Automated control of robots in the work zone: a YOLO-Based approach for optimized maintenance

Abstract

Ensuring the safe and reliable operation of industrial robots requires effective monitoring systems capable of detecting abnormal situations within robotic workspaces. In the context of industry 4.0, computer vision techniques offer promising solutions for real-time supervision of industrial environments. This study proposes a vision-based monitoring framework designed to detect the position of an industrial robot relative to predefined operating and restricted zones. A custom dataset was created from a camera installed above the robot workspace, including images captured under different operational conditions. The models were trained to classify robot positions and detect their status in order to alert maintenance personnel when robots leave predefined zones and optimize interventions. The proposed system relies on deep learning object detection and instance segmentation models based on recent YOLO architectures. A comparative evaluation between YOLOv8 and YOLOv11 was first conducted to identify the most suitable architecture for the considered monitoring task. Experimental results show that both models achieve high detection accuracy, with mAP50 values exceeding 98%; while YOLOv8 demonstrates slightly better localization performance and training stability. Consequently, YOLOv8 was selected for further experiments. To evaluate the robustness and generalization capability of the proposed system, a cross-validation protocol was designed using multiple camera viewpoints and varying illumination conditions. The results indicate that multi-view training significantly improves detection robustness, while strict cross view experiments reveal the strong influence of viewpoint variation on model performance. In contrast, illumination changes have a comparatively smaller impact detection accuracy. Overall, the proposed monitoring framework highlights the potential of the models to significantly reduce unplanned downtime and boost the vigilance of maintenance operators, demonstrating their relevance in the context of Industry 4.0.

1 Introduction

Preventive maintenance remains an essential means of preventing costly breakdowns and unplanned stoppages. The integration of artificial intelligence (AI) into industry has revolutionized maintenance strategies, bringing a new dimension to them. In modern industrial environments, AI techniques are widely deployed for process optimization, predictive maintenance, quality improvement, anomaly detection and equipment monitoring (Soudani et al., 2024a; Soudani et al., 2024b). By leveraging machine learning, deep learning, and data analytics, AI based maintenance can extract actionable insights from heterogeneous data source, thereby improving decision making, operational reliability and overall production efficiency (Zonta et al., 2020; Carvalho et al., 2019).

In this context, the monitoring of industrial equipment, particularly robots and automated systems, is becoming increasingly accessible and effective, in order to prevent collisions and unforeseen malfunctions in defined areas. Computer vision approaches thus represent a major technological advance, guaranteeing safe, supervised operation in real time (Guo et al., 2021). Computer vision-based approaches have therefore emerged as a key technological enabler for intelligent monitoring, offering non-intrusive flexible and real-time supervision of complex industrial workspaces.

Industrial robots are in complex and dynamic environments and are exposed to various types of failures, including mechanical, electrical and software related faults. Maintenance strategies increasingly rely on failure prediction and early anomaly detection to anticipate malfunction and enable proactive interventions (Dui et al., 2023; Soudani et al., 2025; Dhanda et al., 2025). The integration of computer vision techniques offers a promising opportunity to strengthen early detection capabilities and improve the responsiveness of maintenance team.

Among deep learning approaches for visual monitoring, object detection models have become particularly effective for identifying and tracking objects within complex scenes. Convolutional Neural Networks (CNNs) have demonstrated remarkable performance in visual recognition tasks and have become the foundation of modern computer vision systems (LeC et al., 2015). A lot of recent real time object detection frameworks such as You Only Look Once (YOLO) family of models have gained significant attention due to their ability to simultaneously perform object localization and classification within a single Neural Network architecture (Redmon et al., 2016a). The high detection speed and accuracy of YOLO architectures make them particularly suitable for industrial monitoring applications where real time decision making is required.

In industrial robotics, monitoring whether a robot remains within its designated operational workspace is a critical safety requirement. Deviations from predefined work zones may indicate abnormal robot behavior or potential safety risks that require immediate interventions. Vision based monitoring can provide an efficient and flexible solution for detecting such anomalies by analyzing continuously the spatial position of the robot relative to its operation zone.

Our study is part of a maintenance approach aimed at reinforcing the vigilance of teams in charge of interventions and robot parameters setting. The detection of zone exits is a critical anomaly that can generate unexpected risks. In addition, AI-based solutions offer a flexible and cost-effective alternative for monitoring these anomalies. This article adapts YOLO models to robot monitoring in industrial environments, enabling real-time anomaly detection and prevention. The proposed approach is primarily a vision-based monitoring system rather than a full predictive maintenance framework. The contribution of our work lies in the real -time detection of spatial anomalies; such as robot deviations from predefined work zones which can serve as early indicators of abnormal behavior.

