Pumped storage power station operation key target detection method and device

Abstract

The invention discloses a pumped storage power station operation key target detection method and device, and the method comprises the steps: inputting a collected image into a trained first-stage region detection model, and outputting a region detection result image; inputting a region detection result image into the trained two-stage key target detection model, and outputting a pumped storage power station operation key target detection result; wherein the first-stage region detection model is used for performing region division on an image of a region where the pumped storage power station is located to obtain a plurality of sub-images of a set category; and the two-stage key target detection model is used for carrying out key target detection on each sub-image by using windows with different scales. The method improves the accuracy and efficiency of detection, can effectively achieve the intellectualization of hidden danger detection, has the advantages of being wide in coverage range, free of contact, rapid, efficient, high in economical efficiency and the like, provides technical support for intelligent and digital management of the pumped storage power station, and has important practical significance.

Description

Technical Field

[0001] The present invention belongs to the technical field of safety monitoring and hidden danger identification of pumped storage power stations, and in particular relates to a method and device for detecting key targets in the operation of a pumped storage power station. Background Art

[0002] As the operating time increases, more and more pumped-storage power stations will enter the safety inspection and maintenance stage, and the safety monitoring and management tasks of power stations will become increasingly arduous in the future. Traditional hidden danger detection is mainly completed manually, and inspectors need to go to the site for close observation and measurement. However, since pumped-storage power stations are usually located in mountainous areas with complex terrain and a vast inspection area, manual inspection faces many challenges such as low efficiency, high cost, and high safety risks. In addition, some areas are difficult to reach, making it difficult to ensure the comprehensiveness and accuracy of the inspection results. This method of relying on manual inspection can no longer meet the needs of future large-scale pumped-storage power station safety inspections. Summary of the invention

[0003] In order to address the deficiencies in the prior art, the present invention provides a method and device for detecting key targets in the operation of a pumped-storage power station, which improves the accuracy and efficiency of detection, can effectively realize the intelligent detection of hidden dangers, and has the advantages of wide coverage, non-contact, fast and efficient, and strong economy. It provides technical support for the intelligent and digital management of pumped-storage power stations and has important practical significance.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is: In a first aspect, a method for detecting key targets in the operation of a pumped-storage power station is provided, comprising: collecting images of an area where the pumped-storage power station is located; inputting the collected images into a trained first-stage area detection model, and outputting an area detection result image; inputting the area detection result image into a trained second-stage key target detection model, and outputting a detection result of key targets in the operation of the pumped-storage power station; wherein the first-stage area detection model is used to divide the image of the area where the pumped-storage power station is located into regions to obtain a plurality of sub-images of set categories; and the second-stage key target detection model is used to detect key targets in each sub-image using windows of different scales.

[0005] Furthermore, the one-stage region detection model is a Mask2Former model, and the training method of the one-stage region detection model includes: constructing a regional data set; configuring training parameters of the one-stage Mask2Former model; training the Mask2Former model using the constructed regional data set and the configured training parameters to obtain a trained one-stage region detection model.

[0006] Furthermore, a regional data set is constructed, including: collecting images of the area where the pumped-storage power station is located; annotating the collected images, including: the dam body, the water surface of the reservoir area and the mountainous area, to obtain an original annotated data set 1; and using data enhancement technology to expand the original annotated data set 1 to obtain a regional data set.

[0007] Furthermore, configuring the training parameters of the first-stage Mask2Former model includes: configuring the operating environment parameters, initial training parameters and training control parameters of the Mask2Former model; the operating environment parameters include a GPU server and a deep learning framework; the initial training parameters include the number of GPUs, the number of images processed simultaneously by each GPU and the recognition categories; the training control parameters include the number of batches, the total number of iterations, the learning rate and the weight decay coefficient.

[0008] Furthermore, configuring the training parameters of the first-stage Mask2Former model also includes: determining the loss function: , in, is the total loss, is the cross entropy loss, is the dice loss, is the bipartite matching loss.

[0009] Furthermore, the two-stage key target detection model is a Mask2Former model, and the training method of the two-stage key target detection model includes: constructing a key target data set; configuring training parameters of the two-stage Mask2Former model; training the Mask2Former model using the constructed key target data set and the configured training parameters to obtain a trained two-stage key target detection model.

[0010] Furthermore, a key target data set is constructed, including: collecting images of the area where the pumped-storage power station is located; annotating the collected images with key targets, including structural cracks, geological landslides and floating objects on the water surface, to obtain an original annotated data set 2; and using data enhancement technology to expand the original annotated data set 2 to obtain a key target data set.

