Fan bolt fault identification method based on lightweight deep learning target detection

Abstract

The invention discloses a fan bolt fault identification method based on lightweight deep learning target detection, and the method comprises the steps: collecting a video stream through a common industrial camera, carrying out model training, carrying out the detection and identification of a bolt based on a computer vision target detection technology, and finally achieving the fault judgment of the bolt through confidence analysis. An expensive 3D point cloud camera is replaced by a common industrial camera, hardware cost is reduced, lightweight model design is adopted, computing resource occupation is reduced, flexible deployment in various computing power processors can be realized, hardware cost and computing resource consumption are remarkably reduced, management software compatible with the camera is configured, and the system is convenient to use. Complete workflow closed-loop management is achieved, the fault recognition accuracy is high, and an economical, practical and wide-adaptability bolt fault detection solution is provided for a wind power scene.

Description

Technical Field

[0001] The present invention relates to the technical field of fan bolt fault identification, and in particular to a fan bolt fault identification method based on lightweight deep learning target detection. Background Art

[0002] Bolts are key connecting components in wind turbines, used to fix important structures such as blades, towers, and nacelles. However, wind turbines operate for a long time in complex environments such as strong winds, vibrations, and temperature differences, which can cause bolts to loosen, corrode, or break due to fatigue. Loosening or breaking of bolts may cause equipment damage and safety accidents. Therefore, real-time monitoring and fault diagnosis of bolt status are crucial.

[0003] Traditional bolt detection methods rely on manual inspections, which have defects such as low efficiency, high cost, strong subjectivity and inability to monitor in real time. Intelligent detection technology based on video monitoring uses non-contact detection methods combined with deep learning algorithms to not only achieve 7×24 hours of automated high-precision monitoring, but also accurately identify fault characteristics such as loosening and breakage. At the same time, it fully preserves monitoring data to support fault tracing, providing a safe, efficient and intelligent solution for wind turbine operation and maintenance.

[0004] In the existing technology for video monitoring of bolt loosening, 3D cameras are usually used to collect images, point clouds are used instead of traditional RGB images, and the bolt looseness is determined by means of angle difference calculation. However, the existing monitoring technology has the following difficulties: (1) Expensive equipment: The current acquisition equipment for bolt loosening detection technology is mostly 3D point cloud cameras, which analyze the collected point information. Compared with ordinary cameras, this equipment is more expensive. If point cloud cameras are equipped in large quantities, high hardware costs are required; (2) High resource usage: The monitoring algorithm uses segmentation technology, which consumes a lot of computing resources at the station. The computing equipment at the station has already deployed other application systems, which will cause other services to be stuck or abnormal; (3) Difficulty in modifying special equipment: Special sensor bolts are used to achieve fault detection. Although this method has certain technical advantages, the bolts are already fixed when the wind turbine is installed, and subsequent modification is relatively difficult. Based on this, the present invention proposes a low-cost, low-computing-power intelligent detection solution. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and propose a wind turbine bolt fault identification method based on lightweight deep learning target detection. It uses an ordinary camera to collect video streams, and based on computer vision target detection technology, it realizes fault judgment through confidence analysis. It not only significantly reduces hardware costs and computing resource consumption, but also is equipped with camera-compatible management software and supports flexible deployment on various processors, providing an economical, practical and widely adaptable bolt fault detection solution for wind power scenarios.

[0006] The purpose of the present invention is achieved through the following technical solution: a fan bolt fault identification method based on lightweight deep learning target detection, comprising the following steps:

[0007] S1. Collecting video image data of wind turbine bolts in different time periods and weather conditions using a camera device;

[0008] S2. Divide the image data obtained in step S1 into two groups, A and B, and annotate them. The image data in group A is annotated with normal bolts and dropped bolts, while the image data in group B is only annotated with normal bolts and no annotation is made for dropped bolts.

[0009] S3. Build datasets for the two sets of labeled image data, A and B, respectively. Divide each dataset into a training set and a validation set at a preset ratio, so that the image size in the dataset remains within a preset range. Use the datasets as input to train fan bolt fault recognition models A and B.

