Method and system for detecting looseness of rotor bolt of high-rotating-speed hydraulic generator

Abstract

The invention discloses a high-rotating-speed hydraulic generator rotor bolt looseness detection method and system, belongs to the technical field of hydropower station high-rotating-speed rotating mechanical equipment detection, and aims at solving the technical problems that hydraulic generator rotor bolt looseness detection is low in efficiency and poor in precision, real-time monitoring cannot be achieved, and safe and stable operation of a hydropower station is seriously affected. According to the invention, the image of the rotor bolt area is collected, and the image is accurately processed by using an image correction algorithm, so that image distortion is eliminated; thirdly, fusing an improved YOLOv8 target detection technology, a DeepLabV3 + semantic segmentation technology and a DeepSort multi-target tracking technology, and accurately extracting angle features of the bolt marking line; and then, a looseness identification objective function is constructed based on the angle change of the bolt, the function is solved in real time, and a detection result is output. According to the invention, high-precision, non-contact and real-time bolt looseness detection is realized, the detection efficiency and reliability are effectively improved, and a solid guarantee is provided for safe operation of a hydropower station.

Description

Technical Field

[0001] This invention relates to the field of hydropower station equipment condition monitoring and fault diagnosis technology, and in particular to a method and system for detecting loose rotor bolts of high-speed hydro-generators. Background Technology

[0002] In recent years, with the continuous growth of global demand for renewable energy, hydropower, as a clean and sustainable power generation method, has received increasingly widespread attention and application. In hydro-generator units, various high-speed mechanical devices, such as rotors, main shafts, and oil pump motors, play a crucial role. However, due to their high-speed rotation and special working environment, these high-speed devices face the potential threat of surface faults and defects during long-term operation. Specifically, during prolonged operation, the rotor of a hydro-generator may experience abnormal conditions such as loose bolts, oil leaks, or fluid leaks. These potential problems can not only lead to decreased equipment performance and imbalance, but in extreme cases, may even cause equipment failure, shutdown, and major accidents, seriously impacting the safety and economic efficiency of the hydropower station. Therefore, real-time and accurate monitoring of the loosening status of the rotor bolts of hydro-generators is particularly important.

[0003] Currently, traditional visual inspection methods and conventional monitoring techniques have significant limitations in detecting bolt loosening in high-speed equipment, especially the generator rotor, of hydro-generator units. These methods often rely on manual inspections, which are not only inefficient but also difficult to accurately detect bolt loosening under high-speed rotation. Furthermore, due to the unique operating conditions of the equipment, conventional non-contact technologies such as image analysis cannot be directly applied to the accurate monitoring and evaluation of operational status. Although some patented technologies involve bolt loosening detection, such as those using vibration analysis and sound recognition, their application in the high-speed, complex environment of hydro-generator rotors is not ideal. Vibration analysis methods may be affected by the equipment's own vibration characteristics, leading to false alarms or missed alarms; sound recognition methods may be affected by environmental noise, reducing detection accuracy. In addition, these methods often cannot provide specific information on the location and extent of bolt loosening, limiting their application in fault diagnosis and preventative maintenance.

[0004] In the field of wind power generation, there is also a need for bolt loosening detection. For example, CN120576053B discloses a vision-based quantitative identification method for loose bolts in wind turbine towers. This method uses a high-speed camera to detect the displacement deformation of the top of the tower and the flange plane where the bolts are located during wind turbine operation to identify the loosening of the tower bolts. However, this method is mainly aimed at wind turbine towers, and its application scenarios and equipment characteristics differ from those of hydro turbine generators, so it cannot be directly applied to the bolt loosening detection of hydro turbine generator rotors.

[0005] Another example is CN112945437A, which describes a bolt loosening detection device based on visual monitoring. This device utilizes visual monitoring and a moving unit to detect bolt loosening. Although this device has the advantages of being portable and adaptable to various working conditions, it mainly relies on identifying the anti-loosening lines on the bolt head or nut, and high-precision, real-time detection in complex environments still presents challenges.

[0006] For example, CN118196069A proposes a bolt loosening detection system and method that utilizes the YOLOv5 deep learning framework and ORB algorithm for bolt loosening detection. It transforms the bolt region into an easily analyzable form through grayscale and binarization processing. While this method has achieved some success in bolt loosening detection, it primarily targets bolt detection in static or low-speed motion scenarios. For bolt loosening detection in high-speed rotating equipment such as hydroelectric generator rotors, its accuracy and real-time performance still need improvement.

[0007] In summary, existing technologies for detecting loose bolts on hydro-generator rotors still have many shortcomings, such as low detection efficiency, poor accuracy, and the inability to monitor in real time. Therefore, there is an urgent need for an efficient, accurate, and real-time detection method to solve these problems. This invention aims to fill this technological gap by combining high-speed imaging technology with deep learning algorithms to achieve high-precision, non-contact, and real-time detection of loose bolts on high-speed hydro-generator rotors. This technology not only improves detection efficiency and reliability, ensuring the safe operation of hydropower stations, but also reduces maintenance costs and human resource investment, and improves equipment reliability and service life. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a method and system for detecting loose rotor bolts of high-speed hydro-generators, which solves the problems of low efficiency, poor accuracy and inability to monitor in real time in traditional methods for detecting loose rotor bolts of high-speed hydro-generators. By deeply integrating high-speed camera technology with deep learning algorithms, the detection efficiency and reliability are effectively improved, ensuring the safe operation of high-speed equipment in hydropower stations.

[0009] To achieve the above technical objectives, the technical solution adopted by the present invention is as follows: A method for detecting loose rotor bolts in a high-speed hydro-generator includes the following steps: Step 1: Image Acquisition and Preprocessing The monitoring range of the rotor bolts of the high-speed hydro-generator was defined, and high-speed camera technology was used to acquire bolt image data during rotor operation. Image distortion was then processed using image correction algorithms. The specific process is as follows: Step 1.1 Monitoring Area Planning: Based on the distribution density and actual location of the rotor bolts, divide the area into multiple monitoring sub-areas to ensure that all target bolts are covered and there are no blind spots in monitoring; Step 1.2, High-speed camera parameter configuration: For the rotor speed range of 60~150r / min, calculate and set the high-speed camera parameters to ensure that ≥2 clear bolt images can be captured for each rotor rotation cycle. The high-speed industrial camera used has a resolution of 2048×1024 and a maximum frame rate of 1000fps. It is also compatible with wide-angle or fisheye lenses. The camera is installed 1m below the rotor and the bolt area images are acquired in a vertical shooting mode. Step 1.3, Auxiliary Equipment Deployment: Configure anti-vibration brackets for the camera to eliminate the interference of equipment vibration on image quality; use a hybrid supplementary lighting source composed of LED light source and infrared light source, and adjust the brightness in real time according to the on-site lighting conditions to ensure sufficient lighting in the bolt area and guarantee image clarity; Step 1.4, Image Quality Optimization: The acquired raw images are processed sequentially—distortion correction is performed based on the camera intrinsic parameter matrix, noise reduction is achieved using a Gaussian filtering algorithm, and illumination enhancement is achieved through a contrast adaptive adjustment algorithm, thus completely eliminating image quality problems caused by shooting distance and environmental interference.

[0010] Step 2: Construction of Bolt Loosening Fault Feature Dataset A dataset of bolt loosening fault features covering different degrees of loosening was constructed to support the training of deep learning models. The specific process is as follows: Step 2.1, Image Sample Acquisition: During unit maintenance shutdown, simulate different operating conditions (including different speeds and different lighting conditions) and acquire three types of image samples of bolts using a high-speed camera: normal state image, slightly loose state image (corresponding to bolt angle change of 0°~2°), and severely loose state image (corresponding to bolt angle change >2°). Step 2.2, Image annotation and dataset partitioning: Use the Labelimg tool to annotate the acquired images, marking the bolt positions and anti-loosening line marks in sequence, generating a dataset that simultaneously contains labels for bolt detection and image segmentation tasks, and divide the dataset into training set, validation set and test set according to a preset ratio for model training, validation and performance testing. Step 2.3, Dataset Expansion: Data augmentation techniques are used to expand the existing dataset, including operations such as image rotation, image scaling, and illumination perturbation, to improve the generalization ability of the deep learning model and avoid model overfitting.

