Introduction

With the rapid technological development in the agricultural industry, where advanced machinery plays a key role in increasing efficiency, using artificial intelligence in monitoring [1] and diagnosing their technical condition [2] is becoming increasingly important. AI offers innovative solutions, enabling continuous analysis of data from sensors and vision systems to identify subtle anomalies and predict potential failures before they occur. This proactive approach, supported by machine learning algorithms, allows optimization of maintenance schedules, minimizing field downtime and reducing repair costs. Thus, AI-based diagnostic systems [19] are crucial for advancing automation and improving agricultural machinery reliability.

To enhance the process of agricultural machinery damage detection, the implementation of computer tools based on image detection and classification [3] is being considered. The utilization of deep learning algorithms, with a particular focus on convolutional neural networks (CNNs) [20], has emerged as a promising methodology. The following paper [4] provides a detailed description of the application of a CNN network tasked with the recognition of household items and their surroundings, including bicycles, chairs, and deer. The study, which was based on a database of 6,000 images, found that the CNN algorithm achieved a detection accuracy of (74%) for the class "bicycle," (89%) for "chair," and (87%) for "deer" after a training process. The authors of the study did not implement additional modifications to the input data. However, a common challenge in the process of acquiring samples for training algorithms is the scarcity of available data. Examples of data augmentation techniques employed to enhance the performance of models that recognize objects using images are presented in the paper [5].

In their experiment, the researchers employed the RVL-CDIP database to train the CNN model, implementing image operations such as resolution change, Gaussian blur, mirroring, and image rotation. The findings indicated a (1%) enhancement in the efficacy of image recognition algorithms. A frequently employed contemporary model that is founded upon CNN algorithms [20] for object detection is the YOLO model [14], along with methodologies that utilize SSD detection blocks [6]. The application of the aforementioned model can be seen in the work of two scientific teams. The first team, described in the work [7], undertook the task of detecting corn seedlings in the natural environment. The use of detection using the YOLO architecture was aimed at reducing the pollution factor and promoting sustainable agricultural development. To train the algorithm, the team initially collected 1,000 images of seedlings. Then, through the use of digital augmentation involving a fourfold 90-degree rotation, the database size was increased to 2,400 samples. The implementation of the YOLO algorithm achieved a seedling detection efficiency of (86.66%).

The YOLO model is used in the analysis of images acquired from cameras on unmanned aerial vehicles to monitor forest areas to identify pine wilt disease [8]. The model's detection efficiency of disease outbreaks was achieved by the research team at (88%). The model allows both analysis of static images, and video and real-time detection of camera images. Another application of the YOLO model in image analysis is presented in paper [9]. The authors conducted an experiment to train an algorithm to detect objects from the environment of working agricultural machinery using the YOLOv8 architecture. The training process employed the COCO database and drew parallels between examples from seven different databases. The dataset comprised examples of objects such as people, cars, agricultural machinery, and other environmental elements that can be found in the environment of a working agricultural machine. The author trained two variants of the model: nano and small, with both models undergoing a training process comprising 80 epochs. The findings demonstrated a high level of detection efficiency for the "combine" (97%) and tractor (94%) classes. The efficiency of the algorithms was found to be lower in detecting objects such as humans (71-75%), cars (83%), tractors (82%) and agricultural machinery (84%).

In the extant literature concerning the YOLO family of models, comparative analyses of different architectural variants in terms of their effectiveness are commonplace. Of particular interest to researchers are the YOLOv8 and YOLOv11 versions, due to their similar structure and mechanism of operation. An example of such a comparison is furnished by the studies presented in papers [10, 11]. The paper [10] presents a study focusing on the detection of seedlings at different stages of development and the evaluation of their resistance to herbicides. In the course of the experiment, the researchers made use of a collection of 2,348 colour photographs. Furthermore, a range of variants of the YOLO model were employed in the analysis, including YOLOv11 and YOLOv8. The findings of the experimental trials demonstrated comparable object recognition accuracies, attaining (91%) and (92%), respectively. Furthermore, the objective of the study was to ascertain the time taken to detect the presence of the virus. Within the conditions of the natural plant growth environment, YOLOv11 demonstrated a superior object detection speed, outperforming the YOLOv8 model by 10.5 milliseconds. The findings suggest that, while the accuracy rates of the two versions of the model are comparable, the newer version exhibits substantially reduced detection times.

In this article, a comparative analysis of two neural network (NN) architectures, YOLOv11x and YOLOv8x, is conducted. The evaluation of the algorithms was based on the task of detecting specific defects in an agricultural rotary tedder machine. The training of the models was facilitated by a visual dataset that had been acquired under conditions of actual machine operation. The article provides a comprehensive account of the process of preparing the database, which involves the division of the database into three separate sets. In two group, digital data augmentation methods were used to increase the number of training samples. The process of acquiring knowledge of both models using the prepared sets is presented, and an analysis of the results is obtained. A detailed comparison of the performance of the two YOLO models was made based on the confusion matrix.

