Bearing surface defect detection method and system based on improved RMB-YOLOv8

Abstract

The application provides a bearing surface defect detection method and system based on improved RMB-YOLOv8, and relates to the technical fields of industrial visual detection and deep learning target detection, and comprises the following steps: collecting bearing surface images and pre-processing the collected images; inputting the pre-processed images into a previously trained improved RMB-YOLOv8 neural network model to obtain a detection result; wherein the improved RMB-YOLOv8 neural network model replaces the original C2f module in the YOLOv8 backbone network with an RCS-OSA module, introduces a bidirectional feature pyramid network into the neck fusion layer to replace the original PAN-FPN structure, and introduces a multi-scale hollow attention mechanism into the feature output layer. The application realizes full-process automatic detection from image collection input, multi-scale feature extraction, cross-scale bidirectional fusion, attention enhancement output to defect type and position accurate positioning, significantly improves the detection accuracy and robustness of multiple types of bearing surface defects such as scratches, depressions and scratches under the premise of ensuring real-time inference efficiency.

Description

Technical Field

[0001] This invention relates to the field of industrial visual inspection and deep learning target detection technology, and in particular to a bearing surface defect detection method and system based on an improved RMB-YOLOv8. Background Technology

[0002] Bearings, as critical and vulnerable components in industrial equipment, are widely used in high-end fields such as aerospace, automotive manufacturing, agricultural machinery, and shipbuilding. During the manufacturing and assembly of bearings, process errors, improper assembly, and minute foreign objects can easily lead to defects such as dents, indentations, scratches, and abrasions on the bearing surface, directly affecting its performance and reliability, and in severe cases, even causing equipment malfunctions. Existing bearing surface defect detection technologies mainly face the following three typical types of defects: Defect 1: Manual inspection is inefficient and lacks standardized procedures, failing to meet the real-time quality inspection needs of industrial production lines. Traditional manual visual inspection relies on the experience of inspectors. Faced with the complex scenario of multiple types and scales of defects coexisting on bearing surfaces, the judgment standards of different personnel vary significantly, resulting in a high rate of missed detections. The cycle time of bearing industrial production lines is typically on the order of seconds, far exceeding the speed of manual inspection, which has severely hindered the widespread adoption of automated quality inspection.

[0003] Defect 2: Traditional machine vision methods have weak generalization ability and insufficient adaptability to complex backgrounds and small-scale defects. Existing bearing defect detection systems based on traditional image processing methods such as threshold segmentation and contour extraction rely on manually designed features and require the construction of separate detection processes for different defect types. When the defect size is extremely small, the contrast is low, or there is interference from industrial environments such as oil stains, dust, and reflections, the recognition ability of these methods decreases significantly, and both the false detection rate and the false negative rate are difficult to meet engineering requirements. For new defect types, the system cannot adapt, requiring the redesign of features and rules, resulting in high maintenance costs.

[0004] Defect 3: When mainstream deep learning object detection models are directly transferred to bearing defect detection scenarios, they suffer from high false negative rates for small targets, insufficient robustness in multi-target overlapping scenarios, and limited inference efficiency. Although mainstream detection frameworks such as YOLOv8 perform well in general object detection tasks, bearing surface defects are characterized by varying scales (coexistence of tiny scratches and large pits), irregular shapes, low feature contrast, and dense coexistence of multiple defects. YOLOv8's native C2f feature extraction module has a fixed receptive field, making it difficult to fully capture local details of tiny defects. The traditional PAN-FPN structure suffers from dissipation of small target feature information during deep downsampling. The redundant attention allocation of the detection head in complex texture backgrounds leads to significant blind spots in the recognition of tiny defects in dark areas at the bearing edges, making it prone to false negatives and false negatives in multi-target overlapping scenarios. Summary of the Invention

[0005] One of the objectives of this invention is to provide a bearing surface defect detection method and system based on an improved RMB-YOLOv8. The method and system are improved in three aspects: feature extraction, feature fusion and feature output. It realizes fully automated detection from image acquisition input, multi-scale feature extraction, cross-scale bidirectional fusion, attention-enhanced output to precise location of defect type and position. While ensuring real-time inference efficiency, it significantly improves the detection accuracy and robustness of various types of bearing surface defects such as scratches, dents and abrasions.