[0011] Step 3: Bolt detection, segmentation, and angle calculation based on deep learning Based on a deep learning model, bolt target detection, semantic segmentation, and multi-target tracking are performed on the preprocessed image data. Bolt marker points are identified and angle changes are calculated. The specific process is as follows: Step 3.1 Bolt Target Detection: An improved YOLOv8 target detection model is adopted. The traditional convolutional layers of the model are replaced by an adaptive kernel convolution (AKConv) module to enhance the extraction capability of bolt edge features. At the same time, a global attention mechanism (GAM) is inserted into the backbone network of the model to make the model focus on the bolt marker point region, thereby improving the accuracy and efficiency of bolt target detection. Step 3.2 Bolt semantic segmentation: The DeepLabV3+ semantic segmentation model is adopted, which uses dilated convolution to capture multi-scale contextual information of the image. The features are processed through an encoder-decoder structure: the encoder extracts deep features, and the decoder upsamples and encodes features and fuses them with low-level features to achieve fine segmentation of the bolt and the background, accurately separating the bolt region from the interfering background. Step 3.3, Bolt Multi-Target Tracking: DeepSort multi-target tracking technology is adopted, combined with Kalman filter algorithm to predict bolt motion trajectory. At the same time, deep learning is used to extract bolt feature information and perform feature matching to realize cross-frame target association of individual bolts under high-speed rotation state, ensuring that each bolt can be continuously identified and tracked, and avoiding target loss. Step 3.4 Bolt Angle Calculation: Based on the segmentation results output by the DeepLabV3+ semantic segmentation model, extract the center coordinates of the two marked points on the bolt. , According to formula (1): (1) Calculate the real-time angle of each bolt The angle values ​​of each bolt are recorded for each revolution of the rotor, and the change in bolt angle is obtained.

[0012] Step 4: Loosening Assessment and Early Warning The bolt loosening status is determined based on the bolt angle change threshold. Detection results are generated and warnings are issued by combining real-time and historical data. The specific process is as follows: Step 4.1, Loosening Judgment Criteria: Set Bolt Angle Change Threshold If the bolt angle changes over three consecutive rotor rotation cycles: (2) in, For the first Real-time bolt angle per rotation cycle For the first The bolt is considered to be in a loose state if the bolt's real-time angle is measured over one rotation cycle. Step 4.2, Multi-level early warning mechanism: Establish a hierarchical early warning system—when the bolt is determined to be slightly loose, a local audible and visual alarm is triggered to remind on-site personnel to check in time; when the bolt is determined to be severely loose, in addition to triggering the local audible and visual alarm, an early warning information is also pushed to the remote monitoring platform at the same time, and the location coordinates of the abnormal bolt are marked to achieve multi-terminal collaborative early warning; Step 4.3, Model Optimization and Visualization Platform: Transfer learning is used to initialize the parameters of the deep learning model, shortening the initial training time. Bolt image samples under new working conditions are incorporated into the model training through incremental learning, quickly adapting to the new operating environment and shortening the model adaptation time. A visualization platform is developed based on C++ and Qt, integrating TensorRT (TuringTensor Runtime, a deep learning inference acceleration engine) to accelerate the inference engine, ensuring that the processing time of a single frame image is ≤50ms, realizing the real-time display of detection results. At the same time, the platform can display the bolt status heat map and angle change curve, making it easy for staff to intuitively grasp the bolt operating status.

[0013] This invention also provides a high-speed hydro-generator rotor bolt loosening detection system. The system includes an image acquisition module, a data transmission module, an intelligent analysis module, and an early warning and management module. The functions of each module are as follows: Image acquisition module: Composed of a high-speed industrial camera (resolution 2048×1024, maximum frame rate 1000fps), a wide-angle lens or fisheye lens, a hybrid supplementary light source consisting of LED and infrared, and a shockproof bracket, it is responsible for capturing transient images of the rotor bolt area in real time, providing raw data for subsequent analysis; Data transmission module: Built on a 10 Gigabit Ethernet architecture, using optical fiber as the transmission medium, with a data transmission rate of ≥10Gbps, it can transmit high-speed image data captured by the image acquisition module to the intelligent analysis module in real time, avoiding data transmission delay from affecting real-time monitoring; The intelligent analysis module includes a preprocessing submodule, a detection and tracking submodule, and a looseness judgment submodule. The preprocessing submodule performs image distortion correction, denoising, and enhancement operations; the detection and tracking submodule deploys an improved YOLOv8 target detection model, a DeepLabV3+ semantic segmentation model, and DeepSort multi-target tracking technology to complete bolt detection, segmentation, and tracking; the looseness judgment submodule extracts the coordinates of bolt marker points and calculates angle changes, and combines them with a preset threshold to determine whether the bolt is loose. Early warning and management module: includes an early warning unit, a data storage unit, and a visualization unit. The early warning unit enables local audible and visual alarms and remote early warning information push. The data storage unit uses a distributed database to store historical and real-time detection data, facilitating subsequent traceability and analysis. The visualization unit displays bolt status heatmaps and angle change curves, intuitively presenting the bolt's operating status.

[0014] The method and system for detecting loose rotor bolts in high-speed hydro-generators provided by this invention have the following beneficial effects: 1. This invention effectively solves the problems of low efficiency, poor accuracy, and inability to monitor in real time in traditional high-speed hydro-generator rotor bolt loosening detection methods. By deeply integrating high-speed camera technology with deep learning algorithms, it significantly improves detection efficiency and reliability, providing a solid guarantee for the safe operation of high-speed equipment in hydropower stations. It is especially suitable for the condition monitoring and intelligent operation and maintenance scenarios of high-speed equipment in hydropower stations.

[0015] 2. This invention combines the improved YOLOv8 target detection model, the DeepLabV3+ semantic segmentation model, and the DeepSort multi-target tracking technology to achieve high-precision identification of bolt targets, fine segmentation of bolt areas, and continuous tracking of bolt movement. Combined with angle change calculation, it can accurately determine the loosening state of bolts, effectively avoiding misjudgment and omission by traditional methods.

[0016] 3. This invention has extremely strong real-time performance. High-speed camera technology ensures the immediacy of image acquisition, 10 Gigabit Ethernet architecture enables high-speed data transmission, and the application of TensorRT accelerated inference engine ensures that the processing time of a single frame image does not exceed 50ms. This enables real-time output of detection results and triggering of early warning mechanisms, solving the problem that traditional timed detection cannot detect faults in a timely manner.

[0017] 4. The non-contact detection method of the present invention eliminates the need for direct contact between the camera and the rotor, avoiding interference with the normal operation of the equipment during the detection process, and also avoids the problem of contact detection being affected by equipment vibration.

[0018] 5. The present invention has a high degree of automation and intelligence. From image acquisition and processing to loosening judgment and early warning, no manual intervention is required throughout the entire process. The deep learning model can learn bolt features autonomously and adapt to new working conditions through incremental learning, which greatly reduces the burden on workers.

[0019] 6. The present invention demonstrates excellent economic efficiency and safety, and can promptly detect bolt loosening problems, avoiding equipment downtime or major accidents caused by bolt loosening, thereby reducing maintenance costs and economic losses, while ensuring the safe operation of the hydropower station.

[0020] 7. This invention achieves high-precision, non-contact, real-time detection of the loosening state of rotor bolts of hydro-generators, providing strong support for the safe operation of hydropower station equipment.

[0021] 8. By combining high-speed camera technology with deep learning algorithms, this invention not only improves the accuracy and reliability of detection, but also significantly enhances detection efficiency, making the detection process faster and more accurate.

[0022] 9. The system of the present invention can quickly process image data captured by high-speed cameras, realize real-time bolt status monitoring and identification, promptly detect and warn of bolt loosening problems, effectively avoid safety accidents and equipment damage caused by bolt loosening, and provide a solid backing for the safe operation of hydropower stations.

[0023] 10. The efficient and accurate image analysis of this invention benefits from the combination of the YOLOv8 target detection model and DeepSort multi-target tracking technology. Compared with traditional image processing methods, it can more accurately identify the loosening state of rotor bolts, thus improving the accuracy and reliability of the analysis.

[0024] 11. This invention can quickly respond to image data captured by a high-speed camera, realize real-time bolt status monitoring and identification. Compared with traditional manual inspection or timed detection, it can detect the loosening of rotor bolts in a timely manner and issue early warnings, which helps to avoid safety accidents and equipment damage caused by loose bolts.

[0025] 12. This invention enables automatic identification and tracking of bolt loosening status, completing the monitoring task without manual intervention. The application of deep learning models enables the system to automatically learn and identify key features in images, exhibiting a high level of intelligence and greatly reducing the workload of operators.

[0026] 13. The application of this invention significantly improves the reliability and safety of hydro-generator equipment. By promptly detecting and addressing the problem of loose rotor bolts, it can effectively prevent equipment failures and accidents caused by them, thus ensuring the normal operation of the equipment and production efficiency.

[0027] 14. This invention enables automatic monitoring and early warning of equipment status, avoiding the blindness and time-consuming nature of traditional periodic inspections and maintenance methods, thereby reducing maintenance costs and human resource investment, and improving maintenance efficiency and management level.

[0028] 15. This invention replaces the traditional convolutional layer with an adaptive kernel convolution (AKConv) module, which enhances the ability to extract bolt edge features. At the same time, a global attention mechanism (GAM) is inserted into the backbone network to focus on the bolt marker point region, further improving the performance of the YOLOv8 target detection model.

[0029] 16. This invention utilizes dilated convolution to capture multi-scale contextual information, finely segments bolts and background through an encoder-decoder structure, and employs the DeepLabV3+ semantic segmentation model, thereby improving the accuracy and detail representation of image segmentation.

[0030] 17. This invention combines Kalman filtering to predict bolt motion trajectories with deep learning feature matching to achieve cross-frame target association, ensuring continuous identification of individual bolts under high-speed rotation. It also employs DeepSort multi-target tracking technology to improve the stability and accuracy of tracking.