From maintenance perspective this monitoring capability supports preventive maintenance strategies by enabling early detection of unsafe zone or non-compliant robot states. Therefore, the proposed system should be considered as a decision support tool that enhances maintenance operations rather than directly performing failure prediction.

2 Related works

Among computer vision approaches, the “You Only Look Once” (YOLO) algorithm, introduced by Redmon in 2016 (Redmon et al., 2016b). Which has revolutionized real-time object detection thanks to its unified architecture, combining bounding boxes and classification in a single pass (Lirong et al., 2025). Unlike traditional two stage detection framework, YOLO formulates object detection as a regression problem, enabling fast inference and making it particularly suitable for real time industrial monitoring applications.

Initially designed for generic tasks (YOLOv1), this algorithm has evolved to meet increased requirements in terms of accuracy and speed. Successive versions (from YOLOv2 to YOLOv11) have incorporated various improvements, including dimensional anchors (YOLOv2), finer spatial segmentations (YOLOv3), and optimization techniques such as the “Bag of Freebies” (YOLOv4) (Zou, 2019). Recent versions including YOLOv5, YOLOv8 and YOLOv11 further enhance the architecture by introducing modular designs, improved feature fusion mechanisms and more efficient backbone networks, allowing these models to achieve high detection accuracy while maintaining real-time inference capability.

Due to these advantages, YOLO based approaches have been widely applied in various industrial monitoring and inspection tasks. In the field of infrastructure inspection, YOLO models have been successfully used for road maintenance application, enabling accurate mapping of internal road defects (Guo et al., 2021) and YOLO has also optimized the detection of potholes using drones equipped with YOLOv7 (Saranya et al., 2024). In industrial maintenance, a variant of YOLO, YOLODFT (Wang K. et al., 2024), has been used for monitoring electrical equipment. While (Zhanfang and Tuo, 2025) has developed an optimized YOLOv8n framework that improves the detection of surface defects on wind turbines by 2.3% compared to reference models.

Recent studies have further demonstrated the effectiveness of yolo architecture for intelligent industrial monitoring and proactive maintenance. For example, (Wang L. et al., 2024) proposed an improved YOLOv8 architecture for industrial surface defect detection, achieving significant improvements in inspection accuracy in manufacturing environments (Hütten et al., 2024). Investigated deep learning–based visual inspection systems for Industry 4.0 manufacturing, highlighting the role of computer vision models in automated quality control. In addition (Soumyadeep et al., 2023), emphasize that real-time object detection has become an important research area in computer vision.

More recent works have also explored improvement in YOLO architecture for defect detection and safety monitoring (Mehadjbia and Slaoui-Hasnaoui, 2025), has demonstrated that the improved YOLOv11 architecture, which incorporates components such as the C3k2 block, improves the efficiency and accuracy of defect detection compared to previous models. In terms of fire safety, YOLO helps improve anomaly detection and maintenance of firefighting equipment in Saudi Arabia (Alayed et al., 2024). In the same context (Zhao et al., 2024), used YOLO for autonomous vehicle security by combining the strengths of YOLO architectures to ensure better detection of road signs and increased reliability. In the agricultural sector, it is a valuable tool that enables farmers to detect plant diseases early and take preventive action (Rajamohanan and Latha, 2023). Also (Song et al., 2025), presents the RDW-YOLO algorithm based on YOLOv11, which improves methods for detecting pests in agriculture RDW-YOLO. Finally, in environmental protection, YOLOv6 and YOLOv7 are being used in Indonesia to detect and eliminate plastic waste, particularly bottles floating on the surface of the water (Kirana et al., 2024).

Overall, these studies highlight the versatility and robustness of YOLO architectures in real time monitoring and maintenance related tasks, motivation this adaption to industrial robot monitoring and anomaly detection in complex environments.

3 Research methodology

Our study aims to exploit YOLO techniques to facilitate robot preventive maintenance in a collaborative environment, where robots perform automated tasks. To achieve this goal, our approach is based on five main steps as shown in Figure 1: hardware definition, dataset construction, model training and solution installation. Our study also compares the performance of the YOLOv8 and YOLOv11 models.