[0006] This invention provides a bearing surface defect detection method and system based on an improved RMB-YOLOv8, comprising: Acquire images of the bearing surface and preprocess the acquired images; The preprocessed image is input into a pre-trained improved RMB-YOLOv8 neural network model to obtain the detection results; Among them, the improved RMB-YOLOv8 neural network model replaces the original C2f module in the YOLOv8 backbone network with the RCS-OSA module, introduces a bidirectional feature pyramid network in the neck fusion layer to replace the original PAN-FPN structure, and introduces a multi-scale dilated attention mechanism in the feature output layer.

[0007] Preferably, the improved RMB-YOLOv8 neural network model uses the SGD optimizer during training, with an initial learning rate of 0.01, a learning rate momentum of 0.937, and a weight decay coefficient of 0.0005; the training period is 150 epochs, the batch size is set to 16, and the input image size is uniformly 640×640 pixels; the dataset is divided into training set, validation set, and test set in a 7:2:1 ratio.

[0008] Preferably, the detection results include: visualized labeled images and / or structured data.

[0009] Preferably, the preprocessing includes one or more of the following: color transformation, grayscale conversion, contrast enhancement, and geometric transformation.

[0010] Preferably, acquiring images of the bearing surface includes: An industrial camera, in conjunction with a spherical integrating light source and a light source controller, is used to acquire images of the bearing surface at multiple preset angles and under various preset lighting conditions.

[0011] The present invention also provides a bearing surface defect detection system based on the improved RMB-YOLOv8, comprising: a data acquisition module and a detection module; The acquisition module acquires images of the bearing surface and preprocesses the acquired images; the detection module inputs the preprocessed images into a pre-trained improved RMB-YOLOv8 neural network model to obtain detection results. Among them, the improved RMB-YOLOv8 neural network model replaces the original C2f module in the YOLOv8 backbone network with the RCS-OSA module, introduces a bidirectional feature pyramid network in the neck fusion layer to replace the original PAN-FPN structure, and introduces a multi-scale dilated attention mechanism in the feature output layer.

[0012] Preferably, the improved RMB-YOLOv8 neural network model uses the SGD optimizer during training, with an initial learning rate of 0.01, a learning rate momentum of 0.937, and a weight decay coefficient of 0.0005; the training period is 150 epochs, the batch size is set to 16, and the input image size is uniformly 640×640 pixels; the dataset is divided into training set, validation set, and test set in a 7:2:1 ratio.

[0013] Preferably, the detection results include: visualized labeled images and / or structured data.

[0014] Preferably, the preprocessing includes one or more of the following: color transformation, grayscale conversion, contrast enhancement, and geometric transformation.

[0015] Preferably, the acquisition module performs the following operations: An industrial camera, in conjunction with a spherical integrating light source and a light source controller, is used to acquire images of the bearing surface at multiple preset angles and under various preset lighting conditions.

[0016] The beneficial effects of this invention are as follows: To address the challenge of detecting minute defects on bearing surfaces, an RMB-YOLOv8 detection model is proposed. Utilizing the RCS-OSA module and structural reparameterization to replace the C2f structure, multi-path convolution enhances the learning of diverse features during training, while lightweight compression in the inference phase expands the receptive field without increasing inference overhead. This effectively enhances the feature extraction capability for small targets such as 0.3mm-level scratches and abrasions, reducing the false negative rate and significantly improving the accuracy of single-class defects across all three defect types. By employing a BiFPN bidirectional fusion structure and learnable weights, the multi-scale feature fusion effect is optimized, reducing the loss of minute defect information caused by downsampling, resulting in a significant improvement in overall mAP@0.5. The MSDA multi-scale dilated attention mechanism is introduced to suppress attention redundancy, balancing local details and global semantics with only a slight increase in computation, while maintaining inference speed that meets the requirements of real-time industrial inspection. The model combines detection accuracy with robustness against interference in complex scenarios, adapting to conditions such as multiple defects coexisting, reflection, and low light. Furthermore, it features easy data expansion and flexible hardware and software deployment, enabling efficient adaptation to different bearing production lines and providing reliable and practical technical support for intelligent quality inspection of bearing surfaces.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a bearing surface defect detection method based on an improved RMB-YOLOv8 according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a bearing surface defect detection system based on the improved RMB-YOLOv8 in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] This invention provides a bearing surface defect detection method based on an improved RMB-YOLOv8, such as... Figure 1 As shown, it includes: Step 1: Acquire images of the bearing surface and preprocess the acquired images; The process of acquiring bearing surface images includes: using an industrial camera in conjunction with a spherical integrating light source and a light source controller to acquire images of the bearing surface at multiple preset angles and under various preset lighting conditions. Preprocessing includes one or more of the following: color transformation, grayscale conversion, contrast enhancement, and geometric transformation.

