Using AI to Reduce False Positives and Line Stoppages for the Automotive Industry

By Dongyan Wang, Mahesh Bhatia, Quinn Killough

Surface inspection is one of the major tasks to ensure quality in the automotive sector, wherein a wide variety of automotive systems, subsystems and parts need to be inspected for defects such as auto-body scratches and part defects like panel splits, cracks, inner-fin defects, defective automotive vents, etc. While the automotive industry has adopted traditional machine-vision technology widely over the last few decades, it suffers from limited accuracy when there are variances in how a specific type of surface defect manifests itself. This leads to a high number of false positives such that defects that don’t exist are flagged.

These shortcomings of traditional machine-vision systems have significant cost implications for the automotive industry. Every so-called defective part that the machine-vision system flags must be either re-inspected by a human worker or scrapped (depending on the economics of the situation), which invariably adds to the total cost of quality (or non-conformance.) Further, auto OEMs and their Tier 1 suppliers may have instituted standard operating procedures that recommend stopping an assembly line if the number of defects that the machine-vision system flags exceed a predetermined threshold over a specific time period. These stoppages result in lower throughput, lost productivity and understandably frustrated workers.

We have demonstrated to our customers that AI-powered vision performs much better than the traditional rule-based machine vision in drastically lowering (up to 90% in some cases) the number of false positives generated. In turn, this reduces the need for unnecessary manual inspections, speeds up the inspection process, reduces the number of line stoppages and improves throughput and productivity.

Powered by our AI visual inspection platform, we have two categories of products for automotive OEMs and their Tier 1 suppliers: (a) Visual Inspection – For detection of cracks, dents, holes, and real-time video inference to identify defects on glossy auto bodies or parts, and (b) Scalable personalized coaching – For assembly-line workers to ensure adherence to standard operating procedures or increase productivity. These products have several key advantages:

  • Example-based Learning – This is the kind of learning that AI-powered vision applications use to successfully tackle some of the key challenges that routinely stymie traditional machine-vision systems. For example, AI in vision really shines when one needs to account for the acceptable variability in surfaces, textures, defects, etc., which are virtually impossible to codify into rules. This is also the key reason why traditional rule-based, machine-vision systems generate a high number of false positives.
  • Realize Return on Previous Investments – Realize return on investments in custom-designed, image-acquisition systems by merely replacing or augmenting the system’s rule-based, machine-vision software with AI models trained with examples of good and defective parts.
  • Overcome Challenges of Data Availability – With the widespread adoption of Lean Six Sigma practices in the auto industry over the last couple of decades, most OEMs and their Tier 1 suppliers strive to have less than 3-4 defects per million parts. This poses a huge challenge in terms of having sufficient defect data on which to train AI models. However, our proprietary end-to-end platform for developing, deploying and maintaining AI models for vision applications has many powerful capabilities. One of them is the ability to request synthetically generated data to augment the sparse defect data that is available, which helps improve AI-model performance.
  • Analytics and Traceability – These are essential features since the automotive industry continuously strives to gain transparency into all aspects of its operations. These include the querying of aggregated performance metrics (daily, monthly, quarterly, etc.) and an ability to access archived images or videos of defective parts with a time stamp, characteristics and location of defects, etc.
  • System and Environment Monitoring – 24-7 remote monitoring of the system and its environment is necessary to ensure optimal operation of the AI system. Audio/visual alarms warn factory floor operators of sub-optimal conditions that might require their attention and action.

With significant technological advances in AI-powered vision over the last few years, powerful visual inspection capabilities are now available to the automotive industry. These can help make a meaningful impact on the total cost of nonconformance (CONC) and enhance the speed of quality inspections.