[0010] S4. For the trained fan bolt fault recognition model B, manually remove the image data containing dropped bolts from its validation set, and then input the validation set after the removal into the fan bolt fault recognition model B to obtain a basis matrix for comparison;

[0011] S5. Deploy fan bolt fault identification model A and fan bolt fault identification model B to the processor;

[0012] S6. Input the real-time collected fan bolt video into the fan bolt fault identification model A for detection; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is greater than a preset threshold, it is considered that the fan bolt video is abnormal, and an abnormality alarm is directly output; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is not greater than the preset threshold, proceed to step S7;

[0013] S7. Input the real-time collected fan bolt video into the fan bolt fault recognition model B for detection, and the fan bolt fault recognition model B outputs a corresponding matrix; based on the base matrix obtained in step S4, the matrix is ​​sorted using IoU in deep learning and the confidence of the two is compared. If the confidence of a certain position in the matrix is ​​less than 3% of the confidence of the same position in the base matrix, and the ratio of the number of image frames at this position to the number of all image frames is greater than a preset threshold, it is considered that there is an abnormality in the fan bolt video and an abnormality alarm is output; otherwise, it is considered that the real-time fan bolt has no fault.

[0014] Furthermore, the step S1 includes:

[0015] Fixed cameras in wind farms collect video image data, which includes images of various time periods in a 24-hour period and images of various weather conditions. The images of the various time periods include daytime images and nighttime images, and the images of various weather conditions include cloudy, sunny, rainy, snowy, and dusty weather.

[0016] Further, the step S3 includes:

[0017] Data sets were constructed for the annotated image data sets A and B respectively, and the data sets A and B were divided into training set and validation set in a ratio of 8:2, so that the image size in the data set was kept in the range of [640*640,960*960,1080*1080]. The data sets were used as input, Pytorch was selected as the training framework, and the yolov11-s architecture was used as the target detector. The fan bolt fault recognition model A and the fan bolt fault recognition model B were obtained through training.

[0018] Further, the step S4 includes:

[0019] For the trained fan bolt fault recognition model B, the image data with fallen bolts in its validation set were manually removed, and then the validation set after removal was input into the fan bolt fault recognition model B to obtain a basis matrix [N,5] for comparison, where N represents the number of detections, the first four columns [N,4] represent the four-point coordinates of each bolt, and the last column [N,1] is the mean confidence value of each bolt in the entire test set.

[0020] Further, the step S6 includes:

[0021] The real-time collected fan bolt video is input into the fan bolt fault identification model A for detection; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is greater than 30%, it is considered that there is an abnormality in the fan bolt video, and an abnormal alarm is directly output; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is not greater than 30%, step S7 is entered.

[0022] Further, the step S7 includes:

[0023] The real-time collected fan bolt video is input into the fan bolt fault recognition model B for detection, and the fan bolt fault recognition model B outputs a corresponding matrix; based on the base matrix obtained in step S4, the matrix is ​​sorted using the IoU in deep learning and the confidence of the two is compared. If the confidence of a certain position in the matrix is ​​less than 3% of the confidence of the same position in the base matrix, and the ratio of the number of image frames at this position to the number of all image frames is greater than 30%, then it is considered that there is an abnormality in the fan bolt video and an abnormality alarm is output; otherwise, it is considered that the real-time fan bolt has no fault.

[0024] The fan bolt fault identification system based on lightweight deep learning target detection is used to implement the above-mentioned fan bolt fault identification method based on lightweight deep learning target detection, including:

[0025] The intelligent camera management module has built-in management software for implementing the RTSP protocol video stream pulling function, providing a visual interface to configure camera parameters, set the storage path and detection time interval operating parameters, and has a camera connection test function;

[0026] The lightweight algorithm analysis module implements a bolt fault identification model based on target detection technology, automatically performs video stream analysis at preset time intervals, and has a built-in intelligent early warning trigger mechanism that automatically records abnormal events and generates alarms, supporting visual display of results.

[0027] Furthermore, the lightweight algorithm analysis module includes:

[0028] A dataset was constructed for the labeled image data sets A and B. The datasets were divided into training and validation sets at a ratio of 8:2, so that the image sizes in the datasets were kept within the range of [640*640, 960*960, 1080*1080]. The datasets were used as input, the Pytorch training framework was selected, and the yolov11-s architecture was used as the target detector. The fan bolt fault recognition model A and the fan bolt fault recognition model B were trained.