[0022] Step 2: Input the preprocessed image into the pre-trained improved RMB-YOLOv8 neural network model to obtain the detection results; The improved RMB-YOLOv8 neural network model, after training, employs a decoupled detection head to simultaneously perform category prediction and location regression for three types of bearing surface defects (scrape, scratch, and dent), outputting the category label, confidence score, and bounding box coordinates for each defect. The detection results include: visualized annotated images and / or structured data; that is, the detection results are output in the form of visualized annotated images and structured data, providing real-time defect type distribution and location information for use by downstream quality inspection decision-making systems.

[0023] Among them, the improved RMB-YOLOv8 neural network model replaces the original C2f module in the YOLOv8 backbone network with the RCS-OSA module, introduces a bidirectional feature pyramid network in the neck fusion layer to replace the original PAN-FPN structure, and introduces a multi-scale dilated attention mechanism in the feature output layer.

[0024] The improved RMB-YOLOv8 neural network model includes a collaborative backbone network, a neck fusion layer, and a feature output layer. These modules are not independent replacements, but rather form an organic whole to achieve a complete technical logic of small feature extraction, cross-scale preservation, and context awareness. Specifically: The backbone network adopts a reparameterized channel shuffling and one-time aggregation (RCS-OSA) structure to enhance the feature extraction capability for the boundary texture of small defects while expanding the receptive field. The neck fusion layer adopts a bidirectional feature pyramid network (BiFPN) structure, receives multi-scale features output by the backbone network, and adaptively fuses them through learnable weights to suppress information dissipation of small defect features during deep downsampling and cross-scale transmission. The feature output layer employs a multi-scale hole attention (MSDA) mechanism to enhance the context-awareness of the fused features output by the neck fusion layer, thereby improving the defect localization accuracy in complex industrial reflective and low-light backgrounds.

[0025] During the training data acquisition phase, an industrial camera (CMOS industrial camera MV-CS050-10GC) combined with a spherical integrating light source (DBS-WD240) and a light source controller was used to acquire images of the bearing surface from multiple angles and under various lighting conditions, obtaining raw image samples containing three types of defects: scratches, dents, and abrasions. Preprocessing operations such as adding noise, brightness and contrast color transformation, and random angle rotation were performed on the acquired images to reduce redundant background information and enhance the expression of local defect features, thus constructing a high-quality training dataset. During the training phase, a Mosaic data augmentation strategy was introduced, where four images were scaled to the same size and then recombined at random cut points, effectively enriching the samples of small-target defects and improving the model's generalization ability to dense defect scenes. The original C2f module in the YOLOv8 backbone network was replaced with the RCS-OSA (One-Time Feature Aggregation Based on Reparameterized Channel Shuffling) module, fundamentally solving the problem of insufficient ability of a fixed receptive field to capture local features of small-scale defects. During training, the RCS module adopts a multi-path parallel convolutional structure—running identity channels, 1×1 convolutions, and 3×3 convolutions in parallel, integrating them into an equivalent 3×3RepConv through structural reparameterization. Combined with the channel shuffling mechanism, this significantly enhances feature diversity and expressive power without increasing inference overhead. During inference, the multi-path structure is equivalently compressed into a single 3×3RepConv module, achieving lightweight computation. The OSA (One-Time Feature Aggregation) strategy acquires diverse features through multiple receptive fields and integrates all outputs at the end, avoiding the linear expansion of memory access overhead (MAC) caused by DenseNet-style dense connections. While effectively expanding the receptive field and strengthening the ability to model spatial features of small defects, it significantly reduces inference latency. A Bidirectional Feature Pyramid Network (BiFPN) is introduced in the neck fusion layer to replace the original PAN-FPN structure, specifically addressing the problem of information dissipation during deep downsampling of minute bearing defects. BiFPN constructs a bidirectional cross-scale information interaction path from top to bottom and bottom to top, enabling full bidirectional interaction between shallow detail features (edge ​​textures of small target defects) and deep semantic features (defect category semantics). Unlike the direct, weighted accumulation of traditional FPNs, BiFPN assigns learnable fusion weights (supporting scalar, vector, or multi-dimensional tensor forms) to each input feature. It adaptively adjusts the contribution of features at each scale through a normalized weighted fusion formula, effectively reducing parameter redundancy while maintaining low computational overhead. This ensures that shallow detail features are fully transmitted to the detection output, significantly improving the ability to represent minute defects. In this bidirectional, cross-scale information interaction process, to ensure the stable convergence of the network model, the specific data transformation logic of the normalized weighted fusion is as follows: First, the system assigns a non-negative, learnable initial weight parameter to each input feature node participating in the fusion (its non-negativity is guaranteed by a linear rectified activation function). During node fusion operations, the ratio of the initial weight of the current input feature to the sum of the initial weights of all input features participating in the fusion of that node is calculated. Simultaneously, a very small anti-zero outlier constant is added to the sum of the denominator terms, thereby calculating the normalized relative weight of that input feature. Finally, the output feature tensor of that node is the sum of the products of each input feature tensor and its corresponding normalized relative weight. Through the above data operator logic, the system can automatically learn the optimal allocation ratio of features at each scale during backpropagation without hard coding.

