A Reliability-Aware Measurement Framework for Broken Glass Insulator Detection Using YOLOv8n and ANFIS-LFBA

Suspension insulators on high-voltage transmission towers are critical dielectric components whose damage can compromise electrical clearance and increase flashover risk. Glass-disc insulators offer an operational advantage because a broken shell is visually detectable during aerial inspection [1, 2]. However, once broken discs accumulate on the same string, timely maintenance becomes essential. In large transmission networks, this creates a substantial inspection burden and makes automated and operationally trustworthy defect screening operationally important [3].

Unmanned aerial vehicle (UAV)-based inspection offers a scalable alternative to manual patrol and manned aerial survey, enabling broader coverage at lower cost and with sufficient spatial resolution to reveal broken glass-disc geometry [4]. Yet a major barrier to operational deployment is often not raw detection sensitivity, but the reliability of the maintenance decision triggered by each alarm. Even a modest FP rate can translate into repeated unnecessary dispatches, reduced operator trust, and avoidable maintenance overhead [5-7].

Despite the progress of deep-learning detectors for insulator inspection, an important methodological gap remains. Most existing studies optimize the detector itself, yet they do not provide a dedicated decision-reliability layer between raw detections and maintenance action [8]. In practice, detector confidence alone is not equivalent to operational trustworthiness: two detections with similar confidence may differ substantially in geometric plausibility and maintenance relevance [9, 10]. Existing approaches, therefore remain limited in three key respects. They do not explicitly separate detection confidence from decision reliability, they do not provide an interpretable and tunable uncertainty-aware acceptance policy, and they rarely treat FP suppression as a first-class operational objective [11, 12].

To address this gap, this paper proposes a reliability-aware two-stage framework that augments a YOLOv8n detector with an Adaptive Neuro-Fuzzy Inference System (ANFIS) reliability scorer and a Logic-Fuzzy Banding Algorithm (LFBA) [13, 14]. The ANFIS maps four candidate-level geometric and confidence features to a continuous reliability score [9, 10, 15]. Because its Gaussian membership functions remain directly interpretable, the scorer can be inspected and audited by maintenance engineers, which is valuable in safety-critical decision support [16, 17]. The ANFIS score is then embedded within the LFBA, which partitions the detector-confidence axis into three adaptive decision zones: automatic acceptance, automatic rejection, and an intermediate band governed by reliability-aware screening. The corresponding zone boundaries are optimized per fold through Lévy-flight bat search. This design further exposes two runtime operating modes-High-Precision (HP) and HR-allowing operators to select the precision-recall trade-off most appropriate to the current maintenance context without retraining the detector. From a measurement perspective, the central objective is not merely to detect damage, but to quantify the reliability with which a candidate alarm can be accepted as operationally actionable. The main contributions of this work are as follows:

(1) A reliability-aware post-detection framework for broken glass insulator inspection that separates detector confidence from operational decision trustworthiness;

(2) An interpretable ANFIS scorer that maps candidate-level geometric and confidence descriptors to a continuous reliability estimate;

(3) A three-zone LFBA decision policy that enables tunable HP and HR operating modes; and

(4) A fold-wise evaluation on the public BGI dataset showing substantial FP reduction relative to a fixed-threshold baseline, together with explicit identification of the methodological limitations that motivate stricter future validation.

The system pipeline is illustrated in Figure 1.

The proposed framework consists of four sequential stages: (i) stratified cross-validation partitioning of the BGI dataset; (ii) per-fold fine-tuning of a YOLOv8n primary detector; (iii) geometric–confidence feature extraction from all candidate detections; and (iv) ANFIS reliability scoring governed by LFBA zone-boundary optimization. The complete pipeline is illustrated in Figure 1. Each stage is described in detail in the following sub-sections.

3.1 Dataset

The experiments were conducted on the BGI dataset [44]. The dataset contains 604 aerial RGB images of 110 kV transmission towers captured using UAV and helicopter platforms under realistic field inspection conditions. Images exhibit substantial variability in viewing angle, illumination, and background environment, reflecting the operational diversity encountered during aerial inspection missions.