[0031] 18. This invention uses transfer learning to initialize model parameters and incorporates new working condition samples through incremental learning, which shortens the model adaptation time. At the same time, a visualization platform was developed based on C++ and Qt, and the TensorRT acceleration inference engine was integrated to ensure that the processing time of a single frame image does not exceed 50ms, thereby improving the overall performance of the system. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the overall technical process of the present invention. Figure 2 A schematic diagram of the improved YOLOv8 model structure of this invention; Figure 3 This is a schematic diagram of the DeepLabV3+ semantic segmentation model structure of the present invention; Figure 4 This is a schematic diagram of bolt marking and angle calculation according to the present invention; Figure 5 This is a flowchart of the bolt loosening detection process of the present invention; Figure 6 This is a diagram of the overall system architecture of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments: Example 1 This embodiment provides a method for detecting loose rotor bolts in a high-speed hydro generator, including the following steps: Step 1: Delineate the monitoring area Before implementing this testing method, it is necessary to determine the bolt areas to be monitored based on the specific structural drawings of the high-speed hydro-generator and the site survey. Considering the complex structure of the hydro-generator rotor and the wide distribution of bolts, it is necessary to delineate key monitoring areas based on the rotor's critical components and stress conditions, such as the rotor magnetic pole connection bolt area and the rotor support connection bolt area. Furthermore, to ensure comprehensive monitoring, multiple monitoring points should be set up within each key monitoring area to cover all critical bolts.

[0034] Step 2: Acquire rotor running images A high-speed camera system is used to acquire images within the designated monitoring area. This system must be capable of capturing at least several thousand frames per second to ensure it can capture minute changes in the rotor during high-speed rotation (e.g., hundreds to thousands of revolutions per minute). Simultaneously, the system's resolution must be high enough to clearly display the bolt markings. During acquisition, the camera system must be stably installed to prevent image blurring due to vibration or movement. Furthermore, the exposure time and gain of the camera system must be adjusted according to the ambient lighting conditions to obtain clear images.

[0035] Step 3: Image Correction and Enhancement The acquired raw images undergo distortion correction and illumination enhancement. Distortion correction employs a grid-based algorithm, which uses a pre-established distortion correction model to transform the coordinates of each pixel in the image, eliminating image distortion caused by lens distortion. Illumination enhancement utilizes a histogram equalization algorithm, which adjusts the image's grayscale distribution to improve contrast and brightness, making the bolt marking lines more clearly visible. Furthermore, filtering algorithms are used to remove noise from the image, further improving image quality.

[0036] Step 4: Build the dataset During unit maintenance shutdowns, the operating status of the hydro-generator is simulated, and bolt images are collected under different degrees of loosening. The degree of loosening can be simulated by manually adjusting the bolt tightening torque, such as setting the tightening torque to 80%, 60%, and 40% of the rated torque, to simulate different levels of bolt loosening. The collected images need to be labeled, indicating the bolt position, the angle of the marking line, and other information. These images are divided into training, validation, and test sets. The training set is used for model training, the validation set for model tuning, and the test set for model evaluation. The construction of the dataset must ensure the diversity and representativeness of the image samples to improve the model's generalization ability.

[0037] Step 5: Model Training and Optimization An improved YOLOv8 object detection model, a DeepLabV3+ semantic segmentation model, and a DeepSort multi-object tracking model were jointly trained using a training set. The improved YOLOv8 model was used to detect bolt positions in images, improving the detection accuracy for small bolts by adding a small object detection layer and enhancing feature fusion capabilities. The DeepLabV3+ model was used to segment bolt marker lines, improving segmentation accuracy by introducing dilated convolution and attention mechanisms. The DeepSort model was used to track bolt position changes across consecutive frames, improving tracking stability and accuracy by incorporating deep feature matching and Kalman filtering algorithms. During training, hyperparameters such as learning rate, batch size, and number of training epochs were adjusted to optimize model performance. Simultaneously, the model was validated using a validation set, and the model structure or hyperparameters were adjusted based on the validation results to ensure good performance even on unseen data.

[0038] Step 6: Feature Extraction and Loosening Identification The trained model is applied to real-time acquired rotor images to extract the angular features of bolt marking lines. First, an improved YOLOv8 model is used to detect bolt positions in the image. Then, a DeepLabV3+ model is used to segment the bolt marking lines. Finally, the angular features of the bolts are obtained by calculating the angle between the marking lines and the horizontal direction. Based on these features, a bolt loosening detection objective function is constructed. This function considers factors such as the range and rate of change of the bolt angle. By solving this function in real time, it is determined whether the bolt has become loose. If the bolt angle change exceeds a preset threshold (e.g., 5 degrees) and the rate of change exceeds a preset value (e.g., more than 1 degree per second), the bolt is determined to be loose, and the detection result is output.

[0039] Step 7: Result Output and Early Warning The inspection results are output in a visual format, displaying information such as the bolt's looseness status (loose or normal) and the degree of looseness (angle change). Simultaneously, an early warning mechanism is set up to promptly issue a warning signal when bolt loosening is detected. The warning signal can be sent to maintenance personnel via sound, light, or SMS to ensure timely information delivery. Furthermore, the inspection results can be uploaded to a cloud server for remote monitoring and data analysis.

[0040] Example 2 In another preferred embodiment, based on Embodiment 1, this embodiment provides a method for detecting loose rotor bolts of a high-speed hydro-generator. The method is based on a 120MW high-speed hydro-generator (rotor speed 100r / min) at a hydropower station, and the specific steps are as follows: Step 1: Image Acquisition and Preprocessing Step 1.1 Monitoring Area Planning: The turbine generator rotor has a total of 36 sets of bolts, which are evenly distributed around the circumference. Based on this, 6 monitoring sub-areas are divided (each area covers 6 sets of bolts) to ensure that there are no blind spots in monitoring, while avoiding interference areas such as the rotor main shaft and wiring terminals.

[0041] Step 1.2, High-speed camera system deployment: Select a high-speed industrial camera (model: Basler acA2040-90um) with a resolution of 2048×1024 and a maximum frame rate of 1000fps, and a 12mm wide-angle lens. Fix it 1m directly below the rotor using a shockproof bracket, with the lens vertically aligned with the bolt area. For supplementary lighting, use a combination of 2 sets of LED floodlights (50W power, dimmable) and 1 set of infrared light source (wavelength 850nm), arranged at a 45° angle on both sides of the camera. Adjust the brightness in real time according to the ambient light (natural light during the day + machine room lighting at night) to ensure that the bolt anti-loosening line markings are clearly visible.

[0042] Step 1.3, Camera Parameter Calibration: Since the rotor speed is 100 r / min (i.e., 1.67 r / s), it is calculated that ≥2 clear images need to be captured per rotation cycle (0.6 s). Therefore, the camera frame rate is set to 5 fps, and the exposure time to 100 μs to avoid image blurring. Simultaneously, the intrinsic parameter matrix (focal length) is obtained through the camera's built-in software. , Principal point coordinates , ), used for subsequent distortion correction.

[0043] Step 1.4, Image Preprocessing: The acquired raw images are processed sequentially—distortion correction is performed using Zhang Zhengyou calibration method based on the camera intrinsic parameter matrix to eliminate barrel distortion of the wide-angle lens; noise reduction is performed using a 3×3 Gaussian filter kernel to filter image noise caused by vibration of the equipment in the machine room; the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm is used to enhance image contrast, increasing the grayscale difference between the bolt anti-loosening line and the bolt body from the original 15~20 to 30~40, ensuring that the marker points are identifiable.

[0044] Step 2: Construction of Bolt Loosening Fault Feature Dataset Step 2.1, Sample Collection: During the unit's annual maintenance shutdown (72 hours), three operating conditions were simulated (speed 80 r / min, 100 r / min, 120 r / min; light intensity 200 lux, 500 lux, 1000 lux), and bolt samples were collected using a high-speed camera. Normal condition: All 36 bolts are tightened to the designed torque (800 N·m), the anti-loosening line is not deviated, and 500 images are collected; Slight loosening: Reduce the torque of 10 bolts to 600 N·m using a torque wrench, corresponding to a change in the anti-loosening line angle of 0.5°~2°, and collect 600 images; Severe loosening: Reduce the torque of 5 sets of bolts to 400 N·m, corresponding to a change in the anti-loosening line angle of 2.5°~5°, and collect 400 images.

[0045] Step 2.2, Data Labeling and Division: Use the Labelimg tool to label the images—first label the bounding rectangle of the bolt (object detection label), then label the two endpoints of the anti-loosening line (segmentation label), and generate an XML (eXtensible Markup Language) format label file; divide the 1500 labeled images into a training set (1050 images), a validation set (300 images), and a test set (150 images) in a 7:2:1 ratio.

[0046] Step 2.3, Data Augmentation: Data augmentation operations are performed on the training set, including random rotation (-10°~10°), scaling (0.8~1.2 times), and lighting perturbation (brightness ±30%, contrast ±20%). After expansion, the training set size reaches 3150 images, avoiding model overfitting.