[0029] For the trained fan bolt fault recognition model B, we manually removed the image data containing dropped bolts from its validation set. The validation set after removal was then input into the fan bolt fault recognition model B to obtain a basis matrix [N, 5] for comparison. N represents the number of detections, the first four columns [N, 4] represent the four-point coordinates of each bolt, and the last column [N, 1] is the mean confidence score for each bolt in the entire test set.

[0030] The real-time collected fan bolt video is input into the fan bolt fault recognition model A for detection; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is greater than 30%, it is considered that the fan bolt video is abnormal, and an abnormal alarm is directly output; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is not greater than 30%, the real-time collected fan bolt video is input into the fan bolt fault recognition model B for detection, and the fan bolt fault recognition model B outputs a corresponding matrix; based on the base matrix obtained in step S4, the IoU in deep learning is used to sort the matrix and compare the confidence of the two. If the confidence of a certain position in the matrix is ​​less than 3% of the confidence of the same position in the base matrix, and the ratio of the number of image frames at the position to the number of all image frames is greater than 30%, it is considered that the fan bolt video is abnormal, and an abnormal alarm is output; otherwise, it is considered that the real-time fan bolt has no fault.

[0031] A non-transitory computer-readable medium storing instructions, when the instructions are executed by a processor, performs the steps of the above-mentioned wind turbine bolt fault identification method based on lightweight deep learning target detection.

[0032] A computing device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the above-mentioned fan bolt fault identification method based on lightweight deep learning target detection is implemented.

[0033] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0034] 1. The present invention replaces expensive 3D point cloud cameras with ordinary industrial cameras, reducing hardware costs;

[0035] 2. The present invention adopts a lightweight model design, which reduces computing resource usage and can be flexibly deployed in various computing processors;

[0036] 3. The present invention is equipped with an intelligent camera management module, which supports rapid implementation in existing wind farms, realizes 7×24 hours automated monitoring, has complete closed-loop management of work processes, and has high fault identification accuracy, providing an economical, practical, efficient and reliable intelligent solution for wind turbine bolt status monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Flowchart of the fan bolt fault identification method based on lightweight deep learning target detection.

[0038] Figure 2 Schematic diagram of the dropped bolts marked in group A image data.

[0039] Figure 3Schematic diagram of normal bolt condition annotated for group A image data.

[0040] Figure 4 For the image data of group B, only the schematic diagram of the normal bolt condition is marked.

[0041] Figure 5 This is the architecture diagram of the yolov11-s architecture.

[0042] Figure 6 It is a schematic diagram of software configuration based on the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described below with reference to specific embodiments.

[0044] Example 1

[0045] See also Figure 1 As shown, the fan bolt fault identification method based on lightweight deep learning target detection provided by this embodiment includes the following steps:

[0046] S1. Fixed cameras in the wind farm collect video image data. The video image data includes images of various time periods in 24 hours and images of various weather conditions. The images of each time period include daytime images and nighttime images. The images of various weather conditions include cloudy days, sunny days, rainy days, snowy days, and dusty days. The images in the dataset cover various actual scenes as much as possible, so that the model can focus on learning the data distribution characteristics of bolt positions and minimize the impact of factors such as lighting on model performance.

[0047] S2. Divide the image data obtained in step S1 into two groups, A and B, and mark them. Figures 2 to 3 As shown, the normal bolts and dropped bolts are marked on the image data of group A; see Figure 4 As shown, for the image data of group B, only the normal bolt situation is marked, and the dropped bolt situation is not marked.

[0048] S3. Build datasets for the labeled image data sets A and B respectively, and divide the datasets into training and validation sets at a ratio of 8:2. Keep the image size in the range of [640*640, 960*960, 1080*1080]. Random selection will be made during training. If the image size is too small, the data feature loss of bolts, which are small targets, will be more serious. If the input size is too large, the computing power burden of training and inference equipment will be greater. Therefore, this range is selected as the final model input size range. Use the dataset as input, select Pytorch as the training framework, and use the yolov11-s architecture for the target detector. Figure 5As shown, the fan bolt fault identification model A and the fan bolt fault identification model B are obtained through training.

