Hydraulic turbine maintenance tool recognition method based on improved YOLOv8

Abstract

The application discloses a kind of based on improved YOLOv8's water turbine overhaul tool identification method, it is related to target detection field.The method includes: collecting overhaul tool image and forming data set by data processing;Improved YOLOv8 model is constructed, GAM attention module is additionally arranged between each C2f module of neck network and Detect module, and multi-scale feature fusion enhancement layer is introduced, 160x160 detection head is newly added in head network, and EDSR super-resolution reconstruction module is additionally arranged, simultaneously, K-Means dynamic anchor frame generation algorithm is used;Loss function fuses EIoU, alpha-balancedFocal and BCE loss, and introduces tool posture change factor to dynamically adjust the height-width ratio constraint coefficient;After training, generate identification model.The application effectively suppresses complex background interference, improves small size tool detection capability, enhances the adaptability to tool posture change, provides high-precision, high-stability technical support for hydropower station overhaul intelligentization.

Description

Technical Field

[0001] This invention relates to the field of intelligent target detection technology, specifically to a method for identifying turbine maintenance tools based on an improved YOLOv8. Background Technology

[0002] Hydropower stations, as a crucial component of the power system, undertake critical tasks such as frequency regulation, peak shaving, and backup power. Their operational safety directly impacts the stability and reliability of the power grid. During hydropower station maintenance, personnel must use tools of various sizes and specifications, including wrenches, screwdrivers, feeler gauges, and dial indicators, for equipment maintenance and repair. Accurate tool identification and inventory are critical to ensuring maintenance quality and preventing tools from being missed or left inside equipment, which could lead to safety accidents. However, traditional manual tool management relies on the experience and subjective judgment of operators. In the complex environment of turbine maintenance sites (oil stains, metal reflections, changing lighting), this approach is not only inefficient but also prone to misjudgment or omission due to the large variety of tools and the indistinct characteristics of small-sized tools, increasing potential safety hazards in equipment operation.

[0003] In existing technologies, deep learning-based equipment condition detection methods have been widely studied in industrial scenarios. For example, Chen Tiehua et al. proposed a hydropower plant maintenance tool identification system based on the YOLOv5 algorithm, which achieved preliminary automated identification by performing target detection on tool images. Zhou Kun et al. introduced a CoordinateAttention (CA) module into the YOLOv5 feature extraction network for hydropower station equipment defect detection, which improved detection accuracy while reducing the number of model parameters. However, the above technical solutions still have the following shortcomings when applied to hydropower station maintenance tool identification: First, existing methods mainly focus on the detection accuracy of single-frame static images, without fully considering the problem that maintenance tool features are gradually obscured by complex background noise (such as oil stains and metal reflections) as the depth of the convolutional network increases. This leads to a significant increase in the false negative rate of small tools (such as feeler gauges and dial indicators) at a slightly longer distance or against complex backgrounds. Second, existing solutions are mostly based on static image detection paradigms, and lack effective modeling for the inter-frame detection jitter and frequent identity switching problems caused by the dual coupling of handheld shooting equipment and free tool movement in video streams. Third, while lightweight improvement schemes can reduce the number of model parameters, they usually come at the cost of sacrificing the accuracy of small target detection, making it difficult to achieve high-precision, high-stability real-time recognition under the constraints of computing power and power consumption on edge computing devices. Therefore, designing a lightweight target detection system that can solve the problem of feature annihilation of small-sized tools in complex industrial maintenance scenarios, overcome inter-frame instability caused by dual motion coupling, and achieve high-precision real-time recognition under the constraints of edge computing resources has become a pressing technical challenge. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for identifying turbine maintenance tools based on the improved YOLOv8, so as to solve the problems in the prior art where maintenance tools are prone to losing small target features in complex backgrounds, unstable positioning due to changing postures, and insufficient real-time performance of edge deployment.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for identifying turbine maintenance tools based on an improved YOLOv8 includes the following steps: Collect images of maintenance tools and process the data to form a maintenance tool dataset; A turbine maintenance tool identification model based on an improved YOLOv8 was constructed. The model includes: Backbone networks are used to extract multi-scale basic features of images; The neck network adds a GAM attention module between each C2f module and the Detect module of the original neck network, and introduces a multi-scale feature fusion enhancement layer to further refine the fused features. The head network adds a new detection head that connects to the highest resolution feature map on the basis of the original three detection heads, and adds a lightweight super-resolution reconstruction EDSR module in front of the new detection head. At the same time, it adopts a dynamic anchor box generation algorithm based on the size distribution of K-Means clustering tool to replace the fixed anchor boxes. The loss function integrates EIoU loss, α-balanced Focal loss and BCE loss, and introduces a tool attitude change factor to dynamically adjust the aspect ratio constraint coefficient of EIoU loss; The improved YOLOv8 model was trained using the aforementioned maintenance tool dataset to generate a trained recognition model for recognizing turbine maintenance tools.