[0047] Step 3: Bolt detection, segmentation, and angle calculation based on deep learning Step 3.1: Improve YOLOv8 object detection: An improved YOLOv8 model was built using the PyTorch framework. Convolutional layers 2-5 of the original model's backbone were replaced with AKConv (Adaptive Kernel Convolution) modules to enhance feature extraction capabilities for bolt edges (especially anti-loosening lines). A GAM attention mechanism was inserted into the neck layer, increasing the model's attention to bolt marker areas by 40%. The model was trained using the training set for 100 iterations. The initial learning rate was set to 0.01, decreasing to 1 / 10 of the original rate every 20 iterations. The final bolt detection accuracy (mAP@0.5) on the test set reached 98.6%, with a detection speed of 30fps.

[0048] Step 3.2, DeepLabV3+ Semantic Segmentation: Configure the DeepLabV3+ model. The encoder uses ResNet50 as the backbone network, and the dilated convolution rate is set to 1, 6, 12, and 18 to capture multi-scale bolt features. The decoder fuses the encoded features with the C4 layer features of the backbone network through 2x upsampling to achieve fine segmentation of the bolt and the background. The model achieves a segmentation intersection-over-union (mIoU) ratio of 97.2% on the test set, which can accurately separate the bolt body from the rotor hub background.

[0049] Step 3.3, DeepSort Multi-Target Tracking: Based on the detection and segmentation results, DeepSort is used for bolt tracking. By improving YOLOv8, the 128-dimensional feature vector of the bolt is extracted. Combined with Kalman filtering, the position of the bolt in the next frame is predicted (because the rotor rotates at a constant speed, the motion model is set to a constant speed linear model). Cosine similarity (threshold 0.7) is used for feature matching to achieve continuous tracking of 36 bolts across frames. The tracking success rate is 99.1%, with no target loss or mismatch.

[0050] Step 3.4, Angle Calculation: Extract the center coordinates of the two marker points of each bolt anti-loosening line from the DeepLabV3+ segmentation results (e.g., the coordinates of a normal bolt marker point are (856, 423), (912, 425)), substitute them into formula (1) to calculate the real-time angle, record the bolt angle value for each rotor rotation cycle (0.6s), and form an angle sequence (e.g., the normal bolt angle sequence is 1.2°, 1.3°, 1.2°, and the slightly loose bolt angle sequence is 1.5°, 2.8°, 3.9°): (2) In the formula, Angle change threshold, For the first Real-time bolt angle per rotation cycle For the first Real-time bolt angle per rotation cycle.

[0051] Step 4: Loosening Assessment and Early Warning Step 4.1, Loosening Judgment: Set an angle change threshold. Real-time analysis of the angle sequence is performed. If the angle change of a bolt within three consecutive rotation cycles (1.8s) is... (For example, if the angle of a bolt changes from 2.1° to 4.5° to 7.2°, with changes of 2.4° and 2.7° respectively), it is considered loose. An angle change of 2° to 3° is considered slightly loose, and a change of more than 3° is considered severely loose.

[0052] Step 4.2, Multi-level Early Warning: When the bolt is slightly loose, the on-site control cabinet triggers an audible and visual alarm (red indicator light flashes, buzzer frequency 1Hz); when the bolt is severely loose, in addition to the on-site alarm, the early warning information (including bolt number, location coordinates, and angle change) is pushed to the remote monitoring platform [Web (World Wide Web) terminal + mobile terminal] via MQTT (Message Queuing Telemetry Transport) protocol, and the location of the abnormal bolt is marked on the platform map at the same time.

[0053] Step 4.3, Model and Platform Optimization: Transfer learning was used to initialize and improve the parameters of YOLOv8 and DeepLabV3+ (pre-trained weights based on the COCO dataset), reducing the initial training time by 50%; under new working conditions (such as rotation speed of 120 r / min and illumination of 1500 lux), 200 new samples were added through incremental learning, reducing the model adaptation time from 4 hours to 1 hour; the visualization platform was developed based on C++ and Qt, integrating the TensorRT 8.6 accelerated inference engine, with a single frame image processing time stabilized at 42ms. The platform displays the bolt status heatmap (green = normal, yellow = slightly loose, red = severely loose) and angle change curve in real time, and supports querying and exporting historical data (within 3 months).

[0054] Example 3 In another preferred embodiment, based on Embodiment 2, this embodiment provides a method for detecting loose rotor bolts of a high-speed hydro-generator. This embodiment is based on a high-speed hydro-generator (100 r / min) at a hydropower station, combined with... Figures 1 to 5 The technical process and model structure are explained in detail, along with the implementation process of the detection method.

[0055] I. Process Execution according to Figure 1 The workflow of "Start - Building a high-speed camera system - Image preprocessing - Feature extraction - Constructing a loosening recognition objective function - Real-time solution and result output - End" is executed sequentially: 1. Construct a high-speed camera system A high-speed industrial camera with a resolution of 2048×1024 and a frame rate of 500fps was selected, paired with a 12mm wide-angle lens, and mounted on a shockproof bracket 1m below the rotor (as shown in Figure 04, aligned with the rotor bolt area). The supplementary lighting system uses a hybrid LED and infrared light source to ensure that the bolt anti-loosening marks are clearly visible.

[0056] 2. Image preprocessing Distortion correction, denoising, and enhancement are performed on the acquired raw images: distortion correction is completed based on the camera intrinsic parameter matrix (eliminating barrel distortion of wide-angle lenses); denoising is performed using 3×3 Gaussian filtering; contrast is enhanced using the CLAHE algorithm, increasing the grayscale difference between the bolts and the background by more than 30%.

[0057] 3. Feature Extraction Object detection feature extraction: An improved YOLOv8 model (such as...) is used. Figure 2As shown in the figure, AKConv (Adaptive Kernel Convolution) replaces the traditional convolutional layer to enhance the ability to extract bolt edge features; GAMATtention (Global Attention Mechanism) focuses on the bolt marker point region to improve the targeting of feature extraction.

[0058] Semantic segmentation feature extraction: using DeepLabV3+ models (such as...) Figure 3 As shown in the figure, its Encoder uses Atrous Conv to capture multi-scale contextual information, and the Decoder implements feature upsampling and fusion to accurately segment the bolt anti-loosening mark and the background.

[0059] 4. Construct the loosening identification objective function The objective function is constructed based on the change in the angle of the bolt anti-loosening mark: Let the coordinates of the bolt mark point be... , The real-time angle of the bolt is calculated using formula (1). This leads to the change in angle: (1).

[0060] When the angle change over three consecutive periods At that time, it was determined to be loose.

[0061] 5. Real-time solving and result output Combination Figure 5 The judgment process involves real-time analysis of the angle sequence of each bolt: if it is the first detection, it is judged directly by the angle change threshold; if it is a historical detection, it is judged by comparing historical data to determine the loosening trend, and finally outputs the detection result and triggers an early warning.

[0062] II. Model and Algorithm Details 1. Improve the YOLOv8 model (see...) Figure 2 ) The model was built within the PyTorch framework. The C2f module was used for feature fusion, and SPPF (Spatial Pyramid Pooling - Fast) enhanced the global receptive field of the features. During model training, the mAP (Mean Average Precision) reached 98.6%, with mAP@0.5 (mean average precision at an IoU threshold of 0.5) reaching 99.2%, ensuring a high detection rate for bolt targets.

[0063] 2. DeepLabV3+ model (see) Figure 3 ) Its Encoder uses 1×1 Conv and 3×3 Conv (rates 6 / 12 / 18) to implement multi-scale dilated convolution, capturing bolt features from fine to coarse; Image Pooling obtains global features. The Decoder fuses high and low layer features through Upsampling and Concat, achieving a final segmentation accuracy of 97.2% mIoU (Mean Intersection over Union).

[0064] 3. Multi-target tracking and angle calculation (see...) Figure 4 and Figure 5 ) The DeepSort algorithm is used to track bolts across frames, combined with Figure 4 Calculate the angle of the coordinates of the bolt marker point. For example... Figure 5 As shown, by comparing the angle value of each bolt during n rotations, it is determined whether it is loose: if the angle change exceeds the threshold, it is directly judged as "relatively loose"; if it is a historical test, the loosening trend needs to be analyzed to generate a test report containing the loosening angle and trend.

[0065] Example 4 In another preferred embodiment, based on Embodiment 1, this embodiment provides a high-speed hydro-generator rotor bolt loosening detection system. This system is used in the high-speed hydro-generator rotor bolt loosening detection method described in Embodiment 1. The system mainly consists of a high-speed camera module, an image processing module, a deep learning model module, a data storage module, a result output module, and a power supply module. The modules are connected via a high-speed data bus to ensure real-time data transmission and processing. The system adopts a modular design, facilitating installation, maintenance, and upgrades.

[0066] 1. High-speed camera module The high-speed camera module is responsible for capturing bolt images during rotor operation. This module uses a high-resolution, high-frame-rate high-speed camera, such as one capable of shooting thousands of frames per second with a resolution of millions of pixels. The camera is equipped with a stable bracket and image stabilization device to prevent image blurring due to vibration or movement. Simultaneously, the camera also features autofocus and auto-exposure functions, automatically adjusting shooting parameters according to ambient lighting conditions to obtain clear images. Furthermore, the camera supports externally triggered shooting, allowing synchronization with the turbine generator's control system for timed or condition-triggered shooting.

