A method, device, equipment and medium for detecting internal cavity faults of wind turbine blades

Abstract

The present invention discloses a method, device, equipment and medium for detecting the inner cavity fault of a wind turbine blade, which is applied to the field of intelligent detection, including: obtaining the blade features corresponding to the inner cavity image of the blade to be detected, and taking the blade features as shallow features; extracting the deep features of the blade features through a reverse bottleneck module; the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, and the heterogeneous convolution module is a module for extracting features using a large parallel convolution kernel and a plurality of small convolution kernels; using a surface fusion module and a high-level fusion module, the shallow features and the deep features are fused to obtain fused features; the fused features are spatially enhanced and the deep features are semantically fused, the obtained target detection features are identified, and the target fault category is determined. Compared with current manual detection, the present invention retains more in-depth and comprehensive information based on the surface fusion module and the high-level fusion module, thereby improving the accuracy and efficiency of fault detection.

Description

Technical Field

[0001] The present invention relates to the field of intelligent detection technology, and in particular to a method, device, equipment and medium for detecting internal faults of blades of a wind turbine generator set. Background Art

[0002] During long-term operation, wind turbine blades are easily affected by factors such as wind loads and temperature changes, resulting in cracks, debonding, peeling and other damage. If they cannot be detected and repaired in time, it will seriously affect the safety and power generation efficiency of wind turbines. Due to the complex structure and narrow space of the blade cavity, traditional detection methods usually rely on manual inspections, which have technical problems such as low efficiency, high cost and poor reliability.

[0003] Therefore, how to improve the efficiency and accuracy of wind turbine blade inner cavity detection is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0004] In view of this, an object of the present invention is to provide a method, device, equipment and medium for detecting internal faults of wind turbine blades, which solves the technical problem of low efficiency and accuracy of internal fault detection of wind turbine blades in the prior art.

[0005] In order to solve the above technical problems, the present invention provides a method for detecting internal cavity faults of blades of a wind turbine generator set, comprising:

[0006] Acquire the leaf features corresponding to the inner cavity image of the leaf to be detected, and use the leaf features as shallow features;

[0007] The deep features of the leaf features are extracted by a reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, the heterogeneous convolution module is a module for extracting features using a large convolution kernel and a plurality of small convolution kernels in parallel, and the size of the large convolution kernel is larger than the size of the small convolution kernel;

[0008] The shallow features and the deep features are fused by using a surface fusion module and a high-level fusion module to obtain fused features; wherein the surface fusion module fuses the shallow features with the deep features when fusing, and the high-level fusion module performs weighted fusion of the shallow features with the deep features when fusing;

[0009] The spatial feature of the fused feature is enhanced, the enhanced fused feature is fused with the deep feature for semantic feature to obtain the target detection feature, and the target detection feature is identified to determine the target fault category.

[0010] Optionally, the surface fusion module is a module designed based on a bidirectional link mechanism, a feature enhancement strategy and a preliminary feature fusion method, and the preliminary feature fusion method is a method for fusing shallow features with deep features; the high-level fusion module is a module designed based on multi-directional links, multi-resolution feature output and a weighted feature fusion method, and the weighted feature fusion method is a method for integrating the shallow features with the deep features layer by layer through a weighted feature fusion strategy.

[0011] Optionally, the fused features are spatially enhanced, the enhanced fused features are semantically fused with the deep features to obtain target detection features, and the target detection features are identified to determine the target fault category, including:

[0012] Get the lightweight head layer based on the separation convolution and pruning method;

[0013] Based on the multi-level feature aggregation of the lightweight head layer, the spatial feature of the fused feature is enhanced, and the enhanced fused feature is semantically fused with the deep feature to obtain the target detection feature;

[0014] The multi-scale detection head based on the lightweight head layer detects the target detection features to obtain the target fault category.

[0015] Optionally, identifying the target detection feature to determine the target fault category includes:

[0016] The target detection features are identified to determine the target fault category and blade fault location.

[0017] Optionally, after identifying the target detection feature and determining the target fault category and the blade fault position, the method further includes:

[0018] determining a fault level based on the target fault category and the blade fault location;

[0019] A processing priority corresponding to each fault is determined based on the fault level, so that the fault is processed based on the processing priority.

[0020] Optionally, before obtaining the leaf features corresponding to the inner cavity image of the leaf to be detected and using the leaf features as shallow features, the method further includes:

[0021] Acquire a deep neural network including the reverse bottleneck module, the surface fusion module and the high-level fusion module;

[0022] The deep neural network is trained based on the large data set fault training data to obtain a pre-trained blade fault detection model; wherein the large data set fault training data is a data set whose amount of fault training data is greater than the minimum amount of training data;

[0023] The wind turbine blade cavity fault training data is obtained to train the pre-trained blade fault detection model to obtain a target blade fault detection model, so as to perform the above process of determining the target fault category based on the target blade fault detection model.

[0024] Optionally, before obtaining the leaf features corresponding to the inner cavity image of the leaf to be detected and using the leaf features as shallow features, the method further includes:

[0025] The image of the inner cavity of the blade to be inspected is acquired in real time based on a wind turbine blade inner cavity inspection robot; wherein the wind turbine blade inner cavity inspection robot is a robot including wheeled crawler automatic adjustment function, active obstacle avoidance function and adaptive path planning function.

[0026] The present invention also provides a wind turbine blade inner cavity fault detection device, comprising:

[0027] A first feature acquisition module is used to acquire the leaf features corresponding to the inner cavity image of the leaf to be detected, and use the leaf features as shallow features;

[0028] A second feature acquisition module is used to extract deep features of the leaf features through a reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, the heterogeneous convolution module is a module that uses a parallel large convolution kernel and a plurality of small convolution kernels to extract features, and the size of the large convolution kernel is larger than the size of the small convolution kernel;

[0029] A fusion module, used to fuse the shallow features and the deep features using a surface fusion module and a high-level fusion module to obtain fusion features; wherein the surface fusion module fuses the shallow features with the deep features when fusing, and the high-level fusion module performs weighted fusion of the shallow features with the deep features when fusing;

[0030] The fault detection module is used to perform spatial feature enhancement on the fused features, perform semantic feature fusion on the enhanced fused features and the deep features to obtain target detection features, and identify the target detection features to determine the target fault category.

[0031] The present invention also provides a wind turbine blade inner cavity fault detection device, comprising:

[0032] Memory for storing computer programs;

[0033] The processor is used to execute the computer program to implement the steps of the above-mentioned method for detecting internal cavity faults of wind turbine blades.

[0034] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned method for detecting internal cavity faults of blades of a wind turbine are implemented.

[0035] The present invention also provides a computer program product, comprising a computer program / instruction, wherein the computer program / instruction, when executed by a processor, implements the steps of the above-mentioned method for detecting internal cavity faults of blades of a wind turbine generator set.