[0049] S4. For the trained fan bolt fault recognition model B, manually remove the image data with fallen bolts in its validation set, and then input the validation set after removal into the fan bolt fault recognition model B to obtain a basis matrix [N,5] for comparison, where N represents the number of bolts detected, the first four columns [N,4] represent the four-point coordinates of each bolt, and the last column [N,1] is the mean confidence value of each bolt in the entire test set.

[0050] S5. Deploy fan bolt fault identification model A and fan bolt fault identification model B to the processor. The deployment methods include onnx, rknn, engine, and cann. The deployment indicator effects are shown in Table 1:

[0051]

[0052]

[0053] Table 1 Deployment indicator effects

[0054] Click "Add Device" on the software interface, configure the camera information, camera IP, account, password, video save path, inference video save path, etc. After configuration, click "Test". If the configuration is successful and the communication is normal, the program will output a 1-minute test video of each camera position. Figure 6 As shown in the figure, after the configuration is completed, click the "Start" button, and the background algorithm will perform analysis at a fixed time based on the configuration information.

[0055] S6. Input the real-time collected fan bolt video into the fan bolt fault identification model A for detection; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is greater than the preset threshold, it is considered that there is an abnormality in the fan bolt video, and an abnormal alarm is directly output; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is not greater than the preset threshold, enter step S7.

[0056] S7. Input the real-time collected fan bolt video into the fan bolt fault recognition model B for detection, and the fan bolt fault recognition model B outputs a corresponding matrix; based on the base matrix obtained in step S4, the matrix is ​​sorted using IoU in deep learning and the confidence of the two is compared. If the confidence of a certain position in the matrix is ​​less than 3% of the confidence of the same position in the base matrix, and the ratio of the number of image frames at this position to the number of all image frames is greater than a preset threshold, it is considered that there is an abnormality in the fan bolt video and an abnormality alarm is output; otherwise, it is considered that the real-time fan bolt has no fault.

[0057] Example 2

[0058] This embodiment discloses a fan bolt fault identification system based on lightweight deep learning target detection, which is used to implement the above-mentioned fan bolt fault identification method based on lightweight deep learning target detection, including:

[0059] The intelligent camera management module has built-in management software for implementing the RTSP protocol video stream pulling function, providing a visual interface to configure camera parameters, set the storage path and detection time interval operating parameters, and has a camera connection test function;

[0060] The lightweight algorithm analysis module implements a bolt fault identification model based on target detection technology, automatically performs video stream analysis at preset time intervals, and has a built-in intelligent early warning trigger mechanism that automatically records abnormal events and generates alarms, supporting visual display of results.

[0061] The lightweight algorithm analysis module includes:

[0062] A dataset was constructed for the labeled image data sets A and B. The datasets were divided into training and validation sets at a ratio of 8:2, so that the image sizes in the datasets were kept within the range of [640*640, 960*960, 1080*1080]. The datasets were used as input, the Pytorch training framework was selected, and the yolov11-s architecture was used as the target detector. The fan bolt fault recognition model A and the fan bolt fault recognition model B were trained.

[0063] For the trained fan bolt fault recognition model B, we manually removed the image data containing dropped bolts from its validation set. The validation set after removal was then input into the fan bolt fault recognition model B to obtain a basis matrix [N, 5] for comparison. N represents the number of detections, the first four columns [N, 4] represent the four-point coordinates of each bolt, and the last column [N, 1] is the mean confidence score for each bolt in the entire test set.

[0064] The real-time collected fan bolt video is input into the fan bolt fault recognition model A for detection; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is greater than 30%, it is considered that the fan bolt video is abnormal, and an abnormal alarm is directly output; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is not greater than 30%, the real-time collected fan bolt video is input into the fan bolt fault recognition model B for detection, and the fan bolt fault recognition model B outputs a corresponding matrix; based on the base matrix obtained in step S4, the IoU in deep learning is used to sort the matrix and compare the confidence of the two. If the confidence of a certain position in the matrix is ​​less than 3% of the confidence of the same position in the base matrix, and the ratio of the number of image frames at the position to the number of all image frames is greater than 30%, it is considered that the fan bolt video is abnormal, and an abnormal alarm is output; otherwise, it is considered that the real-time fan bolt has no fault.