FIGURE 1

3.1 Dataset creation pipeline

3.1.1 Experimental context

To evaluate a visual monitoring system capable of verifying whether an industrial robot is operating within its authorized work area, an experimental setup was established around an ABB IRB 1100 robotic manipulator installed in the M2S2I laboratory at ENSET Mohammedia.

The visual capture system relies on a fixed RGB camera mounted on a rigid support, positioned to observe the robot’s work area from different perspectives. To improve the robot’s robustness with respect to geometric variations and acquisition conditions, data was collected using several camera configurations corresponding to different observation angles of the scene.

FIGURE 2

3.1.2 Multi-angle acquisition

In this configuration, controlled lighting variation has been introduced in order to study the effect of illumination conditions on object detection performances. The experimental conditions are: Reduced lighting and Enhanced lighting.

Details of these configurations will be given in the results Table 1.

3.1.3 Image extraction

This approach allows for the construction of a diverse dataset while limiting excessive temporal correlation between consecutive images.

To prevent data leakage due to temporal correlation between consecutive frames, we extracted frames at regular intervals to ensure sufficient variability between samples. We also filtered out visually similar consecutive frames to reduce redundancy.

Also, a strict separation between training, validation, and test sets was applied at the sequence level. This ensures that frames from the same video segment never appear in different subsets, preventing information leakage and allowing for a proper evaluation of the model’s generalization capabilities.

3.2 Annotation procedure

3.2.1 Purpose of the annotation

FIGURE 3

With this approach, the model doesn't just learn the robot’s state, it also understands the spatial relationship between the robot and its work area. Which is essential for reliable anomaly detection.

Table 2 shows a sample of the images collected.

TABLE 2

Classes	Example of images
Robot in work zone
Robot in restricted zone
Operating zone	The yellow and black rectangle you see in the image represents the safety zone or the operation zone demarcated for the robot

Example of dataset.

3.2.2 Annotation tool

The annotations were performed using the ROBOFLOW platform, which allows for the annotation of bounding boxes, dataset management, and the automatic generation of shapes compatible with YOLO models.

3.2.3 Classes and annotation types

The annotation type used is polygonal segmentation/bounding boxes.

3.3 Quality control

3.4 Training model

Both models were trained under the same conditions to ensure a fair comparison.

For our initial configuration C1, according to the 70/15/15; we used a dataset of 1423 images for data training, 306 for model validation and 305 for test, as shown in Table 4. The same spilt is used for all configurations.

3.5 Evaluation protocol

3.5.1 Standard evaluation (in domain)

3.5.2 Cross-condition evaluation

The conditions are compared to the reference configuration C1, to evaluate the relative degradation.

3.6 Reproducibility and availability

To improve the diversity of the dataset and enhance the robustness and generalizability of the deep learning models, the dataset was augmented using standard data augmentation techniques. These techniques included random rotations, horizontal and vertical translations, scaling, brightness and contrast adjustments, and slight perspective transformations. Data augmentation was applied exclusively to the training set to prevent data leakage and ensure an objective evaluation of the model.

3.7 Presentation of YOLO models used

Model training has been carried out using the yolo versions family, namely, YOLOv8 and YOLOv11, which are designed to provide improved detection accuracy and computational efficiency, making them particularly suitable for real time industrial monitoring applications. YOLO based models follow a unified object detection framework capable of simultaneously performing object localization and classification within a single neural network architecture (Redmon et al., 2016a).

Both YOLOv8 and YOLOv11 adopt a modular architecture composed of three main components: a backbone network, a neck module and a detection head, Figure 4. The backbone is responsible for extracting hierarchical visual features from the input image using convolutional layers and feature extraction blocks. In YOLOv8, the backbone architecture relies on an optimized Cross stage Partial (CSP) structure combined with advanced convolutional modules to improve feature representation while maintaining computational efficiency (Ultralytics, 2026).

FIGURE 4

The neck component aggregates multi sale features through a feature fusion strategy that typically combines a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN). This structure allows the model to integrate both low level spatial information and high-level semantic features, enabling robust object detection across different object scales and complex industrial scenes.