[0036] It can be seen that the present invention obtains the blade features corresponding to the inner cavity image of the blade to be detected, and uses the blade features as shallow features; extracts the deep features of the blade features through the reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, and the heterogeneous convolution module is a module that uses parallel large convolution kernels and multiple small convolution kernels to extract features, and the size of the large convolution kernel is larger than the size of the small convolution kernel; uses a surface fusion module and a high-level fusion module to fuse the shallow features and the deep features to obtain fused features; wherein the surface fusion module fuses the shallow features with the deep features when fusing, and the high-level fusion module performs weighted fusion of the shallow features with the deep features when fusing; performs spatial feature enhancement on the fused features, performs semantic feature fusion on the enhanced fused features and the deep features to obtain target detection features, and identifies the target detection features to determine the target fault category.

[0037] The beneficial effect of the present invention is that compared with the current manual detection that results in low efficiency and low accuracy, the present application obtains deep features and shallow features by dividing the blade features into two tributaries. The deep features are obtained based on dynamic convolution kernels, thereby greatly expanding the perception field of the network, and the shallow features are effectively retained by the surface fusion module, while the high-level fusion module promotes the output layer to retain diverse multi-scale information through enhanced information fusion, thereby improving the depth and comprehensiveness of feature extraction. Therefore, the intelligent wind turbine blade cavity fault detection method provided by the present application can improve the accuracy and efficiency of fault detection.

[0038] In addition, the present invention also provides a wind turbine blade cavity fault detection device, equipment and medium, which also have the above-mentioned beneficial effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0040] Figure 1 A flow chart of a method for detecting internal faults of blades of a wind turbine generator set provided by an embodiment of the present invention;

[0041] Figure 2 An example flow chart of a method for detecting internal cavity faults of a blade of a wind turbine generator system provided by an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of a target blade fault detection model architecture provided by an embodiment of the present invention;

[0043] Figure 4 A schematic structural diagram of a wind turbine blade inner cavity fault detection device provided by an embodiment of the present invention;

[0044] Figure 5 A schematic structural diagram of a wind turbine blade inner cavity fault detection device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Please refer to Figure 1 , Figure 1 A flow chart of a method for detecting a fault in a blade cavity of a wind turbine generator set provided by an embodiment of the present invention. The method may include:

[0047] S101, obtaining blade features corresponding to the inner cavity image of the blade to be detected, and using the blade features as shallow features.

[0048] The executor of this embodiment is an electronic device. The electronic device in this embodiment can be a computer, a mobile phone, etc. This embodiment does not limit the specific method of acquiring the inner cavity image of the blade to be detected. For example, this embodiment can be obtained by a patrol robot, or this embodiment can also acquire the inner cavity image of the blade to be detected by a sensor. The backbone network is the basis of the entire detection network, which is responsible for extracting the basic features of the image and directly affects the performance and efficiency of subsequent detection tasks. In this embodiment, the backbone network is used to divide the blade features into two tributaries. The first tributary uses the blade features as shallow features, retains the original features, and enhances the transmission of feature information. The second tributary is used to extract deep features.

[0049] It should be further explained that, based on any of the above embodiments, in order to improve the efficiency of model training, before obtaining the blade features corresponding to the inner cavity image of the blade to be detected and using the blade features as shallow features, it can also include: obtaining a deep neural network including a reverse bottleneck module, a surface fusion module and a high-level fusion module; training the deep neural network based on the large data set fault training data to obtain a pre-trained blade fault detection model; wherein the large data set fault training data is a data set with a fault training data volume greater than the minimum training data volume; obtaining the wind turbine blade inner cavity fault training data to train the pre-trained blade fault detection model to obtain a target blade fault detection model including a trained reverse bottleneck module, a surface fusion module and a high-level fusion module, so as to perform the above process of determining the target fault category based on the reverse bottleneck module, the surface fusion module and the high-level fusion module in the target blade fault detection model. The large data set fault training data in this embodiment includes other machine inner cavity fault data. In this embodiment, when performing model training, a staged training strategy is adopted, first pre-training the backbone network, and then jointly optimizing the bottleneck layer and the detection head (head layer). At the same time, a transfer learning method is used to transfer the knowledge of the pre-trained model to a specific scenario, effectively improving the detection accuracy.

[0050] S102, extracting deep features of leaf features through a reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, the heterogeneous convolution module is a module that uses a large parallel convolution kernel and multiple small convolution kernels for feature extraction, and the size of the large convolution kernel is larger than the size of the small convolution kernel.

[0051] In this embodiment, the reverse bottleneck module of the backbone network is used to extract the deep features of the leaf features. The specific structure of the reverse bottleneck is to first adjust the dimension through 1x1 convolution, and then use the heterogeneous convolution module and dynamic convolution kernel selection to set the size of the convolution kernel. Heterogeneous convolution module: In each layer of feature extraction, a large convolution kernel (such as 9×9) and multiple small convolution kernels (such as 3×3, 5×5, 7×7) are used in parallel. This design not only expands the perception range, but also captures the detailed information of small targets, while avoiding the high computational cost brought by traditional large convolution. Dynamic convolution kernel selection: Combined with the global heterogeneous kernel selection mechanism, the convolution kernel size is dynamically adjusted according to the resolution of the feature layer. For example, a larger convolution kernel is selected in the low-resolution layer, and a smaller convolution kernel is used in the high-resolution layer. This mechanism effectively expands the receptive field of the network and adapts it to the needs of multi-scale target detection. It is understandable that although larger convolution kernels can improve performance by encoding a wider area, they may inadvertently mask details related to small targets, leaving room for further improvement. Therefore, the present invention transfers the heterogeneous idea of ​​the global architecture into a single convolution, and combines the idea of ​​reparameterization to realize reparameterized non-uniformly correlated convolution. Specifically, the model supplements the detection of small targets by running large and small convolution kernels simultaneously. Convolution kernels of different sizes enhance the diverse representation of features. The reverse bottleneck shows a certain difference between training and reasoning. During training, the network runs n parallel deep convolution operations of different sizes, while during reasoning, these convolutions are merged into one, which does not lead to a decrease in reasoning speed. The output features of the multi-scale feature processing backbone network are multi-scale and can process large and small targets at the same time. The present invention introduces a multi-scale fusion module in the network structure to combine deep features with shallow layers, thereby enhancing adaptability to complex scenes.

