Suppose a housing passes final inspection. All dimensions are within tolerance, the visual inspection finds no issues, and the part is shipped. Weeks later, the customer reports that a screw connection inside has come loose. The torque curve had already indicated this issue during assembly, but no one evaluated it. When it comes to defect detection in production, the first question is therefore: At what point can you actually see a defect?
THE MOST IMPORTANT POINTS AT A GLANCE
|
Many manufacturing operations perform extensive inspections but still detect defects too late. They place inspection stations at the end of the line, train visual inspectors, and maintain control charts. Nevertheless, defects reach the customer because each of these methods detects only certain types of defects at a specific point in time.
This article categorizes common methods based on which defects they detect and when. You’ll learn how reliable manual visual inspection is according to research, what statistical process control can and cannot achieve, and why component-specific process data can reveal defects even when the part has been measured as acceptable.
The article concludes with a phased approach that allows you to determine the earliest practical detection point for each type of defect.
IN SHORT
|
What does error detection mean in manufacturing?
Defect detection in manufacturing refers to all methods that identify a deviation from the target state in a part or process. This should be distinguished from two related concepts. Defect prevention takes place earlier in the process and prevents deviations from occurring—for example, through operator guidance that only releases the next assembly step after proper confirmation. Defect prediction, often called “Predictive Quality,” uses process data to estimate the probability that a part will be defective even before it is finished. You can find more on the distinction in the article “What Is Predictive Quality?”
In practice, a second distinction is more important: the point of detection. A defect can be detected directly at the process step, at a downstream inspection station, during final inspection, or only at the customer’s site. Each subsequent stage means more value has already been added to the defective part, and more parts with the same defect have been produced by the time it is detected.
| Key metric | Value | Source |
|---|---|---|
| Cost increase per phase during which a defect remains undetected | approximately a factor of 10 | Rule of Tens according to VDI 2247, cited by the Design Society |
| Industrial average detection rate for manual visual inspection | about 80 percent | See, Sandia National Laboratories, Human Factors 2015 |
| False alarm rate of a Shewhart control chart with 3-sigma limits in a stable process | on average, one false alarm per 371 data points | NIST/SEMATECH Engineering Statistics Handbook |
| Manufacturing companies that see potential for AI in quality management | 58 percent | Bitkom, May 2025 |
Why does the timing of detection determine the cost of errors?
The rule of ten for defect costs states that the cost of correcting a defect increases by a factor of ten with each phase of product development if the defect remains undetected and progresses through the process. The Design Society traces the rule back to VDI 2247 and distinguishes between the phases of design and development, procurement and manufacturing, and delivery and deployment. The same source estimates that the costs incurred by defect correction can amount to as much as 20 percent of revenue.
The rule is a rule of thumb, not a law of nature. However, it accurately describes what happens in manufacturing. A torque error reported by the control system on the screwdriver requires a re-tightening. The same error detected during final inspection requires disassembly, troubleshooting, and re-inspection. At the customer’s site, this is compounded by a sorting operation, an 8D report, and, in the worst case, a recall. The article “Quality Control in Manufacturing: Methods, Costs, Practice” describes how to evaluate the individual cost categories in quality control.
| Detection point | Typical Response | Relative effort according to the rule of tens |
|---|---|---|
| At the process step | Block the part, repeat the step, or rework it | 1 |
| At the inspection station along the production line | Eject part, rework, investigate cause in the preceding process | 10 |
| At the final inspection | Disassembly, troubleshooting, holding the batch | 100 |
| At the customer’s site | Sorting, complaint handling, warranty, possible recall | 1,000 |
The table suggests a simple rule for planning: Each type of defect should be detected at the earliest point at which it is technically measurable. Final inspection remains useful, but as a last line of defense rather than the sole filter.
What types of defects occur in mass production?
Before you select a method, you need a clear picture of the defects that actually occur in your production. The following six categories provide a practical framework for discrete mass production. The right-hand column is particularly important when choosing a method: When does the defect actually become visible?
| Defect Type | Examples | When Do They Become Visible? |
|---|---|---|
| Dimensional Defects | Diameter, position, or gap dimension outside tolerance | Measurable on the part, immediately after the process step |
| Surface defects | Scratches, pits, burrs, coating defects | Visible on the part, depending on lighting and position |
| Assembly defects | Missing or incorrect part, incorrect sequence, mixed-up variants | Often only visible during assembly; later concealed |
| Process defects | Torque, pressure, temperature, or travel outside the limits | In the process data for the step |
| Hidden process errors | Curve profile appears abnormal, but final values are within limits | Only visible in the process data trend |
| Documentation errors | Missing serial number, unassigned measured values | Only detected during traceability checks or audits |
Two categories deserve special attention. Assembly errors often result from mix-ups or omitted steps and are no longer visible after installation. The article “Human Error: Why Inspection Processes Need Clear Rules ” addresses this in detail. Hidden process errors constitute the second group: The final value of a screw connection falls within the tolerance window, but the curve shows settling behavior or a foreign object in the thread. The article “Screw Connection Curves: Patterns and Causes” explains which curve patterns indicate which causes.