The dataset contains a single object class corresponding to broken glass insulator discs, with 604 bounding-box annotations across 604 images (approximately one annotation per image). Bounding-box sizes vary significantly due to differences in camera stand-off distance and tower geometry, producing a wide distribution of object scales.

To obtain robust generalization estimates, stratified 5-fold cross-validation was implemented using Scikit-learn StratifiedKFold with parameters n_splits = 5, shuffle = True, and random_state = 42. A stratification key was constructed for each image by combining: a box-count category (0, 1, 2, ≥ 3), and the quartile bin of log₁₀ (mean bounding-box area). This joint stratification preserves both annotation density and defect scale distribution across folds. Each fold therefore contains 483 training images and 121 validation images. All preprocessing operations—including image resizing, letterboxing, and label preparation—were performed independently within each fold. Care was taken to ensure that no information leakage occurred between training and validation partitions.

3.2 Primary detector: YOLOv8n

The YOLOv8n (nano) architecture [45-48] was selected as the primary detector because it provides a favorable balance between detection accuracy and computational efficiency for UAV-based inspection tasks. The model contains approximately 3.2 million parameters, which is sufficient to achieve high performance on the single-class BGI detection problem while maintaining lightweight inference suitable for aerial deployment.

For comparison, a simple confidence-threshold baseline (single scalar comparison) adds negligible overhead — effectively zero beyond the detector itself. The ANFIS stage (96 parameters, pure NumPy forward pass) adds approximately 0.3–0.8 ms per image, representing less than 20% additional time over the threshold baseline. The LFBA optimization is performed offline once per deployment context and requires less than 8 seconds per fold on CPU, making it operationally negligible. Total online inference time is dominated by the YOLOv8n forward pass (2–5 ms), and the full pipeline remains well within the frame budget of a UAV inspection system operating at standard capture rates of 2–5 Hz.

For evaluation, a separate YOLOv8n model was fine-tuned for each cross-validation fold using COCO-pretrained weights. This ensures that no validation image contributes to the training of the model used for its evaluation. The complete training configuration is summarized in Table 2.

Table 2. YOLOv8n training configuration used in all cross-validation folds

Training was performed for a maximum of 80 epochs with early stopping (patience = 20) based on validation box loss. The learning rate followed a cosine annealing schedule from lr₀ = 0.01 to a final fraction lr_f = 0.01.

Standard Ultralytics data augmentation strategies were applied during training, including mosaic composition, horizontal flipping, HSV color jitter, and random erasing. Validation was performed at a single scale without test-time augmentation. The final model checkpoint (best.pt) was selected automatically as the epoch achieving the highest validation mAP@50.

All experiments were conducted on a Tesla T4 GPU (14 GB VRAM) using CUDA 12.6, PyTorch 2.9, and Ultralytics YOLOv8 with deterministic execution enabled.

During inference, all candidate detections with confidence c≥0.01 are forwarded to the ANFIS reliability scorer. No upper confidence threshold is applied at this stage; the final decision boundaries are determined by the LFBA optimization described in Section 3.5. Data leakage prevention is ensured as follows. The StratifiedKFold split was fixed with random_state = 42, making the partition fully reproducible. All preprocessing operations — image resizing to 640 × 640 px, letterbox padding, and label normalization — were performed independently within each fold using only training-fold data. No statistics or normalization parameters from the validation fold were used to determine any preprocessing parameter.The BGI dataset contains aerial images of approximately uniform resolution; letterboxing pads shorter edges with grey fill to reach 640 × 640 without distorting aspect ratios. Feature normalization for the ANFIS inputs (f1–f4) is similarly computed on training-fold candidates only and applied to validation candidates, preventing any information leakage across the fold boundary. The computational efficiency of these operations and the overall pipeline performance are detailed in Table 3, which provides a breakdown of inference time.

