Fan blade fault detection method and system based on improved YOLOv5

Abstract

The present invention relates to a fan blade fault detection method based on improved YOLOv5, and relates to the field of fan detection. The method comprises: collecting a fan blade data set; using an AutoAugment method to perform data enhancement on sample images in the data set to obtain a first data enhanced data set; using a Retinex algorithm to perform data enhancement on images in the first data enhanced data set to obtain a second data enhanced data set; using the second data enhanced data set to train an improved YOLOv5 to obtain a fan blade fault detection model; the improved YOLOv5 comprises adding an adaptive attention module at a jump connection of a Neck part, and adding a feature enhancement module at an upsampling part of the Neck part; and using the fan blade fault detection model to perform on a fan blade to be detected. The present invention improves the robustness of the detection model.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine detection, and in particular to a wind turbine blade fault detection method and system based on improved YOLOv5. Background Art

[0002] With the rapid economic development, people's demand for energy and changes in the ecological environment force people to constantly explore new energy. People refer to the technology of using hydropower, solar energy, nuclear energy and wind energy to generate electricity as new energy power generation. China has abundant water resources, but less development. Compensation for farmers who have been relocated due to the construction of hydropower stations and ecological environmental protection are the main reasons for curbing development. China's total nuclear power generation is 1.8%, which is lower than the world average. Nuclear safety is a stumbling block to the development of nuclear power. Solar energy is sufficient and pollution-free, but the energy density is uneven and the intensity varies greatly due to the influence of seasons, which has become a stumbling block to the development and utilization of solar energy. Wind energy is widely distributed in nature, with a large storage capacity, and wind power generation is not limited by fuel and water. At the same time, because the power generation process is safe, reliable and pollution-free, it has not only become the vanguard of new energy power generation, but also truly become an important part of modern energy supply.

[0003] The most important component in a wind turbine that converts wind energy into mechanical energy is the blades that rotate under wind force. However, the working environment of the blades is very unfriendly, the load is complex and changeable, and the damage rate is extremely high. As an important equipment of the entire wind turbine, the cost of the blades accounts for 15%-20% of the total equipment cost of the unit, and the price is expensive.

[0004] The blades are very heavy and located at high altitudes, so it is difficult to detect early damage to the blades with the naked eye. At the same time, manual maintenance is difficult, and once an accident occurs, the situation is very serious; if the small defects of the blades are not discovered and handled in time, the blades may break and the unit may be shut down, causing huge losses in manpower and material resources.

[0005] For large wind farms, the inspection task of blades is very tedious and difficult. If regular inspections are carried out, the unit will be inspected, fault analysis will be carried out, and faults will be handled after the problem is found. These small defects may cause some major faults, resulting in huge damage and serious consequences. In order to ensure that wind turbines generate electricity in a safe state, reduce the funds required for maintenance, and improve the utilization rate of wind energy, it is of vital importance to detect early defects of blades in a timely and accurate manner.

[0006] The commonly used manual inspection method at present needs to be carried out when the wind turbine is stopped, and the blade surface is observed through a telescope to see if there are cracks, damage or lightning strikes. This method makes the inspection personnel have a large workload, which easily causes fatigue of the inspection personnel, increases the insecurity of the work, and also reduces the efficiency of blade inspection. At the same time, the inspection process has strict requirements on the inspection precision instruments and the professional technical level of the inspection personnel.

[0007] Manual visual inspection of wind turbine blade surface defects is inefficient and cannot guarantee the safety of inspectors. At the same time, the accuracy of inspection cannot be guaranteed. The installation of various sensors is expensive and has high requirements for the inspection environment. Currently, wind turbine blade fault detection based on machine vision has problems of poor robustness or poor detection efficiency. Summary of the invention

[0008] The purpose of the present invention is to provide a fan blade fault detection method and system based on improved YOLOv5, thereby improving the robustness of the detection model.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A fan blade fault detection method based on improved YOLOv5, comprising:

[0011] Collect wind turbine blade data sets;

[0012] Using the AutoAugment method to perform data enhancement on the sample images in the data set to obtain a first data enhanced data set;

[0013] Using the Retinex algorithm to perform data enhancement on the images in the first data enhancement dataset to obtain a second data enhancement dataset;

[0014] The improved YOLOv5 is trained using the data set after the second data enhancement to obtain a fan blade fault detection model; the improved YOLOv5 includes adding an adaptive attention module at the jump connection of the Neck part, and adding a feature enhancement module at the upsampling part of the Neck part;

[0015] The fan blade fault detection model is used to detect the fan blade to be detected.

