How explanation-based anomaly detection prioritizes infrastructure inspection review, what its insulator tests show, and which errors it still misses.
Select a figure to open it at full size.
A drone can photograph thousands of power-line components faster than an inspection team can review them. An image classifier helps process that volume, but its confident answers create another problem: which predictions should an engineer double-check?
Research by Giovanni Floreale, Piero Baraldi, Enrico Zio and Olga Fink investigates a useful middle ground. Instead of asking people to inspect every AI explanation, a second model searches those explanations for unusual patterns and sends selected images for review.
The study concerns photographs of electrical insulator shells. Its contribution is a way to prioritize review of a classifier—not an autonomous maintenance system or proof that equipment failures can be prevented. That distinction matters to infrastructure teams deciding where human attention belongs.
An inspection model may label a damaged component correctly because it recognizes the damage. It may also reach the same answer because a label, background or lighting pattern happens to correlate with damaged examples in its training data. Both predictions count as correct in an ordinary accuracy report, but the second may fail when the surrounding conditions change.
The researchers examine this problem with GradCAM, which produces a heat map of image regions influencing a neural network’s classification. A bright region on the damaged surface is more reassuring than a bright region on an identification tag, although a heat map alone does not establish causation or correctness.
Their next step is to learn what explanations usually look like for each predicted fault class. An unusual explanation then becomes a reason to investigate, even when the prediction itself appears plausible.
The original workflow shows the separation of responsibilities: the first model classifies the image, the explanation module highlights influential regions, and the anomaly detector decides whether that explanation needs attention.
The first model assigns a cropped insulator image to healthy, broken or flashover-damaged. GradCAM then explains the predicted class. The researchers use a separate Deep Semi-Supervised Anomaly Detection model, or Deep SAD, for each evaluated fault-class/classifier combination.
Deep SAD converts a heat map into a compact numerical representation. Training encourages the main body of explanations to cluster together, while explanations associated with known classification errors are pushed away. At evaluation time, distance from the cluster’s center supplies the review signal.
A subtle training choice is important. The authors do not label all explanations of correctly classified images as unquestionably normal. They leave them unlabeled, because some correct predictions may depend on irrelevant clues. Explanations of known misclassifications provide the examples of abnormal behavior.
An engineer then selects the distance threshold using validation data. A restrictive threshold creates a smaller queue containing more likely errors. A permissive threshold catches more problems but also sends more correct predictions for review. This is a workload decision, not merely a model parameter.
The paper’s worked cases are more informative than a single improvement figure. In one case, the explanation emphasizes actual flashover damage. In another, shell color or reflected sunlight contributes to an incorrect damage prediction.
Most revealing is example d below. The classifier reaches the correct fault label, but its explanation emphasizes an identification tag rather than the damaged area. The explanation looks unusual relative to other flashover predictions. Reviewing that case can uncover a training-data weakness even though changing its label would not improve accuracy.
Example f shows the opposite failure. A healthy shell has a stain resembling damage. The classifier is wrong, but its heat map resembles the usual fault explanations, so the second model does not identify the error. An explanation can be familiar and still support a mistaken conclusion.
The study uses the public Insulator Defect Image Dataset: 13,336 healthy shells, 2,564 with flashover damage and 1,204 broken shells. Images from the same insulator remain together when the classifier’s training and evaluation sets are separated, reducing the chance of testing on near-duplicates of training components.
Two classifiers, MobileNetV3 Small and EfficientNet-B0, achieve roughly 91% overall accuracy. The healthy class dominates the dataset, so that overall number obscures weaker performance on the fault classes. The explanation-review method is evaluated on predictions assigned to the broken and flashover classes, using five cross-validation folds.
Table 4 reports review workloads of roughly 11–19% in those evaluated subsets. The corresponding accuracy improvements are about 6–10 percentage points assuming experts correctly reclassify every flagged image. These are not observed improvements from a deployed maintenance crew, and they are not an 8% reduction in infrastructure failures.
The workload reduction also has a cost. The table’s error-detection recall ranges from 24% to 44%: many incorrect predictions remain outside the queue. The chosen threshold deliberately favors a manageable workload. Results vary substantially across folds, reflecting the limited number of error examples available to train and calibrate the second model.
This makes the evidence useful for review prioritization, but insufficient for automatically clearing everything the filter leaves untouched. It also does not establish coverage of faults incorrectly labeled healthy, since the reported review experiments focus on the two predicted fault classes.
For an infrastructure inspection team, the first application is a parallel review queue alongside existing procedures. Keep the model’s proposed fault label, its heat map and the reason for referral together. Ask reviewers to distinguish three outcomes: an incorrect label, a correct label supported by a suspicious clue, and an acceptable prediction.
Those outcomes lead to different actions. Incorrect labels may require case correction. Recurring attention to tags or backgrounds may require changes to the training images. Uninformative heat maps may require a different explanation method rather than another classifier.
The benchmark also suggests a useful acceptance test. Compare the explanation-based queue with the same number of cases selected by low classifier confidence and by random sampling. Measure errors found per hour of expert review, missed faults in an independently checked sample, and recurring shortcut patterns. Retain some random review so the system cannot make its own blind spots invisible.
This complements the broader problem of checking the system around an AI model. Reliability depends on how uncertain outputs reach people, not simply on adding an explanatory picture.
For a PyTorch classifier, Captum’s LayerGradCam provides the explanation component. Select the relevant convolutional layer and target the predicted class, then check that resizing the attribution back to the image preserves alignment. It supplies attributions; it does not decide which inspection images are wrong.
The authors’ method uses Deep SAD for that second role. The Deep SAD research implementation is a starting point for the semi-supervised detector, but connecting it to inspection heat maps requires dataset preparation and class-specific training. It is not a ready-made maintenance product.
Start with one stable classifier and one fault class. Preserve the original photographs and confirmed labels, generate consistent heat maps, separate training and validation by physical asset, and fit the explanation detector. Select the referral threshold against a fixed review budget. Only then evaluate a fresh group of assets, including an independently reviewed sample that was not referred.
Version the classifier, explanation settings and detector together. Retraining the first model can change its heat maps, making an old explanation threshold unreliable even if the classifier’s headline accuracy improves.
The strongest idea here is that inspection quality includes the clues a model relies on, not just whether its label happened to match. Automated explanation review can make that problem manageable at scale.
We would adopt it as an additional way to direct expert attention. We would not use it to justify removing review from all unflagged predictions. The decision to expand should rest on faults found, shortcuts uncovered and residual risk measured on fresh assets—not on the size of the queue alone.
Automated Processing of eXplainable Artificial Intelligence Outputs in Deep Learning Models for Fault Diagnostics of Large Infrastructures, Giovanni Floreale, Piero Baraldi, Enrico Zio and Olga Fink. arXiv version 1, March 19, 2025.
Explore visual defect detection when examples of normal products outnumber labeled faults.
Explore how confidence scores can support escalation or abstention instead of automatic acceptance.