[0011] Furthermore, configuring the training parameters of the two-stage Mask2Former model includes: configuring the operating environment parameters, initial training parameters and training control parameters of the two-stage Mask2Former model; the operating environment parameters include a GPU server and a deep learning framework; the initial training parameters include the number of GPUs, the number of images processed simultaneously by each GPU and the recognition category; the training control parameters include the number of batches, the total number of iterations, the learning rate and the weight decay coefficient.

[0012] Furthermore, configuring the training parameters of the two-stage Mask2Former model also includes: determining the loss function: , in, is the total loss, is the cross entropy loss, is the dice loss, is the bipartite matching loss.

[0013] In a second aspect, a pumped-storage power station operation key target detection device is provided, including: a data acquisition module, used to collect images of the area where the pumped-storage power station is located; a first-stage detection module, used to input the collected images into a trained first-stage regional detection model, and output a regional detection result image; a second-stage detection module, used to input the regional detection result image into a trained second-stage key target detection model, and output the pumped-storage power station operation key target detection result; wherein, the first-stage regional detection model is used to divide the image of the area where the pumped-storage power station is located into regions, and obtain a number of sub-images of set categories; the second-stage key target detection model is used to detect key targets in each sub-image using windows of different scales.

[0014] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: the present invention inputs the collected image into a trained one-stage area detection model to output an area detection result image; inputs the area detection result image into a trained two-stage key target detection model to output a pumped-storage power station operation key target detection result; wherein, the one-stage area detection model is used to divide the image of the area where the pumped-storage power station is located into regions to obtain a number of sub-images of set categories; the two-stage key target detection model is used to detect key targets in each sub-image using windows of different scales; the detection accuracy and efficiency are improved, and the intelligent hidden danger detection can be effectively realized, with the advantages of wide coverage, non-contact, fast and efficient, strong economy, etc., providing technical support for the intelligent and digital management of pumped-storage power stations, and having important practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a main flow chart of a method for detecting key targets in operation of a pumped storage power station provided by an embodiment of the present invention; Figure 2 are key targets marked in the embodiments of the present invention, where (a) is crack one, (b) is crack two; Figure 3 Schematic diagram of the structure of the Mask2Former model in an embodiment of the present invention; Figure 4These are the key targets detected in the embodiment of the present invention, where (a) is crack one, (b) is crack two, (c) is crack three, (d) is floating object one, (e) is floating object two, and (f) is floating object three. DETAILED DESCRIPTION

[0016] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0017] Embodiment 1

[0018] A method for detecting key targets in the operation of a pumped-storage power station comprises: collecting images of an area where the pumped-storage power station is located; inputting the collected images into a trained first-stage area detection model to output an area detection result image; inputting the area detection result image into a trained second-stage key target detection model to output a pumped-storage power station operation key target detection result; wherein the first-stage area detection model is used to divide the image of the area where the pumped-storage power station is located into regions to obtain a plurality of sub-images of set categories; and the second-stage key target detection model is used to detect key targets in each sub-image using windows of different scales.

[0019] The main process of the key target detection method for pumped storage power station operation according to the present invention is as follows: Figure 1 shown.

[0020] Step 1: Build a regional dataset.

[0021] Step 1.1: Collect high-resolution images: Use drones to collect a large amount of high-resolution image data of the pumped-storage power station area.

[0022] Step 1.2, regional annotation: perform manual regional annotation on the acquired images, including the dam body, reservoir water surface and mountainous area, to construct the original annotation dataset 1.

[0023] Step 1.3, data enhancement: Expand the data scale through a variety of data enhancement techniques, including geometric transformation (such as rotation, flipping), and optical transformation (such as brightness contrast adjustment, noise injection, etc.).

[0024] Step 1.4: Construct regional dataset D1: Integrate the processed data to construct a large-scale, high-quality regional dataset D1.

[0025] Step 2: Configure the training parameters of the one-stage Mask2Former model, and use the constructed regional dataset to train the Mask2Former model with the configured training parameters to obtain a trained one-stage regional detection model.

[0026] Step 2.1. Configure the operating environment parameters of the first-stage Mask2Former model: Configure the hardware and software environment required for the first-stage Mask2Former model to run, including a high-performance GPU server and a deep learning framework.