[0016] Optionally, the adaptive attention module is used to obtain multiple context features of different scales through an adaptive average pooling layer.

[0017] Optionally, the feature enhancement module includes a multi-branch convolutional layer and a branch pooling layer, the multi-branch convolutional layer is used to provide receptive fields of different sizes for the input feature map, and the branch pooling layer is used to perform feature fusion on the output of the multi-branch convolutional layer.

[0018] Optionally, before the AutoAugment method is used to perform data enhancement on the sample images in the data set to obtain the first data-enhanced data set, the method further includes:

[0019] The sample images in the data set are enhanced by spatial geometric transformation and RGB color perturbation respectively; the spatial geometric transformation includes horizontal or vertical flipping, random cropping, rotation and radial transformation of the sample images.

[0020] Optionally, the output of the wind turbine blade fault detection model includes fault type and blade position, and the fault type includes corrosion, scratches, cracks, coating peeling, surface spots and surface sand holes.

[0021] Optionally, it also includes: deploying the wind turbine blade fault detection model on a mobile terminal of the drone, and the wind turbine blade fault detection model deployed on the mobile terminal of the drone is used to perform real-time fault detection on the wind turbine blade images collected by the drone.

[0022] The present invention discloses a fan blade fault detection system based on improved YOLOv5, comprising:

[0023] A data set acquisition module, used to acquire wind turbine blade data sets;

[0024] A first data enhancement module, used to perform data enhancement on the sample images in the data set by using an AutoAugment method to obtain a first data enhanced data set;

[0025] A second data enhancement module, configured to perform data enhancement on the images in the first data enhancement dataset by using a Retinex algorithm to obtain a second data enhancement dataset;

[0026] An improved YOLOv5 training module is used to train the improved YOLOv5 using the data set after the second data enhancement to obtain a fan blade fault detection model; the improved YOLOv5 includes adding an adaptive attention module at the jump connection of the Neck part, and adding a feature enhancement module at the upsampling part of the Neck part;

[0027] The fan blade fault detection module is used to detect the fan blade to be detected using the fan blade fault detection model.

[0028] Optionally, the adaptive attention module is used to obtain multiple context features of different scales through an adaptive average pooling layer.

[0029] Optionally, the feature enhancement module includes a multi-branch convolutional layer and a branch pooling layer, the multi-branch convolutional layer is used to provide receptive fields of different sizes for the input feature map, and the branch pooling layer is used to perform feature fusion on the output of the multi-branch convolutional layer.

[0030] Optionally, the output of the wind turbine blade fault detection model includes fault type and blade position, and the fault type includes corrosion, scratches, cracks, coating peeling, surface spots and surface sand holes.

[0031] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0032] The present invention discloses a fan blade fault detection method based on improved YOLOv5, which adopts AutoAugment method and Retinex algorithm to perform data enhancement on data set images, enriches the data set by AutoAugment method to make it more suitable for actual scenes, and performs image enhancement preprocessing by Retinex algorithm to eliminate the influence of uneven illumination, improve the visual effect of the image, and thus improve the robustness of the model. In addition, the improved model uses adaptive attention module (AAM) and feature enhancement module (FEM) to reduce information loss in the process of feature map generation and enhance representation ability, thereby improving the accuracy of fault detection of the detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] 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 will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 The present invention is a flow chart of a method for detecting fan blade faults based on improved YOLOv5;

[0035] Figure 2 This is a structural schematic diagram of a fan blade fault detection system based on improved YOLOv5 in the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0037] The purpose of the present invention is to provide a fan blade fault detection method and system based on improved YOLOv5, thereby improving the robustness of the detection model.