[0052] It should be further explained that, based on any of the above embodiments, in order to improve the accuracy of acquiring the inner cavity image of the blade to be detected, before acquiring the blade features corresponding to the inner cavity image of the blade to be detected and using the blade features as shallow features, it may also include: acquiring the inner cavity image of the blade to be detected in real time based on the wind turbine blade inner cavity inspection robot; wherein the wind turbine blade inner cavity inspection robot is a robot including wheeled crawler automatic adjustment function, active obstacle avoidance function and adaptive path planning function. The specific design of the wind turbine blade inner cavity inspection robot in this embodiment includes: lightweight and modular design. The core concept of the inspection robot design is lightweight and modular to adapt to the complex inner cavity environment of wind turbine blades. The design focuses on the following aspects: flexible movement mechanism; (1) multi-axis wheeled crawler system: the robot is equipped with a multi-axis wheeled crawler system, and the wheeled crawler can be adaptively adjusted according to the different materials and inclination angles of the inner cavity surface, thereby improving the overall flexibility of movement. The crawler adopts high-strength and lightweight materials to ensure the service life in a high-wear environment. (2) Active obstacle avoidance: By integrating lidar and ultrasonic sensors, the robot can scan the surrounding environment in real time, build a three-dimensional environmental model, and automatically plan a path to avoid obstacles. (3) Adaptive path planning algorithm: Path planning uses The algorithm is combined with the deep learning algorithm to optimize the inspection path in real time, reducing repeated movements and time waste. Modular structure: Camera module: supports the rapid replacement of different types of cameras (such as wide-angle cameras and macro cameras) to meet the needs of different tasks. Light source module: The LED light source module has a variety of lighting modes (such as strong light and soft light). Through the plug-in design, it is convenient to upgrade in different tasks. Sensor module: Integrates a variety of environmental sensors (such as temperature and humidity sensors, vibration sensors) to support the expansion of new functions. Data transmission module: The wireless module supports a variety of communication protocols (such as Wi-Fi and 5G) to ensure the stability and security of real-time data transmission. High-resolution imaging and intelligent light source adjustment; wide-angle and high-definition imaging: (1) Equipped with a 5-megapixel wide-angle camera, it supports 4K resolution video recording and can capture the panoramic view and local close-up of the inner cavity of the wind turbine blade. (2) The dynamic scene optimization algorithm automatically adjusts the camera parameters, such as exposure, focus and white balance, to ensure clear imaging under different lighting conditions. Intelligent light source system: (1) Automatic light source adjustment: The light source combines the light sensor data to adjust the brightness and color temperature in real time to ensure clear shooting in low light or high reflective environment. (2) Energy saving mode: The light source module is designed with an energy saving algorithm to dynamically control power consumption according to the ambient light intensity to extend the robot's battery life. (3) Anti-vibration and data collection. Anti-vibration design: (1) Double-layer shock-absorbing structure: The combination of elastic materials and mechanical stabilizers significantly reduces the vibration of the robot when it moves. (2) Gyro stabilization technology: By real-time detection of the robot's angular velocity and acceleration, the camera's posture is adjusted to reduce image blur caused by vibration. Edge computing capabilities: (1) Real-time data preprocessing: The integrated high-performance edge processor can process the collected data in real time, including image denoising, contrast enhancement, etc. (2) Optimize the amount of transmitted data: Compress video data and extract key frames to reduce redundant information transmission and improve transmission efficiency.

[0053] S103, using the surface fusion module and the high-level fusion module to fuse the shallow features and the deep features to obtain fused features; wherein the surface fusion module fuses the shallow features with the deep features when fusing, and the high-level fusion module performs weighted fusion of the shallow features with the deep features when fusing.

[0054] The surface fusion module and the high-level fusion module in this embodiment belong to the bottleneck layer. In this embodiment, the bottleneck layer is a key component connecting the backbone network and the detection head, and its design directly determines the network's ability to detect multi-scale targets. It is understandable that the goal of the surface fusion module is to enhance the expression ability of shallow features and improve the network's detection performance for small targets. Retaining shallow spatial information in the backbone is crucial to enhancing the detection ability of smaller targets. However, the information provided by the backbone is relatively elementary and easily disturbed. Therefore, we incorporate shallow information into a deeper network as an auxiliary branch to ensure the stability of subsequent layer learning. The goal of the high-level fusion module is to integrate cross-layer features (deep features, shallow features) to enhance the global expression ability of the network and its adaptability to multi-scale targets. In order to further enhance the interactive use of feature layer information, a high-level fusion module is used in the deeper level of the bottleneck layer for multi-scale information integration. When this embodiment weightedly fuses shallow features with deep features, the weights of deep features and shallow features can be set according to demand.

[0055] It should be further explained that, based on any of the above embodiments, in order to improve the accuracy of the design of the surface fusion module and the high-level fusion module, the surface fusion module is a module designed based on a bidirectional link mechanism, a feature enhancement strategy and a preliminary feature fusion method, and the preliminary feature fusion method is a method for fusing shallow features with deep features; the high-level fusion module is a module designed based on multi-directional links, multi-resolution feature outputs and a weighted feature fusion method, and the weighted feature fusion method is a method for integrating shallow features with deep features layer by layer through a weighted feature fusion strategy. It can be understood that the specific design of the surface fusion module is: Bidirectional link mechanism: retain shallow backbone network information, and at the same time enable shallow features to share information with the bottleneck layer through a bidirectional link structure. Feature enhancement strategy: Strengthen shallow features through lightweight convolution operations (such as depth-separable convolution) so that they have stronger discrimination while retaining the original information. Preliminary feature fusion: Preliminary fusion of shallow features with deep features of the backbone network to generate richer feature representations. Specifically, the main goal of the surface fusion module is to combine deep information with features of the same level and high-resolution shallow layers in the backbone, aiming to preserve rich local details to enhance the spatial representation of the network. In addition, we use 1×1 convolution to control the number of channels in the shallow information, ensuring that it occupies a smaller proportion in the splicing operation without affecting subsequent learning. The specific design of the high-level fusion module is: Multi-directional link: Through a dense connection mechanism, the feature maps of each layer are collected from bottom to top and fused into high-level features. This design can effectively alleviate the gradient vanishing problem and promote feature sharing. Multi-resolution feature output: In the high-level fusion module, a multi-resolution output strategy is adopted to enable the detection head to receive information from different scales, thereby improving the robustness of detection. Feature fusion method: Through a weighted feature fusion strategy, multi-scale features are integrated layer by layer to avoid information loss or excessive redundancy. Weighted feature fusion strategy: Through empirical assignment, it is generally 0.7 and 0.3. Specifically, the high-level fusion module is specifically embodied in the aggregation and fusion of information from multiple layers. The high-level fusion module comprehensively improves the expressiveness of features by integrating the rich detail information provided by the shallow high-resolution layer, the wide-area semantic information transmitted by the shallow low-resolution layer, the features of the current layer, and the features of the previous layer. In order to reasonably control the impact of features at different levels on the final result, the high-level fusion module uses 1×1 convolution to adjust the channels and dynamically allocate the contribution of features at each layer. In addition, experiments show that setting the number of shallow channels to half the number of deep channels in the surface fusion module will lead to a slight decrease in performance. To this end, the number of channels in each layer is balanced in the bottleneck layer (balanced through the fully connected layer) to ensure that the model can balance the feature outputs of different layers while embedding the initial guidance information in the shallow features. Finally, the overall performance of the model is significantly improved through the high-level fusion module.The surface fusion module and the high-level fusion module provided in this embodiment make it possible to enrich the detail features based on the surface fusion module and to enhance the global expression capability of the network and its adaptability to multi-scale targets based on the high-level fusion module.

[0056] S104, performing spatial feature enhancement on the fused features, performing semantic feature fusion on the enhanced fused features and the deep features to obtain target detection features, and identifying the target detection features to determine the target fault category.