Table 3. Inference time breakdown of the proposed pipeline (Tesla T4 GPU, single image)

3.3 Feature extraction for the Adaptive Neuro-Fuzzy Inference System reliability scorer

The ANFIS reliability scorer is designed to complement rather than replicate the information already contained in the detector confidence score. Candidate descriptors were initially drawn from a larger pool of fourteen geometric and photometric variables derived from the detector output. Feature selection followed three criteria:

1. Discriminative power, measured as separability between TP and FP candidates in pooled validation data.
2. Computational availability, requiring that features be derived directly from bounding-box geometry and detector confidence without additional model inference.
3. Physical interpretability, ensuring that each feature corresponds to a meaningful geometric property of the inspection scene.

Based on the analysis of 703 validation candidates pooled across the five folds (586 TP and 117 FP), four features satisfied these criteria. Their formal definitions and motivations are summarized in Table 4.

The selected features describe complementary aspects of candidate detections: detector certainty, object scale, geometric shape, and spatial context within the image. Together they form the feature vector $x=\left(f_1, f_2, f_3, f_4\right) \in[0,1]^4$ which is used as the input to the ANFIS reliability scorer.

All features are normalized to the unit interval to ensure consistent gradient scaling during ANFIS training. This normalization is required because the Gaussian membership functions used in the fuzzy inference layer operate in a shared feature space where comparable input magnitudes improve optimization stability.

A preliminary separability analysis confirms the relevance of the selected features. Detector confidence f₁ provides the strongest discrimination, with TP mean = 0.8199 ± 0.118 and FP mean = 0.4268 ± 0.167. The center-distance feature f₄ provides complementary information: genuine broken-disc detections tend to appear near the image center because UAV operators typically frame the insulator string during inspection, whereas FP detections are more frequently located in peripheral regions of the image.

The remaining descriptors—log-normalized area f₂ and aspect ratio f₃—provide weaker individual separation but improve discrimination when combined with the other features within the ANFIS rule structure. To verify detection quality, the mean IoU between TP detections and the nearest ground-truth annotation is 0.824, confirming accurate localization, whereas FP candidates show negligible overlap (mean IoU = 0.0317). This contrast confirms that FP detections arise from background structures rather than mislocalized ground-truth objects.

Table 4. Feature definitions, formulas, and physical motivation

3.4 Adaptive Neuro-Fuzzy Inference System reliability scorer

The reliability scoring stage employs an ANFIS implementing a first-order Sugeno-type fuzzy model with trainable parameters [13, 36]. Given the input feature $x=\left(f_1, f_2, f_3, f_4\right) \in[0,1]^4$ the model produces a scalar reliability score $s \in[1]$ through five sequential layers: fuzzification, rule firing, normalization, consequent combination, and output activation. The overall architecture and training configuration are summarized in Table 5.

Table 5. Adaptive Neuro-Fuzzy Inference System (ANFIS) reliability scorer architecture and training parameters

Layer 1 — Fuzzification: Each input feature f_i is evaluated using two Gaussian membership functions (MFs)

$\mu_{i k}\left(f_i\right)=\exp \left(-\frac{\left(f_i-\mu_{i k}\right)^2}{2 \sigma_{i k}^2}\right)$          (1)

where, μ_ik and σ_ik denote the center and width of the k-th MF associated with feature f_i. Initial MF centers are obtained using k-means clustering (k = 2) on TP and FP training candidates of each fold, while all widths are initialized to σ = 0.3. This data-driven initialization accelerates convergence by placing initial MF centers close to the natural TP–FP feature clusters.

Layer 2 — Rule firing: With two MFs per input, the system produces 2⁴ = 16 fuzzy rules.

The firing strength of rule r is:

$w_r=\prod_{i=1}^4 \mu_{i, k_r(i)}\left(f_i\right)$           (2)

where, k_r denotes the MF index selected for feature i in rule r.

Layer 3 — Normalization: Rule firing strengths are normalized as:

$\bar{w}_r=\frac{w_r}{\sum_{j=1}^{16} w_j}$          (3)

Ensuring that the contributions of all rules sum to unity.