[0038] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] Figure 1 FIG. 1 is a flow chart of a fan blade fault detection method based on improved YOLOv5 according to the present invention. Figure 1 As shown, a fan blade fault detection method based on improved YOLOv5 includes:

[0040] Step 101: Collect a wind turbine blade data set.

[0041] The experimental objects collected by the present invention are fan blades. The working environment of the blades is particularly harsh, and they are subject to the superposition of aerodynamic loads, gravity loads and inertial loads. Moreover, they are constantly running in the natural environment for a long time. The external climate will have a great impact on the operation of the blades, especially typhoons, thunderstorms, ice and snow, sandstorms and other severe weather conditions that may cause damage to the blades at any time, thereby causing economic losses to wind power generation. For example, the dust on the scene erodes the surface of the blades over time, causing surface sand holes on the blades; the operation of the blades inevitably produces torsion, bending, vibration, etc., causing cracks or fractures in the fan blades; the fan blades are installed too high, causing the blades to be struck by lightning, etc. The surface defects of the fan blades generally include corrosion, scratches and cracks, coating peeling, surface spots, surface sand holes, etc., and the background includes gray-yellow wilderness, agricultural land, industrial land, sky without precipitation, etc. The original images are all images of fan blades taken using Hikvision industrial color cameras, totaling 1,000 images. After the data collection is completed, the 1,000 collected wind turbine blade images are annotated, and the annotated parts are mainly the blades of the wind turbine.

[0042] Step 102: using the AutoAugment method to perform data enhancement on the sample images in the data set to obtain a first data-enhanced data set.

[0043] Before step 102, the method further includes:

[0044] The sample images in the data set are enhanced by spatial geometric transformation and RGB color perturbation respectively; the spatial geometric transformation includes horizontal or vertical flipping, random cropping, rotation and radial transformation of the sample images.

[0045] The RGB color perturbation method converts the image from the RGB color space to another color space, increases or decreases the color parameters, and then returns to the RGB color space.

[0046] Improve the generalization ability of the model by increasing the size and quality of the dataset.

[0047] The present invention adopts AutoAugment, a method for automatically searching for a suitable data enhancement strategy, to effectively enhance the data set of wind turbine blades and increase the quantity and diversity of training image data.

[0048] AutoAugment is a method proposed by researchers at Google Brain in 2018 for automatically searching for suitable data augmentation strategies. This method creates a search space for data augmentation strategies and uses a search algorithm to select a data augmentation strategy suitable for a specific data set.

[0049] Step 103: Use the Retinex algorithm to perform data enhancement on the images in the first data enhancement dataset to obtain a second data enhancement dataset.

[0050] The Retinex algorithm is used to perform image enhancement preprocessing on the wind turbine blade image after data enhancement in step 102, thereby reducing the influence of complex environments such as external fog and darkness on the identification of wind turbine blade faults, making the wind turbine blade image easier to identify with the improved YOLOv5 model.

[0051] The data set after the second data augmentation is input into the Retinex algorithm and processed in the following three steps:

[0052] 1) Get R(x,y) based on the original image, and calculate the image L(x,y) after the original image is blurred at the specified scale;

[0053] 2) According to the following formula (2), the value of logR(x,y) is obtained;

[0054] 3) Quantize logR(x,y) to pixel values ​​from 0 to 255 to obtain the final output, which is the processed image.

[0055] The Retinex algorithm works as follows:

[0056] According to the principle of image generation, an image can be regarded as composed of an incident image and a reflected image. The incident light shines on the reflective object, and the reflected light enters the human eye through the reflective object, which is the image seen by humans.