[0057] In this embodiment, the head layer is an output module responsible for predicting the bounding box of the target and its category. The present invention introduces a multi-level fusion strategy in the head layer design to further improve the detection performance. The spatial feature enhancement of this embodiment refers to the spatial feature enhancement of the features input to the bottleneck layer, such as introducing a self-attention mechanism to improve the adaptability to complex scenes; the semantic feature fusion in this embodiment can combine the features of the head layer with the deep features of the backbone network through a context fusion module to further improve the semantic expression ability.

[0058] It should be further explained that, based on any of the above embodiments, in order to reduce the computing cost and improve the reasoning speed, the above-mentioned spatial feature enhancement of the fused feature, the semantic feature fusion of the enhanced fused feature and the deep feature to obtain the target detection feature, and the target detection feature is identified to determine the target fault category, which may include:

[0059] S1041, obtaining a lightweight head layer based on a separation convolution and pruning method;

[0060] S1042, based on the multi-level feature aggregation of the lightweight head layer, the fused feature is spatially enhanced, and the enhanced fused feature is semantically fused with the deep feature to obtain the target detection feature;

[0061] S1043, detecting the target detection features based on the multi-scale detection head of the lightweight head layer to obtain the target fault category.

[0062] This embodiment performs lightweight processing on the head layer based on the separation convolution and pruning method, so that the network maintains high reasoning efficiency while improving performance.

[0063] It should be further explained that, based on any of the above embodiments, in order to improve the accuracy of fault location positioning, the above identification of target detection features and determination of target fault categories may include: identifying target detection features, determining target fault categories and blade fault locations. This embodiment can mark the specific coordinates of the damaged area by integrating the image segmentation and positioning results. This embodiment can analyze multiple types such as cracks, corrosion, and shedding in combination with a predefined damage classification system. It is understandable that this embodiment can mark the location of the fault when training the model, so that when performing subsequent detection, not only the target fault category can be determined but also the fault location can be determined. The fault level in this embodiment can be determined based on the scoring results, and the scoring results are used to dynamically adjust the inspection plan to ensure that high-risk damage is handled first. For example, for equipment with a score of "high risk", an emergency maintenance task can be triggered immediately to avoid greater economic losses or safety hazards caused by delays.

[0064] It should be further explained that, based on any of the above embodiments, in order to improve the efficiency of fault processing, after identifying the target detection features and determining the target fault category and blade fault location, it can also include: determining the fault level based on the target fault category and blade fault location; determining the processing priority corresponding to each fault based on the fault level, so as to process the fault based on the processing priority. For example, the fault level at the web position of the blade is the highest, the fault level at the leading and trailing edge areas of the blade is the second highest, and the fault level at the main beam position of the blade is the lowest; and for the target fault category, the fault level of cracking is the highest, the fault level of bulging is the second highest, and the fault level of whitening of the inner cavity is the lowest.

[0065] A method for detecting a fault in a blade cavity of a wind turbine provided by an embodiment of the present invention may include: S101, obtaining blade features corresponding to a blade cavity image to be detected, and using the blade features as shallow features; S102, extracting deep features of the blade features through a reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, and the heterogeneous convolution module is a module for extracting features using a large convolution kernel and a plurality of small convolution kernels in parallel, and the size of the large convolution kernel is larger than the size of the small convolution kernel; S103, using a surface fusion module and a high-level fusion module to fuse shallow features and deep features to obtain fused features; wherein the surface fusion module fuses shallow features with deep features when fusing, and the high-level fusion module performs weighted fusion of shallow features with deep features when fusing; S104, performing spatial feature enhancement on the fused features, performing semantic feature fusion on the enhanced fused features and the deep features to obtain target detection features, and identifying the target detection features to determine the target fault category. Compared with the current manual detection that results in low efficiency and accuracy in blade cavity fault detection, the present application divides the blade features into two tributaries to obtain deep features and shallow features. The deep features are features obtained based on dynamic convolution kernels, thereby greatly expanding the perception field of the network, and are based on the shallow features effectively retained by the surface fusion module, while the high-level fusion module promotes the output layer to retain diverse multi-scale information through enhanced information fusion, thereby improving the depth and comprehensiveness of feature extraction. Therefore, the intelligent wind turbine blade cavity fault detection method provided by the present application can improve the accuracy and efficiency of fault detection.

[0066] The existing wind turbine blade cavity damage detection technology has great limitations in efficiency and accuracy, especially in the face of complex environments and highly concealed cavity structures, traditional detection methods are difficult to meet the requirements of automation and real-time. To address this problem, the present invention proposes a blade cavity image fault detection method for the first time.

[0067] The present invention aims to solve the following key technical problems: obtain real-time video data of the blade cavity through a patrol robot, upload it to the cloud, and use the target blade fault detection model of a deep neural network to automatically identify and classify damage defects in the video. In view of the complex background characteristics of the blade cavity and the diverse defect types, the target blade fault detection model is designed and optimized to improve the accuracy, efficiency and robustness of the detection. Ultimately, the detection results are automatically fed back to the operation and maintenance management system, and a work order is generated to support the efficient maintenance and management of wind turbines, thereby improving the automation and reliability of blade damage detection.

[0068] In order to make the present invention easier to understand, please refer to Figure 2 , Figure 2A flow chart of a method for detecting a fault in a blade cavity of a wind turbine generator set provided by an embodiment of the present invention may specifically include:

[0069] S201, using a fan blade inner cavity inspection robot to collect an inner cavity image of a blade to be inspected.

[0070] This embodiment can adopt a distributed architecture. In order to meet the needs of massive detection tasks, the cloud platform has designed an efficient distributed computing architecture. Through task slicing and dynamic scheduling mechanisms, computing tasks are assigned to different computing nodes. Each node is equipped with a high-performance processor and a large-capacity memory, which can process multiple tasks in parallel, ensuring that the detection efficiency increases linearly with the increase of computing resources. In addition, high-speed network interconnection is adopted between nodes, combined with data caching technology, to further reduce the delay of data transmission. During the calculation process, the platform applies an adaptive load balancing algorithm. The algorithm dynamically adjusts the task allocation according to the real-time resource utilization rate of the node to avoid overloading or idleness of some nodes and maximize resource utilization. At the same time, a fault-tolerant mechanism is introduced during task execution. Even if some nodes fail, the stability and reliability of the system can be ensured by task redistribution. In response to the common large-scale data processing needs in the industrial field, the platform adopts image slicing and batch processing technology to divide ultra-high-resolution images into multiple small fragments of fixed size, which are processed separately and then integrated into a complete result. This method effectively reduces the memory usage of a single processing task and improves the parallel processing capability. In addition, through batch operations, the time for data loading and preprocessing can be significantly shortened, thereby optimizing the efficiency of overall data processing. The platform also supports a variety of data acceleration solutions, such as GPU (graphics processing unit) parallel computing and TPU (tensor processing unit) optimization computing, to meet the real-time processing needs in more complex scenarios. For example, when processing drone inspection images, the platform can use deep learning algorithms to quickly identify feature points in the image, combined with cloud distributed storage to achieve efficient access, and significantly reduce data processing latency. A dedicated wind turbine blade cavity inspection robot is used in the inspection process. The robot is equipped with a high-definition camera and light source system, which can move freely in the narrow blade cavity and collect high-definition images. Image acquisition standard: One frame of image is collected every 10 cm to ensure that the entire cavity surface is covered, while increasing the image acquisition density in complex areas. Environmental adaptability: The robot has a built-in automatic light source adjustment function to ensure that high-quality images can be obtained even in a cavity environment with insufficient light.