How do the various error detection methods compare?
The methods differ in three ways: what they check, when they check, and whether they cover every part or only a sample. The matrix compares seven common methods side by side. Worker guidance is intentionally included in the list, even though it prevents defects rather than detecting them, because it closes the biggest gap in the other methods: assembly errors.
| Method | Detects primarily | Time | Coverage | Limit |
|---|---|---|---|---|
| Manual visual inspection | Surfaces and major assembly defects | After the process or at the end of the line | Depending on the inspection plan, every part or a random sample | Fatigue, daily performance, subjective evaluation |
| Measurement technology at the inspection station | Dimensional errors, functional errors | After the process step | Each part or sample | Only inspected characteristics, no cause |
| Camera-based image processing | Surfaces, presence, position | Inline with the process | Every part | Only visible features, training effort |
| Statistical Process Control (SPC) | Trends, shifts, increased variation | Ongoing via sampling | Sample | Evaluates the process, not the individual part |
| Process data monitoring per component | Process defects, hidden process defects | During the process step | Every part | Only measurable process variables |
| AI anomaly detection | Unknown patterns in curves and images | During or immediately after the process step | Every part | Data foundation, explainability, approval process |
| Digital Worker Guidance | Omitted steps, incorrect parts, and variants | During assembly | Every part | Applies only to guided manual tasks |
The matrix shows why a single method is never sufficient. Visual inspection and measurement technology evaluate the finished part. SPC evaluates the process over time. Only process data monitoring per component and operator guidance capture the exact moment when the error occurs.
How reliable is a manual visual inspection?
Sandia National Laboratories, a research laboratory of the U.S. Department of Energy, has investigated the reliability of visual inspection of precision parts. In the study, published in 2015 in the journal *Human Factors*, 82 inspectors examined 140 parts for eight different types of defects. They correctly rejected 85 percent of the defective parts but also incorrectly rejected 35 percent of the defect-free parts.
The author herself interprets the results: The accuracy rate was not much higher than the industry average of 80 percent and came at the cost of a high rejection rate. The study also demonstrates that visual inspection is a mentally demanding task that requires significant mental effort.
| Metric | Result |
|---|---|
| Inspectors Participating | 82 |
| Parts inspected | 140 |
| Types of defects | 8 |
| Defective parts correctly rejected | 85 percent |
| Flawless parts incorrectly rejected | 35 percent |
| Industry average hit rate according to the study | 80 percent |
|
Practical error: a second round of inspection based on intuition An obvious idea is to have questionable parts inspected a second time. The Sandia study tested exactly that: a tiered inspection process guided by the inspectors’ self-assessment. The result was neither effective nor efficient. If you want to validate visual inspections, replace the judgment in cases where a measured value is available, rather than adding more judgments in sequence. |
This does not mean that visual inspection is unnecessary. For surface characteristics without a measurable substitute parameter, it often remains the most practical solution. It just shouldn’t be the sole method on which your defect detection in production relies.
How does SPC detect defects before scrap is produced?
Statistical Process Control, or SPC for short, monitors a process using samples and control charts. The NIST/SEMATECH Engineering Statistics Handbook, a standard reference published by the U.S. National Institute of Standards and Technology, describes action limits at a distance of three standard deviations from the centerline as the industry-accepted standard.
SPC can predict defects before defective parts are produced because it responds to changes in the process rather than to instances where specifications are exceeded. A process can slowly drift toward a tolerance limit without a single part falling outside it. The control chart shows the drift as long as there is still a margin to the tolerance.
| Rule | What it indicates | False alarm rate for a stable process |
|---|---|---|
| A point outside the 3-sigma limits (Shewhart) | Large, sudden shift | On average, one out of every 371 points (NIST) |
| Additional Western Electric rules, such as eight consecutive points on one side of the centerline | Small, persistent shifts and trends | On average, about one per 92 points (Accendo Reliability) |
The table illustrates the trade-off. Additional rules make the map more sensitive to drift, but the NIST handbook explicitly points out that they increase the number of false alarms. Therefore, select the rules in advance based on process knowledge, rather than waiting until you look at the map.