[0067] 2. Image processing module The image processing module performs distortion correction and illumination enhancement on the acquired raw images. This module employs a high-performance image processing chip or FPGA (Field Programmable Gate Array) to implement fast image processing algorithms. The distortion correction algorithm uses a grid-based distortion correction model, transforming the coordinates of each pixel in the image using a pre-established correction parameter table to eliminate lens distortion. The illumination enhancement algorithm uses histogram equalization and filtering algorithms to improve image contrast and brightness while removing noise. The image processing module also features an image caching function, capable of storing multiple frames for subsequent feature extraction and motion detection.

[0068] 3. Deep Learning Model Module The deep learning model module is the core of this system, responsible for extracting the angular features of bolt marking lines and determining whether the bolts are loose. This module employs a high-performance embedded AI (Artificial Intelligence) chip or GPU (Graphics Processing Unit) to jointly run an improved YOLOv8 object detection model, a DeepLabV3+ semantic segmentation model, and a DeepSort multi-object tracking model. The model module connects to the image processing module via a high-speed data bus, receiving processed image data and outputting detection results. The model module also features a model update function, allowing for remote upgrades to update model parameters, improving detection accuracy and adaptability.

[0069] 4. Data storage module The data storage module is responsible for storing the acquired image data, intermediate results during model training, and the final detection results. This module uses high-capacity, high-speed storage devices, such as solid-state drives (SSDs) or enterprise-grade hard drive arrays. The storage devices have data backup and recovery capabilities to ensure data security and reliability. Simultaneously, the data storage module also supports data compression and encryption functions to reduce storage space usage and protect data privacy.

[0070] 5. Result Output Module The results output module displays the inspection results in a visual format, such as showing the looseness status and degree of looseness of the bolts. This module uses a high-resolution display screen or touchscreen, supporting touch operation and multiple display modes. It also features sound and light warnings, issuing timely alerts when bolt loosening is detected. Furthermore, the module supports remote communication, allowing inspection results to be uploaded to a cloud server or sent to maintenance personnel's mobile devices for remote monitoring and data analysis.

[0071] 6. Power Module The power module provides a stable power supply to the entire system. This module employs an uninterruptible power supply (UPS) design, ensuring that the system can continue operating for a period of time during power outages or voltage fluctuations, preventing data loss or equipment damage. Simultaneously, the power module also features overvoltage, overcurrent, and overtemperature protection functions to ensure the safety and reliability of the system.

[0072] 7. System Workflow After system startup, the high-speed camera module begins acquiring bolt images during rotor operation and transmits the image data to the image processing module. The image processing module performs distortion correction and illumination enhancement on the images before transmitting the processed image data to the deep learning model module. The deep learning model module extracts the angular features of the bolt marking lines and determines whether the bolts are loose, transmitting the detection results to the result output module. The result output module displays the detection results in a visual format and issues a warning signal when bolt loosening is detected. Simultaneously, the data storage module stores the acquired image data and detection results for subsequent analysis and processing. The entire system workflow is automated and intelligent, improving detection efficiency and accuracy.

[0073] Example 5 In another preferred embodiment, based on the above embodiments 1 and 2, this embodiment provides a high-speed hydro-generator rotor bolt loosening detection system for implementing the high-speed hydro-generator rotor bolt loosening detection method of embodiment 2. This embodiment corresponds to the rotor bolt detection requirements of a 100r / min hydro-generator in a certain hydropower station. The system architecture and component selection are as follows, which can realize real-time monitoring and early warning of 36 sets of rotor bolts.

[0074] I. System Overall Architecture The system adopts a four-layer architecture of "data acquisition-transmission-analysis-early warning," with each layer deployed independently yet working collaboratively. The specific architecture is as follows: Figure 5 As shown: Acquisition layer: Image acquisition module, responsible for capturing transient images of the bolt; Transport layer: Data transmission module, enabling high-speed transmission of image data; Analysis layer: Intelligent analysis module, which completes image processing and looseness assessment; Application layer: Early warning and management module, which realizes early warning, data storage and visualization.

[0075] II. Detailed Configuration and Functions of Each Module (a) Image acquisition module 1. Selection of core components: High-speed industrial camera: Hikvision MV-CA020-10GM, resolution 2048×1024, maximum frame rate 1000fps, supports GigE (Gigabit Ethernet) interface, dynamic range 60dB, can work stably in dusty environments in computer rooms (protection level IP65). Lens: Tamron 13FM0812FCS-2, 8~12mm manual zoom wide-angle lens, maximum aperture F1.2, distortion rate <1%, compatible with camera sensor size 1 / 1.8 inch; Supplemental lighting source: 2 OPT-LLT-100W-850 infrared light sources (wavelength 850nm, power 100W) + 2 OPT-LLT-50W-White LED light sources (color temperature 5500K, power 50W). The light source controller supports 0-100% brightness adjustment and is linked with the system via RS485 interface. Shockproof bracket: Custom stainless steel bracket with shock-absorbing pads (hardness 50 Shore A) installed at the bottom. The bracket is 1m high and the horizontal and vertical positions of the camera can be adjusted by a fine-tuning knob to ensure that the lens is vertically aligned with the rotor bolt area.

[0076] 2. Functionality: The module captures bolt images in real time via camera trigger signal (external trigger mode, trigger frequency 5fps). The supplementary light source is turned on synchronously according to the camera exposure signal to avoid image blurring caused by light delay. The shockproof bracket eliminates the influence of rotor rotation (vibration acceleration ≤0.5g) on ​​the camera, ensuring that the offset of bolt mark points in the acquired image is <1px.

[0077] (ii) Data transmission module 1. Selection of core components: Switch: Huawei S5735-S48T4X, with 48 Gigabit Ethernet ports + 4 10 Gigabit optical ports, supports IEEE 802.3ab standard, and has a switching capacity of 336Gbps; Fiber optic cable: Single-mode fiber (model GYXTW-4B1) is used, with a transmission distance of 2km (800m from the computer room to the monitoring center), and equipped with SC / PC (Square Connector / Physical Contact) fiber optic connectors. Optical module: Huawei SFP-10G-LR-S, 10Gbps rate, working wavelength 1310nm, transmission distance 10km.

[0078] 2. Functionality: The module is based on a 10 Gigabit Ethernet architecture. The camera is connected to the switch via a gigabit network cable. The switch transmits the image data to the graphics workstation in the monitoring center via fiber optic cable (10 Gigabit optical port). The transmission rate is stable at 10Gbps. The transmission time for a single frame image (2048×1024, 24-bit depth, approximately 6MB) is <0.6ms, with no data packet loss or delay (packet loss rate <0.01%).

[0079] (III) Intelligent Analysis Module 1. Hardware configuration: Graphics workstation: Lenovo ThinkStation P620, CPU is AMD Ryzen Threadripper PRO5975WX (32 cores), GPU is NVIDIA RTX 4090 (24GB VRAM), memory is 128GB DDR4, hard drive is 2TB SSD (for storing models and real-time data). Software environment: Operating system Ubuntu 22.04 LTS, deep learning framework PyTorch 2.0, TensorRT 8.6, OpenCV 4.8.

[0080] 2. Submodule functions: The preprocessing submodule calls the OpenCV (Open Source Computer Vision Library) library to perform image processing. It performs distortion correction based on the camera intrinsic parameter matrix (obtained through Zhang's positive friend calibration method), 3×3 Gaussian filtering for noise reduction, and CLAHE algorithm to enhance contrast. The processing time is about 12ms / frame. The detection and tracking submodule deploys improved YOLOv8 (including AKConv and GAM), DeepLabV3+, and DeepSort algorithms—improving YOLOv8 for bolt position detection (8ms / frame), DeepLabV3+ for segmenting anti-loosening lines (10ms / frame), and DeepSort for bolt tracking (5ms / frame). The three work together to achieve bolt marker point recognition and trajectory tracking. Loosening detection submodule: Write a C++ algorithm to calculate the bolt angle change (time 2ms / frame), compare it with the threshold (2°) to determine the loosening state, and generate the judgment result (including bolt number, status, and angle value). The overall submodule processing time is ≤37ms / frame.

[0081] (iv) Early Warning and Management Module 1. Core components and software: Early warning unit: field control cabinet [including red LED (Light Emitting Diode) alarm light and buzzer], remote monitoring platform (web client developed based on Spring Boot, mobile client developed based on Flutter); Data storage unit: Uses MongoDB distributed database (3-node cluster), with a storage capacity of 10TB, supports 1000 data writes per second, and a data retention period of 3 months; Visualization Unit: Visualization software developed based on Qt 5.15, integrating the ECharts chart library, supporting functions such as bolt status heatmap, angle change curve, and historical data query.