[0027] Step 2.2, configure the initial training parameters of the first-stage Mask2Former model, including: GPU_COUNT = 1 (the number of GPUs is 1), IMAGES_PER_GPU = 24 (each GPU processes 24 images simultaneously), NUM_CALSSES ​​= 4 (there are four recognition categories, including background, dam body, reservoir water surface and mountainous area).

[0028] Step 2.3, determine the loss function; there are three loss functions in Mask2Former, and the total loss is defined as follows: , in, is the total loss, is the cross entropy loss, is the dice loss, is the bipartite matching loss.

[0029] Step 2.4. Configure the training control parameters of the first-stage Mask2Former model, including: BATCH_SIZE=GPU_COUNT*IMAGES_PER_GPU (number of batches, 24 here), MAX_ITERS=20000 (total number of iterations 20000 times), LEARNING_RATE = 0.0001 (learning rate is set to 0.0001), WEIGHT_DECAY = 0.05 (weight decay coefficient is 0.05).

[0030] Step 2.5: Perform model training based on the regional dataset D1, and adjust the network weight parameters by continuously optimizing the loss function until the performance index of the model on the validation set reaches the preset threshold or completes the established training rounds, and finally obtains the one-stage region detection Mask2Former model M1 that can be used for region segmentation.

[0031] The present invention is based on the Mask2Former deep learning algorithm. First, a large amount of high-resolution image data is collected by drones, and the pumped storage power station area (such as the dam body, the water surface of the reservoir area, and the mountainous area) is manually labeled to construct the original labeled data set 1, and the data scale is expanded through data enhancement technology. Subsequently, for the regional detection task, the key parameters of the algorithm (including target recognition type, training rounds, learning rate, etc.) are optimized to train the first-stage regional detection Mask2Former model M1. By continuously optimizing the loss function, a model that can accurately divide the region is finally obtained to ensure accurate recognition and segmentation of the dam, reservoir area and mountainous area.

[0032] Step 3: Build key target datasets.

[0033] Step 3.1: Collect high-resolution image data of the area where the pumped-storage power station is located for key targets such as structural cracks, geological landslides, and floating objects on the water surface.

[0034] Step 3.2: Differentiated cutting: Differentiated cutting is performed according to the characteristic scales of different key targets. Microscopic hidden dangers are cut at a small scale, and macroscopic hidden dangers are cut at a large scale.

[0035] Step 3.3, manual labeling: Perform accurate manual labeling on the cropped image samples to construct a special original labeling dataset 2.

[0036] Step 3.4, data enhancement: Expand the data scale through data enhancement technology, including geometric transformation, optical transformation and noise injection methods.

[0037] Step 3.5: Integrate the processed data to finally obtain a more comprehensive and diversified key target data set D2.

[0038] The present invention first prepares the data set required for key target detection, collects more than 10,000 images containing various hidden dangers, and different key target images have different scales and RGB channels. The images are preprocessed, and the images are enlarged, rotated, cropped, grayed, etc., to remove useless information and make the key target image features more prominent.

[0039] Then, the Labelme image annotation tool is used to annotate the images in the dataset, where the annotation data of two cracks in the crack sample are as follows: Figure 2 shown.

[0040] Divide the data set: First, divide the collected images into a test set in an 8:2 ratio; then use the k-fold cross-validation method to divide the training set and the validation set. The data set used by the model during the training process is called the training set of the model. After the model training is completed, the prediction effect of the model needs to be tested on some new samples. The new samples used at this time are the test set of the model. The model validation set can be used to adjust the hyperparameters.

[0041] The cross-validation method first randomly divides the data into K non-overlapping subsets: , where each subset The datasets contain the same number of samples. In order to better preserve the characteristics of sample distribution in the dataset, stratified sampling is used to divide the dataset subsets. For the dataset subsets, K-1 data subsets are used as training sets for each training, and the remaining subset is used as the test set. Then K training and testing are performed, and the average of the K test results is output as the final result.

[0042] The present invention adopts the Mask2Former model for key target detection. As an innovative improvement to the unified image segmentation framework, Mask2Former can not only handle instance segmentation, semantic segmentation, and panoptic segmentation tasks simultaneously, but also has significant improvements in performance and efficiency. Compared with traditional methods, Mask2Former has the following characteristics.

[0043] 1) Learnable query features are introduced to replace previous pixel or region proposal-based methods, which can capture objects and features in images more flexibly.

[0044] 2) The Masked Attention mechanism is used to limit the calculation scope of attention to a local area, which not only improves the calculation efficiency but also enhances the accuracy of feature extraction.