Fixed thresholds apply in the automotive industry for assessing process capability. A technical paper from the University of Waterloo cites the AIAG PPAP Manual: During initial sampling, a Ppk of at least 1.67 should be achieved; during ongoing production, a Cpk of at least 1.33. To learn how to calculate and interpret both metrics, see “Cpk and Ppk: Calculating and Interpreting Process Metrics.”
The limitation of SPC lies in the sample size. The control chart indicates whether the process is stable. It does not indicate whether the part with serial number X was an outlier between two samples.
How does process data monitoring detect defects per component in real time?
Component-specific process data monitoring fills exactly this gap. Each process step provides its own measured values and curves—such as torque and angle of rotation during screwing, or force and displacement during press-fitting. The software assigns the data to the individual part based on its serial number and immediately evaluates it against limit values, tolerance windows, or envelope curves.
The advantage lies in the timing. The evaluation takes place while the part is still at the station. A part with anomalous data is blocked before assembly continues, and the cause can be investigated while the process is still running. The article “Process Data Analysis in Quality Assurance: Detecting Defects Earlier” describes how to derive patterns and causes from this data.
| Characteristic | Sampling and Final Inspection | Component-Based Process Monitoring |
|---|---|---|
| Time of Evaluation | After the process is completed or at the end of the production line | During the process step |
| Coverage | Selected parts or selected characteristics | Every part, every monitored step |
| Hidden process defects | Remain undetected if the final characteristic meets the requirement | Visible in the curve |
| Traceability | By lot or batch | Via the part’s serial number |
| Narrowing down the issue in case of a complaint | The entire lot must be sorted | Affected parts can be specifically identified |
One point is often underestimated: Limit values based on final values are not sufficient. A bolted joint can reach its target torque within the permitted angular window and still be defective—for example, if a chip in the thread alters its profile. Only by evaluating the entire curve can such cases be identified.
“Many defects that reach the customer were previously visible in a data set. Whether they are detected depends on whether this data set was assigned to a part and evaluated in a timely manner.”
Key message of this article
What can AI do in terms of error detection, and what are its limitations?
In defect detection, you’ll encounter two forms of AI. Computer vision analyzes camera images to detect surface defects and determine the presence and location of components. Anomaly detection based on process data analyzes time series and curves and flags trends that deviate from the learned normal behavior, even if they fall within all specified limits.
In a May 2025 survey by the digital industry association Bitkom, 58 percent of industrial companies saw potential for AI in quality management. Only 24 percent stated that they were already making good use of AI’s capabilities. There is thus a significant gap between expectations and implementation.
The boundaries are specific and should be clarified before a project begins:
- Data foundation: A model can only evaluate what is contained in the data. If there is no mapping between a curve and a serial number, the foundation is missing.
- Explainability: An alert without a clear reason is quickly ignored on the production line. The anomaly must be visible to the process manager, for example, as a highlighted section of the curve.
- Approval and Audit: A model that helps determine whether parts are blocked or approved requires documented approval, version control, and a clear rule regarding who makes the final decision.
AI therefore replaces neither SPC nor limit monitoring. It complements both in cases where defects exhibit unknown patterns. AI provides an overview of useful applications in manufacturing quality assurance. Curve Anomaly AI demonstrates exactly how outlier detection in control charts works.
How do you develop a multi-stage defect detection strategy?
A robust approach to defect detection in production begins with a clear definition: What type of defect is detected where, and by what means? Adding another inspection station does not answer this question. The following six steps build on one another.
- Document the defect profile: Analyze customer complaints, internal scrap reports, and 8D reports from the past twelve months, and classify each case as one of the six defect types.
- Determine the detection point: For each defect type, identify the earliest process step at which it can be technically measured. This is your target.
- Record data by component: Ensure that measured values and curves are assigned to the part via its serial number. Without this assignment, any subsequent analysis will remain a batch-based assessment.
- Set up limit values and control charts: Monitor final values against limits, curves against envelope curves, and critical characteristics using control charts. Define the signal rules in advance.
- Address residual risk with AI: Use anomaly detection where errors occur despite limits being met, typically in screw-fastening and joining processes.
- Feedback for prevention: Trace recurring assembly errors back to the worker’s instructions so that the error doesn’t occur in the first place.
Each type of detected error shows where prevention would be less costly than detection. To see how digital operator guidance reduces assembly errors and training time, see “Digital Operator Guidance: Reduce Errors, Shorten Training Time.”
How does the Manufacturing OS support error detection?
