Intelligent online inspection and closed-loop quality control: creating a zero-defect production line for stamped parts
Introduction: From Human Vision to AI Full Inspection
Stamping parts quality inspection has long relied on manual sampling, visual inspection, touch and caliper measurement. However, there are three major flaws in manual inspection: strong subjectivity leads to missed inspection, sampling scheme cannot cover all parts, and response lag causes batch defects. With the stamping speed increasing to more than 800 strokes/min, and the automotive and electronics industries' pursuit of zero defects, intelligent online inspection has become inevitable.
This paper describes the three pillars of modern stamping quality assurance system: optical vision and structured light detection, indirect monitoring based on sensors and force waveforms, and closed-loop feedback control and deep integration of SPC.
I. Typical defects of stamped parts and their physical characteristics
Defect Category Appearance/Geometric Feature Generation Mechanism Online Detectable Means
Cracking/necking local blackening, light transmission, thinning rate > 25% tensile stress metamaterial limit high resolution camera + transmitted light/thermal imaging
Wrinkling and undulating, insufficient blank holder force for material accumulation, or multi-laser triangulation scanning of materials
Burr excessive punching edge protruding metal spike punch wear, excessive gap edge backlight projection + sub-pixel
Concave and convex injury point or strip surface indentation foreign body or mold defect multi-angle bright field + dark field combination
2D/3D Visual Measurement of the Position Deviation, Bending Angle Deviation and Springback, Feeding Error
Few holes/porous punching Missing or excess holes Punch break or malfunction Transmitted light + photoelectric array
AI visual inspection: core technologies and engineering deployment
2.1 Imaging system design
The imaging system for on-line inspection of stamped parts faces challenges such as high-speed motion, high metal reflectivity, and environmental oil pollution. Typical configurations include:
Line array camera: with continuous scanning in the direction of motion, suitable for large size sheets.
Area camera + strobe: trigger the dead spot on the press slider to take pictures and capture the parts in a static state.
Telecentric lens: eliminates perspective errors for precision dimensional measurement.
Multispectral lighting: Red ring light highlights surface undulations, blue coaxial light eliminates reflections.
For complex three-dimensional stamping parts, a single camera cannot cover them completely. It is necessary to arrange 4 to 8 cameras to shoot from different angles and synchronize the space through the calibration plate.
2.2 Training and application of deep learning models
Traditional image processing (threshold segmentation, edge detection) can only identify simple defects. For texture defects such as wrinkling and necking, convolutional neural networks must be used. Typical process:
Data collection and labeling: Collect tens of thousands of images of stamping parts, and label the defect type and location one by one by quality inspection experts.
Model training: Adopt advanced architectures such as YOLOv8, EfficientNet, or Swin Transformer to accelerate convergence through transfer learning.
Model optimization: Use TensorRT or OpenVINO inference engine to compress the detection time of a single image to within 10ms.
Deployment and incremental learning: Edge computing devices (such as NVIDIA Jetson) infer in real time while uploading new false positives or false positives to the cloud to regularly update the model.
After a large-scale stamping plant deployed an AI vision inspection system, the detection rate of defects above 0.1mm ² was as high as 99.97%, the false alarm rate was only 0.3%, and it could simultaneously detect four defects: cracking, scratching, bumping, and wrinkling.
2.3 Online size measurement
The plane size (hole position, outline) can be extracted by subpixel edge with backlighting and telecentric lens, and the accuracy can reach ±0.02mm. However, for the three-dimensional angle and drop of the curved part, a laser profiler or a structured light 3D sensor must be used. The latter can obtain the point cloud model of the entire part surface in 0.5 seconds by projecting the fringe pattern and solving the phase, and compare it with the CAD model to generate a color difference map.