[0045] 3) The Transformer Decoder structure is designed to refine the prediction results through multi-layer interactions, which significantly improves the quality of the segmentation mask. The overall architecture of Mask2Former is shown in the figure. Figure 3 As shown in the figure, the innovative design adopted by the model enables it to demonstrate excellent performance in detection tasks of fine structures such as hidden dangers, and can more accurately identify and segment various complex hidden danger forms.

[0046] Mask2Former uses Swin Transformer as the backbone network for feature extraction. After basic preprocessing, the input image is passed through the backbone network to generate a multi-scale feature map. The network design adopts a feature pyramid structure, which can effectively extract and fuse image features of different scales. This multi-scale feature extraction method enables the model to focus on local details and global semantic information in the image at the same time.

[0047] During the decoding process, the model uses a masked attention mechanism to allow queries to gradually locate and understand different semantic regions in the image. This query-based approach is different from the traditional pixel-by-pixel classification method and can better handle the relationship between objects and improve segmentation accuracy.

[0048] After feature extraction and attention processing, the model predicts the corresponding semantic category and segmentation mask for each query. The loss calculation takes into account both mask quality and classification accuracy. This end-to-end prediction method simplifies post-processing steps while ensuring the accuracy of the results.

[0049] Step 4: Configure the training parameters of the two-stage Mask2Former model, and use the constructed key target dataset to train the Mask2Former model with the configured training parameters to obtain a trained two-stage key target detection model.

[0050] Step 4.1: Configure the training parameters of the two-stage Mask2Former model: Configure the hardware and software environment required for running the two-stage Mask2Former model, including a high-performance GPU server and a deep learning framework. Step 4.2, configure the initial training parameters of the two-stage Mask2Former model, including: GPU_COUNT = 1 (the number of GPUs is 1), IMAGES_PER_GPU = 24 (each GPU processes 24 images simultaneously), NUM_CALSSES ​​= 4 (there are four recognition categories, including background and target objects, such as structural cracks, geological landslides, floating objects on the water surface, etc.).

[0051] Step 4.3, determine the loss function; there are three loss functions in Mask2Former, and the total loss is defined as follows: , in, is the total loss, is the cross entropy loss, is the dice loss, is the bipartite matching loss.

[0052] Step 4.4. Configure the training control parameters of the two-stage Mask2Former model, including: BATCH_SIZE=GPU_COUNT*IMAGES_PER_GPU (number of batches, 24 here), MAX_ITERS=20000 (total number of iterations 20000 times), LEARNING_RATE = 0.0001 (learning rate is set to 0.0001), WEIGHT_DECAY = 0.05 (weight decay coefficient is 0.05).

[0053] Step 4.5: Perform model training based on the key target dataset D2, and adjust the network weight parameters by continuously optimizing the loss function until the performance index of the model on the validation set reaches the preset threshold or completes the established training rounds, and finally obtain the two-stage key target detection Mask2Former model M2 that can be used for key target detection.

[0054] Enter the following command in the terminal to perform training: 1. python train.py configs / configs / mask2former-crack.py; The test can be run using jupyter notebook. You can also run run.py in the terminal or pychram.

[0055] Python needs to install third-party packages: 1. pytorch; 2. scipy; 3. pyyaml; 4. mmdet; 5. yapf; 6. opencv-python; 7. mmcv; 8. mmsegmentation; 9. Timm.

[0056] After training, a two-stage key target detection Mask2Former model M2 for key target detection is obtained.

[0057] The present invention collects high-resolution images containing key targets (such as structural cracks, geological landslides, and floating objects on the water surface) and performs differential cropping according to target features to construct a special key target dataset (original annotated dataset 2). The dataset is trained with deep learning, and relevant parameters (such as training rounds, learning rate, batch size, etc.) are optimized to finally obtain a two-stage key target detection Mask2Former model M2 for key target detection. The model can accurately detect key hidden dangers in drone images and ensure efficient inspection and hidden danger discovery of pumped storage power stations.

[0058] Step 5: For the high-resolution image data obtained by the drone inspection, the one-stage region detection Mask2Former model M1 is used to complete the division of the region (dam body, reservoir water surface and mountainous area), and sub-images are obtained using windows of different scales in different regions.

[0059] Step 5.1. Input image data: Obtain high-resolution image data of drone inspection.

[0060] Step 5.2, region division: Use the one-stage region detection Mask2Former model M1 to complete the division of the dam body, reservoir water surface and mountainous area in the image.