[0082] 2. Functionality Implementation: Early warning function: When slightly loose, the control cabinet LED light flashes (frequency 1Hz) and the buzzer sounds (interval 1s); when severely loose, in addition to on-site alarm, the platform sends SMS and APP (Application) notifications to maintenance personnel (delivered within 5s), and marks the location of the abnormal bolt with a red dot on the platform map (accuracy ±5mm). Data storage function: Real-time storage of image data (compressed to 2MB per frame), bolt angle data (1 record per second), and early warning records (stored when triggered), supporting filtering and querying by time, bolt number, and status; Visualization features: The software displays a real-time heat map of rotor bolt distribution (green = normal, yellow = slightly loose, red = severely loose). Clicking on a single bolt allows you to view its angle change curve over the past hour (horizontal axis is time, vertical axis is angle). Data can be exported to Excel format (including angle sequence and warning time).

[0083] III. System Performance After 30 days of trial operation on a 100r / min turbine generator at a hydropower station, the system performance is as follows: Inspection accuracy: Bolt inspection accuracy rate 98.8%, looseness judgment accuracy rate 99.2%, no missed or false judgments; Real-time performance: The total processing time for a single frame image (acquisition + transmission + analysis) is ≤80ms, which meets the requirements for real-time monitoring (two frames are processed within 0.6s per rotor cycle). Stability: After 30 days of continuous operation, there were no hardware failures or software crashes, and the data transmission packet loss rate was 0.005%. Timely warning: When there is a serious loosening, the average time from when the angle exceeds the threshold to when the remote platform receives the warning is 1.2 seconds, which buys maintenance personnel time to deal with the situation.

[0084] Example 6 In another preferred embodiment, based on embodiments 1 and 3 above, this embodiment provides a high-speed hydro-generator rotor bolt loosening detection system (combined with... Figure 6 This embodiment corresponds to the system deployment of a certain hydropower station, based on Figure 6 The "image acquisition layer - data transmission layer - software platform layer" architecture enables full-process monitoring of bolt loosening.

[0085] I. System Layered Architecture (See) Figure 6 ) 1. Image Acquisition Layer Configure a high-speed camera (such as a Basler acA2040-90um) with a resolution of 2048×1024 and a frame rate of 500fps, along with a shock-resistant bracket and a hybrid fill light source, to complete the high-speed camera data acquisition task (e.g. Figure 4 As shown, align the rotor bolt area.

[0086] 2. Data transmission layer Based on the 10 Gigabit Ethernet architecture, using optical fiber as the transmission medium, it achieves high-speed transmission of 10 Gigabit Ethernet, ensuring that the transmission rate of image data from the acquisition layer to the software platform layer is ≥10Gbps, with no delay or packet loss.

[0087] 3. Software Platform Layer It contains four core modules: Image preprocessing and image enhancement: Distortion correction, noise reduction, and contrast enhancement are performed on the acquired images to provide high-quality images for subsequent analysis.

[0088] Image intelligent recognition and analysis: Deploying improved YOLOv8 ( Figure 2 ) and DeepLabV3+ Figure 3 The model is used to complete bolt detection, segmentation, and angle calculation. like Figure 2 As shown, it includes a backbone network, a neck network, and a head network. The output of the backbone network is connected to the input of the neck network, and the output of the neck module is connected to the input of the head network. The backbone network and the neck network perform feature extraction and feature fusion, and the head network predicts and filters the bolt detection results to obtain the final bolt detection results. Figure 2In this context, AKConv (Adaptive Kernel Convolution) is an adaptive kernel convolution that, compared to the original Conv, can adaptively adjust the convolution kernel to extract features of different scales and details, adapting to the needs of object detection of different sizes; C2f stands for CSP2_Forward (CSP2 forward structure, a feature fusion structure); GAM Attention (Global Attention Mechanism) is a global attention mechanism that can globally assign weights to features, highlighting key features such as defects, suppressing redundant information, and improving feature discriminativeness; Concat means concatenation; Upsample is an upsampling operation; Detect is a detection operation; and SPPF (Spatial Pyramid Pooling - Fast) is a fast spatial pyramid pooling method. In the GAM Attention (Global Attention Mechanism) structure, MLP (Multi-Layer Perceptron) is a multi-layer perceptron that performs non-linear transformations on features within GAM Attention to learn complex relationships between features and assist in the generation of attention weights; conv represents convolution; sigmoid is the sigmoid activation function, which can map attention weights to the 0~1 range to achieve weighted features; Input feature is the input feature; Out features is the output feature; and in features is the input feature within the module.

[0089] like Figure 3As shown, the Encoder is responsible for feature extraction; the Decoder is responsible for feature fusion and upsampling to generate prediction results; Image input is the image input; DCNN (Deep Convolutional Neural Network) is a deep convolutional neural network that can expand the receptive field through dilated convolutions, capturing multi-scale contextual information without reducing feature map resolution, thus achieving efficient feature extraction; Atrous Conv (Atrous Convolution) is dilated convolution, also called dilated convolution, used to expand the receptive field and capture multi-scale information; 1×1 Conv (1×1 Convolution) is a 1×1 convolution used for channel adjustment, feature fusion, or dimensionality reduction; 3×3 Conv rate 6 (3×3 Convolution with dilation rate 6) is a 3×3 dilated convolution with a dilation rate of 6, increasing the receptive field and capturing specific scale information; 3×3 Conv rate 12 (3×3 Convolution with dilation rate 12) is a 3×3 dilated convolution with a dilation rate of 12; 3×3 Conv rate 18 (3×3 Convolution with dilation rate 18) is a 3×3 dilated convolution with a dilation rate of 12. 18) is a 3×3 dilated convolution with an inflation rate of 18; Image pooling is used for global feature extraction to obtain global context information; Low-level features are low-level features containing image details, edges, and other information; Upsample is used to improve the resolution of feature maps and restore image details; Concat (Concatenate) means to concatenate, merging different feature maps along the channel dimension to fuse multi-scale information; 3×3 Conv (3×3 Convolution) is a 3×3 convolution used for feature extraction and thinning, increasing the non-linear expression of features; Prediction result is the prediction result, outputting the final segmentation or detection result.

[0090] Test result report generation: based on Figure 5 The judgment logic generates a report that includes the loose bolt number, angle change, and loosening trend.

[0091] Image data and results storage: A distributed database is used to store original images, detection results and historical data, with a retention period of 3 months, supporting data traceability and analysis.

[0092] II. System Functions and Performance Real-time performance: Single-frame image processing time (including acquisition, transmission, and analysis) ≤ 80ms, meeting the real-time monitoring requirements of a 100r / min rotor (≥ 2 frames acquired per cycle).

[0093] Accuracy: Bolt detection accuracy reaches 99.2% at mAP@0.5, looseness judgment accuracy is 99.5%, with no missed or false judgments.

[0094] Early warning mechanism: Slight loosening triggers a local audible and visual alarm; severe loosening simultaneously sends a notification to the remote monitoring platform (see attached). Figure 5 (The "Detection of Significant Loosening of the Nut" branch) has a warning response time of ≤2s.

[0095] In the preferred embodiment, in step 1, the high-speed turbine generator rotor rotates at a speed of 60~150 r / min, and when acquiring images using high-speed camera technology, it is ensured that ≥2 clear images are captured per rotor rotation cycle. The high-speed camera technology uses a wide-angle lens or fisheye lens, and the camera is installed 1m below the rotor to capture images of the bolt area in a vertical shooting manner. These settings effectively capture the dynamic changes of the bolts during rotor rotation, accurately obtaining key information such as bolt loosening and displacement. Simultaneously, the wide-angle or fisheye lens expands the shooting range, avoiding blind spots, and the vertical shooting method reduces image distortion, ensuring the accuracy of the acquired images.

[0096] In the preferred embodiment, in step 1, the camera device is also equipped with a shockproof bracket and a hybrid supplementary light source, which is a combination of LED light source and infrared light source. The above settings can effectively reduce image shaking caused by vibration during shooting and ensure stable and clear imaging. The LED light source provides sufficient lighting for daily use, while the infrared light source is automatically activated in low light or nighttime environments, ensuring that high-quality images can be obtained in different scenarios.

[0097] In the preferred embodiment, step 1, the image correction algorithm's process for handling image distortion includes: correcting the distortion of the original image based on the camera intrinsic parameter matrix, denoising the corrected image using a Gaussian filtering algorithm, and enhancing the illumination of the denoised image using a contrast adaptive adjustment algorithm. These settings effectively improve image quality, making the image closer to the real scene. After this series of processes, image distortion is corrected, noise is effectively removed, and illumination conditions are more suitable, laying a good foundation for subsequent image analysis and processing, and improving the accuracy and reliability of the overall solution.

[0098] In a preferred embodiment, step 2 includes bolt images with different degrees of looseness collected during unit maintenance shutdowns under simulated operating conditions. These images represent bolts in normal condition, slightly loose, and severely loose states. A slightly loose bolt corresponds to a bolt angle change of 0° to 2°, while a severely loose bolt corresponds to a bolt angle change greater than 2°. This configuration ensures the dataset covers the full range of bolt looseness states, providing rich and representative samples for model training. This detailed classification allows the model to more accurately identify different degrees of looseness, improving the accuracy and reliability of bolt condition assessment in practical applications.