YOLOv11 represents a further evolution of the YOLO architecture and introduces several improvements in feature extraction and feature fusion mechanisms. In particular enhanced convolutional blocks such as the C3K2 module improve the efficiency of feature representation and enable more accurate detection of objects and structural details in complex environments. Furthermore, YOLOv11 integrates instance segmentation capabilities, allowing the model not only to detect objects using bounding boxes but also to generate pixel-level segmentation masks for each detected object. This enables more precise scene interpretation and facilitates advanced applications such as detailed object localization, spatial analysis and industrial monitoring tasks (Mehadjbia and Slaoui-Hasnaoui, 2025).

Finally, the detection head of YOLOv8 and YOLOv11 performs the final prediction stage by estimating class probabilities, bounding bx coordinates and confidence scores based on the extracted features/in segmentation mode, the model additionally predicts object masks, enabling fine grained representation of objects within the monitoring environment.

3.8 Decision rule for zone violation detection

In our approach, zone violation detection relies not only on the predicted class label but also on spatial reasoning derived from the segmentation results. Thus, our decision mechanism is based on the geometric relationship between the robot and the operating zone. To this end, in our study, a violation is detected when the spatial overlap between the robot’s segmentation mask and the operational area mask is less than a predefined threshold τ. Thus, we empirically set the value of τ to 0.7 which offers a balanced compromise between sensitivity to area violations and robustness to segmentation noise.

In practice, the system evaluates whether the end effector, its axis, and a significant portion of the robot’s structure lie outside the predefined zone polygon. If this condition is met, an alert indicating a potential anomaly requiring intervention is generated.

4 Results

The objective of our research is to present a new monitoring framework based on real-time monitoring for proactive maintenance of robots in industrial environments. Thus, based on YOLO architectures, the system studied enables real-time detection of the position of industrial robots and verifies their spatial compliance with the recommended work areas. This approach offers visual geolocation, enabling immediate alerts to be triggered when a robot is located outside its designated workspace. Figure 7 explains the approach used in more detail.

4.1 Standard evaluation in domain

4.1.1 Yolov8 results

Initially, the model training is based on YOLOv8. The results of this training are shown in the Table 6 and Figure 5 which provide the model values for four parameters: precision, recall, mAP50, and mAP50-95 for bounding boxes and segmentation masks.

FIGURE 5

The results show that the YOLOv8 model used achieves excellent results in bounding box detection and good performance in mask segmentation.

In fact, for bounding boxes, the model’s accuracy is 96.85%, which shows that the vast majority of detections are correct and with very few false positives. For recall, it reached 95.87%, which means that the model manages to detect almost all objects present. In addition, the mAP50 metric is high, recording a value of 98.91%, which means a good ability to locate objects at an IoU threshold of 0.5. Even with stricter criteria (mAP50-95 at 87.10%), performance remains very solid, confirming accurate and reliable localization.

On the other hand, with regard to segmentation masks, the model’s accuracy is 89.29% and its recall is 87.40%, which shows that the model correctly segments most objects while covering a large part of their actual surface area. In addition, the mAP50 is 89.52%, which is lower than the values obtained for bounding boxes. However, the mAP50-95 is lower (71.43%), revealing a drop in performance when segmentation must be very precise on the contours or fine details of objects.

4.1.2 Yolov11 results

The second phase of our research consists of using the YOLOv11 model in segmentation mode and adjusting the hyperparameters.

To this end, we have chosen two hyperparameters that are likely to influence the performance and results of the model. These two hyperparameters are: the number of training epochs and the learning rate.

Thus, the number of epochs is the number of times the entire training dataset is completely traversed by the model. We selected 100, 200, and 300 epochs.

On the other hand, the learning rate determines the extent of the adjustments made to the weights of the neural network at each iteration of the optimization. In this case, we chose the following values: (lr = 0.01, 0.001, and 0.0001).

Thus, Tables 7, 8 show samples of images illustrating the detection of robots in their work area and in the restricted area obtained by the two models at different training times.

TABLE 7

Detection model	100 epoch	200 epoch	300 epoch
YOLOv8
YOLOv11

Sample of robot results in the work area.

TABLE 8

Detection model	100 epoch	200 epoch	300 epoch
YOLOv8
YOLOv11

Sample of robot results in the restricted area.

Table 9 and Figure 6 summarize the values of the mAP50, mAP50-95, precision, recall, F1-score, and FPS metrics for the two models according to the different epochs and learning rates.

FIGURE 6

An overall reading of the results shows that both the YOLOv8 and YOLOv11 models perform very well in all configurations, with mAP50 values above 98% and F1 scores between 96% and 97%. This performance indicates excellent accuracy and very good recall. It is also worth noting that the differences in performance are small, indicating that the two architectures are competitive.