[0065] Example 3

[0066] This embodiment discloses a non-transitory computer-readable medium storing instructions. When the instructions are executed by a processor, the steps of the fan bolt fault identification method based on lightweight deep learning target detection described in Example 1 are performed.

[0067] The non-transitory computer-readable medium in this embodiment can be a disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), a USB flash drive, a mobile hard disk, or other media.

[0068] Example 4

[0069] This embodiment discloses a computing device, including a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the fan bolt fault identification method based on lightweight deep learning target detection described in Example 1 is implemented.

[0070] The computing device described in this embodiment may be a desktop computer, a laptop computer, a smart phone, a PDA handheld terminal, a tablet computer, a programmable logic controller (PLC), or other terminal devices with a processor function.

[0071] The embodiments described above are only preferred embodiments of the present invention and are not intended to limit the scope of implementation of the present invention. Therefore, any changes made based on the shape and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fan bolt fault identification method based on lightweight deep learning target detection is characterized by: The following steps are involved: S1. Collecting video image data of wind turbine bolts in different time periods and weather conditions using a camera device; S2. Divide the image data obtained in step S1 into two groups, A and B, and annotate them. The image data in group A is annotated with normal bolts and dropped bolts, while the image data in group B is only annotated with normal bolts and no annotation is made for dropped bolts. S3. Build datasets for the two sets of labeled image data, A and B, respectively. Divide each dataset into a training set and a validation set at a preset ratio, so that the image size in the dataset remains within a preset range. Use the datasets as input to train fan bolt fault recognition models A and B. S4. For the trained fan bolt fault recognition model B, manually remove the image data containing dropped bolts from its validation set, and then input the validation set after the removal into the fan bolt fault recognition model B to obtain a basis matrix for comparison; S5. Deploy fan bolt fault identification model A and fan bolt fault identification model B to the processor; S6. Input the real-time collected fan bolt video into the fan bolt fault identification model A for detection; if the ratio of the number of abnormal image frames of the fan bolt video to the total number of image frames is greater than a preset threshold, the fan bolt video is considered to be abnormal, and an abnormality alarm is directly output; If the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is not greater than the preset threshold, the process proceeds to step S7; S7. Input the real-time collected fan bolt video into the fan bolt fault recognition model B for detection. The fan bolt fault recognition model B outputs a corresponding matrix. Based on the base matrix obtained in step S4, the matrix is ​​sorted using the IoU method in deep learning and the confidence levels of the two are compared. If the confidence level of a certain position in the matrix is ​​less than 3% of the confidence level of the same position in the base matrix, and the ratio of the number of image frames at that position to the total number of image frames is greater than a preset threshold, it is considered that the fan bolt video is abnormal and an abnormality alarm is output. Otherwise, it is considered that there is no fault in the real-time fan bolts.

2. The fan bolt fault identification method based on lightweight deep learning target detection according to claim 1 is characterized in that: The step S1 comprises: Fixed industrial cameras in wind farms collect video image data, which includes images of various time periods in a 24-hour period and images of various weather conditions. The images of the various time periods include daytime images and nighttime images, and the images of various weather conditions include cloudy, sunny, rainy, snowy, and dusty weather.

3. The fan bolt fault identification method based on lightweight deep learning target detection according to claim 1 is characterized in that: The step S3 comprises: Data sets were constructed for the annotated image data sets A and B respectively, and the data sets A and B were divided into training set and validation set in a ratio of 8:2, so that the image size in the data set was kept in the range of [640*640,960*960,1080*1080]. The data sets were used as input, Pytorch was selected as the training framework, and the yolov11-s architecture was used as the target detector. The fan bolt fault recognition model A and the fan bolt fault recognition model B were obtained through training.

4. The fan bolt fault identification method based on lightweight deep learning target detection according to claim 1 is characterized in that: The step S4 comprises: For the trained fan bolt fault recognition model B, the image data with fallen bolts in its validation set were manually removed, and then the validation set after removal was input into the fan bolt fault recognition model B to obtain a basis matrix [N,5] for comparison, where N represents the number of bolts detected, the first four columns [N,4] represent the four-point coordinates of each bolt, and the last column [N,1] is the mean confidence value of each bolt in the entire test set.