[0071] It should be further explained that the collected images of the inner cavity of the blade to be inspected can be transmitted to the cloud processing platform in real time through the 5G module. In order to ensure the efficiency and stability of data transmission, compression and encryption can be performed: lossless compression of the image before transmission, and AES (Advanced Encryption Standard) encryption technology is used to ensure data security. Transmission log: Each image is accompanied by location information and acquisition timestamp, which is convenient for subsequent positioning of the damage location. The original image transmitted to the cloud may have inconsistent resolution or distortion problems, so the following adjustments need to be made: Image scaling (Resize): The image is uniformly adjusted to the model input size (such as 512×512 pixels) to ensure compatibility in the inference stage. Image cropping: Remove irrelevant parts of the image edge and retain information in the key area (blade inner cavity surface). In order to improve the detection performance of the model, the image is enhanced, including: contrast stretching: adjust the image contrast to make the damaged area more prominent; CLAHE equalization: apply adaptive histogram equalization to optimize the texture details of the blade surface; gamma correction: enhance dark details while preventing overexposure. Image rotation, flipping, scaling, mosaic, brightness adjustment, etc. The augmented images are saved in a standardized format (such as PNG) and fed into the model inference phase.

[0072] S202, uploading the inner cavity image of the wind turbine blade to be detected to the target blade fault detection model based on wireless transmission technology.

[0073] The design of the wireless transmission technology in this embodiment may include: First: wireless data transmission; stability and security; (1) Adopt 5G high-speed wireless transmission technology to support low-latency and high-bandwidth data communication to ensure real-time data transmission. (2) Use end-to-end encryption protocols (such as TLS1.3) to prevent data from being intercepted or tampered with during transmission. Adaptive bit rate transmission: Dynamically adjust the bit rate of video transmission according to the network environment to ensure that key data can still be stably transmitted in areas with weak signals. Second: Edge preprocessing; (1) Data compression: Use H.265 encoding to compress the video in real time. Compared with traditional H.264, the compression efficiency is improved by about 30%, which greatly reduces the bandwidth requirement for data transmission. (2) Image enhancement: Integrate image processing algorithms, including filtering, CLAHE (adaptive histogram equalization) and gamma correction, to improve image quality and lay the foundation for subsequent target detection.

[0074] S203, using the backbone network of the target blade fault detection model to obtain blade features corresponding to the inner cavity image of the blade to be detected, using the blade features as shallow features, and extracting deep features of the blade features through a reverse bottleneck module.

[0075] Target detection is an important task in the field of computer vision, which aims to identify the category and precise location of targets in images or videos. In order to solve the problems of insufficient detection ability of traditional algorithms for small targets, limited feature expression ability and low reasoning efficiency, the present invention proposes an improved target detection algorithm. The improved target detection algorithm in this embodiment is a target blade fault detection model. The architecture diagram of the target blade fault detection model is shown in the figure. Figure 3 As shown, Figure 3 A schematic diagram of a target blade fault detection model architecture provided by an embodiment of the present invention, the target blade fault detection model mainly includes a feature extraction layer based on a backbone network; a bottleneck layer based on a surface fusion module (surface fusion module) and a high-level fusion module, and a head layer based on a detection head. This embodiment aims at the limitations of the traditional YOLO (You Only Look Once, an efficient target detection algorithm) method. The present invention proposes a network based on adaptive convolution kernel settings to change the receptive field, which includes two key components: a surface fusion module and a high-level fusion module. The surface fusion module is used to effectively retain the shallow information in the backbone, and the high-level fusion module promotes the output layer to retain diverse multi-scale information through enhanced information fusion. In addition, we integrate adaptive dynamic convolution kernel adjustment into the network, which can dynamically expand the convolution kernels in the entire architecture to greatly expand the perception field of the network. In addition, a new bottleneck layer design is introduced, which greatly enhances the multi-scale characterization capability by using re-parameterized heterogeneous convolutions and then parallel feature aggregation. Therefore, the network of the present invention shows excellent overall performance while maintaining a considerable number of parameters.

[0076] S204, using the surface fusion module and the high-level fusion module of the bottleneck layer of the target blade fault detection model, the shallow features and the deep features are fused to obtain fused features.

[0077] The goal of the surface fusion module in this embodiment is to enhance the expressiveness of shallow features and improve the network's detection performance for small targets. The goal of the high-level fusion module in this embodiment is to integrate cross-layer features (deep features, shallow features) to enhance the network's global expressiveness and adaptability to multi-scale targets.

[0078] S205, using the head layer of the target blade fault detection model to perform spatial feature enhancement on the fused features, performing semantic feature fusion on the enhanced fused features and the deep features to obtain target detection features, and identifying the target detection features to determine the target fault category and target fault location.

[0079] This embodiment performs multi-level feature aggregation on the output of the detection head in the head layer, aiming to enhance the model's detection ability for the target by fusing feature information at different levels. This embodiment can be achieved by performing spatial feature enhancement on the feature map of the detection head and performing semantic feature fusion. In order to reduce the computational cost, some lightweight designs such as deep separable convolution and pruning technology can be introduced in the head layer, so that the network can maintain a high reasoning efficiency while improving performance. It should be noted that in order to further improve the network performance of the target blade fault detection model, the present invention performs global optimization on the entire deep neural network, mainly including the following aspects: 1. Improvement of loss function. Traditional target detection networks usually use a fixed loss function and cannot be adaptively optimized for specific scenarios. The present invention introduces a dynamic loss weighting mechanism to adaptively adjust the loss weights of classification and regression according to the difficulty of the sample, and enhance the learning ability of difficult samples. 2. Data enhancement strategy. In order to improve the generalization ability of the network, the present invention designs a set of diversified data enhancement strategies for target detection, including random cropping, multi-scale transformation, color enhancement, etc. While retaining the target information, the diversity of training samples is increased. 3. Training strategy optimization. A phased training strategy is adopted to pre-train the backbone network first, and then jointly optimize the bottleneck layer and the detection head. At the same time, the transfer learning method is used to transfer the knowledge of the pre-trained model to specific scenarios, effectively improving the detection accuracy.

[0080] S206, determining a risk level based on the target fault category and the target fault location, and processing the fault based on the risk level of each fault.