[0026] A multi-scale dilated attention (MSDA) mechanism is introduced at the feature output layer. Using a sliding window expanded attention (SWDA) strategy, differentiated expansion rates are set for different attention heads (corresponding to different receptive field ranges such as 3×3, 5×5, and 7×7). Self-attention computation is performed on other image patches sparsely selected within the neighborhood of the query image patch. Each attention head extracts local detail features (small expansion rate head) and long-range contextual semantics (large expansion rate head) at its corresponding scale. The outputs of all heads are then concatenated and integrated through a linear layer to achieve simultaneous modeling of local details and global contextual information of bearing surface defects. This mechanism effectively suppresses self-attention redundancy without introducing complex additional computations, comprehensively improving feature representation quality and inference efficiency without significantly increasing computational overhead, and enhancing the model's robustness to complex lighting, reflective interference, and densely distributed multi-defect scenarios. The improved RMB-YOLOv8 neural network model employs the SGD optimizer during training, with an initial learning rate of 0.01, a learning rate momentum of 0.937, and a weight decay coefficient of 0.0005. The training period is 150 epochs, the batch size is 16, and the input image size is uniformly 640×640 pixels. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Combined with data augmentation strategies in the data preprocessing stage, overfitting is effectively prevented while ensuring sufficient model convergence.

[0027] This invention preferably employs a specific set of training hyperparameters to ensure that the improved RMB-YOLOv8 model structure can be trained sufficiently and stably, avoiding gradient explosion or feature loss. Specifically, the initial learning rate is set to 0.01 with a momentum of 0.937. This value setting has been experimentally verified to effectively adapt to the gradient propagation characteristics of the multi-path parallelism in the training phase of the aforementioned RCS-OSA module, overcoming the potential training instability introduced by structural reparameterization. Simultaneously, the batch size is set to 16, which is an optimized trade-off for the accurate updating of learnable weight parameters in the BiFPN network structure, taking into account the hardware memory limitations of industrial quality inspection equipment, ensuring the effectiveness of cross-scale feature fusion. 150 training epochs ensure that the model can converge sufficiently on datasets containing a large number of small defect samples, enabling the MSDA mechanism to ultimately learn more discriminative sparse attention weights. The coordinated configuration of the above hyperparameters is not only a standard experimental choice but also a crucial operating condition to ensure that the core architecture of this invention achieves the expected technical results.

[0028] In addition, a Task-Aligned Assigner is introduced to dynamically adjust the positive and negative sample allocation strategy, thereby improving the model's adaptability to multi-scale and multi-type defect data distribution.