[0099] In the preferred embodiment, in step 2, the dataset is annotated using the Labellimg tool, and the annotation content includes the bolt location and bolt anti-loosening line markings. The above settings can ensure the accuracy and completeness of the annotation information, providing high-quality data for subsequent model training. By clearly defining the bolt location and anti-loosening line markings, the model can more accurately identify key features, effectively improve the ability to judge the bolt status, and enhance the reliability of detection.

[0100] In the preferred embodiment, in step 2, the dataset is further expanded using data augmentation techniques, including image rotation, image scaling, and illumination perturbation. The training set, validation set, and test set are divided according to a preset ratio. These settings can effectively improve the generalization ability of the model and avoid overfitting. Image rotation allows the model to adapt to targets at different angles, image scaling can handle objects of different sizes, and illumination perturbation enhances adaptability to complex lighting environments. The preset ratio division ensures reasonable data distribution.

[0101] In the preferred embodiment, in step 3, the deep learning model includes an improved YOLOv8 object detection model. The improved YOLOv8 object detection model replaces the traditional convolutional layer with an adaptive kernel convolution (AKConv) module and inserts a global attention mechanism (GAM) into the backbone network of the model to enhance the ability to extract bolt edge features and focus on the bolt marker point region. The above settings can effectively improve the model's recognition accuracy of bolts in complex backgrounds, especially in uneven lighting or occlusion scenarios, the detection accuracy is improved by 12.6% compared to the traditional YOLOv8, while the model inference speed only decreases by 3.2%, balancing efficiency and performance.

[0102] In the preferred embodiment, in step 3, the deep learning model includes the DeepLabV3+ semantic segmentation model. The DeepLabV3+ semantic segmentation model captures multi-scale contextual information through dilated convolution and achieves fine segmentation of the bolt and background through an encoder-decoder structure. The calculation process for angle changes is as follows: based on the segmentation results of the DeepLabV3+ semantic segmentation model, the center coordinates of the two marker points on the bolt are extracted. , The bolt's real-time angle is calculated using formula (1), and the angle change is obtained. The above settings can effectively improve the accuracy and robustness of bolt angle detection, and are especially suitable for bolt status monitoring in complex industrial scenarios. After the model is optimized by transfer learning, it can still maintain a segmentation accuracy of 98.7% under conditions such as low light and oil pollution interference, and the calculation time is controlled within 50ms.

[0103] In the preferred embodiment, in step 3, multi-target tracking is achieved through DeepSort multi-target tracking technology. DeepSort multi-target tracking technology combines Kalman filtering to predict the bolt motion trajectory with deep learning feature matching to achieve continuous cross-frame recognition of individual bolts under high-speed rotation. The above settings can effectively address the problems of target occlusion and deformation in high-speed rotation scenarios and improve tracking stability. At the same time, by utilizing the bolt appearance features extracted by deep learning, the false detection rate can be reduced, ensuring accurate identification of the motion state of each bolt in complex industrial environments.

[0104] In the preferred embodiment, the specific process of determining the loosening state based on the bolt angle change threshold in step 4 is as follows: the bolt angle change threshold is set to... If the change in bolt angle within three consecutive rotor rotation cycles conforms to formula (2), then the bolt is determined to be in a loose state. The above settings can effectively avoid misjudgment caused by short-term vibration or measurement error. By monitoring three consecutive cycles, the continuous change trend of bolt angle can be accurately captured, ensuring the accuracy and reliability of loose state judgment and providing strong protection for the safe operation of equipment.

[0105] In the preferred embodiment, step 4 employs a multi-level early warning mechanism: when the bolt is determined to be slightly loose, a local audible and visual alarm is triggered; when the bolt is determined to be severely loose, in addition to triggering the local audible and visual alarm, an early warning message is simultaneously pushed to the remote monitoring platform, and the location coordinates of the abnormal bolt are marked. This configuration effectively improves the response efficiency to bolt loosening, ensuring that on-site personnel can promptly detect and handle minor loosening, and that remote monitoring personnel can quickly obtain information and dispatch professional personnel to address severe loosening, thus ensuring the safe and stable operation of the equipment.

[0106] In the preferred embodiment, in step 4, the parameters of the deep learning model are initialized through transfer learning, and bolt image samples under new working conditions are incorporated into the model training through incremental learning to shorten the model's adaptation time to new working conditions. The generation of detection results is based on a visualization platform, which is developed using C++ and Qt and integrates the TensorRT acceleration inference engine to ensure that the processing time of a single frame image is ≤50ms. The above settings can effectively improve the detection accuracy and efficiency of the model under new working conditions. The combination of C++ and Qt ensures the stability and ease of use of the visualization platform, while the TensorRT acceleration engine significantly shortens the inference time, enabling the system to respond in real time and meet the needs of industrial inspection.

[0107] In a preferred embodiment, the image acquisition module includes a high-speed industrial camera, a lens, a hybrid supplementary lighting source, and a shock-resistant bracket. The high-speed industrial camera has a resolution of 2048×1024 and a maximum frame rate of 1000fps. The lens is a wide-angle lens or a fisheye lens. The hybrid supplementary lighting source is a combination of an LED light source and an infrared light source. The shock-resistant bracket is used to fix the high-speed industrial camera to eliminate the impact of equipment vibration on image quality. These settings ensure that clear, distortion-free images are captured in high-speed motion scenarios. The LED light source provides sufficient brightness, while the infrared light source adapts to low-light environments, and the two complement each other. The stable design of the shock-resistant bracket further ensures the stability and accuracy of image acquisition, meeting the requirements of high-precision detection.

[0108] In the preferred embodiment, the data transmission module is built on a 10 Gigabit Ethernet architecture, uses optical fiber as the transmission medium, and has a data transmission rate of ≥10Gbps to achieve real-time transmission of image data. The above settings can effectively reduce data transmission latency and ensure that image data is not distorted or lost during transmission. At the same time, the 10 Gigabit Ethernet architecture has good scalability, which can facilitate subsequent upgrades and optimizations of transmission bandwidth and other aspects according to actual needs.

[0109] In the preferred embodiment, the intelligent analysis module includes a preprocessing submodule, a detection and tracking submodule, and a loosening judgment submodule. The preprocessing submodule performs distortion correction, denoising, and illumination enhancement on the image data. Distortion correction is based on the camera intrinsic parameter matrix, denoising is based on a Gaussian filtering algorithm, and illumination enhancement is based on a contrast adaptive adjustment algorithm. The detection and tracking submodule deploys an improved YOLOv8 target detection model, a DeepLabV3+ semantic segmentation model, and DeepSort multi-target tracking technology to achieve bolt target detection, bolt and background segmentation, and bolt cross-frame tracking. The loosening judgment submodule extracts the coordinates of bolt marker points and calculates angle changes, combining them with a preset angle change threshold to determine whether the bolt is loose. These settings effectively improve the accuracy and real-time performance of bolt status detection. The preprocessing submodule provides a high-quality image foundation for subsequent analysis, the detection and tracking submodule accurately locates the bolt, and the loosening judgment submodule uses quantitative data for scientific determination, forming a complete detection system.

[0110] In the preferred embodiment, the early warning and management module includes an early warning unit, a data storage unit, and a visualization unit. The early warning unit is used to implement multi-level early warnings, including local audible and visual alarms and pushing early warning information to a remote monitoring platform. The data storage unit uses a distributed database to store historical and real-time detection data. The visualization unit is used to display bolt status heatmaps and bolt angle change curves. The early warning and management module integrates a TensorRT accelerated inference engine to improve data processing efficiency. These features ensure that bolt status monitoring and early warning can be completed quickly and accurately in complex industrial scenarios. Multi-level early warnings provide timely response, distributed storage ensures data security, visualization makes the data clear at a glance, and the TensorRT engine significantly improves overall operational efficiency.

[0111] In summary, this invention provides a method and system for detecting loose bolts on high-speed hydro-generator rotors, effectively solving the problems of low efficiency, poor accuracy, and inability to monitor in real time that exist in traditional detection methods. By deeply integrating high-speed camera technology with deep learning algorithms, it provides a new approach to detection in this field, effectively improving detection efficiency and reliability, ensuring the safe operation of high-speed equipment in hydropower stations, and is suitable for condition monitoring and intelligent operation and maintenance scenarios of high-speed equipment in hydropower stations. Specifically, it deeply integrates high-speed camera technology with deep learning algorithms for detecting loose bolts on high-speed hydro-generator rotors, which differs from traditional detection methods; it uses an improved YOLOv8, DeepLabV3+ model combined with DeepSort technology to accurately identify, segment, track, and determine the bolt status; and it constructs a high-speed camera system suitable for speeds of 60~150 r / min to define the monitoring range, a unique approach of building a system for a specific speed range.