Furthermore, the results of the configuration comparison are as follows:

First, after training for 100 epochs with lr = 0.01, the results show that YOLOv8 achieves a slightly higher mAP50 (99.13% vs. 98.81% for YOLOv11) and a higher mAP50-95. However, YOLOv11 compensates with a stronger recall (97.50% vs. 96.19%), indicating that it detects slightly more objects with a few more false positives. The FPS remains the same (∼92), showing that speed is not impacted by the model.

Next, at 200 epochs and Lr = 0.001, we see that YOLOv11 takes the lead on mAP50 (99.17%) and maintains a good balance between precision and recall. Meanwhile, YOLOv8 remains very close with an mAP50 of 99.05%, but has slightly higher accuracy. For FPS and F1-score, the two models are almost identical.

Finally, at 300 epochs and lr = 0.0001, we observe that performance is stable, which shows that increasing the number of epochs at a low learning rate does not bring any notable gain. The two models are almost equivalent. Also, FPS remains the same (∼88), confirming that speed depends more on hardware configuration than on architecture or the number of epochs.

The analysis of results and taking into account the objective is to monitor whether a robot is within its working area. The main risk is triggering a false alarm (saying that the robot is out of range when it is still within range). In this type of mission, we are looking for fewer false alarms, which means higher accuracy. It is preferable to be sure before triggering an alarm, even if this sometimes means missing a borderline case. According to the results, YOLOv8 always has a slightly higher or equivalent accuracy to YOLOv11, especially at 100 epochs (+2.33% difference).

When it comes to a security system, a false positive can cause unnecessary shutdowns, waste time, and even disrupt production. YOLOv8 is therefore more reliable for confirming an out-of-zone event, as it makes fewer false alarms.

In summary, the results show that YOLOv8 and YOLOv11 display consistent performance for all cases, where there are slight differences with regards to the values of metrics used for evaluation. From a broader view, In the case of YOLOv8, the model shows better localization while YOLOv11 performs similarly well in terms of recall. Generally speaking, Considering the negligible difference in performance and lack of statistical test between YOLOv8 and YOLOv11, the choice of using YOLOv8 for the upcoming experiment can be attributed to its slight consistency advantage.

Based on the comparative evaluation of YOLOv8 and YOLOv11, YOLOv8 trained for 200 epochs was selected as the final model for the cross-validation experiments.

4.2 Cross -view and cross-condition evaluation

In this section, the objective is to rigorously evaluate the performance and generalization capability of the proposed monitoring system. So, we conducted a series of cross-view experiments based on the experimental protocol defined in the previous section. While the dataset construction and training pipeline were previously described, the objective of the following experiments is to assess the robustness of the trained model under different operational conditions.

To address this challenge, a set of controlled experiments was designed using multiple combinations of training and testing datasets collected from different configurations viewpoints (C1, C2, C3, and C4). These experiments allow us to evaluate several aspects of the model performance, including baseline in-domain accuracy, robustness to multi-view conditions, generalization to partially or fully unseen viewpoints and sensitivity to illumination changes. Tables 10–12 give a sample of these datasets.

TABLE 10

C1	C2

Example of dataset for C1 and C2.

TABLE 11

C3	C4

Example of dataset for C3 and C4.

TABLE 12

C5: Enhanced lighting	C5: Reduced lighting

Example of dataset for C5 enhanced lighting. And reduced lighting.

The different experiments for cross-view and cross-condition evaluation are summarized in Table 13.

We organized the experiments according to increasing levels of difficulty as follows: first a basic evaluation in the domain (E1), followed by multi-view training scenarios (E2–E4), and finally inter-view generalization tests and evaluation of lighting robustness (E5–E6a-E6b).

The datasets for each experiment are given in the Tables 14.

4.2.1 Overall analysis of results obtained

The results in Table 15, the experiments E1-E6; allow us to rigorously evaluate the performance of the YOLO V8 model in several representative industrial deployment scenarios, including intra-domain validation, Multiview learning, generalization to partially observed views, generalization to a completely unobserved view, and robustness to photometric variations:

Using formula 1, we calculated the performance loss experimentally.

In general, the result s in Tables 15–17 highlight three major findings:

TABLE 17

E2
E3
E4
E6a/E6b

Sample of cross-view and cross-condition evaluation results.