[0007] Furthermore, the aforementioned multi-scale feature fusion enhancement layer consists of a 1×1 convolutional layer, a 3×3 convolutional layer, and a lightweight channel attention (SE) module connected in sequence; the 1×1 convolutional layer is used for channel compression, the 3×3 convolutional layer is used for spatial feature enhancement, and the SE module is used for channel recalibration to focus on tool features.

[0008] Furthermore, the aforementioned lightweight super-resolution reconstruction EDSR module includes: a 3×3 convolutional layer for shallow feature extraction, multiple cascaded residual blocks for deep residual feature learning, a 3×3 convolutional layer for global feature fusion, and a sub-pixel convolutional layer for resolution enhancement; wherein, each residual block has its batch normalization layer removed.

[0009] Furthermore, the number of the aforementioned multiple cascaded residual blocks is 8. Each residual block consists of a 3×3 convolutional layer, a ReLU activation function, and a 3×3 convolutional layer connected in series, which are then residually connected to the input.

[0010] Furthermore, the above dynamic anchor frame generation algorithm specifically involves: performing K-Means clustering on the width and height of all labeled boxes in the dataset before training, with the number of clusters k=3, to obtain the anchor frame size representing the actual size distribution of the turbine maintenance tools, which replaces the preset fixed anchor frame size.

[0011] Furthermore, the tool pose change factor δ_pose is defined as the absolute difference between the predicted bounding box aspect ratio and the true bounding box aspect ratio, expressed as:

[0012] Where w and h are the predicted bounding box width and height, and w^gt and h^gt are the true bounding box width and height; the above tool pose change factor is used to dynamically scale the coefficient of the aspect ratio penalty term V in the EIoU loss, and the scaling function is expressed as:

[0013] Where μ is the adjustment coefficient.

[0014] Furthermore, the above EIoU loss function is expressed as:

[0015] Where IoU is the intersection-union ratio, ρ²(b,b^gt) is the Euclidean distance between the center of the predicted box and the center of the ground truth box, c is the diagonal length of the minimum bounding rectangle of the two boxes, and V is the aspect ratio penalty term.

[0016] Furthermore, during training, the batch size was set to 16, the total number of training rounds was 300, the initial learning rate was 0.001, and a cosine annealing learning rate scheduling strategy was adopted. The Adam optimizer was selected, and Dropout layers with a probability of 0.1 were introduced in the C2f and GAM modules.

[0017] Furthermore, the above data processing includes data augmentation, data size normalization, and data annotation; the above data augmentation includes horizontal flipping, vertical flipping, random rotation, random cropping, brightness adjustment, and contrast adjustment; the above data size normalization is to uniformly scale all images to 640×640 pixels.

[0018] Furthermore, the aforementioned maintenance tool dataset was divided into training, validation, and test sets in a 7:2:1 ratio. Stratified sampling was used during the division process to ensure that the proportion of each type of tool in each dataset remained consistent with the original dataset.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By introducing a GAM attention module and a feature fusion enhancement layer into the neck network, the interference from complex backgrounds is effectively suppressed, and the key features of the tool are enhanced; 2. A small target detection head is added to the head network and combined with the EDSR super-resolution reconstruction module, which significantly improves the ability to capture small-sized tools; 3. A dynamic anchor frame generation algorithm is used to adapt the anchor frame size to the actual distribution of the tools; 4. Construct a composite loss function that includes attitude factors to enhance the model's robustness to changes in tool attitude. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the overall process of the turbine maintenance tool identification method based on the improved YOLOv8 provided in this embodiment of the invention; Figure 2 A diagram of a hydropower station maintenance tool identification network based on an improved YOLOv8 provided in an embodiment of the present invention; Figure 3 The network structure diagram of the lightweight super-resolution reconstruction module based on the EDSR architecture provided in the embodiments of the present invention is shown. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The method in this embodiment is executed by a terminal, which can be a mobile phone, tablet computer, PDA, laptop or desktop computer, etc. Of course, it can also be other devices with similar functions, and this embodiment does not limit them.