[0029] A specific implementation of the bearing surface defect detection method based on the improved RMB-YOLOv8: An online visual inspection station is deployed at the end of the bearing outer ring production line. When the bearing outer ring is conveyed to the inspection station, a CMOS industrial camera MV-CS050-10GC, in conjunction with a spherical integrating light source DBS-WD240, completes 360° toroidal image acquisition, with a single acquisition taking approximately 8ms. After the image is input into the RMB-YOLOv8 inference engine, the RCS-OSA backbone network performs multi-scale extraction of the toroidal features of the bearing outer ring. For tiny scratch defects with a size of only about 0.3mm in this scenario, the RCS module, through multi-path parallel convolution, learns equivalent 3×3 RepConv weights during the training phase, which can accurately capture its edge texture details without increasing the inference computation, overcoming the problem of the original YOLOv8 missing bearing edge scratches in low-light environments. BiFPN integrates scratch texture details from shallow high-resolution feature maps with deep semantic features through adaptive up-adjustment of learnable fusion weights. This effectively preserves key spatial information of scratch defects during multiple downsampling processes and avoids the feature dissipation of small targets caused by the original PAN-FPN structure. The MSDA mechanism performs multi-scale sparse attention enhancement on the input features of the detection head: a small expansion rate head (3×3 receptive field) accurately focuses on the local texture of the scratch boundary, while a large expansion rate head (7×7 receptive field) perceives the overall surface uniformity context of the bearing. The collaboration between the two enables the detection head to accurately locate the scratch area even in complex reflective backgrounds. The final detection result outputs the scratch defect category label "Scrape", confidence score, and bounding box coordinates in real time, with an inference latency of approximately 8ms (approximately 125 frames / second), meeting the production line cycle time requirements.

[0030] The dataset used during training directly affects the applicability of the improved RMB-YOLOv8 neural network model after training. Therefore, in one embodiment, the bearing surface defect detection method based on the improved RMB-YOLOv8 further includes: an application environment analysis module to obtain defective detection records from historical production records corresponding to the configured detection line; analyze the defective detection records to determine the proportion of each defect item in the detection results; and configure the sample proportion of each defect type in the dataset based on the determined proportion.

[0031] In practical applications of online inspection systems, production line defects are typically caused by one or more defects occurring repeatedly. To facilitate adaptive updates, a triggering device is configured at the online inspection end. Operators can use this device to retrain the model. During this retraining process, a backup inspection model is configured. While retraining the commonly used model, the system takes over inspection of the production line. When triggering the system, operators must provide one or more defective parts, which are then captured by the image acquisition device of the triggering module. Based on the captured image, the sample proportions corresponding to each defect type in the dataset are reconfigured. After retraining is complete, the retrained model takes over production line inspection, and the inspection results are statistically analyzed within a preset time period. When the statistical results meet pre-configured conditions, the backup model is set as the primary model. Furthermore, since there is more than one production line in the factory, before retraining, we can communicate with the online detection systems of other production lines to determine if there are already trained models that can be directly copied and called, thus reducing retraining time. Specifically, we can construct the identifier vector based on the reconfiguration of the sample ratio. We can then match the identifier vector of the model that needs to be retrained with the identifier vector of the model used in other online systems, and directly copy and call the matched model used in other online systems.

[0032] To enable direct copying and calling of multi-production line models, the system incorporates a multi-production line model matching mechanism. The specific data conversion and judgment logic is as follows: First, an identifier vector is constructed: the percentage of each preset defect type (such as scratches, dents, etc.) in the latest dataset is arranged according to the fixed defect category order predetermined by the system and transformed into a one-dimensional continuous floating-point feature vector (e.g., [scratch percentage, scratch percentage, dent percentage]). This one-dimensional vector is used as an identifier vector to represent the current defect data distribution characteristics of the production line. Next, model matching and invocation are performed: the identifier vector of the model to be trained on the current production line is compared with the historical identifier vectors of models already saved in other online systems in the communication network to calculate spatial similarity. The system uses a judgment operator to calculate the spatial distance between the two vectors (preferably using cosine similarity or Euclidean distance) to generate a quantified similarity score. When the calculated similarity score of a certain historical model reaches a preset matching threshold (e.g., cosine similarity greater than or equal to 0.95), it is determined that the data distribution is highly consistent and the match is successful. At this time, the automated control path is triggered, and the current online system is exempt from the process of retraining from scratch, directly copying and taking over the invocation of the successfully matched existing model across production lines. Through the above quantified matching operator, the engineering execution problem of adaptive synchronous model changeover across multiple production lines is effectively solved.

[0033] This invention also provides a bearing surface defect detection system based on the improved RMB-YOLOv8, such as... Figure 2As shown, it includes: acquisition module 1 and detection module 2; wherein, acquisition module 1 acquires images of the bearing surface and preprocesses the acquired images; detection module 2 inputs the preprocessed images into a pre-trained improved RMB-YOLOv8 neural network model to obtain detection results; Among them, the improved RMB-YOLOv8 neural network model replaces the original C2f module in the YOLOv8 backbone network with the RCS-OSA module, introduces a bidirectional feature pyramid network in the neck fusion layer to replace the original PAN-FPN structure, and introduces a multi-scale dilated attention mechanism in the feature output layer.