[0057] The image S of the object seen by the observer is obtained by reflecting the incident light L from the surface of the object. The reflectivity R is determined by the object itself and is not affected by the incident light L. The original image S is the product of the illumination image L and the reflectivity image R, which can be expressed as follows:

[0058] S(x,y)=R(x,y)·L(x,y) (1)

[0059] Among them, L(x,y) represents the incident light image, which directly determines the dynamic range that the pixels in the image can achieve. R(x,y) represents the reflection property image of the object, that is, the intrinsic property of the image. S(x,y) represents the reflected light image that can be received by the human eye.

[0060] The purpose of image preprocessing based on Retinex is to estimate the illumination L from the original image S, thereby decomposing R and eliminating the influence of uneven illumination to improve the visual effect of the image, just like the human visual system. In processing, the image is usually converted to the logarithmic domain, so as to convert the product relationship into a sum relationship, that is:

[0061] logS(x,y)=logR(x,y)+logL(x,y) (2)

[0062] The core of the Retinex method is to estimate the illumination L, estimate the L component from the image S, and remove the L component to obtain the original reflection component R. Let s = logS(x, y), r = logR(x, y), l = logL(x, y), and set l = f(s), that is, s = r + l = r + f(s), we can get:

[0063] r=sf(s) (3)

[0064] Assume r = f(x), the function f(x) estimates the illuminance L to obtain the r component.

[0065] Step 104: Use the second data-enhanced data set to train an improved YOLOv5 to obtain a wind turbine blade fault detection model; the improved YOLOv5 includes adding an adaptive attention module at the jump connection of the Neck part, and adding a feature enhancement module to the upsampling part of the Neck part.

[0066] The adaptive attention module is used to obtain multiple context features of different scales through an adaptive average pooling layer.

[0067] The feature enhancement module includes a multi-branch convolution layer and a branch pooling layer, the multi-branch convolution layer is used to provide receptive fields of different sizes for the input feature map, and the branch pooling layer is used to perform feature fusion on the output of the multi-branch convolution layer.

[0068] The improved YOLOv5 is trained using the data set after the second data enhancement, wherein the image input into the improved YOLOv5 is a manually annotated image of a wind turbine blade, and the annotated part is the blade of a wind turbine. The network training parameters including batch size, learning rate, and number of iterations are set, and the network model (improved YOLOv5) is trained using the annotated data set (the data set after the second data enhancement). When the network training converges and the loss function is less than a given value, the network training is completed.

[0069] The improved YOLOv5 training evaluation indicators include precision and recall.

[0070] Precision: (correct positive class / all positive classes found);

[0071]

[0072] Recall: (positive classes found correctly / all positive classes that should have been found correctly);

[0073]

[0074] Among them, FN: FalseNegative, is judged as a negative sample, but in fact it is a positive sample. FP: FalsePositive, is judged as a positive sample, but in fact it is a negative sample. TP: TruePositive, is judged as a positive sample, but in fact it is also a positive sample.

[0075] In a data set detection, four types of detection results are generated: TP, TN, FP, FN:

[0076] T——true means correct

[0077] F——false indicates an error

[0078] P——positive means positive, regarded as a positive example

[0079] N——negative means negative, regarded as a negative example

[0080] TP——Predicted as P (positive class), the prediction is correct, it was originally a positive sample, and the detection is a positive sample (true positive).

[0081] TN——The prediction is N (negative class), the prediction is correct, it was originally a negative sample, and it was detected as a negative sample (true negative).

[0082] FP——Predicted as P (positive class), the prediction is wrong, it was originally a negative sample, but detected as a positive sample (false positive).

[0083] FN——The prediction is N (negative class), the prediction is wrong, it was originally a positive sample, but it was detected as a negative sample (false negative).

[0084] In the present invention, the positive class is the improved YOLOv5 model used to detect defects on the blades of the wind turbine. If the detection result is consistent with the actual situation, it is a positive class; if the detection result is inconsistent with the actual situation, it is a negative class.

[0085] The basic framework of the improved YOLOv5 can be divided into 4 parts:

[0086] (1) Input: AutoAugment data enhancement and adaptive anchor box calculation;

[0087] (2) Backbone: Focus structure, CSP structure (Cross Stage Paritial Network);

[0088] (3) Neck: FPN (Feature Pyramid Network) (integration of AAM and FEM) + PAN (Region Proposal Network) structure;

[0089] (4) Prediction: IOU_Loss (Intersection over Union_Loss).

