What the evidence says
Several sources describe how the accuracy of automated monitoring and predictive-maintenance systems changes between narrow, controlled test conditions and broader, generalized deployment conditions. In hand-hygiene monitoring for healthcare workers, personalized models trained on a specific worker reach 95% recognition accuracy in a controlled ICU setting, but accuracy decreases to 56% when the same class of model is built to generalize across users and institutions, a difference attributed to variation in hand size and washing habits between individuals. A separate radar/radio-frequency detection approach shows a comparable pattern: accuracy decreases to a range of 68% to 75% when multiple individuals are present simultaneously, compared to single-person scenarios, reflecting interference in the sensing signal under more realistic, less-controlled conditions.
In HVAC fault prediction, feature engineering choice changes predictive accuracy. Dissimilarity-based cohort prediction features produced a higher average AUC than baseline statistical features across all four alarm types tested. Combining dissimilarity-based and baseline features produced the best AUC in three of the four alarm types, reaching 0.899 for one alarm type (AlmSP) and 0.956 for another (IntProt). Separately, certain fault types could be predicted one week in advance with a True Positive Rate as high as almost 100% and a False Positive Rate lower than 5%, though the source does not specify which fault types achieved this combination.
For infrastructure maintenance more broadly, a systematic review of 67 peer-reviewed studies reports that IoT-based smart monitoring decreases maintenance costs by 30% and extends service life by 25% in AI-modular systems relative to traditional methods, including traditional cast-in-place construction for the service-life comparison. The same review reports that AI predictive-maintenance systems, processing continuous multi-channel sensor data, predict pavement failure modes with 85–92% accuracy up to 6 months before the deterioration is observable at the surface.
Manufacturing quality control shows a similar pattern of automation outperforming manual or condition-dependent processes. Computer vision quality control reaches 97% accuracy, compared to traditional quality control whose performance is described as dependent on weather and labor conditions, a reported increase of 27 percentage points. Automated precast pavement manufacturing is associated with an 80% reduction in defect rates compared to baseline production, drawn from a study of automated precast concrete plants cited within the review.
Across these sources, evidence on what happens when systems move from controlled or single-context conditions to generalized, multi-user, or field conditions is present only for the two accuracy-degradation findings (hand-hygiene recognition and radar/RF detection). None of the maintenance-cost, service-life, fault-prediction, or quality-control findings report a matching controlled-versus-generalized comparison; those figures come from single reported conditions rather than paired before/after or controlled/uncontrolled contrasts.
Findings
Cross-user generalization decreases hand-hygiene recognition accuracy in personalized AI models
confidence — direction: highconfidence — magnitude: highconsensus: thin — one scoping review synthesis
- mechanism
- models personalized to a specific healthcare worker's hand size and washing habits do not transfer to other users; generalizable, cross-user models must accommodate that variation
- effect
- accuracy decreases from 95% (setting-specific ICU model) to 56% (generalizable cross-user model)
- conditions / moderators
- reported for ICU hand-hygiene recognition; whether this pattern holds across different clinical environments and AI architectures beyond the ICU context is not established
Simultaneous multi-person presence decreases radar/RF hand-hygiene recognition accuracy
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — one scoping review synthesis
- mechanism
- interference in the radar/radio-frequency signal when multiple individuals are present simultaneously
- effect
- accuracy decreases to a 68–75% range, compared to single-person scenarios
- conditions / moderators
- reported for clinical settings; whether this range is stable across different clinical environments and radar implementations is not established
Dissimilarity-based and combined features increase HVAC fault-prediction accuracy over baseline features
confidence — direction: highconfidence — magnitude: medium to highconsensus: thin — single study
- mechanism
- dissimilarity-based cohort comparison, alone or combined with baseline statistical features, changes which patterns the prediction model detects
- effect
- dissimilarity-based features produced a higher average AUC than baseline features across all four alarm types tested; the combination of dissimilarity-based and baseline features produced the best AUC in three of four alarm types, reaching 0.899 (AlmSP) and 0.956 (IntProt)
- conditions / moderators
- validated on historical data from 17 HVAC systems; whether the comparison gave both methods equal training/tuning effort is not stated
Week-ahead HVAC fault prediction achieves high sensitivity and low false-alarm rate for select fault types
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — single study
- mechanism
- not stated beyond the predictive model producing a one-week-advance forecast
- effect
- True Positive Rate as high as almost 100%, with a False Positive Rate lower than 5%, one week before fault occurrence
- conditions / moderators
- applies to "certain kinds of faults"; which specific fault types achieve this performance is not established
IoT-based smart monitoring decreases maintenance costs in AI-modular systems
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — review-level synthesis, underlying study designs unclear
- mechanism
- continuous sensor-based monitoring in place of traditional maintenance methods
- effect
- maintenance costs decrease by 30% compared to traditional methods
- conditions / moderators
- drawn from a systematic review of 67 peer-reviewed studies; whether underlying studies used experimental or observational designs, and how contexts varied, is not established
AI-modular systems increase service life compared to traditional cast-in-place methods
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — review-level synthesis
- mechanism
- not stated beyond IoT-based smart monitoring within AI-modular systems
- effect
- service life increases by 25% compared to traditional cast-in-place methods
- conditions / moderators
- drawn from the same 67-study systematic review; consistency across underlying studies and study design are not established
AI predictive maintenance predicts pavement failure modes months before observable deterioration
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — review-level synthesis
- mechanism
- AI analytics engines process continuous multi-channel sensor streams
- effect
- 85–92% prediction accuracy up to 6 months before observable surface deterioration
- conditions / moderators
- consistency of this accuracy range across model types (ANN, SVM, ensemble learning) and validation in practice are not established
Computer vision quality control increases inspection accuracy over traditional quality control
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — review-level synthesis
- effect
- computer vision quality control reaches 97% accuracy; traditional quality control is described as weather- and labor-dependent, with computer vision an increase of 27 percentage points over the traditional baseline
- conditions / moderators
- whether the traditional baseline figure is representative across all conditions is not established
Automated precast pavement manufacturing decreases defect rates compared to baseline production
confidence — direction: highconfidence — magnitude: mediumconsensus: thin — single cited study within a review
- mechanism
- not stated beyond automation of the precast concrete production process
- effect
- defect rates decrease by 80% in automated precast concrete plants, per the Zhang et al. (2021) findings cited in the review
- conditions / moderators
- whether the underlying study used an experimental or observational design, and applicability across plant types and geographies, are not established