[0081] In industrial equipment inspections, the risk levels of different damage types vary greatly. Traditional manual decision-making may lead to waste of resources or omission of hidden dangers due to subjective judgment. To solve this problem, the platform introduces a dynamic scoring model that generates a scientific risk level for each inspection by comprehensively analyzing the damage type, severity, and historical data. The core factors of the scoring model are: (1) Damage type: For example, cracks usually have a higher priority than surface stains; (2) Severity: Grading based on quantitative indicators such as damage area, depth, and expansion trend; (3) Historical data: Combined with the historical damage records of the equipment or area, predict its future risk level. The scoring results are used to dynamically adjust the inspection plan to ensure that high-risk damage is handled first. For example, for equipment with a "high level" score, an emergency maintenance task can be triggered immediately to avoid greater economic losses or safety hazards caused by delays. High-risk damage immediately generates a maintenance work order and pushes it to the mobile device of the relevant maintenance personnel. The work order content includes: Damage details: damage type, location, and severity score; Maintenance suggestions: recommended maintenance methods and required tools. At the same time, the detection data and work order execution status can be synchronized to the cloud operation and maintenance platform in real time to form a closed-loop management. The accumulation of long-term data supports the following analyses: (1) Trend prediction: Analyze the expansion trend of blade damage and optimize maintenance plans; (2) Health scoring: Score the overall operating status of the wind turbine to assist enterprises in formulating equipment management strategies; (3) Group analysis: Cluster analysis of the health status of wind turbines in different regions to identify potential common problems.

[0082] S207: Generate a fault detection report based on the target fault category and the target fault location.

[0083] The fault detection report content of this embodiment may include: damage location: by integrating the image segmentation and positioning results, the specific coordinates of the damaged area are marked; damage type: combined with the pre-defined damage classification system, multiple types such as cracks, corrosion, and shedding are analyzed; damage severity: the damage area, depth and other indicators are quantified through the scoring model, and the severity classification is provided. At the same time, the platform supports a variety of visual presentation methods, such as intuitively displaying the detection results through heat maps, distribution maps, etc., so that users can quickly understand the equipment status. In order to adapt to the needs of different users, the automated report supports the generation of multiple formats (such as PDF, Excel, JSON, etc.). Users can not only download reports, but also push the detection results to other business systems through the API interface to achieve efficient information interaction. For example, enterprises can directly integrate reports into their asset management platform to automatically generate relevant maintenance plans or risk assessments. In addition, the platform supports custom template functions, and users can adjust the report content and format according to their own needs, including adding corporate logos, adjusting field order, etc., to provide users with personalized service experience. The report can be presented in the following forms: Visual charts: Generate heat maps or distribution maps to intuitively display the damage distribution; File export: Support PDF and Excel formats for easy archiving and sharing; API push: Push the report content to the company's asset management system to achieve seamless information connection.

[0084] It should be further explained that the entire data acquisition and detection process can be concentrated on one platform, which synchronizes the detection data and work order information to the operation and maintenance data management platform in real time through the cloud interface. The platform not only records the detailed data of each detection, but also forms a closed-loop feedback based on the execution of the work order. For example, users can view all the detection records and maintenance logs of a certain device in the past year to understand its overall operating status. Based on the long-term accumulated historical data, the platform supports multi-dimensional trend analysis and prediction: (1) Damage development trend: Through the time series analysis model, the expansion speed and possible impact range of equipment damage are predicted; (2) Equipment health score: Combined with all inspection records and maintenance history, the equipment health status is comprehensively scored to assist enterprises in formulating more reasonable asset management strategies; (3) Inspection strategy optimization: According to the historical inspection and maintenance results, the inspection frequency and path are adjusted to improve the inspection efficiency and accuracy. The platform also supports the analysis of the group health status of equipment, such as clustering by dimensions such as equipment type and installation area, to find high-risk equipment under specific conditions and provide scientific risk warnings for enterprises.

[0085] Beneficial effects of the technical solution of the present invention:

[0086] Compared with the prior art, the present invention has significant advantages and positive effects in detecting damage in the inner cavity of wind turbine blades, which are specifically reflected in the following aspects:

[0087] (1) High degree of automation, improving detection efficiency:

[0088] The present invention realizes the automatic inspection and damage detection of the blade cavity through the combination of inspection robots and cloud-based target detection algorithms, completely getting rid of the reliance on traditional manual inspections. Compared with time-consuming and labor-intensive manual inspections, the automated solution significantly improves the detection efficiency, can greatly shorten the inspection time, and provides strong support for the daily maintenance of wind turbines.

[0089] (2) High precision and high robustness:

[0090] The improved target detection algorithm designed by the present invention combines the multi-scale feature extraction and attention mechanism in deep learning, which can not only cope with the complex background features of the blade cavity, but also accurately identify various defect types such as cracks, delamination, fatigue damage, etc. Experimental data show that the defect detection accuracy of the algorithm of the present invention in complex scenes can reach more than 95%, which is significantly improved compared with the traditional rule-based method.

[0091] (3) Real-time and intelligent feedback:

[0092] The video data collected by the inspection robot can be uploaded to the cloud for processing in real time. The cloud computing platform completes defect detection and classification in a short time and directly feeds back the results to the operation and maintenance management system. The automatically generated work orders further optimize the maintenance process, making the system have intelligent closed-loop management capabilities. This real-time and intelligent feature significantly improves the operation and maintenance efficiency of wind turbines.

[0093] (4) Economical and scalable:

[0094] Compared with the traditional method that relies on expensive high-precision sensors or complex hardware, the present invention mainly relies on software algorithms and general hardware equipment, which reduces equipment and operation costs. At the same time, the system has good scalability and can adapt to different types of inspection robots and wind turbines to provide support for diverse detection needs.

[0095] (5) Functional differences from existing technologies:

[0096] Unlike the existing technology that focuses on external damage detection, the present invention focuses on the automated detection of blade cavity damage, filling the gap in the existing technology. Traditional cavity detection relies on manual inspection, which has problems such as strong subjectivity and low detection efficiency. However, the present invention, through the introduction of deep learning algorithms, not only realizes the automation of the detection process, but also greatly improves the objectivity and reliability of the detection results. In addition, the present invention also integrates the functions of inspection and operation and maintenance management to form a complete detection-feedback closed loop, which significantly enhances the practicality of the system.

[0097] (6) Experimental data support:

[0098] In the experimental verification, the solution of the present invention was tested in a variety of typical scenarios, including poor lighting conditions, contamination of the inner cavity surface, and complex background interference. The experimental results show that the detection algorithm of the present invention has high accuracy (average accuracy exceeds 95%) and maintains a low missed detection rate (less than 15%). Compared with the existing detection method that relies on manual labor, the detection speed of the present invention is greatly improved, and it performs better in terms of consistency and stability.

[0099] In summary, the present invention effectively solves the problem of difficulty in detecting blade cavity damage through the innovative combination of inspection robots and deep learning target detection algorithms, realizes an efficient, accurate and intelligent detection process, and provides a strong guarantee for the safety and reliability of wind turbines.

[0100] The following is an introduction to a wind turbine blade cavity fault detection device provided by an embodiment of the present invention. The wind turbine blade cavity fault detection device described below and the wind turbine blade cavity fault detection method described above can be referenced to each other.