[0090] Input terminal function:

[0091] ①AutoAugment data enhancement, which effectively enhances the data set of wind turbine blades and increases the quantity and diversity of training image data.

[0092] ② Adaptive anchor box calculation. In the YOLO algorithm, for different data sets, there will be anchor boxes with initial length and width. In YOLOv5, this function adaptively calculates the optimal anchor box value in different training sets each time the training is calculated.

[0093] Backbone (backbone network) function:

[0094] ①Focus structure plays a role of downsampling and improves the detection speed. For example, the original 608*608*3 image is input into the Focus structure, and the slicing operation is used to first convert it into a 304*304*12 feature map, and then a convolution operation with 32 convolution kernels is performed to finally convert it into a 304*304*32 feature map.

[0095] ②CSP structure (Cross Stage Paritial Network) mainly solves the problem of large amount of computation in reasoning from the perspective of network structure design. Therefore, the CSP module is used to first divide the feature map of the base layer into two parts, and then merge them through the cross-stage hierarchical structure, which can reduce the amount of computation while ensuring accuracy.

[0096] Neck function:

[0097] ①FPN (Feature Pyramid Network), that is, feature pyramid, realizes the restoration of feature image resolution and the combination of shallow features.

[0098] ②RPN (Region Proposal Network), that is, the region generation network, is specifically used to extract candidate boxes.

[0099] Prediction (prediction end), through IOU_Loss (Intersection over Union_Loss) to continuously adjust the weights, so as to achieve the purpose of training the network.

[0100] The improved YOLOv5 is as follows: In the traditional feature pyramid (FPN) of the Neck part of the original YOLOv5, an adaptive attention module (AAM) is added to the skip connection to alleviate the context loss in the high-level feature graph caused by the reduction of feature channels. A feature enhancement module is added to the upsampling part to enhance the representation of the feature pyramid and improve the inference speed. For the specific added modules, see the Neck part in the figure above.

[0101] Specifically, Backbone includes Focus, CBL, csp1, CBL, csp1, CBL, CBL, csp1, CBL, CBL, CBL, SPP, and CBL, which are connected in sequence. An adaptive attention module is added to the jump connection at the output of the third CBL. A feature enhancement module is added to the upsampling part of the output of the fifth CBL.

[0102] The operation of the adaptive module can be divided into two steps:

[0103] First, multiple context features of different scales are obtained through an adaptive average pooling layer. The pooling layer coefficient is [0.1, 0.5], which changes adaptively according to the target size of the dataset.

[0104] Secondly, a weight map is generated for each feature map through the spatial attention mechanism. The context is fused through the weight map to generate a new feature map containing multi-scale up and down information.

[0105] The feature enhancement module (FEM) mainly uses dilated convolution to adaptively learn different receptive fields in each feature map according to the different scales of the detected wind turbine blades, thereby improving the accuracy of multi-scale target detection and recognition. The module is divided into a multi-branch convolution layer and a branch pooling layer. The multi-branch convolution layer provides receptive fields of different sizes for the input feature map through dilated convolution, and the branch pooling layer is used to fuse traffic information from the receptive fields of the three branches, thereby improving the multi-scale precision prediction.

[0106] The receptive field formula of the dilated convolution is as follows:

[0107] r1=d*(k-1)+1 (8)

[0108] r n =d*(k-1)+r n-1 (9)

[0109] Where k represents kernel-size (convolution kernel size), d represents the stride of convolution, and r n Represents the expansion rate.

[0110] The branch pooling layer is used to fuse information from different parallel branches to avoid introducing additional parameters. During training, the average operation is used to balance the representation of different parallel branches so that a single branch can achieve reasoning during testing. The expression is as follows:

[0111]

[0112] Among them, B represents the number of convolutional layer branches in the feature enhancement module, and y i Represents the output of each branch convolution layer, y p Represents the output of the branch pooling layer.

