A manufacturer of power electronic components was already using CT to inspect solder joints on heat sinks, but the evaluation macros were generating too many false rejects. Microvista has developed CT software featuring AI models that automatically recognises five component types and provides a pass or fail decision for each component ID. In the attributive MSA involving 120 assessments, no defective part was approved.
Background: CT inspection of solder joints in a production line setting
The project began with an enquiry from a manufacturer of power electronic components. Solder joints on heat sinks are inspected using industrial computed tomography. The CT scanner was already providing the necessary volumetric data: five heat sinks are scanned simultaneously within seven minutes. The solder quality assessment was carried out using evaluation macros, but this generated too many false rejects. As a result, many results had to be checked manually. This took time, tied up staff and prevented the components from being sorted directly in line with the production cycle.
Test criteria for pore detection and solder quality
The crux of the enquiry was therefore clear: the CT inspection was technically feasible, but the evaluation process was not yet aligned with the production cycle. The software was intended to reduce the manual inspection workload, minimise false rejects and ultimately provide a decision on each component that could be used for release.
Customer specifications were used to establish specific rules for each inspection area: relevant soldered areas, permissible pore content, critical material content and limits for ‘pass’ or ‘fail’. The database consisted of actual ‘pass’ and ‘fail’ parts and covered positional deviations, noise, fluctuating scan quality, sheet metal bending and varying solder patterns. The validation plan was also finalised.
From the manual testing process to software logic
At the same time, real CT data and existing assessments were examined. The aim was not simply to replicate the existing inspection process. The software had to reflect the technical logic behind the decision: How is the component aligned? Where are the relevant solder areas? Which areas should be included in the assessment? Where do particular challenges arise as a result of varying scan quality or complex geometries? It is precisely these questions that determine whether the automated assessment will prove viable in subsequent series production.
Even before each CT scan began, the component ID and component position were clearly matched. This ensured the traceability of each individual system throughout the entire test campaign.
After the scan, the components were automatically separated and the belt pulley was aligned for analysis. The alignment was a crucial step in this process. The relevant spalling was located on the upper edge of the belt pulley. In order to detect this reliably, the entire circumference of the pulley had to be precisely registered in each data set.
Based on this registration, a stack of images was generated that extended along the axis of rotation of the belt pulley. This made it possible to systematically assess the entire circumference of the component. Spalling could be identified quickly and reproducibly.
Automatic CT analysis using AI models
On this basis, a multi-stage analysis chain was developed for automatic CT analysis. The software isolates the five components within the CT volume, registers each component using reference points, and extracts the relevant soldered areas, including sheet metal bends. AI models were trained for different inspection regions to segment the soldered areas and analyse for pores or material defects. The end result is an inspection result for each component ID, including an assessment, measured values and the basis for the decision.
The application has been designed for five component types. Following the scan, the CT system transfers the relevant system settings to the software. The analysis module then identifies the component type and automatically loads the appropriate analysis chain, including test ranges, limit values and AI models.
- 1Scan and reconstruction
five heat sinks at once in seven minutes
- 2Identify part type
system setting loads the matching analysis chain
- 3Separate and register
each part in the volume, aligned via reference points
- 4AI analysis
segment solder areas, evaluate pores and material regions
- 5OK or NOK
result per part ID with measured values and reasoning
- 6InspectHub
results, viewer and reports, changes logged
Connection to the CT system, data analysis and the InspectHub portal
To support production-oriented workflows, the software has been integrated with the existing CT infrastructure. Following scanning and reconstruction, the volumetric data is automatically transferred to the CT analysis system. The operator views the results on the analysis screen, checks them in the viewer if necessary, and transfers them – with optional follow-up assessment – to the InspectHub portal. There, inspection results, the viewer and quality reports are all stored in one place. Quality staff can search for components, check for anomalies and trace decisions made for each component.
Roles and perspective on the results
The machine operator needs a clear view of each individual component at each cycle: bounding boxes for anomalies, size templates, segmentation and the rationale for the pass or fail decision. The display must make it possible to identify the fault within a few seconds.
The quality manager begins by reviewing the statistics. They can view affected parts, defect types and borderline cases by batch, time period or inspection area. The individual part remains accessible so that the reasons for non-conformities and any manual investigations can be traced.
A phased approach to the roll-out
The roll-out begins with an assisted assessment. The software carries out an automatic preliminary assessment; the machine operator checks the results in the viewer and confirms them. This feedback highlights borderline cases and helps to further refine the evaluation.
Once validation has been successfully completed, the software carries out the assessment independently. Only random samples or defined special cases are checked manually. The assisted assessment becomes an inline check, which provides a reliable ‘IO’ or ‘NIO’ decision immediately after the CT scan.
The software pre-assesses, the operator checks the results in the viewer and confirms them. The feedback reveals borderline cases.
120 assessments
The software assesses independently. Only random samples and defined special cases are checked manually.
Validation of automated CT assessment using attributive MSA
For the acceptance test, it had to be clear what decision the software was supposed to make. It must clearly distinguish between good and defective parts. Critical parts must not be incorrectly approved.
The validation of automated CT assessment is carried out using an attributive MSA. The process checks whether the software’s output corresponds to the defined reference decision. Although the attributive MSA assesses only the classification as IO or NIO, this decision is based on specific measured values which are recorded by the software and documented in the test result.
The test scope comprised 60 components, each with two CT scans, resulting in 120 assessments. No NIO component was approved as IO. In each scan run, one IO component was assessed as NIO. This resulted in an overall agreement rate of 98.3 per cent and an error rate of 1.7 per cent. The rate of ‘NIO components rated as IO’ was 0 per cent, whilst the rate of ‘IO components rated as NIO’ was 5 per cent. The software therefore generated pseudo-rejects in two cases, but did not allow any defective components to pass.
Result: documented in-line inspection for mass production
The end result is a digital audit trail stretching from the CT scan right through to the documented decision in mass production. Each inspection result is linked to a component ID. The areas relevant to the decision are clearly displayed for the operator in the viewer and documented in the InspectHub portal.
Threshold values, test rules and software version are stored in the results dataset. Manual changes are logged along with the user and the time. This enables the customer to use the CT software from a technical perspective and to approve it in terms of quality.