[0101] Please refer to Figure 4 , Figure 4 A schematic structural diagram of a wind turbine blade cavity fault detection device provided by an embodiment of the present invention may include:

[0102] The first feature acquisition module 100: acquires the leaf features corresponding to the inner cavity image of the leaf to be detected, and uses the leaf features as shallow features;

[0103] A second feature acquisition module 200 is used to extract deep features of the leaf features through a reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, the heterogeneous convolution module is a module for extracting features using a large convolution kernel and a plurality of small convolution kernels in parallel, and the size of the large convolution kernel is larger than the size of the small convolution kernel;

[0104] A fusion module 300 is used to fuse the shallow features and the deep features using a surface fusion module and a high-level fusion module to obtain fusion features; wherein the surface fusion module fuses the shallow features with the deep features when fusing, and the high-level fusion module performs weighted fusion of the shallow features with the deep features when fusing;

[0105] The fault detection module 400 is used to perform spatial feature enhancement on the fused features, perform semantic feature fusion on the enhanced fused features and the deep features to obtain target detection features, and identify the target detection features to determine the target fault category.

[0106] Further, based on the above embodiments, the surface fusion module is a module designed based on a bidirectional link mechanism, a feature enhancement strategy and a preliminary feature fusion method, and the preliminary feature fusion method is a method for fusing shallow features with deep features; the high-level fusion module is a module designed based on multi-directional links, multi-resolution feature output and a weighted feature fusion method, and the weighted feature fusion method is a method for integrating the shallow features with the deep features layer by layer through a weighted feature fusion strategy.

[0107] Further, based on any of the above embodiments, the fault detection module 400 may include:

[0108] A lightweight head layer acquisition unit, used to acquire a lightweight head layer based on a separation convolution and pruning method;

[0109] A feature enhancement unit, configured to perform spatial feature enhancement on the fused feature based on multi-level feature aggregation of the lightweight head layer, and perform semantic feature fusion on the enhanced fused feature with the deep feature to obtain a target detection feature;

[0110] A fault type determination unit is used to detect the target detection feature based on the multi-scale detection head of the lightweight head layer to obtain the target fault category.

[0111] Further, based on any of the above embodiments, the fault detection module 400 may include:

[0112] The category and fault determination unit is used to identify the target detection features and determine the target fault category and blade fault location.

[0113] Further, based on the above embodiment, the above wind turbine blade cavity fault detection device may further include:

[0114] A fault level determination module, configured to determine a fault level based on the target fault category and the blade fault position;

[0115] The fault processing module is used to determine the processing priority corresponding to each fault based on the fault level, so as to process the fault based on the processing priority.

[0116] Further, based on any of the above embodiments, the above wind turbine blade cavity fault detection device may further include:

[0117] A deep neural network building module, used to obtain a deep neural network including the reverse bottleneck module, the surface fusion module and the high-level fusion module;

[0118] A first training module is used to train the deep neural network based on large data set fault training data to obtain a pre-trained blade fault detection model; wherein the large data set fault training data is a data set whose amount of fault training data is greater than the minimum amount of training data;

[0119] The second training module is used to obtain the wind turbine blade cavity fault training data to train the pre-trained blade fault detection model to obtain a target blade fault detection model including the trained reverse bottleneck module, the surface fusion module and the high-level fusion module.

[0120] Further, based on any of the above embodiments, the above wind turbine blade cavity fault detection device may further include:

[0121] The inspection robot-based image acquisition unit is used to acquire the inner cavity image of the blade to be inspected in real time based on the wind turbine blade inner cavity inspection robot; wherein the wind turbine blade inner cavity inspection robot is a robot including wheeled crawler automatic adjustment function, active obstacle avoidance function and adaptive path planning function.

[0122] It should be noted that the order of the modules and units in the above-mentioned wind turbine blade cavity fault detection device can be changed without affecting the logic.

[0123] The wind turbine blade cavity fault detection device provided by the embodiment of the present invention may include: a first feature acquisition module 100, used to acquire blade features corresponding to the blade cavity image to be detected, and use the blade features as shallow features; a second feature acquisition module 200, used to extract deep features of the blade features through a reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, and the heterogeneous convolution module is a module for extracting features using a parallel large convolution kernel and a plurality of small convolution kernels, and the size of the large convolution kernel is larger than the size of the small convolution kernel; a fusion module 300 , used to use the surface fusion module and the high-level fusion module to fuse the shallow features and the deep features to obtain fused features; wherein the surface fusion module fuses the shallow features with the deep features when fusing, and the high-level fusion module performs weighted fusion of the shallow features and the deep features when fusing; the fault detection module 400 is used to perform spatial feature enhancement on the fused features, perform semantic feature fusion on the enhanced fused features and the deep features to obtain target detection features, and identify the target detection features to determine the target fault category. Compared with the current manual detection that results in low efficiency and accuracy in blade cavity fault detection, the present application divides the blade features into two tributaries to obtain deep features and shallow features. The deep features are features obtained based on dynamic convolution kernels, thereby greatly expanding the perception field of the network, and are based on the shallow features effectively retained by the surface fusion module, while the high-level fusion module promotes the output layer to retain diverse multi-scale information through enhanced information fusion, thereby improving the depth and comprehensiveness of feature extraction. Therefore, the intelligent wind turbine blade cavity fault detection method provided by the present application can improve the accuracy and efficiency of fault detection.

[0124] A wind turbine blade cavity fault detection device provided by an embodiment of the present invention is introduced below. The wind turbine blade cavity fault detection device described below and the wind turbine blade cavity fault detection method described above can refer to each other.

[0125] Please refer to Figure 5 , Figure 5 A schematic structural diagram of a wind turbine blade cavity fault detection device provided by an embodiment of the present invention may include:

[0126] A memory 10, used for storing computer programs;

[0127] The processor 20 is used to execute a computer program to implement the above-mentioned method for detecting internal cavity faults of blades of a wind turbine generator set.

[0128] The memory 10 , the processor 20 , and the communication interface 30 all communicate with each other via a communication bus 40 .

[0129] In the embodiment of the present invention, the memory 10 is used to store one or more programs, and the program may include program code, and the program code includes computer operation instructions. In the embodiment of the present invention, the memory 10 may store programs for implementing the following functions:

[0130] Obtain the leaf features corresponding to the inner cavity image of the leaf to be detected, and use the leaf features as shallow features;

[0131] The deep features of the leaf features are extracted through the reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, and the heterogeneous convolution module is a module for extracting features using a large convolution kernel and multiple small convolution kernels in parallel, and the size of the large convolution kernel is larger than the size of the small convolution kernel;

[0132] The surface fusion module and the high-level fusion module are used to fuse the shallow features and the deep features to obtain fused features; wherein the surface fusion module fuses the shallow features with the deep features when fusing, and the high-level fusion module performs weighted fusion of the shallow features with the deep features when fusing;

[0133] The spatial features of the fused features are enhanced, and the enhanced fused features are semantically fused with the deep features to obtain the target detection features. The target detection features are then identified to determine the target fault category.