[0113] Step 105: Use the fan blade fault detection model to detect the fan blade to be detected.

[0114] The output of the wind turbine blade fault detection model includes fault type and blade position, and the fault type includes corrosion, scratches, cracks, coating peeling, surface spots and surface sand holes.

[0115] A fan blade fault detection method based on improved YOLOv5 also includes: deploying the fan blade fault detection model on a mobile terminal of an unmanned aerial vehicle, and the fan blade fault detection model deployed on the mobile terminal of the unmanned aerial vehicle is used to perform real-time fault detection on the fan blade images collected by the unmanned aerial vehicle.

[0116] The present invention uses the JETSON TX2 development board. The JETSON series is an embedded platform launched by Nvidia for the field of unmanned intelligence. It can process complex data on edge devices and realize artificial intelligence. JETSON TX2 is a modular AI supercomputer that uses NVIDIA Pascal TM Architecture. It is powerful, yet compact and energy-efficient, making it ideal for intelligent edge devices such as robots, drones, smart cameras, and portable medical devices. It supports all the features of the JetsonTX1 module while enabling larger and more complex deep neural networks.

[0117] The trained fan blade fault detection model is deployed on the TX2 edge detection device, and the trained network is used to recognize the fan blade images collected by the industrial camera, and the fan blade defects identified in the image are obtained. The trained network is used to recognize the images collected by the camera, and the fan blade defects identified in the image are obtained and classified to achieve fan blade fault detection.

[0118] The present invention designs a fan blade fault detection method based on the improved YOLOv5 algorithm. Firstly, the fault detection technology is introduced into the wind turbine unit drone inspection, and the fan blades are photographed in real time using a Hikvision industrial camera. The image is then input into the Retinex image enhancement algorithm, and finally the processed image is input into the improved YOLOv5 detection network to obtain defective images and classify the defects.

[0119] The advantage is that the AutoAugment data enhancement algorithm is added to increase the quantity and diversity of training image data. The Retinex image enhancement algorithm is introduced to reduce the impact of complex external environments such as fog and darkness on the accuracy of artifact recognition. The YOLOv5 detection model is improved, and an adaptive attention module and a feature enhancement module are added to the traditional feature map pyramid. Through adaptive feature fusion and receptive field enhancement, channel information is largely retained during feature transfer, and different receptive fields in each feature map are adaptively learned to enhance the representation of the feature pyramid, effectively improving the accuracy of multi-scale target recognition. Unlike the existing YOLOv5 network, the current version has been improved to reduce the impact of scale invariance. At the same time, it can be deployed on the mobile terminal of the drone to perform real-time detection and recognition of wind turbine blades.

[0120] Figure 2 FIG. 1 is a schematic diagram of a fan blade fault detection system based on improved YOLOv5 according to the present invention. Figure 2 As shown, a fan blade fault detection system based on improved YOLOv5 is characterized by comprising:

[0121] The data set acquisition module 201 is used to acquire wind turbine blade data sets.

[0122] The first data enhancement module 202 is used to perform data enhancement on the sample images in the data set by using the AutoAugment method to obtain a first data-enhanced data set.

[0123] The second data enhancement module 203 is used to perform data enhancement on the images in the first data enhancement data set by using a Retinex algorithm to obtain a second data enhancement data set.

[0124] The improved YOLOv5 training module 204 is used to train the improved YOLOv5 using the second data augmented data set to obtain a wind turbine blade fault detection model; the improved YOLOv5 includes adding an adaptive attention module at the jump connection of the Neck part, and adding a feature enhancement module at the upsampling part of the Neck part.

[0125] The fan blade fault detection module 205 is used to detect the fan blades using the fan blade fault detection model.

[0126] The adaptive attention module is used to obtain multiple context features of different scales through an adaptive average pooling layer.

[0127] The feature enhancement module includes a multi-branch convolution layer and a branch pooling layer, the multi-branch convolution layer is used to provide receptive fields of different sizes for the input feature map, and the branch pooling layer is used to perform feature fusion on the output of the multi-branch convolution layer.

