Engine Block & Cylinder Head Defect Inspection: Deep-Learning Vision for Casting Defects

Overview

Engine blocks and cylinder heads can develop defects at different stages of manufacturing. Casting defects such as sand holes, porosity, shrinkage defects and inclusions originate during casting, while machining and handling defects such as impact marks and scratches can occur during subsequent production processes.

Big Bird Industrial's vision inspection solution combines 20 MP industrial cameras, high-brightness lighting, robotic imaging and deep-learning algorithms to inspect both machined-surface defects and visible casting defects on engine blocks and cylinder heads.

The referenced system has a published inspection cycle of 4.5 minutes, with the final configuration and takt customizable according to the workpiece and inspection requirements.

1. Casting Defects and Machining Defects Are Different Inspection Problems

Engine blocks and cylinder heads may contain defects originating from fundamentally different manufacturing stages.

Casting Defects

Typical casting-related defects include:

• Sand holes
• Gas porosity
• Shrinkage defects
• Inclusions

These defects are formed during metal filling, solidification or other casting processes. When they are exposed on the surface, they may appear as dark spots, pits, irregular depressions or areas with different surface texture.

Machining and Handling Defects

After casting, engine blocks and cylinder heads undergo machining, transportation, clamping and other production operations.

Typical defects from these stages include:

• Impact marks
• Scratches
• Surface damage
• Local dents

These defects may have a substantially different visual appearance from casting defects.

This difference is important for machine vision because the two defect groups may require different imaging and algorithm strategies.

2. Why Casting Defect Detection Is Difficult

Casting surfaces are naturally more complex than highly uniform machined surfaces.

A machined surface may have relatively consistent texture and reflectivity, while a casting surface can contain:

• Irregular texture
• Natural surface variation
• Roughness changes
• Different lighting responses
• Material-related visual features

Casting defects such as sand holes and porosity may also vary considerably in shape and size.

A simple threshold-based vision system may therefore struggle to distinguish a true defect from normal casting texture.

For this reason, the referenced system uses deep-learning-based machine vision.

Instead of relying only on fixed grayscale or geometric thresholds, a trained model can learn visual characteristics from representative OK and NG samples.

However, deep learning does not eliminate the need for appropriate imaging. Lighting, optical resolution, sample quality and inspection-region definition remain fundamental to system performance.

3. Inspection Solution: Deep Learning + High-Resolution Vision

The referenced equipment integrates machining-defect and casting-defect inspection into a common inspection workflow.

Machined-Surface Defects

The system can inspect specified machined areas for defects such as:

• Impact marks
• Scratches
• Surface damage

Casting Defects

The inspection scope also includes:

• Sand holes
• Gas porosity
• Shrinkage defects
• Inclusions

The actual detectable defect size and acceptance criteria depend on the workpiece surface, imaging conditions and customer-defined quality standards.

Deep-Learning Inspection

The system uses deep-learning algorithms to analyze captured images.

This approach is particularly useful when defect morphology is difficult to describe using simple fixed rules.

For example, sand holes and porosity may differ in:

• Shape
• Size
• Edge characteristics
• Contrast
• Surface texture
• Distribution

A deep-learning model can be trained using representative samples to identify these characteristics.

4. Hardware Configuration

The referenced system uses high-resolution imaging together with robotic positioning.

Vision System

• 20 MP industrial camera
• Extra-large high-brightness light source

The high-resolution camera provides detailed image information, while the large lighting system is designed to provide suitable illumination across the inspection area.

Robotic Imaging

A FANUC robot is used for automated camera positioning and image acquisition.

This configuration allows the system to inspect multiple areas of complex engine components without requiring a separate fixed camera for every inspection position.

Control and Data Functions

The system integrates:

• Code reading
• Machine vision inspection
• Deep-learning algorithms
• Inspection traceability
• Data analysis
• Remote data access

This allows inspection results to be associated with individual workpieces and reviewed later for quality analysis.

5. Technical Specifications

The following specifications are based on the publicly available product information for the referenced system.

Item	Specification
Application	Commercial vehicles, passenger vehicles, including NEVs
Workpieces	New-energy powertrain components; diesel and gasoline powertrain components
Inspection scope	Machined-surface and casting defects
Casting defects	Sand holes, porosity, shrinkage defects, inclusions
Machined defects	Impact marks, scratches and related surface damage
Published inspection cycle	4.5 minutes
Industrial camera	20 MP
Lighting	Extra-large high-brightness light source
Robot	FANUC
Control	Siemens / Advantech industrial PC
Inspection technology	Machine vision + deep learning
Functions	Code reading, inspection, traceability, data analysis and remote access

The 4.5-minute cycle is a published application value. Actual inspection takt should be evaluated according to the number of inspection regions, workpiece geometry, robot path, image-processing requirements and production mode.

6. Key Selection Considerations

6.1 Build the Sample Database Before Algorithm Development

Deep-learning performance depends heavily on sample quality.

The project should collect representative samples covering:

• Normal products
• Clear NG products
• Different defect types
• Different defect sizes
• Boundary or borderline samples
• Different surface conditions

Boundary samples are particularly important because real production defects are not always as obvious as laboratory examples.

6.2 Define Quantitative Defect Criteria

Terms such as "small porosity" or "minor sand hole" are difficult for a vision system to interpret without engineering criteria.

The quality specification should define, where applicable:

• Minimum defect size
• Maximum acceptable defect size
• Number of defects
• Defect density
• Defect location
• Acceptable distance between defects

The algorithm can then be trained and validated against a clearly defined inspection target.

6.3 Consider the 4.5-Minute Cycle in Production Planning

The published cycle time is relatively moderate compared with high-speed inline vision inspection.