First, the model exhibits excellent performance under conditions close to the training domain; Second, the diversity of viewpoints during training significantly improves robustness; Third, generalization to a completely unobserved view remains the main limiting factor, even more so than variations in lighting.

4.2.2 Analysis by experiment

4.2.2.1 The intradomain reference performance E1

These results show that the model correctly learns the geometric structure of the scene, the robot’s position, and the boundary of the operating area. The model does not suffer from an excess of false positives or a major detection deficit effect in this context.

4.2.2.2 Observed multiview learning E2

These results show that introducing a second view does not degrade learning, but on the contrary, improves the model’s ability to produce more robust localizations across several IoU thresholds. This experiment therefore validates a key hypothesis: that Multiview learning improves robustness without compromising nominal performance.

4.2.2.3 Generalization to partially observed views

Experiments E3 and E4 evaluate the model in an intermediate situation. Views C3 and C′ are not completely unknown; a small proportion of images from these views are included in the training.

–E3, test on C3

This indicates that view C3 introduces a more significant visual break compared to C1 and C2, encompassing several factors: perspective variation, modification of the robot’s relative appearance, and a change in proportion between the robot and the area.

–E4 test on C4

Thus, the contrast between E3 and E4 is scientifically very interesting. While both experiments are based on the same methodological principle, C4 is clearly closer to the learning domain than C3. This means that not all variations in perspective have the same impact on performance.

4.2.2.4 Strict generalization to a completely unobserved view E5

This decrease demonstrates that the model, when trained on a single perspective, develops a representation that is too specific from a training perspective. In practice, this means that the excellent performance of E1 does not guarantee robust deployment capabilities in a variable environment.

Thus, E5 highlights the structural limitations of single-view learning and reinforces the credibility of the experimental protocol we established.

4.2.2.5 Robustness to photometric variations

These results show that the model is significantly more robust to photometric variation than to a strong geometric viewpoint, as the E6 results are considerably superior to the E3 and especially the E5 results.

Thus, this result is important, meaning that the system could maintain acceptable performance in an environment with reduced or strong lighting, provided that the geometric configuration remains close to that known during training.

4.2.3 Conclusion

5 Discussion

5.1 Results summary

The experimental analysis first compared two recent object detection architectures namelyYOLOv8 and YOLOv11, to identify the most suitable model for the proposed robotic monitoring system. Both models achieved very high detection performance. With mAP50 values exceeding 98%. However, using the more demanding mAP50-95 metric, YOLOv8 exhibits slightly higher localization accuracy (88.34%) than YOLOv11 (87.29%). Furthermore, YOLOv8 demonstrated greater stability across training configurations while maintaining a similar inference speed. Therefore, YOLOv8, trained for 200 epochs, was selected for subsequent experiments.

Cross-view and cross-condition evaluation experiments were conducted to evaluate the robustness and generalizability of the proposed monitoring framework under different viewpoints and lighting conditions. The baseline experiment E1 confirms excellent performance in this area, with a mAP50 of 0.991 and a mAP50-95 of 0.86. When multiple viewpoints were included in the training E2, the model maintained stable performance and slightly improved accuracy indicating that Multiview learning enhanced robustness.

Experiments involving partially observed viewpoints E3 and E4 showed that model performance depends on the geometric similarity between viewpoints, with better results obtained on viewpoint C4 than on C3.

The strict viewpoint comparison experiments where the model was trained on C1 and tested on the unknown viewpoint C2, revealed a significant drop in performance, highlighting the strong influence of viewpoint bias in vision surveillance systems.

Finally, the lighting experiment E6 demonstrated that the model remains relatively robust in low-light conditions. Although segmentation accuracy decreases slightly, overall detection performance remains high. These results indicate that variations in viewpoint have a greater impact on system performance than variations in lighting.

Overall, the proposed monitoring system achieves high detection accuracy while maintaining reasonable robustness in the face of diverse viewpoints and environmental conditions.

5.2 Scientific and research contributions

FIGURE 7

The experimental results highlight the strong influence of viewpoint variations on detection performance. In particular, the strict cross view experiment reveals a significant performance drop when the model is trained on a single viewpoint, emphasizing the importance of multi-view training for reliable deployment in industrial environments.

Overall, these contributions provide practical and methodological insights for the development and evaluation of vision-based monitoring systems for industrial robotic applications.