The main novelty of this study aims to evaluate the feasibility of using AI-based models to detect potential damage in our specific use case. Therefore, the primary contribution of this work is a practical, empirical comparison of two state-of-the-art, off-the-shelf YOLO architectures (YOLOv8x and YOLOv11x) for a specific, data-constrained agricultural fault detection task. We provide a detailed analysis of their performance under data scarcity, the impact of progressive dataset augmentation, and practical insights for implementing such systems in real-world conditions. In particular, machine component recognition using vision systems has the potential to allow for the identification of damage types in real time and during machine maintenance. It is possible that this could reduce operating costs and ultimately mitigate potentially harmful environmental impacts.

Methodology

The task was accomplished by utilizing the YOLOv11x and YOLOv8x models to detect the elements of the teeth of an agricultural machine. Models were trained using the PyTorch framework [16] and the Ultralytics YOLO library [14], utilizing physical hardware equipped with CUDA technology [17] version 12.6. The process incorporated the NVIDIA GeForce RTX 3060 Laptop GPU, the Windows 10 Pro operating system, and the Visual Studio development environment.

A dedicated database has been developed for the training of neural networks, with the networks being extracted from video recordings of mowing operations. The parameters of the footage are shown in Tab. 1. The footage was recorded from the vantage point of a standing individual observing the operation of the machine in a working environment. The extraction of frames from the video into individual images in graphic format was made with DaVinci Resolve 17 software. In addition, 291 photos of the SaMASZ P10-1100 rotary tedder machine were taken to mitigate the dataset imbalance. The focus was mainly on potential damage to the teeth. This allowed for the expansion of the input data for model learning.

The acquired materials were added and pre-processed before they were given to the input of the YOLO network model. Based on the acquired materials, three datasets used to train the neural network will be prepared. Assumptions for the use of the effects for the databases are shown in Tab. 2. Effects imposed on the materials are as follows:

- Rotate +45* and -45* to achieve up to a full 360-degree rotation;

- changing the colour palette;

- changing the resolution;

- pixelation;

- black-white contrast;

- changing threshold values.

Modifications applied to the database

The training of neural networks will be performed based on two models: YOLOv11x and YOLOv8x. Each model must undergo a learning process using Datasets. The machine on which the learning process will be carried out is shown in Tab. 3.

The network's learning assumptions will be established for all cases as follows, as demonstrated in Tab. 4. In the YOLOv8x architecture (see Fig. 1), the image detection process involves the input of a normalised and rescaled input image of predefined illustrative dimensions 640 x 640 pixels into the backbone module. This fundamental component of the network, predicated on C2f blocks, sequentially extracts hierarchical feature maps of varying resolutions, ranging from higher resolution maps characterised by their richness in spatial detail to lower resolution maps.

As illustrated in Fig. 1, the architecture of YOLOv11x [12] demonstrates evolutionary advancements in comparison to YOLOv8x, while concurrently preserving substantial similarities. As in YOLOv8x, the scaling of YOLOv11x variants is implemented using depth_multiple, width_multiple, and max_channels parameters, while the initial stem component remains identical. In the cervical region, both architectures incorporate the SPPF block, in addition to merging and zooming operations. A substantial modification in YOLOv11x is the integration of an advanced C3k2 block within the backbone modules, which replaces the C2f from YOLOv8x [13]. This results in a reduced number of parameters and the potential for enhanced efficiency.

Neck of model YOLOv11x, following the SPPF block, the integration of a C2PSA self-attention module was implemented to augment the model's capacity to discern global relationships within the analysed image. The down-sampling process in YOLOv11x is implemented through convolution layers with a 3x3 kernel and a step of 2. The structure of head YOLOv11x shows some similarities with the solutions from YOLOv8x in terms of an anchor-free approach. These modifications indicate the focus of YOLOv11x and development on increasing efficiency and improving contextual modelling while using proven solutions from previous generations. The YOLO versions were selected based on their comparable features. The YOLOv11x model represents an endeavour to enhance the architecture of its previous version. The diagrams are intended to provide a visual representation of the process by which the YOLO algorithm learns to recognise objects. The algorithm uses samples based on data collected from the actual object, and then compresses and analyses it.

Dataset and preprocessing

The implementation of the algorithms in this study involved the creation of a dedicated database for model learning. The source data was obtained from video footage recorded in the operating environment of an agricultural machine, considering the damage sustained by a broken tooth and a missing tooth, see Fig. 2. The video footage was decomposed into individual frames, which were then subjected to analysis. A selected portion of these frames was then used in a further process. The implementation of this procedure resulted in the acquisition of a total of 2,519 frames. From this pool, the frames that met the specified criteria were selected for further processing. The selection process focused on three key elements of the machine: healthy tooth, missing tooth, and broken tooth. It is evident that the visual patterns exhibited by the majority of the frames are repetitive, and the machine's operation is characterized by a cyclical pattern. Consequently, the data set under consideration consists of a representative sample of 34 images, with a resolution of 1980x1080. The acquired image database was then subjected to further processing to increase the number of samples required for the training process. The database was divided into three data sets. The initial training set for the algorithm was rotated by 45 degrees using Roboflow tools [15]. The rotation operation was performed on 233 samples.