[0023] Please see Figure 1 This embodiment provides a method for identifying turbine maintenance tools based on an improved YOLOv8, including the following steps: S1. Collect images of maintenance tools and process the data to form a maintenance tool dataset.

[0024] In this embodiment, 10 types of tools commonly used in turbine maintenance are selected as research objects, including box wrenches, torque wrenches, hammers, feeler gauges, screwdrivers, Allen wrenches, dial indicators, levels, jacks, and locating pins, covering key maintenance links such as disassembly, measurement, and assembly. Industrial cameras were used to collect images at a turbine maintenance site in a hydropower station, including images with different lighting conditions, different backgrounds (clean tools and tools with oil stains), and different tool states (individual placement and combined use). The collected images were processed to expand the dataset through data augmentation, data size normalization, and data annotation. Data augmentation included horizontal flipping, vertical flipping, random rotation, random cropping, brightness adjustment, and contrast adjustment; data size normalization involved scaling all images to 640×640 pixels to meet the input requirements of the YOLOv8 model; data annotation used the LabelImg annotation tool to annotate tool categories and bounding boxes in the images, outputting an XML intermediate format file. A custom Python script then converted the XML file to the txt format required for YOLOv8 training, ensuring that the annotated data was directly adapted for model training.

[0025] S2. Divide the maintenance tool dataset into training, validation, and test sets in a ratio of 7:2:1.

[0026] The training set is used for model parameter learning and feature extraction, the validation set is used to monitor the training effect in real time during training, and hyperparameters are dynamically adjusted by calculating metrics such as mAP and recall, and the test set is used to finally evaluate the model's recognition performance. A stratified sampling method is used during the partitioning process to ensure that the proportion of various tools in the training, validation, and test sets is consistent with the original dataset, avoiding the impact of uneven sample distribution on the accuracy of model training and evaluation.

[0027] S3. Construct a turbine maintenance tool identification model based on the improved YOLOv8.

[0028] like Figure 2 As shown, the original YOLOv8 network is used as the basic framework, retaining its three-order architecture of backbone network-neck network-head network, and the neck network, head network and loss function are optimized respectively.

[0029] S301. The backbone network is retained.

[0030] The original YOLOv8 backbone network consists of an input layer, multiple CBS modules (convolution + batch normalization + activation function), a C2f module (feature fusion structure with residual connections), and an SPPF module (spatial pyramid pooling). The CBS module uses 3×3 convolutional kernels for initial feature extraction; the C2f module enhances the fluidity and expressiveness of features by splitting the input features into multiple branches and performing residual connections; the SPPF module aggregates features through pooling operations of different sizes (1×1, 5×5, 9×9, 13×13) to improve adaptability to targets of different scales.

[0031] S302. Improved neck network.

[0032] The neck network consists of multiple C2f modules, ConCat operations, and UpSample operations, used to fuse and propagate multi-scale features output from the backbone network. In this embodiment, a GAM attention module is added between the output of each C2f module in the original neck network and the input of the subsequent Detect module. Each GAM module includes a channel attention subnet and a spatial attention subnet. The channel attention subnet performs global average pooling on the input feature map to generate channel descriptors, learns the dependencies between channels through a multilayer perceptron, generates channel attention weight vectors, and multiplies them with the input features to achieve channel-dimensional feature recalibration, highlighting important channels such as tool texture and color, while suppressing unimportant channels such as background and noise. The spatial attention subnet applies lightweight convolutional layers to the channel-weighted feature map to integrate spatial information, generates a spatial attention weight map, and multiplies it with the feature map to achieve spatial-dimensional feature focusing, emphasizing tool region features and suppressing background region features.

[0033] After completing a feature fusion operation (upsampling + concatenation + C2f), a feature fusion enhancement layer is added. This layer works in conjunction with the GAM module to further refine the fused features. The feature fusion enhancement layer consists of three core components connected in series: (1) 1×1 convolutional layer: The channel dimension of the fused feature map is compressed, and the number of output channels is set to half the number of input channels to reduce the amount of computation and achieve preliminary feature crossing and information integration.