Layer 4 — Consequent combination: Each rule has a first-order Takagi–Sugeno consequent $z_r=c_{0 r}+c_{1 r} f_1+c_{2 r} f_2+c_{3 r} f_3+c_{4 r} f_4$.  

The aggregated output is therefore:

$z=\sum_{r=1}^{16} \bar{w}_r \cdot z_r$           (4)

Layer 5 — Output activation: The final reliability score is obtained through a sigmoid activation:

$s=\sigma(z)=\frac{1}{1+e^{-z}} \in(0,1)$           (5)

Values close to 1 indicate high-reliability detections (likely TP), whereas values near 0 indicate low-reliability detections (likely FP).

Parameterization: Each input feature has two Gaussian membership functions, each defined by parameters μ and σ.

For four inputs this yields 4 × 2 × 2 = 16 premise parameters.

The consequent layer contains 16 rules × 5 coefficients = 80 parameters.

The model therefore contains 96 trainable parameters in total.

This compact parameterization allows reliable training on the 130–155 candidate samples per fold, a regime where larger neural re-ranking models would be difficult to train reliably.

Training procedure: All parameters are optimized jointly using the Adam optimizer with learning rate 0.01 for 160 training epochs, minimizing binary cross-entropy loss. Adam was preferred over SGD because its adaptive step sizes improve convergence on the relatively small candidate sets available in each fold.

To mitigate class imbalance, class balancing was applied during ANFIS training so that the scorer would not be dominated by the majority class. A separate ANFIS model is trained for each cross-validation fold, ensuring that the scorer is never evaluated on images used during its training.

Training curves confirmed stable convergence in all folds, typically reaching a plateau well before epoch 160.

3.5 Logic-Fuzzy Banding Algorithm

The LFBA is a post-processing decision mechanism operating in the joint space defined by the detector confidence c and the ANFIS reliability score s. The algorithm partitions the detector confidence axis into three decision zones using two thresholds conf_low and conf_high, while a third parameter score_thr controls the acceptance rule within the intermediate band.

This structure reflects an empirical decomposition of the candidate population into three regions with distinct statistical behavior.

Zone I — Auto-accept (c ≥ conf_high): Candidates with very high detector confidence are overwhelmingly true detections. In this region, the ANFIS score provides negligible additional discrimination; therefore, candidates are accepted directly.

Zone II — ANFIS-governed band (conf_low ≤ c < conf_high): The intermediate region contains a mixture of uncertain true detections and FP structures that achieve moderate detector confidence. In this region, the ANFIS reliability score provides the strongest discrimination and acts as a secondary decision gate.

Zone III — Auto-reject (c < conf_low): Low-confidence detections are dominated by background noise and spurious activations. Their ANFIS scores are approximately uniformly distributed and therefore provide little useful information. Such candidates are rejected directly.

Decision rule: For a candidate with detector confidence c and ANFIS reliability score s, the LFBA decision rule is:

accept $= \begin{cases}1, & \text { if } c \geq \operatorname{conf}_{\text {high }} \\ 1, & \text { if conf } \operatorname{cow}_{\text {low }} \leq c<\operatorname{conf}_{\text {high }} \text { and } s \geq \operatorname{score}_{\text {thr }} \\ 0, & \text { otherwise }\end{cases}$          (6)

This formulation separates high-confidence acceptance, reliability-based filtering, and low-confidence rejection.

Operating modes: The LFBA exposes two operating modes that allow operators to adjust the precision–recall trade-off depending on inspection priorities. The optimization objective is defined as:

$\operatorname{obj}(\theta)=F_1(\theta)-\lambda \cdot \mathrm{FP}_{\text {rate}}(\theta)$           (7)

where, θ = (conf_low, conf_high, score_thr).

Two values of λ are used:

HP mode — λ = 0.08: Places stronger penalty on false positives and is recommended for routine inspection scenarios where maintenance dispatch cost is high.

HR mode — λ = 0.02: Reduces the FP penalty and prioritizes defect detection, which is preferable after extreme weather events or in ageing infrastructure surveys.