5.3 Limitations

Despite the promising results obtained in this study, several limitations should be acknowledged.

First, the dataset used in this work is limited to a specific robotic workspace and a relatively controlled industrial environment.

Although multiple viewpoints were considered, the diversity of environmental conditions remains limited compared with real-world industrial deployments.

Second, the system relies primarily on visual information captured by fixed cameras. In highly dynamic environments, occlusions or complex background structures may affect detection performance.

Third, although segmentation capabilities were evaluated, the current framework focuses primarily on object detection rather than precise robot pose estimation or motion prediction.

6 Conclusion

This study focuses on the application of AI to optimize robot detection and maintenance interventions in industrial environments. Based on the YOLO v8 and YOLO v11 models, we have developed a real-time detection system that monitors the operational states of robots in predefined zones. The results show that both models deliver excellent performance, with YOLO v8 and YOLO v11 achieving high values of mean accuracy and impressive measures of recall and precision over different epochs. YOLO v8, in particular, shows consistent improvement with increasing training epochs, while YOLO v11 also delivers robust performance despite a slight dip at intermediate epochs. The Cross-validation experiments have also revealed that Multiview training significantly improves detection robustness, while rigorous cross-assessments highlight the significant influence of viewpoint variations on model performance. In contrast, lighting variations have a comparatively smaller impact on detection accuracy.

This research highlights the potential of AI-driven solutions to improve maintenance efficiency; by enabling automated detection of zone violations and abnormal robot positions, the proposed framework helps to reduce safety risks and improve operational efficiency in industrial environments.

Nevertheless, several challenges remain for large-scale deployment, including the diversity of datasets, robustness in the face of complex environmental conditions, and computational constraints for real-time processing. Future research will focus on integrating multisensory data, exploring domain adaptation techniques to improve cross-environment generalization, and extending the framework to predictive maintenance by integrating temporal analysis of robot behavior.

References

• 1

  AlayedA.AlidrisiR.FerasE.AboukozzanaS.AlomayriA. (2024). Real-time inspection of fire safety equipment using computer vision and deep learning. Eng. Technol. Appl. Sci. Res.14 (2), 13290–13298. 10.48084/etasr.6753

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  CarvalhoT. P.SoaresF. A. A. M. N.VitaR.FranciscoR. da P.BastoJ. P.AlcaláS. G. S. (2019). A systematic literature review of machine learning methods applied to predictive maintenance. Comput. Ind. Eng.137, 106024. 10.1016/j.cie.2019.106024

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  DhandaM.RogersB. A.HallS.DekoninckE.DhokiaV. (2025). Reviewing human-robot collaboration in manufacturing: opportunities and challenges in the context of industry 5.0. Robot. Comput.-Integr. Manuf.93, 102937. 10.1016/j.rcim.2024.102937

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  DuiH.XuH.ZhangL.WangJ. (2023). Cost-based preventive maintenance of industrial robot system. Reliab. Eng. Syst. Saf.240, 109595. 10.1016/j.ress.2023.109595

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  GuoB. H. W.ZouY.FangY.GohY. M.ZouP. X. W. (2021). Computer vision technologies for safety science and management in construction: a critical review and future research directions. Saf. Sci.135, 105130. 10.1016/j.ssci.2020.105130

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  HüttenN.Alves GomesM.HölkenF.AndricevicK.MeyesR.MeisenT. (2024). Deep learning for automated visual inspection in manufacturing and Maintenance: a survey of Open- access papers. Appl. Syst. Innov.7 (1), 11. 10.3390/asi7010011

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  KiranaN. L.KurnianingtyasD.Indriati (2024). A deep learning approach to plastic bottle waste detection on the water surface using YOLOv6 and YOLOv7. Eng. Technol. Appl. Sci. Res.14 (6), 18623–18630. 10.48084/etasr.8592

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  LeCunY.BengioY.HintonG. (2015). Deep learning. “Deep Learning,” Nat.521 (7553), 436–444. 10.1038/nature14539

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  LirongL.HaoC.JunweiD.BingM.ZhijieZ.ZhangqinH.et al (2025). High-voltage transmission line inspection based on multi-scale lightweight convolution. Measurement.251. 117193. 10.2139/ssrn.4861628

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  MehadjbiaA.Slaoui-HasnaouiF. (2025). “Structural damage detection of wind turbine based on YOLO11,” in 5th International Conference on Innovative Research in Applied Science, Engineering and Technology (IRASET). Fez, Morocco: IEEE,2025. 1–6. 10.1109/IRASET64571.2025.11008148