This can make the configuration suitable for:

• Full inspection of critical components
• Batch inspection
• End-of-line quality inspection
• Sampling inspection

If the inspection needs to be integrated into a faster production line, the inspection points can potentially be distributed across multiple parallel stations.

The actual solution should be determined through cycle-time analysis rather than assuming that one robotized inspection cell can cover every takt requirement.

6.4 Plan Data Access and Retention

Remote data-query functions are useful for quality traceability, but the system should be designed around the factory's actual IT environment.

The technical specification should define:

• Data-retention period
• Image-storage capacity
• User permissions
• Remote-access method
• Backup requirements
• Workpiece-to-result association

7. Can Vision Inspection Detect Casting Defects?

Machine vision can detect casting defects that are visually exposed on the inspected surface, including suitable examples of sand holes, porosity, shrinkage-related surface defects and inclusions.

However, there is an important limitation.

A defect completely buried inside the casting with no visible surface indication cannot be reliably detected using surface imaging alone.

Such internal defects may require other non-destructive testing methods, depending on the application, such as:

• X-ray inspection
• Ultrasonic testing
• Magnetic particle inspection
• Other application-specific NDT methods

Therefore, vision inspection should be understood as a method for inspecting visible surface and near-surface manifestations, rather than a replacement for every form of casting inspection.

8. Maintaining Deep-Learning Model Performance

A deep-learning model does not normally become inaccurate simply because it has been operating for a period of time.

However, inspection performance can change when production conditions change.

Examples include:

• New casting material or supplier
• New machining process
• Surface-finish changes
• Lighting changes
• New product models
• New defect types
• Camera replacement
• Fixture or robot-position changes

A robust vision project should therefore include a model-maintenance mechanism covering:

1. New defect sample collection
2. Production sample review
3. Model validation
4. Model update
5. Version control
6. Re-validation after changes

These responsibilities and procedures are best defined in the technical agreement before equipment acceptance.

9. Integrating Vision Inspection with Cleaning

Engine blocks and cylinder heads can undergo cleaning before final inspection.

For production lines requiring both cleaning and vision inspection, the two processes can be integrated into a broader quality-control workflow.

A typical sequence may be:

Machining → Cleaning → Drying → Vision Inspection → Traceability → Assembly

This approach can reduce the possibility of new contamination or impact damage occurring during unnecessary intermediate handling.

It can also make the inspection result more representative of the component's condition immediately before assembly.

Big Bird Industrial has both industrial precision cleaning equipment and machine vision inspection capabilities, making combined cleaning and inspection solutions possible for applications where both processes are required.

10. Big Bird Industrial's Machine Vision Inspection Capabilities

Harbin Big Bird Industrial Co., Ltd. (Big Bird Industrial), formerly Harbin Shimada Big Bird Industrial Co., Ltd., provides integrated industrial cleaning and machine vision solutions for automotive powertrain and precision-manufacturing applications.

Its machine vision business covers six specialized areas:

• Orient See — dimensional and geometric inspection
• Surface See — surface defect inspection
• Assembly See — assembly inspection and poka-yoke
• Inner See — internal-hole and inner-wall inspection
• Paint See — paint-surface defect inspection
• Algorithm & Software Platform — machine vision algorithms and software

For engine blocks and cylinder heads, Surface See can be configured for both machined-surface defects and visible casting defects.

Where internal-hole inspection is required, Inner See can address inspection requirements involving internal holes and inner-wall surfaces.

This allows the inspection system to be configured according to the actual defect type, inspection location and production process rather than applying one imaging method to every area.

FAQ

Can machine vision replace manual inspection for sand holes and porosity?

For defects that are visibly exposed on the inspected surface, automated vision can replace or significantly reduce manual visual inspection.

However, defects completely hidden inside the casting and without any surface indication cannot be detected by surface imaging alone. Such defects require appropriate non-destructive testing methods.

Will a deep-learning model become inaccurate after running for a period of time?

Not simply because of operating time.

Performance can change when the product, process, material, lighting or defect population changes. Continuous sample collection and model validation are therefore important for long-term production stability.

Why is the published inspection cycle 4.5 minutes?

The referenced system performs multiple imaging operations across complex engine-component surfaces using robotic positioning and image processing.

The published 4.5-minute cycle is suitable for the referenced application. If a faster production takt is required, inspection points may be divided between parallel stations or the imaging and robot paths optimized.

Can the vision system be connected with a cleaning machine?

Yes.

Big Bird Industrial provides both precision industrial cleaning equipment and machine vision inspection systems. Where the production process requires it, cleaning and inspection can be connected so that components proceed directly from cleaning and drying to automated visual inspection and traceability.

Conclusion

Casting and machining defects represent two different quality-control challenges for engine blocks and cylinder heads.

Sand holes, porosity, shrinkage defects and inclusions originate from casting, while scratches and impact marks may be introduced during machining, handling or production transfer. Their visual characteristics can differ substantially, making a single fixed-threshold inspection strategy difficult to apply reliably.

The referenced solution combines a 20 MP industrial camera, high-brightness lighting, FANUC robotic positioning and deep-learning machine vision to inspect both specified machined surfaces and visible casting defects.

For production applications, however, the effectiveness of deep-learning inspection depends on more than the algorithm itself. Representative samples, quantitative acceptance criteria, stable imaging conditions and a defined model-maintenance process are all essential parts of a reliable vision inspection system.

For manufacturers looking to combine cleaning, surface inspection and traceability, Big Bird Industrial can integrate these processes into a broader automated quality-control workflow, with SHIMADA retained as part of the company's historical brand identity.