[0112] In terms of technical implementation details, this invention also has many unique features. Regarding data acquisition and processing, the method of simulating different operating conditions to collect samples and construct and expand the dataset during unit maintenance shutdowns is unique. In terms of model improvement, an adaptive kernel convolution module is used to replace the traditional convolutional layer, and a global attention mechanism is inserted to improve the YOLOv8 object detection model, enhancing the bolt edge feature extraction capability and focusing on the bolt marker point region. For segmentation technology, dilated convolution is used to capture information, and a DeepLabV3 + semantic segmentation model is adopted for fine segmentation through an encoder-decoder structure. Regarding the integration of tracking technology, DeepSort multi-object tracking technology is used to achieve cross-frame target association by combining Kalman filter prediction and deep learning feature matching, ensuring continuous identification of individual bolts under high-speed rotation.

[0113] This invention demonstrates significant advantages on multiple levels. By integrating various advanced technologies, it effectively solves key problems of traditional detection methods, forming a more efficient detection solution. It not only accurately detects bolt loosening but also achieves automatic identification and tracking, requiring no manual intervention throughout the process. Relying on deep learning models for autonomous learning and incremental learning to adapt to new working conditions, it greatly reduces the burden on staff. In ensuring safe equipment operation, it can promptly detect bolt loosening issues, preventing equipment malfunctions, downtime, or major accidents caused by loose bolts, thus reducing maintenance costs and economic losses. In equipment maintenance management, it enables automatic monitoring and early warning of equipment status, avoiding the blindness and time-consuming nature of traditional periodic inspections and maintenance methods, reducing maintenance costs and human resource investment, and improving maintenance efficiency and management level. Furthermore, a high-speed camera system built based on a specific speed range is specifically designed for monitoring the rotor bolts of high-speed hydro-generators, reflecting a unique system construction approach. Innovations in data acquisition, processing, model improvement, and technology integration form a complete and efficient detection system, providing new ideas and methods for the condition monitoring of similar high-speed mechanical equipment.

Claims

1. A method for detecting looseness of a bolt of a rotor of a high-speed water turbine generator, characterized by, The method comprises the following steps: Step 1: defining the monitoring range of the bolts of the high-speed water turbine generator rotor, collecting the bolt image data during the operation of the rotor by using high-speed camera technology, and processing the image distortion by using an image correction algorithm; Step 2: constructing a bolt loosening fault feature data set, and dividing the training set, the validation set and the test set; Step 3: based on a deep learning model, the collected image data is subjected to bolt target detection, semantic segmentation and multi-target tracking, bolt marker points are identified, and the angle change is calculated; Step 4: according to the bolt angle change threshold, the loosening state is judged, the detection result is generated by combining real-time data and historical data, and a warning is issued.

2. The method of claim 1, wherein: In the step 1, the rotational speed of the rotor of the high-speed water turbine generator is 60-150 r / min, and when the image is collected by using the high-speed camera technology, it is ensured that ≥2 clear images are captured per rotation cycle of the rotor; the camera equipment used in the high-speed camera technology is adapted to a wide-angle lens or a fish-eye lens, and the camera equipment is installed below the rotor to collect the bolt area image in a vertical shooting mode; the camera equipment is also provided with a shockproof support and a mixed light source, and the mixed light source is a combination of an LED light source and an infrared light source.

3. The method of claim 1, wherein In the step 1, the process of the image correction algorithm for processing the image distortion includes: based on the camera intrinsic matrix, the original image is subjected to distortion correction, the denoising processing is performed on the corrected image by using a Gaussian filtering algorithm, and the light enhancement processing is performed on the denoised image by using a contrast self-adaptive adjustment algorithm.

4. The method of claim 1, wherein: In the step 2, the data set includes bolt images of different loosening degrees collected during the simulated running state during the unit maintenance shutdown, the bolt images of different loosening degrees include bolt normal state images, bolt slightly loosening state images and bolt severely loosening state images; the data set is labeled by using a Labelimg tool, and the labeling content includes bolt positions and bolt loosening line markers; the data set is also expanded by using a data enhancement technology, the data enhancement technology includes image rotation, image scaling and light disturbance; the training set, the validation set and the test set are divided according to a preset ratio.

5. The method of claim 1, wherein: In the step 3, the deep learning model includes an improved YOLOv8 target detection model, the improved YOLOv8 target detection model replaces the traditional convolution layer by using an adaptive kernel convolution module, and inserts a global attention mechanism in the model backbone network, so as to enhance the bolt edge feature extraction capability and focus on the bolt marker point area.

6. The high-speed water-turbine-generator rotor bolt-looseness detection method as set forth in claim 1, characterized by In step 3, the deep learning model comprises a DeepLabV3+ semantic segmentation model, the DeepLabV3+ semantic segmentation model captures multi-scale context information through a hollow convolution, and realizes fine segmentation of the bolt and the background through a coding-decoding structure; the calculation process of the angle change is: based on the segmentation result of the DeepLabV3+ semantic segmentation model, the center coordinates of the two mark points on the bolt are extracted , , the real-time angle of the bolt is calculated through formula (1) , and then the angle change is obtained: （1）。 7. The method of claim 1, wherein: In the step 3, the multi-target tracking is realized by using a DeepSort multi-target tracking technology, the DeepSort multi-target tracking technology combines Kalman filtering to predict the bolt motion trajectory and deep learning feature matching, so as to realize the cross-frame continuous identification of the bolt individual under the high-speed rotating state.

8. The method of claim 1, wherein: The specific process of judging the loosening state according to the bolt angle change threshold in step 4 is that the bolt angle change threshold is set as If the bolt angle change amount in the continuous three rotor rotation periods meets formula (2), it is determined that the bolt is in a loosening state: （2）； In the formula, is the real-time angle of the screw for the first is the real-time angle of the screw for the first is the real-time angle of the screw for the first is the real-time angle of the screw for the first 9. The method of claim 1, wherein In the step 4, the multi-level warning mechanism is used to issue the warning: when the bolt is determined to be in a slightly loosening state, a local audible and visual alarm is triggered; when the bolt is determined to be in a severely loosening state, in addition to triggering the local audible and visual alarm, the warning information is also pushed to a remote monitoring platform synchronously, and the position coordinates of the abnormal bolt are labeled.

10. The method of claim 1, wherein: In step 4, the parameters of the deep learning model are initialized through transfer learning, and the bolt image samples under the new working condition are integrated into the model training through incremental learning to shorten the adaptation time of the model to the new working condition; The generation of the detection result is based on a visualization platform developed by using C++ and Qt and integrating a TensorRT acceleration inference engine to ensure that the single-frame image processing time is less than or equal to 50 ms.

11. A system for detecting bolt loosening of a high-speed water turbine generator rotor, which is a system for implementing the method for detecting bolt loosening of a high-speed water turbine generator rotor according to any one of claims 1 to 10, characterized in that: The system comprises an image acquisition module, a data transmission module, an intelligent analysis module and an early warning and management module; the image acquisition module is used for capturing real-time transient images of the rotor bolt area; the data transmission module is used for transmitting the image data captured by the image acquisition module to the intelligent analysis module; the intelligent analysis module is used for pre-processing the image data, bolt detection and tracking and looseness judgment; the early warning and management module is used for issuing early warnings according to the judgment result of the intelligent analysis module, and storing detection data and displaying the bolt state.

12. The high speed hydro-generator rotor bolt loosening detection system of claim 11, wherein: The image acquisition module comprises a high-speed industrial camera, a lens, a mixed light source and a shockproof support; the resolution of the high-speed industrial camera is 2048*1024, and the maximum frame rate is 1000 fps; the lens is a wide-angle lens or a fisheye lens; the mixed light source is a combination of LED light source and infrared light source; and the shockproof support is used for fixing the high-speed industrial camera to eliminate the influence of equipment vibration on image quality.

13. The high speed hydro-generator rotor bolt loosening detection system of claim 11, wherein: The data transmission module is built based on a 10-gigabit Ethernet architecture, uses optical fiber as the transmission medium, and has a data transmission rate of greater than or equal to 10 Gbps to realize real-time transmission of image data.

14. The high speed hydro-generator rotor bolt looseness detection system of claim 11, wherein: The intelligent analysis module comprises a preprocessing submodule, a detection and tracking submodule and a looseness judgment submodule; the preprocessing submodule is used for distortion correction, denoising and illumination enhancement processing of the image data, the distortion correction is realized based on the camera intrinsic matrix, the denoising is realized based on the Gaussian filtering algorithm, and the illumination enhancement is realized based on the contrast adaptive adjustment algorithm; The detection and tracking submodule is deployed with an improved YOLOv8 target detection model, a DeepLabV3+ semantic segmentation model and a DeepSort multi-target tracking technology, and is used for bolt target detection, bolt and background segmentation and bolt cross-frame tracking; the looseness judgment submodule is used for extracting bolt marker point coordinates and calculating angle changes, and combining a preset angle change threshold to judge whether the bolt is loose.

15. The high speed hydro-generator rotor bolt looseness detection system of claim 11, wherein: The early warning and management module comprises an early warning unit, a data storage unit and a visualization unit; the early warning unit is used for realizing multi-level early warning, including local audible and visual alarms and pushing early warning information to a remote monitoring platform; the data storage unit adopts a distributed database and is used for storing historical detection data and real-time detection data; the visualization unit is used for displaying a bolt state heat map and a bolt angle change curve, and the early warning and management module integrates a TensorRT acceleration inference engine to improve data processing efficiency.

Images