[0034] (2) 3×3 convolutional layer: spatial feature fusion and enhancement of compressed feature map, effectively capturing local spatial context information and enhancing key spatial features at the edge.

[0035] (3) Lightweight Channel Attention (SE) Module: The Squeeze-and-Excitation operation is applied to the feature map output by the 3×3 convolution. First, the spatial information is compressed by global average pooling to generate channel descriptors. Then, the nonlinear relationship between channels is learned through a gating mechanism to generate channel weight vectors. Finally, the weight vectors are multiplied with the original feature map to achieve channel recalibration, which strengthens the feature channels that are highly related to the tool and suppresses the feature channels that are related to interference such as oil stains and metal reflections.

[0036] S303. Header network improvements.

[0037] The original YOLOv8 head network contains three detector heads, corresponding to the 80×80, 40×40, and 20×20 feature maps output by the neck network, respectively. This embodiment adds a new detector head to the head network, which is connected to the 160×160 high-resolution feature map output by the neck network and is specifically designed for detecting small objects. A lightweight super-resolution reconstruction module based on the EDSR architecture is inserted before this new detector head, and its structure is as follows... Figure 3 As shown, it includes: (1) Shallow feature extraction: A 3×3 convolutional layer extracts the initial features of the input feature map.

[0038] (2) Deep residual feature learning: 8 cascaded residual blocks. Each residual block consists of a 3×3 convolutional layer, a ReLU activation function and a 3×3 convolutional layer connected in series with the input and residual connection. The batch normalization layer is removed from each residual block to reduce computation and memory consumption, avoid artifacts and allow a larger learning rate.

[0039] (3) Global feature fusion: A 3×3 convolutional layer aggregates the features extracted from all residual blocks.

[0040] (4) Upsampling reconstruction: Sub-pixel convolutional layer, which increases the number of channels through specific convolution, and then converts the channel information into spatial information through pixel rearrangement to generate high-resolution feature maps.

[0041] Meanwhile, a dynamic anchor box generation algorithm based on K-Means clustering tool size distribution is used in the head network to replace the fixed anchor boxes. Specifically, before training, K-Means clustering is performed on the width and height of all labeled boxes in the dataset, with the number of clusters k=3, to obtain the anchor box size representing the actual size distribution of the turbine maintenance tool, which is used to replace the preset fixed anchor box size.

[0042] S304. Loss function optimization.

[0043] To improve the robustness and accuracy of the model in scenarios with complex background interference, difficulty in detecting small targets, and variable tool poses, a scene-aware composite loss function is constructed. This loss function integrates EIoU loss, α-balancedFocal loss, and BCE loss, and introduces a tool pose change factor to dynamically adjust the aspect ratio constraint coefficient of the EIoU loss.

[0044] The composite loss function is expressed as follows:

[0045] in, To improve EIoU loss, For α-balanced Focal loss, This is the cross-entropy loss for binary classification.

[0046] The placement of turbine maintenance tools is not fixed (e.g., horizontal, vertical, tilted), causing variations in the aspect ratio of the actual frame. Traditional EIoU loss may over-penalize aspect ratio differences. Therefore, a tool posture variation factor is defined. The aspect ratio difference between the predicted bounding box and the ground truth bounding box:

[0047] in, , To predict the width and height of the bounding box, , This represents the actual frame width and height. A scaling function is introduced. Dynamically adjust the aspect ratio penalty term in EIoU loss coefficient:

[0048] in The adjustment factor is set to 0.5 in this embodiment. The improved EIoU loss is:

[0049]

[0050] in For intersection, union, and comparison, This represents the Euclidean distance between the center of the predicted bounding box and the center of the ground truth bounding box. This is the length of the diagonal of the smallest bounding rectangle of the two frames. When... When it is large, Decrease, for The penalty for the item is reduced to avoid oversensitivity to reasonable changes in posture; when When smaller, Approaching 1, maintains sufficient penalty for shape error.

[0051] To address the class imbalance issue among commonly used tools in maintenance tools, such as box wrenches with a large sample size and niche tools with a small sample size, α-balanced Focal loss is employed:

[0052] in For real category labels, To predict probabilities, The category weights are 0.2 for commonly used tools and 0.8 for less common tools. These are focusing parameters used to enhance the weights of difficult-to-differentiate samples.