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  RajamohananR.LathaB. C. (2023). An optimized YOLO v5 model for tomato leaf disease classification with field dataset. Eng. Technol. Appl. Sci. Res.13 (6), 12033–12038. 10.48084/etasr.6377

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  RedmonJ.DivvalaS.GirshickR.FarhadiA. (2016a). You only look once: unified, real-time object detection. arXiv arXiv:1506.02640. 10.48550/arXiv.1506.02640

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  RedmonJ.DivvalaS.GirshickR.FarhadiA. (2016b). “You only look once: unified, real–time object detection,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 779–788. 10.1109/CVPR.2016.91

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  SaranyaE.NivethaR.AbiramiS.Mohaideen ArsathM.DharaneeshS. (2024). “Revolutionizing road maintenance: YOLO based pothole detection system,” in 2024 10th International Conference on Advanced Computing and Communication Systems (ICACCS), 1991–1997. 10.1109/ICACCS60874.2024.10716937

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  SongJ.ChengK.ChenF.HuaX. (2025). RDW-YOLO: a deep learning framework for scalable agricultural Pest monitoring and control. Insects16 (5), 545. 10.3390/insects16050545

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  SoudaniM. L.NissabouriS.Ech-ChhibatM. E. H. (2024a). “Maintenance and production optimization using artificial intelligence (AI) tools: a bibliometric analysis and review,” in 2024 4th International Conference on Innovative Research in Applied Science, Engineering and Technology (IRASET), 1–8. 10.1109/IRASET60544.2024.10549204

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  SoudaniM. L.NissabouriS.Ech-ChhibatM. E. H. (2024b). Artificial intelligence for maintenance and production optimization: applications, techniques and challenges. Int. Rev. Mech. Eng. IREME18 (11), 577. 10.15866/ireme.v18i11.25151

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  SoudaniM. L.NissabouriS.Ech-ChhibatM. E. H.HaidouryM.ElhirchI. (2025). “Industrial robots and collaborative robots: a comparative study,” in 2025 5th International Conference on Innovative Research in Applied Science, Engineering and Technology (IRASET), 1–10. 10.1109/IRASET64571.2025.11008132

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  SoumyadeepM.SinghA. P.SharmaA.KumarP. (2023). “Real-time object detection and tracking using deep learning,” in 2023 11th International Conference on Intelligent Systems and Embedded Design (ISED), 1–7. 10.1109/ISED59382.2023.10444534

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  Ultralytics (2026). Explore ultralytics YOLOv8. Available online at: https://docs.ultralytics.com/models/yolov8/(Accessed March 08, 2026).

  • Google Scholar
  • View reference in article

• 21

  WangK.ZhangX.SunY.XuT.LiJ.CaoS. (2024). YOLO-DFT: an object detection method based on cloud data fusion and transfer learning for power system equipment maintenance. IET Collab. Intell. Manuf.6 (2), e12104. 10.1049/cim2.12104

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  WangL.ZhangG.WangW.ChenJ.JiangX.YuanH.et al (2024). A defect detection method for industrial aluminum sheet surface based on improved YOLOv8 algorithm. Front. Phys.12 (Jun), 1419998. 10.3389/fphy.2024.1419998

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  ZhanfangZ.TuoL. (2025). Enhancing wind turbine blade damage detection with YOLO-wind. Sci. Rep.15 (1), 18667. 10.1038/s41598-025-03639-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  ZhaoR.TangS. H.ShenJ.SupeniE. E. B.RahimS. A. (2024). Corrigendum to ‘enhancing autonomous driving safety: a robust traffic sign detection and recognition model TSD-YOLO’. Signal Process.225, 109619. 10.1016/j.sigpro.2024.109619

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  ZontaT.da CostaC. A.da Rosa RighiR.de LimaM. J.da TrindadeE. S.LiG. P. (2020). Predictive maintenance in the industry 4.0: a systematic literature review. Comput. Ind. Eng.150, 106889. 10.1016/j.cie.2020.106889

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  ZouX. (2019). “A review of object detection techniques,” in 2019 International Conference on Smart Grid and Electrical Automation (ICSGEA), 251–254. 10.1109/ICSGEA.2019.00065

  • CrossRef
  • Google Scholar
  • View reference in article