CSP’s Manufacturing OS implements the stages of the concept using the four modules—IPM, QST, PG, and CHRONOS—as well as the AI solution Curve Anomaly AI. The modules operate individually or in combination based on a single dataset.
| Module | Role in Fault Detection |
|---|---|
| IPM (Process Data Management) | Monitors production steps in real time, detects deviations, and immediately alerts the responsible employees. Assigns process data to the component. |
| Curve Anomaly AI | Detects hidden outliers in curves that were previously deemed acceptable. |
| QST (Tool Inspection) | Ensures the inspection of screwdriving tools, thereby laying the foundation for reliable process data. |
| PG (Operator Guidance) | Guides operators step-by-step through the assembly process, prevents errors, and continuously monitors production quality. |
| CHRONOS (Archiving) | Archives process and quality data for the long term so that it remains available in the event of complaints or audits. |
You can find more information about the individual modules on the pages Process Data Management with IPM, Anomaly Detection, and Operator Guidance PG.
Frequently Asked Questions
What is the difference between error detection and error prevention?
Error prevention stops an error from occurring in the first place, for example, through operator guidance that blocks incorrect parts or omitted steps. Error detection identifies an error that has already occurred, such as through measurement technology, visual inspection, or process data. An effective quality strategy combines both.
Which error detection method is the most reliable in mass production?
No single method can detect all types of defects. For process defects, component-specific monitoring of process data is the most direct approach because it evaluates each part at the moment the defect occurs. For surface defects, image processing or visual inspection is suitable; for trends, statistical process control is appropriate.
What is the detection rate for manual visual inspection?
A 2015 study by Sandia National Laboratories published in *Human Factors* cites an industry average of about 80 percent. In the study itself, 82 inspectors correctly rejected 85 percent of defective parts but also rejected 35 percent of defect-free parts.
What does the rule of ten for defect costs state?
The rule of ten states that the cost of correcting a defect increases by a factor of ten for each phase in which it remains undetected. It is based on VDI 2247 and is considered a rule of thumb, not an exact calculation.
Can SPC detect individual defective parts?
Only to a limited extent. Statistical process control works with samples and evaluates whether a process is stable. An outlier between two samples remains undetected. Evaluating each individual part requires component-specific process data monitoring.
How can you detect process defects if the part was measured as acceptable?
By analyzing the process data trend. A screw-in or press-fit curve can reach its final value within the limits and still be suspicious. Envelope curves and AI anomaly detection evaluate the entire trend and reveal such hidden defects.
What are the requirements for AI-based defect detection in production?
The most important prerequisite is a clean database in which measured values and curves are assigned to individual parts via their serial numbers. In addition, there must be traceable messages for the production line as well as documented approval and versioning of the model.
Sources
- See, J. E. (2015): Visual Inspection Reliability for Precision Manufactured Parts. Human Factors 57 (8). Sandia National Laboratories: https://www.osti.gov/biblio/1236203
- Summary of the Sandia study (abstract with key figures), science.gov: https://www.science.gov/topicpages/v/visual+inspection+process
- Sandia National Laboratories: Publication page for the study: https://www.sandia.gov/research/publications/details/visual-inspection-reliability-for-precision-manufactured-parts-2015-12-01/
- Design Society: Bug or Feature? Possibilities and Limitations of Defect Management in Product Development (Rule of Ten according to VDI 2247): https://www.designsociety.org/download-publication/27665/bug_or_feature_m%C3%B6glichkeiten_und_grenzen_des_fehlermanagements_in_der_produktentwicklung
- NIST/SEMATECH Engineering Statistics Handbook, Section 6.3.2 “What Are Variables Control Charts?”: https://itl.nist.gov/div898/handbook/pmc/section3/pmc32.htm
- Accendo Reliability: When Are We Out of Control? (False Alarm Rates of the Western Electric Rules): https://accendoreliability.com/when-are-we-out-of-control/
- Steiner, S. H. et al.: Process Capability Indices, University of Waterloo (citing the AIAG PPAP and APQP Manual): https://sas.uwaterloo.ca/~shsteine/papers/cap.pdf
- Bitkom survey from May 2025, reported by Silicon Saxony: Three-quarters of the industry is missing out on AI opportunities: https://silicon-saxony.de/bitkom-drei-viertel-der-industrie-lassen-ki-chancen-liegen/
- CSP: IPM Real-Time Monitoring of Production Steps (product page on DirectIndustry): https://trends.directindustry.com/csp-gmbh-co-kg/project-4599583-1170975.html
- CSP: Worker Assistance PG (product page on DirectIndustry): https://trends.directindustry.com/csp-intelligence-gmbh/project-4599583-1172854.html
- Wirtschaftsforum: Interview with Korbinian Hermann on Curve Anomaly AI and CSP: https://www.wirtschaftsforum.de/csp-gmbh-co-kg/effiziente-anomalie-erkennung-fuer-die-industrie