[0053] To help eliminate interference from complex backgrounds (oil stains, equipment, etc.), a binary classification cross-entropy loss is used:

[0054] in This is a real label (1 indicates the existence of a tool). To predict the probability of existence.

[0055] S4. Train a turbine maintenance tool recognition model based on the improved YOLOv8.

[0056] Considering the size and feature complexity of the turbine maintenance tool dataset, the training parameters were set as follows: batch size of 16, total training epochs of 300, initial learning rate of 0.001, and a cosine annealing learning rate scheduling strategy, using the Adam optimizer. Simultaneously, Dropout layers with a probability of 0.1 were introduced into the C2f and GAM modules of the network to suppress overfitting risks in complex contexts by randomly discarding neurons, thereby improving the model's generalization ability.

[0057] S5. Generate a turbine maintenance tool identification model based on the improved YOLOv8.

[0058] The trained network weights are saved to obtain a turbine maintenance tool recognition model based on the improved YOLOv8, which is used for real-time recognition of turbine maintenance tools in real-world scenarios.

[0059] Example 2 This embodiment is basically the same as Embodiment 1, except for the data augmentation operation in data processing and the value of the adjustment coefficient μ in the loss function.

[0060] In terms of data augmentation, in addition to the horizontal flipping, vertical flipping, random rotation, random cropping, brightness adjustment, and contrast adjustment used in Example 1, this example adds two more augmentation methods: Gaussian noise and color dithering. Specifically, during the image preprocessing stage, Gaussian noise with a mean of 0 and a standard deviation of 0.01 is randomly added to some images to simulate a sensor noise environment; simultaneously, the hue and saturation of the images are randomly fine-tuned to enhance the model's adaptability to changes in tool color under different lighting conditions. These augmentation operations further expand the diversity of the dataset and help improve the model's robustness under complex lighting and noise conditions.

[0061] Regarding the loss function, in this embodiment, the adjustment coefficient μ of the tool attitude change factor is set to 0.4. This value can also moderately reduce the aspect ratio penalty when the attitude difference is large. Furthermore, experimental verification shows that in the detection tasks of some small-sized tools (such as feeler gauges and dial indicators), μ=0.4 can achieve positioning accuracy comparable to μ=0.5, further demonstrating that the attitude adaptive loss function of this invention has good robustness to changes in the adjustment coefficient μ. Those skilled in the art can reasonably select a value of μ within the range of 0.3 to 0.7 according to the actual tool type and scenario characteristics.

[0062] Example 3 This embodiment is essentially the same as Embodiment 1, except for the specific implementation of K-Means clustering in the dynamic anchor frame generation algorithm. In Embodiment 1, the number of clusters k is 3, resulting in three sets of anchor frame sizes. In this embodiment, considering the significant differences in aspect ratios of some maintenance tools (such as box wrenches and torque wrenches), to further refine the adaptability of the anchor frames, two independent K-Means clustering operations are performed on the width and height of all labeled boxes in the dataset before training: the first clustering is performed uniformly on all tool categories to obtain global anchor frames; the second clustering is performed separately on slender tools with aspect ratios greater than 3 (such as wrenches and screwdrivers) to obtain anchor frames specifically for slender tools. During model training, for detection heads of different scales, appropriate anchor frame groups are selected based on the target scale range they are responsible for. This method further improves the matching degree between the anchor frames and the actual target size, especially significantly improving the detection effect for slender tools.

[0063] It should be noted that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art based on the technical solutions disclosed in the present invention should be included within the scope of protection of the present invention.

[0064] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0065] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, equipment, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0066] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0067] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program instructions, such as USB flash drives, portable hard drives, read-only storage servers, random access storage servers, magnetic disks, or optical disks.