[0128] The output of the wind turbine blade fault detection model includes fault type and blade position, and the fault type includes corrosion, scratches, cracks, coating peeling, surface spots and surface sand holes.

[0129] 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 system 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.

[0130] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A fan blade fault detection method based on improved YOLOv5, characterized in that: include: Collect wind turbine blade data sets; Using the AutoAugment method to perform data enhancement on the sample images in the data set to obtain a first data enhanced data set; Using the Retinex algorithm to perform data enhancement on the images in the first data enhancement dataset to obtain a second data enhancement dataset; The improved YOLOv5 is trained using the data set after the second data enhancement to obtain a fan blade fault detection model; the improved YOLOv5 includes adding an adaptive attention module at the jump connection of the Neck part, and adding a feature enhancement module at the upsampling part of the Neck part; The fan blade fault detection model is used to detect the fan blade to be detected.

2. The fan blade fault detection method based on improved YOLOv5 according to claim 1, characterized in that: The adaptive attention module is used to obtain multiple context features of different scales through an adaptive average pooling layer.

3. The fan blade fault detection method based on improved YOLOv5 according to claim 1, characterized in that: The feature enhancement module includes a multi-branch convolution layer and a branch pooling layer, the multi-branch convolution layer is used to provide receptive fields of different sizes for the input feature map, and the branch pooling layer is used to perform feature fusion on the output of the multi-branch convolution layer.

4. The fan blade fault detection method based on improved YOLOv5 according to claim 1, characterized in that: Before the AutoAugment method is used to perform data enhancement on the sample images in the data set to obtain the first data-enhanced data set, the method further includes: The sample images in the data set are enhanced by spatial geometric transformation and RGB color perturbation respectively; the spatial geometric transformation includes horizontal or vertical flipping, random cropping, rotation and radial transformation of the sample images.

5. The fan blade fault detection method based on improved YOLOv5 according to claim 1, characterized in that: The output of the wind turbine blade fault detection model includes fault type and blade position, and the fault type includes corrosion, scratches, cracks, coating peeling, surface spots and surface sand holes.

6. The fan blade fault detection method based on improved YOLOv5 according to claim 1, characterized in that: Also includes: The fan blade fault detection model is deployed on a mobile terminal of the drone. The fan blade fault detection model deployed on the mobile terminal of the drone is used to perform real-time fault detection on the fan blade images collected by the drone.

7. A fan blade fault detection system based on improved YOLOv5, characterized in that: include: A data set acquisition module, used to acquire wind turbine blade data sets; A first data enhancement module, used to perform data enhancement on the sample images in the data set by using an AutoAugment method to obtain a first data enhanced data set; A second data enhancement module, configured to perform data enhancement on the images in the first data enhancement dataset by using a Retinex algorithm to obtain a second data enhancement dataset; An improved YOLOv5 training module is used to train the improved YOLOv5 using the data set after the second data enhancement to obtain a fan blade fault detection model; the improved YOLOv5 includes adding an adaptive attention module at the jump connection of the Neck part, and adding a feature enhancement module at the upsampling part of the Neck part; The fan blade fault detection module is used to detect the fan blade to be detected using the fan blade fault detection model.

8. The wind turbine blade fault detection system based on improved YOLOv5 according to claim 7, characterized in that: The adaptive attention module is used to obtain multiple context features of different scales through an adaptive average pooling layer.

9. The wind turbine blade fault detection system based on improved YOLOv5 according to claim 7, characterized in that: The feature enhancement module includes a multi-branch convolution layer and a branch pooling layer, the multi-branch convolution layer is used to provide receptive fields of different sizes for the input feature map, and the branch pooling layer is used to perform feature fusion on the output of the multi-branch convolution layer.

10. The wind turbine blade fault detection system based on improved YOLOv5 according to claim 7, characterized in that: The output of the wind turbine blade fault detection model includes fault type and blade position, and the fault type includes corrosion, scratches, cracks, coating peeling, surface spots and surface sand holes.

Images