintenance?
Preventive maintenance follows a fixed schedule, such as every three months, while predictive maintenance repairs a machine when its data says a failure is likely soon. A fixed schedule is simple, but it replaces healthy parts and still lets some failures happen between visits. Predictive maintenance needs sensors, failure history and an up-to-date model, often using anomaly detection or remaining-useful-life estimates from vibration, temperature or current readings.
Why do AI inspection systems flag so many good parts?
When defects are rare, a model's false flags on the many good parts can outnumber its true catches. If 1 percent of 10,000 parts are defective and the model catches 95 percent of them while wrongly flagging 2 percent of good parts, it flags 293 parts, only 95 of them defective. Recall is 95 percent but precision is about 32 percent, and an overall accuracy near 98 percent hides the problem.
How do you calculate a reorder point with safety stock?
A standard reorder point is the expected demand during the supplier's lead time plus safety stock, where safety stock equals z times the daily forecast error times the square root of the lead time. With z of 1.65 for about a 95 percent chance of not running out, demand of 50 units a day, a 4-day lead time and a daily error of 10 units, the reorder point is 200 plus 33, or 233 units.
How is an insurance premium calculated from expected loss?
The premium is the expected loss divided by the share of each premium left after expenses and profit. Expected loss is the chance of a claim times the average claim. With a 4 percent claim chance, a $5,000 average claim and 20 percent of each premium going to expenses and profit, the expected loss is $200, and the premium is $200 divided by 0.8, or $250.

About the Introduction to Artificial Intelligence course

How many college credits is Introduction to Artificial Intelligence?
Introduction to Artificial Intelligence (Computer Science 373) carries a NCCRS credit recommendation for 3 semester credits. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities.
How much does Introduction to Artificial Intelligence cost and what is included?
$250, paid once, with lifetime access: all 69 video lessons, a practice quiz after each, the graded work and final exam, and an official transcript when you finish. Students who want several courses choose UPI Study PRO instead.
Which universities accept UPI Study credits?
UPI Study has signed partner agreements with universities such as Thomas Edison State University, University of the People and Charter Oak State College, and works with universities such as Penn State, SNHU and SUNY. ACE and NCCRS have over 2,100 cooperating colleges that will consider accepting credits; the final decision is with the universities. Use "Find my college" on this page to check yours.
How long does Introduction to Artificial Intelligence take?
It is self-paced with no deadlines. The 69 lessons total about 16 hours of video; most students add reading and practice time and finish in a few weeks to a few months, on their own schedule.
How are the exams done?
Quizzes and assignments are online in the course. The final exam is taken online with proctoring; a person on the UPI Study team checks assignments, usually within two days.
How do my credits reach my college?
When the course shows 100% complete, request the official transcript from your dashboard. It is sent to the college you name, and you can send it to more than one. Check your college first with "Find my college" on this page, or ask Athena.

Ready for the full course? Earn 3 credits in Introduction to Artificial Intelligence.

69 video lessons, a quiz after each one, and the graded work that earns the credit. Recommended for credit by NCCRS.

Enroll in Introduction to Artificial Intelligence →Download the syllabusEvery course with PRO →
Introduction to Artificial Intelligence
3 credits · NCCRSEnroll →