[0034] The improved RMB-YOLOv8 neural network model uses the SGD optimizer during training, with an initial learning rate of 0.01, a learning rate momentum of 0.937, and a weight decay coefficient of 0.0005. The training period is 150 epochs, the batch size is 16, and the input image size is uniformly 640×640 pixels. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio.

[0035] The detection results include: visually labeled images and / or structured data.

[0036] Preprocessing includes one or more of the following: color transformation, grayscale conversion, contrast enhancement, and geometric transformation.

[0037] The acquisition module performs the following operations: using an industrial camera in conjunction with a spherical integrating light source and a light source controller, it acquires images of the bearing surface at multiple preset angles and under various preset lighting conditions.

[0038] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A bearing surface defect detection method based on an improved RMB-YOLOv8, characterized in that, include: Acquire images of the bearing surface and preprocess the acquired images; The preprocessed image is input into a pre-trained improved RMB-YOLOv8 neural network model to obtain the detection results; Among them, the improved RMB-YOLOv8 neural network model replaces the original C2f module in the YOLOv8 backbone network with the RCS-OSA module, introduces a bidirectional feature pyramid network in the neck fusion layer to replace the original PAN-FPN structure, and introduces a multi-scale dilated attention mechanism in the feature output layer.

2. The bearing surface defect detection method based on the improved RMB-YOLOv8 as described in claim 1, characterized in that, The improved RMB-YOLOv8 neural network model uses the SGD optimizer during training, with an initial learning rate of 0.01, a learning rate momentum of 0.937, and a weight decay coefficient of 0.0005. The training period is 150 epochs, the batch size is 16, and the input image size is uniformly 640×640 pixels. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio.

3. The bearing surface defect detection method based on the improved RMB-YOLOv8 as described in claim 1, characterized in that, The detection results include: visually labeled images and / or structured data.

4. The bearing surface defect detection method based on the improved RMB-YOLOv8 as described in claim 1, characterized in that, Preprocessing includes: Color transformation, grayscale conversion, contrast enhancement, and geometric transformation, or a combination of one or more of these techniques.

5. The bearing surface defect detection method based on the improved RMB-YOLOv8 as described in claim 1, characterized in that, The acquisition of bearing surface images includes: An industrial camera, in conjunction with a spherical integrating light source and a light source controller, is used to acquire images of the bearing surface at multiple preset angles and under various preset lighting conditions.

6. A bearing surface defect detection system based on an improved RMB-YOLOv8, characterized in that, include: Acquisition module and detection module; The acquisition module acquires images of the bearing surface and preprocesses the acquired images; the detection module inputs the preprocessed images into a pre-trained improved RMB-YOLOv8 neural network model to obtain detection results. Among them, the improved RMB-YOLOv8 neural network model replaces the original C2f module in the YOLOv8 backbone network with the RCS-OSA module, introduces a bidirectional feature pyramid network in the neck fusion layer to replace the original PAN-FPN structure, and introduces a multi-scale dilated attention mechanism in the feature output layer.

7. The bearing surface defect detection system based on the improved RMB-YOLOv8 as described in claim 6, characterized in that, The improved RMB-YOLOv8 neural network model uses the SGD optimizer during training, with an initial learning rate of 0.01, a learning rate momentum of 0.937, and a weight decay coefficient of 0.0005. The training period is 150 epochs, the batch size is 16, and the input image size is uniformly 640×640 pixels. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio.

8. The bearing surface defect detection system based on the improved RMB-YOLOv8 as described in claim 6, characterized in that, The detection results include: visually labeled images and / or structured data.

9. The bearing surface defect detection system based on the improved RMB-YOLOv8 as described in claim 6, characterized in that, Preprocessing includes: Color transformation, grayscale conversion, contrast enhancement, and geometric transformation, or a combination of one or more of these techniques.

10. The bearing surface defect detection system based on the improved RMB-YOLOv8 as described in claim 6, characterized in that, The data acquisition module performs the following operations: An industrial camera, in conjunction with a spherical integrating light source and a light source controller, is used to acquire images of the bearing surface at multiple preset angles and under various preset lighting conditions.

Images