5. The fan bolt fault identification method based on lightweight deep learning target detection according to claim 1 is characterized in that: The step S6 comprises: The real-time collected fan bolt video is input into the fan bolt fault identification model A for detection; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is greater than 30%, it is considered that there is an abnormality in the fan bolt video, and an abnormal alarm is directly output; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is not greater than 30%, step S7 is entered.

6. The method for identifying fan bolt faults based on lightweight deep learning target detection according to claim 1 is characterized in that: The step S7 comprises: The real-time collected fan bolt video is input into the fan bolt fault recognition model B for detection, and the fan bolt fault recognition model B outputs a corresponding matrix; based on the base matrix obtained in step S4, the matrix is ​​sorted using the IoU in deep learning and the confidence of the two is compared. If the confidence of a certain position in the matrix is ​​less than 3% of the confidence of the same position in the base matrix, and the ratio of the number of image frames at this position to the number of all image frames is greater than 30%, then it is considered that there is an abnormality in the fan bolt video and an abnormality alarm is output; otherwise, it is considered that the real-time fan bolt has no fault.

7. A fan bolt fault identification system based on lightweight deep learning target detection is characterized by: A method for identifying fan bolt faults based on lightweight deep learning target detection according to any one of claims 1 to 6, comprising: The intelligent camera management module has built-in management software for implementing the RTSP protocol video stream pulling function, providing a visual interface to configure camera parameters, set the storage path and detection time interval operating parameters, and has a camera connection test function; The lightweight algorithm analysis module implements a bolt fault identification model based on target detection technology, automatically performs video stream analysis at preset time intervals, and has a built-in intelligent early warning trigger mechanism that automatically records abnormal events and generates alarms, supporting visual display of results.

8. The fan bolt fault identification system based on lightweight deep learning target detection according to claim 7 is characterized in that: The lightweight algorithm analysis module includes: A dataset was constructed for the labeled image data sets A and B. The datasets were divided into training and validation sets at a ratio of 8:2, so that the image sizes in the datasets were kept within the range of [640*640, 960*960, 1080*1080]. The datasets were used as input, the Pytorch training framework was selected, and the yolov11-s architecture was used as the target detector. The fan bolt fault recognition model A and the fan bolt fault recognition model B were trained. For the trained fan bolt fault recognition model B, we manually removed the image data containing dropped bolts from its validation set. The validation set after removal was then input into the fan bolt fault recognition model B to obtain a basis matrix [N, 5] for comparison. N represents the number of bolts detected, the first four columns [N, 4] represent the four-point coordinates of each bolt, and the last column [N, 1] is the mean confidence score for each bolt in the entire test set. The real-time collected fan bolt video is input into the fan bolt fault recognition model A for detection; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is greater than 30%, it is considered that the fan bolt video is abnormal, and an abnormal alarm is directly output; if the ratio of the number of abnormal image frames of the fan bolt video to the number of all image frames is not greater than 30%, the real-time collected fan bolt video is input into the fan bolt fault recognition model B for detection, and the fan bolt fault recognition model B outputs a corresponding matrix; based on the base matrix obtained in step S4, the IoU in deep learning is used to sort the matrix and compare the confidence of the two. If the confidence of a certain position in the matrix is ​​less than 3% of the confidence of the same position in the base matrix, and the ratio of the number of image frames at the position to the number of all image frames is greater than 30%, it is considered that the fan bolt video is abnormal, and an abnormal alarm is output; otherwise, it is considered that the real-time fan bolt has no fault.

9. A non-transitory computer-readable medium storing instructions, characterized in that: When the instruction is executed by the processor, the steps of the fan bolt fault identification method based on lightweight deep learning target detection according to any one of claims 1 to 6 are performed.

10. A computing device comprising a processor and a memory for storing a program executable by the processor, characterized in that When the processor executes the program stored in the memory, it implements the fan bolt fault identification method based on lightweight deep learning target detection described in any one of claims 1 to 6.

Images