[0134] In a possible implementation, the memory 10 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function, etc.; the data storage area may store data created during use.

[0135] In addition, the memory 10 may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include an NVRAM. The memory stores an operating system and operating instructions, executable modules or data structures, or a subset thereof, or an extended set thereof, wherein the operating instructions may include various operating instructions for implementing various operations. The operating system may include various system programs for implementing various basic tasks and processing hardware-based tasks.

[0136] The processor 20 may be a central processing unit (CPU), an application specific integrated circuit, a digital signal processor, a field programmable gate array or other programmable logic device, a microprocessor or any conventional processor, etc. The processor 20 may call a program stored in the memory 10 .

[0137] The communication interface 30 may be an interface of a communication module, and is used to connect to other devices or systems.

[0138] Of course, it should be noted that Figure 5 The structure shown does not constitute a limitation on the wind turbine blade cavity fault detection device in the embodiment of the present invention. In actual applications, the wind turbine blade cavity fault detection device may include: Figure 5 More or fewer components than shown, or combinations of certain components.

[0139] The following is an introduction to a computer-readable storage medium (corresponding to the medium described above) provided in an embodiment of the present invention. The computer-readable storage medium described below and the wind turbine blade cavity fault detection method described above may refer to each other.

[0140] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned method for detecting internal cavity faults of blades of a wind turbine generator set are implemented.

[0141] The computer-readable storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0142] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0143] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0144] Finally, it should be noted that, in this article, relationships such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0145] The above is a detailed introduction to a method, device, equipment and medium for detecting inner cavity faults of wind turbine blades provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for detecting internal cavity faults of wind turbine blades, characterized in that: include: Acquire the leaf features corresponding to the inner cavity image of the leaf to be detected, and use the leaf features as shallow features; The deep features of the leaf features are extracted by a reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, the heterogeneous convolution module is a module for extracting features using a large convolution kernel and a plurality of small convolution kernels in parallel, and the size of the large convolution kernel is larger than the size of the small convolution kernel; The shallow features and the deep features are fused by using a surface fusion module and a high-level fusion module to obtain fused features; wherein the surface fusion module fuses the shallow features with the deep features when fusing, and the high-level fusion module performs weighted fusion of the shallow features with the deep features when fusing; the surface fusion module is a module designed based on a bidirectional link mechanism, a feature enhancement strategy and a preliminary feature fusion method, and the preliminary feature fusion method is a method for fusing shallow features with deep features; the high-level fusion module is a module designed based on multi-directional links, multi-resolution feature output and a weighted feature fusion method, and the weighted feature fusion method is a method for integrating the shallow features with the deep features layer by layer through a weighted feature fusion strategy; The spatial feature of the fused feature is enhanced, the enhanced fused feature is fused with the deep feature for semantic feature to obtain the target detection feature, and the target detection feature is identified to determine the target fault category.

2. The wind turbine blade inner cavity fault detection method according to claim 1, characterized in that: Performing spatial feature enhancement on the fusion feature, performing semantic feature fusion on the enhanced fusion feature and the deep feature to obtain a target detection feature, and identifying the target detection feature to determine the target fault category, including: Get the lightweight head layer based on the separation convolution and pruning method; Based on the multi-level feature aggregation of the lightweight head layer, the spatial feature of the fused feature is enhanced, and the enhanced fused feature is semantically fused with the deep feature to obtain the target detection feature; The multi-scale detection head based on the lightweight head layer detects the target detection features to obtain the target fault category.

3. The method for detecting internal cavity faults of blades of a wind turbine generator set according to any one of claims 1 to 2, characterized in that: Identifying the target detection features and determining the target fault category includes: The target detection features are identified to determine the target fault category and blade fault location.

4. The wind turbine blade cavity fault detection method according to claim 3, characterized in that: After identifying the target detection features and determining the target fault category and the blade fault location, the method further includes: determining a fault level based on the target fault category and the blade fault location; A processing priority corresponding to each fault is determined based on the fault level, so that the fault is processed based on the processing priority.

5. The wind turbine blade inner cavity fault detection method according to claim 1, characterized in that: Before obtaining the leaf features corresponding to the inner cavity image of the leaf to be detected and using the leaf features as shallow features, the method further includes: Acquire a deep neural network including the reverse bottleneck module, the surface fusion module and the high-level fusion module; The deep neural network is trained based on the large data set fault training data to obtain a pre-trained blade fault detection model; wherein the large data set fault training data is a data set whose amount of fault training data is greater than the minimum amount of training data; The wind turbine blade cavity fault training data is obtained to train the pre-trained blade fault detection model, and a target blade fault detection model including the trained reverse bottleneck module, the surface fusion module and the high-level fusion module is obtained.

6. The method for detecting internal cavity faults of wind turbine blades according to claim 1, characterized in that: Before obtaining the blade features corresponding to the inner cavity image of the blade to be detected and using the blade features as shallow features, the method further includes: The image of the inner cavity of the blade to be inspected is acquired in real time based on a wind turbine blade inner cavity inspection robot; wherein the wind turbine blade inner cavity inspection robot is a robot including wheeled crawler automatic adjustment function, active obstacle avoidance function and adaptive path planning function.

7. A wind turbine blade cavity fault detection device, characterized in that: include: A first feature acquisition module is used to acquire the leaf features corresponding to the inner cavity image of the leaf to be detected, and use the leaf features as shallow features; A second feature acquisition module is used to extract deep features of the leaf features through a reverse bottleneck module; wherein the reverse bottleneck module includes a heterogeneous convolution module and a dynamic convolution kernel, the heterogeneous convolution module is a module that uses a parallel large convolution kernel and a plurality of small convolution kernels to extract features, and the size of the large convolution kernel is larger than the size of the small convolution kernel; A fusion module, used for fusing the shallow features and the deep features by using a surface fusion module and a high-level fusion module to obtain fusion features; wherein the surface fusion module fuses the shallow features with the deep features when fusing, and the high-level fusion module performs weighted fusion of the shallow features with the deep features when fusing; the surface fusion module is a module designed based on a bidirectional link mechanism, a feature enhancement strategy and a preliminary feature fusion method, and the preliminary feature fusion method is a method for fusing shallow features with deep features; the high-level fusion module is a module designed based on multi-directional links, multi-resolution feature output and a weighted feature fusion method, and the weighted feature fusion method is a method for integrating the shallow features with the deep features layer by layer through a weighted feature fusion strategy; The fault detection module is used to perform spatial feature enhancement on the fused features, perform semantic feature fusion on the enhanced fused features and the deep features to obtain target detection features, and identify the target detection features to determine the target fault category.

8. A wind turbine blade cavity fault detection device, characterized in that: include: Memory for storing computer programs; A processor is used to execute the computer program to implement the steps of the wind turbine blade cavity fault detection method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for detecting internal cavity faults of a wind turbine blade as claimed in any one of claims 1 to 6 are implemented.

Images