Threshold optimization: The optimal parameter vector $\theta=\left(\operatorname{conf}_{\text {low}}, \operatorname{conf}_{\text {high}}, \operatorname{score}_{\text {thr}}\right) \in[0,1]^3$ is determined using a Lévy-flight Bat Algorithm [14, 49, 50]. The optimization landscape of the objective function is highly non-smooth because F1 and FP rate change discretely with threshold variations. Gradient-based optimization is therefore unsuitable.

The Lévy-flight mechanism introduces occasional long-distance steps following the power-law distribution $P(L) \propto L^{-\beta}, \beta=1.5$, which helps the algorithm escape local optima and explore the global search space.

Bat positions are updated iteratively using: $x_i^{t+1}=x_i^t+v_i^t$ with a local random walk applied with probability equal to the pulse rate r_i. When a candidate solution improves the objective function and satisfies the acceptance condition controlled by the loudness parameter A_i, the position is updated.

To maintain valid thresholds, the constraint conf_low< conf_high is enforced at every iteration.

The algorithm parameters are listed in Table 6. Optimization typically requires less than 8 seconds per fold on CPU, which is negligible compared with detector training time.

Table 6. Lévy-flight bat algorithm parameters for Logic-Fuzzy Banding Algorithm (LFBA) optimization

3.6 Evaluation protocol

All performance metrics are computed on the held-out validation subset of each cross-validation fold. A predicted bounding box is classified as a TP if its IoU with the nearest unmatched ground-truth box in the same image exceeds 0.50; otherwise, it is counted as an FP. Ground-truth boxes not matched by any prediction are treated as false negatives (FN). Box matching follows greedy assignment in descending order of prediction confidence, ensuring that each ground-truth box can be matched at most once.

For each fold, Precision, Recall, and F1-score are computed at the LFBA operating point for both HP and HR modes.

To provide a fair baseline comparison, a confidence-threshold detector baseline is also evaluated. For this baseline, the optimal confidence threshold conf* is selected by maximizing F1 over the grid (0.05,0.10,0.20,0.30,0.40,0.50,0.60,0.70) using the same validation fold. This procedure ensures that both the baseline detector and the LFBA system operate under thresholds optimized on identical validation data.

The ANFIS reliability scorer is evaluated independently from the LFBA decision layer using ROC-AUC and Average Precision (AP) computed from the continuous reliability scores. To quantify statistical uncertainty, 95% bootstrap confidence intervals for fold-level AUC and AP are estimated by resampling the fivefold-level values with replacement (10,000 bootstrap replicates, random seed = 42).

Two types of aggregated statistics are reported:

• Macro-average, defined as the arithmetic mean of per-fold metrics, is used to report mean performance and standard deviation.
• Micro-aggregate, obtained by summing TP, FP, and FN counts across folds, is used to compute global Precision, Recall, and F1 as well as cumulative FP counts.

The complete experimental pipeline—including YOLOv8n training, feature extraction, ANFIS training, and LFBA optimization.

5.1 Why are false positives reduced

Background structures that trigger YOLOv8n at moderate confidence (0.2–0.6) — tower cross-arms, metallic hardware, vegetation — differ systematically along the ANFIS feature axes: (i) lower average confidence (FP mean = 0.427 vs TP mean = 0.820); (ii) more peripheral image location (f₄ higher, since UAV operators center the insulator string during capture); (iii) more elongated bounding boxes (f₃ deviates from near-unity values of genuine disc caps); and (iv) smaller image area (f₂ lower for FPs than TPs at typical inspection distances). The ANFIS 16-rule Takagi-Sugeno structure exploits all four cues simultaneously, enabling discrimination that no single scalar threshold could achieve.