[0068] Furthermore, it should be noted that the combination of the various technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0069] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

[0070] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying turbine maintenance tools based on an improved YOLOv8, characterized in that, Includes the following steps: Collect images of maintenance tools and process the data to form a maintenance tool dataset; A turbine maintenance tool identification model based on an improved YOLOv8 is constructed, the model comprising: Backbone networks are used to extract multi-scale basic features of images; The neck network adds a GAM attention module between each C2f module and the Detect module of the original neck network, and introduces a multi-scale feature fusion enhancement layer to further refine the fused features. The head network adds a new detection head that connects to the highest resolution feature map on the basis of the original three detection heads, and adds a lightweight super-resolution reconstruction EDSR module in front of the new detection head. At the same time, it adopts a dynamic anchor box generation algorithm based on the size distribution of K-Means clustering tool to replace the fixed anchor boxes. The loss function integrates EIoU loss, α-balanced Focal loss and BCE loss, and introduces a tool attitude change factor to dynamically adjust the aspect ratio constraint coefficient of EIoU loss; The improved YOLOv8 model is trained using the aforementioned maintenance tool dataset to generate a trained recognition model for recognizing turbine maintenance tools.

2. The identification method for turbine maintenance tools based on the improved YOLOv8 according to claim 1, characterized in that, The multi-scale feature fusion enhancement layer consists of a 1×1 convolutional layer, a 3×3 convolutional layer, and a lightweight channel attention (SE) module connected in sequence; the 1×1 convolutional layer is used for channel compression, the 3×3 convolutional layer is used for spatial feature enhancement, and the SE module is used for channel recalibration to focus on tool features.

3. The identification method for turbine maintenance tools based on the improved YOLOv8 according to claim 1, characterized in that, The lightweight super-resolution reconstruction EDSR module includes: a 3×3 convolutional layer for shallow feature extraction, multiple cascaded residual blocks for deep residual feature learning, a 3×3 convolutional layer for global feature fusion, and a sub-pixel convolutional layer for resolution enhancement; wherein, each residual block has a batch normalization layer removed.

4. The identification method for turbine maintenance tools based on the improved YOLOv8 according to claim 3, characterized in that, The number of the multiple cascaded residual blocks is 8. Each residual block is composed of a 3×3 convolutional layer, a ReLU activation function, and a 3×3 convolutional layer connected in series and then residually connected to the input.

5. The identification method for turbine maintenance tools based on the improved YOLOv8 according to claim 1, characterized in that, The dynamic anchor frame generation algorithm is as follows: before training, K-Means clustering is performed on the width and height of all labeled boxes in the dataset, with the number of clusters k=3, to obtain the anchor frame size representing the actual size distribution of the turbine maintenance tools, which replaces the preset fixed anchor frame size.

6. The identification method for turbine maintenance tools based on the improved YOLOv8 according to claim 1, characterized in that, The tool pose change factor δ_pose is defined as the absolute difference between the aspect ratio of the predicted bounding box and the aspect ratio of the true bounding box, and is expressed as: Where w and h are the predicted bounding box width and height, and w^gt and h^gt are the true bounding box width and height; the tool pose change factor is used to dynamically scale the coefficient of the aspect ratio penalty term V in the EIoU loss, and the scaling function is expressed as: Where μ is the adjustment coefficient.

7. The identification method for turbine maintenance tools based on the improved YOLOv8 according to claim 6, characterized in that, The EIoU loss function is expressed as follows: Where IoU is the intersection-union ratio, ρ²(b,b^gt) is the Euclidean distance between the center of the predicted box and the center of the ground truth box, c is the diagonal length of the minimum bounding rectangle of the two boxes, and V is the aspect ratio penalty term.

8. The identification method for turbine maintenance tools based on the improved YOLOv8 according to claim 1, characterized in that, The training process is configured with a batch size of 16, a total of 300 training rounds, an initial learning rate of 0.001, and a cosine annealing learning rate scheduling strategy. The Adam optimizer is selected, and Dropout layers with a probability of 0.1 are introduced in the C2f and GAM modules.

9. The identification method for turbine maintenance tools based on the improved YOLOv8 according to claim 1, characterized in that, The data processing includes data augmentation, data size normalization, and data annotation; the data augmentation includes horizontal flipping, vertical flipping, random rotation, random cropping, brightness adjustment, and contrast adjustment; the data size normalization is to uniformly scale all images to 640×640 pixels.

10. The identification method for turbine maintenance tools based on the improved YOLOv8 according to claim 1, characterized in that, The maintenance tool dataset is divided into training, validation and test sets in a ratio of 7:2:

1. Stratified sampling is used in the partitioning process to ensure that the proportion of each type of tool in each dataset is consistent with that in the original dataset.

Images