III. Indirect monitoring based on sensors and stamping curves
3.1 Pressure curve monitoring (tonnage monitoring)
Each press is equipped with a piezoelectric force sensor on the slider to record the force-time curve for each press stroke. During normal stamping, the curve exhibits characteristic peaks (punching penetration, drawing forming, etc.). When the curve area or peak is outside the statistical control range, it indicates that:
Blanking clearance becomes larger (force decrease) or die interference (force surge)
Fluctuations in material properties (increased yield strength causing the force peak to shift to the right)
Waste is not discharged (secondary shear generates additional peaks)
The advanced tonnage monitoring system is equipped with a learning function, capable of self-learning the standard template of each mold and alarming based on the EWMA control chart.
3.2 Mold acoustic emission and vibration detection
Acoustic emission sensors are very sensitive to high-frequency elastic waves generated by material cracking, coating peeling, and microcrack propagation. For example, when punch microcracks occur, energy peaks occur in a specific frequency band (100-300kHz). The source of the anomaly can be located through multiple AE sensors placed at key locations in the mold.
Vibration sensors are focused on the low-frequency range (0-1 kHz) to reflect loose molds or bearing failures.
3.3 Temperature and lubrication status monitoring
An infrared camera or point temperature sensor monitors the temperature in critical areas of the die. An abnormal rise in temperature may indicate excessive friction or clogged cooling channels. Combined with a lubrication system flowmeter, determine whether the nozzle is clogged.
Statistical process control (SPC) and process capacity enhancement
Online detection generates massive data, which must be converted into management actions through SPC. Key steps:
Define key quality characteristics (CTQ) such as burr height, bending angle, and thinning rate.
Real-time computing process capability index (Cpk): when Cpk
Abnormal pattern recognition: Use the discrimination rules of the control chart (e.g. 8 discrimination criteria: one point exceeds the control limit, 7 consecutive points rise, etc.).
Through SPC, companies can distinguish between "random fluctuations" and "special-cause fluctuations", and thus decide whether to shut down and adjust.
Fifth, closed-loop feedback control: from detection to automatic adjustment
The highest level of intelligent quality system is closed-loop control: the online detection device sends the quality deviation to the press PLC or die servo regulator in real time, automatically correcting the process parameters.
Application example 1: Real-time laser scanning of the rebound angle of the curved part, the controller automatically adjusts the wedge pad at the bottom of the mold after calculating the deviation, changes the bending depth, and compensates for the rebound, and controls the angle deviation within ±0.2.
Application example 2: The line force curve detects the downward trend of punching force, the system determines the punch wear, automatically sends a replacement order to the grinding maintenance station, and executes it in the next mold change cycle instead of waiting for the part to appear burr defects.
VI. Implementation challenges and best practices
6.1 Data synchronization and latency
Online detection must be completed within the stamping cycle (usually 0.1 to 0.5 seconds). High-speed data transmission (10GbE industrial camera interface) and real-time processing by edge computing are required, and the cloud is only used for long-term storage and model training.
6.2 Environmental adaptability
There is oil mist, iron filings and vibration in the stamping workshop. The camera needs to be equipped with a protective cover and positive air purge, and the sensor needs IP67 protection level.
6.3 Best Practice Recommendations
Phased implementation: visual inspection of the most critical defects (cracking, lack of holes) is first carried out, and then gradually expanded to size and surface defects.
Create a defect image database: automatically save images and corresponding sensor data every time an alarm is triggered, for continuous optimization of the model.
Regular comparison of manual re-inspection: In the early stage of system operation, personnel need to be arranged to conduct random inspection of the qualified parts judged by the system to verify the missed inspection rate.
Conclusion: The last line of defense for zero-defect targets
Achieving zero defects in stamped parts cannot rely solely on final inspection. Quality control must be embedded in every stamping cycle. The combination of AI vision, force curve monitoring and SPC closed-loop makes it a reality that "every part is detected, every anomaly is traced, and every deviation is corrected". In the future, with the increase of edge AI computing power and the decrease of sensor costs, comprehensive online inspection will become the standard configuration of stamping production lines, and those companies that still rely on manual sampling will not be able to gain trust in the quality-sensitive automotive, medical and aerospace markets.
BQUQ is a professional metal stamping manufacturer, please send us drawings, and our company will quote you within 12 hours.