5.2 Mode selection guidelines

The HP and HR modes serve distinct operational scenarios: HP mode (λ = 0.08) is recommended when: (1) routine scheduled inspection applies, and maintenance dispatch cost is high; (2) infrastructure risk level is moderate, and long inspection intervals are acceptable; or (3) operator trust requires high confidence for every accepted alarm. HR mode (λ = 0.02) is recommended when: (1) post-extreme-weather emergency inspections require maximum defect sensitivity; (2) ageing or previously uninspected infrastructure poses elevated safety risk; or (3) the cost of a missed defect (flashover risk) clearly outweighs the cost of occasional false alarms. For intermediate risk profiles, λ can be tuned between 0.02 and 0.08, and LFBA thresholds can be re-optimized offline in under 8 seconds without detector retraining.

5.3 λ sensitivity analysis

When λ = 0.08 (HP mode), the Bat Algorithm assigns higher penalty to FP rate during optimization, causing conf_high to be set more conservatively (higher), conf_low to be raised, and score_thr to be raised. The net effect is a narrower auto-accept zone and a stricter intermediate screening gate: ΣFP drops from 13 (HR) to 9 (HP) at the cost of ~1.4% mean recall reduction. When λ = 0.02 (HR mode), the FP penalty is four times smaller, allowing the optimization to lower conf_high and score_thr, widening the acceptance band and recovering additional true detections. Figure 10 illustrates this behavior across all five folds. A full grid-based sensitivity analysis across $\lambda \in$ {0.01, 0.02, 0.04, 0.06, 0.08, 0.10, 0.15} was conducted on all five folds; results are summarized in Table 8. The analysis confirms that the F1 score peaks in the range $\lambda \in$ [0.02, 0.04] and degrades monotonically beyond λ = 0.10, where the FP penalty becomes excessive and recall loss outweighs further precision gains. The selected values λ = 0.02 (HR) and λ = 0.08 (HP) sit at the two operational knees of this trade-off curve, justifying their use as the recommended operating points.

5.4 Operational implications

In a large network deployment with 1,000 towers and an estimated 5% defect prevalence, HP mode reduces FPs from 30 to 9 across the 5-fold evaluation — corresponding to approximately two-thirds fewer false dispatch orders per inspection cycle. The LFBA decision policy can be recalibrated efficiently when the detector is updated or when operating conditions change, since LFBA optimization is computationally lightweight (≤ 8 s/fold on CPU). Faster R-CNN, by contrast, requires ~180ms per image, making it unsuitable for real-time airborne inspection at standard UAV capture rates of 2–5 Hz. YOLOv5n, while fast, achieves a 78% higher FP count than the proposed HP mode (41 vs 9), substantially increasing unnecessary maintenance dispatches in operational deployment.

The impact of the sensitivity parameter λ on the overall system performance is illustrated in Figure 10 and summarized numerically in Table 13. These results showcase the trade-off between the mean F1-score and the cumulative false positives ($\Sigma F P$) across seven candidate values of λ. As depicted, the selected values λ = 0.02 (HR mode) and λ = 0.08 (HP mode) are strategically positioned at the operational knees of the trade-off curve, providing the best balance for their respective objectives.

Table 13. λ sensitivity analysis — mean performance across 5 folds for seven values of λ

5.5 Limitations

Several limitations should be stated explicitly. First, the primary evaluation was conducted on a single dataset (BGI, 604 images). Although a cross-dataset generalisation experiment on CPLID (120 images) was added in Section 4.6 — confirming F1 = 0.926 and ↓48% FP reduction without retraining — full multi-dataset validation covering diverse geographic and equipment contexts remains desirable. Second, the BGI and CPLID datasets do not include systematically controlled adverse-weather variation (rain streak, fog attenuation, night-time low-light). The framework addresses viewpoint and illumination variation present in both datasets, but meteorological robustness under rain and fog has not been quantified; this remains a stated future-work priority. Third, LFBA thresholds are fold-specific and require recalibration (≤ 8 s on CPU) when deployed on a new detector or dataset. A full λ sensitivity grid (λ ∈ {0.01–0.15}, Table 8) and a nested CV protocol with outer held-out Fold 0 have been added to address prior evaluation concerns. Future work includes weather-augmented robustness testing, application of the ANFIS-LFBA layer to YOLOv11 and transformer-based backbones, and deployment trials on real UAV inspection platforms.