Theoretical issues in the design of visual inspection systems

Aligning work analysis and modeling with the engineering goals of a Cyber-Physical-Social System for industrial inspection

Human inspectors rely on a significant number of macrocognitive functions, processes, and tacit knowledge to diagnose the condition of aircraft engine components. A deep understanding of inspectors' cognition and actions in the wild may establish the requirements to develop intelligent automation that truly enhances their perceptual, cognitive and social abilities. This paper takes a two-pronged approach to uncover and model the complexity of industrial inspection in a manner that aligns with the technical development stages of a Cyber-Physical-Social System. The findings offer thick descriptions accompanied by four descriptive empirical models that depict inspectors' meaning-making and decision-making processes. It includes how they gather, process, and apply domain-specific knowledge to diagnose a component's condition and how they deal with domain-related factors (norms, institution's rules, standard operating procedures). This study also highlights the support provided by empirical data/models in designers' work packages. It concludes by presenting the design implications of these findings to envision future human-automation work situations.

Some cues are more equal than others: Cue plausibility for false alarms in baggage screening

This study investigated the effects of cue plausibility in a baggage screening task. 120 participants had to indicate whether a prohibited item was present in a series of grey-scaled X-ray images of baggage. They were assisted by a support system, which pointed at the location of a suspicious object. A 2 × 2 × 2 between-subjects design was used. Cue plausibility for false alarms (i.e. how the cued object was similar to a prohibited item) and support system reliability were manipulated at two levels (high/low). Furthermore, half of participants were provided with a rationale about automation failures (RAF) to reduce their negative impact on trust and performance. The results showed lower performance and more compliance with automation suggestions when cues were implausible than plausible. The RAF increased the response time and did not improve detection performance. Overall, this suggests that effective (computer-based) training is needed to reduce the negative effect of plausible cues.