The subsequent stage of the process involved the annotation of the elements to be incorporated into each image. For this purpose, the tool offered by the Roboflow platform was utilised. The labelling process was executed manually, based on assumptions concerning the identified elements. The following labels were entered: healthy tooth, broken tooth, and missing tooth. Following the conclusion of the annotation process, the final step in the augmentation of the training dataset was to rotate each image by 45 degrees. The resultant dataset comprised images for training, testing, and validation. Consequently, four distinct variants were formulated through the implementation of image modification.

Due to the limited number of samples extracted from the videos, additional operations were performed on the selected images to increase the number of samples and introduce more variance into the learning process. Using Dithermark image processing software, which allows manipulation of threshold, saturation, image base colour change, and additional parameters such as the texture of the displayed image. The total number of samples was increased. The image variants are given in Fig. 3.

Expanding the initial dataset mitigated the constraints of a limited sample size, specifically enriching the repository of rotary tedder tooth defects. This enhancement facilitated the development of the third database iteration Tab. 5, which prioritizes the representation of fractured components to improve the model.

Results And Discussion

The subject of this chapter is the analysis of the results of learning two separate instances of the YOLO algorithm, aimed at recognizing three defined states of the working teeth of the rotary tedder:

healthy tooth;

broken tooth;

missing tooth.

Proceeding to a detailed analysis of the learning results, it is necessary to characterize the differences between the datasets used. A quantitative summary of each class of objects for Dataset V1, Dataset V2 and Dataset V3 is shown in Fig. 5. Due to the use of methods involving the use of filters, the set of iterations for the learning process has significantly increased. The time in which both models completed their training shown in Tab. 6 suggests that the increased number of iterations extended the process to finish.

Conclusions And Future Work

In the above article, a comparative analysis of two neural network architectures, YOLOv8x and YOLOv11x, in the task of detecting damage to the teeth of an agricultural machine, using three separate datasets was carried out, providing important conclusions about their relative performance and the impact of training data characteristics. On Dataset V1, the two models achieved identical overall mAP0.5 performance of 0.625 (62%), but revealed differences in the detection precision of each class. YOLOv11x showed higher precision in detecting a missing tooth lesion of 0.670, 17% better than YOLOv8x with mAP0.5 score of 0.501. In contrast, YOLOv8x was more effective in identifying healthy tooth with a precision of 0.729, compared to 0.684 for YOLOv11x by 9% worse. Analysis of the confusion matrix indicated general difficulties with the broken tooth class and with distinguishing objects from the background. Analysis using Dataset V2 showed a slight improvement in the overall mAP0.5 value for YOLOv8x before excluding the problematic missing tooth class. After excluding this class due to suspected overfitting, achieving 99% precision in both models, However, YOLOv11x showed better overall detection statistics at 62% after excluding the missing tooth. The model also achieved better distinction of background from healthy tooth class by 33%. A comparison between the datasets revealed that Dataset V1 did a better job of differentiating between the two models' abilities to detect particular types of damage. Dataset V3 seemed to pose more promising results, achieving 82% and 86% accuracy, precision and recall. That tells us that improving datasets has a significant impact on the end result of models. This demonstrates the potential of implementing YOLO models for damage detection in our application.

In further work, it is planned to increase the sample size in the dataset for each class: healthy tooth, missing tooth, and broken tooth. In addition, it is anticipated that the database will be expanded to include new damage categories for SaMASZ P8-890 and P10-1100 rotary tedders. A key goal is to significantly improve object classification accuracy metrics. It is also necessary to implement methods to reduce visual artifacts from the background. The machine environment generates numerous distortions that negatively affect the visibility of analyzed objects. Effective segmentation of the background and identification of its components is a fundamental aspect of further research to improve the accuracy of damage detection in machine operating conditions. The deployment of the YOLOv11x and YOLOv8x architecture for high-speed applications necessitates a rigorous balance between spatial resolution and temporal frequency. In the interest of maintaining a high mean Average Precision mAP, it is recommended that, when detecting small or distant fast-moving targets, an input resolution of at least 1280x1280 pixels be considered, as lower densities result in feature degradation during down-sampling. While a standard temporal resolution of 30 FPS will be sufficient for static environments such as inspection in a diagnostic station, capturing high-velocity dynamics such as work in a field could require a high-speed acquisition rate of 60 to 120 FPS. It is thought that this increased frame rate is essential for reducing the displacement interval between successive frames, thereby enhancing the stability of object tracking algorithms. Furthermore, to mitigate the risk of motion blur, which can obscure edge gradients critical for neural network feature extraction, it is advisable to synchronise a short exposure time shutter speed of 1/1000s or faster with high-bandwidth interfaces to ensure data integrity at peak resolutions.

This research is supported by Ministry of Science and Higher Education, “VIA CARPATIA Universities of Technology Network named after the President of the Republic of Poland Lech Kaczyński” project no MEiN/2022/DPI/2575, as part of the action “ISKRA – building inter-university research teams”.