[0061] Step 5.3, sub-image acquisition: For different regions, use windows of different scales to acquire sub-images.

[0062] Step 6: Use the two-stage key target detection Mask2Former model M2 to perform accurate key target detection on the sub-image obtained in step 5 and output the key targets in the drone inspection image.

[0063] Step 6.1. Input sub-image: Input the sub-image obtained in step 5 into the two-stage key target detection Mask2Former model M2.

[0064] Step 6.2, precise target detection: Use the two-stage key target detection Mask2Former model M2 to perform precise key target detection on the sub-image, where the key targets include: structural cracks, geological landslides, and floating objects on the water surface.

[0065] Step 6.3, output results: Output the key target detection results in the drone inspection image, including target type, location, size and other information.

[0066] The present invention first uses drones to collect high-resolution images, uses the one-stage regional detection Mask2Former model M1 to divide the pumped-storage power station area, and uses windows of appropriate scales to obtain sub-images according to different areas. Then, the two-stage key target detection Mask2Former model M2 is applied to accurately detect key targets in these sub-images, output key hidden danger information such as structural cracks, geological landslides and floating objects on the water surface, and realize automatic and efficient detection of key targets during the operation of the power station.

[0067] The test results of some key targets using the two-stage key target detection Mask2Former model M2 are as follows: Figure 4 As shown, it can be seen that the present invention can identify potential hidden dangers more accurately and quickly, thus providing a solid foundation for taking preventive measures in a timely manner. Through an effective identification process, not only can risks be reduced, but the overall safety management level can also be improved to ensure the reliability and safety of the working environment. This efficient hidden danger identification capability can help the team better respond to emergencies, reduce the probability of accidents, and protect the lives of employees and the assets of the company.

[0068] Based on the Mask2Former deep learning algorithm, the present invention realizes the automatic identification of various potential hidden dangers in pumped storage power stations. Through the two-stage model training strategy, regional detection and key target detection are optimized respectively, which improves the accuracy and efficiency of detection. The present invention combines the differentiated cropping algorithm, performs multi-scale processing according to different hidden danger types, and applies data enhancement technology to expand the data set, which effectively improves the generalization ability and recognition accuracy of the model. By processing large-size images in different regions and at multiple scales, comprehensive coverage detection of hidden dangers is achieved, ensuring the efficiency and accuracy of detection.

[0069] The present invention uses drone high-altitude scanning technology to replace traditional manual inspection, which has significant advantages, especially in areas with complex terrain and difficult to reach. The present invention realizes remote, contactless and efficient hidden danger detection, which not only reduces the detection cost and personnel safety risks, but also provides strong technical support for the intelligent management and hidden danger detection of pumped storage power stations, and has wide engineering practicality and promotion value.

[0070] Embodiment 2 Based on the method for detecting key operation targets of a pumped-storage power station described in Embodiment 1, this embodiment provides a device for detecting key operation targets of a pumped-storage power station, including: A data acquisition module, used to collect images of the area where the pumped storage power station is located; The one-stage detection module is used to input the collected image into the trained one-stage region detection model and output the region detection result image; The two-stage detection module is used to input the regional detection result image into the trained two-stage key target detection model and output the key target detection result of the pumped storage power station operation; Among them, the first-stage regional detection model is used to divide the image of the area where the pumped-storage power station is located to obtain several sub-images of set categories; the second-stage key target detection model is used to detect key targets in each sub-image using windows of different scales.

[0071] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0072] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0073] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0074] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0075] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A method for detecting key targets in the operation of a pumped storage power station, characterized in that: include: Collect images of the area where the pumped storage power station is located; Input the collected image into the trained one-stage region detection model and output the region detection result image; The regional detection result image is input into the trained two-stage key target detection model, and the key target detection result of the pumped storage power station operation is output; Among them, the first-stage regional detection model is used to divide the image of the area where the pumped-storage power station is located to obtain several sub-images of set categories; the second-stage key target detection model is used to detect key targets in each sub-image using windows of different scales.

2. The method for detecting key operation targets of a pumped storage power station according to claim 1, characterized in that: The one-stage region detection model is a Mask2Former model, and the training method of the one-stage region detection model includes: Construct regional datasets; Configure the training parameters of the one-stage Mask2Former model; The Mask2Former model is trained using the constructed regional dataset and the configured training parameters to obtain a trained one-stage region detection model.

3. The method for detecting key operation targets of a pumped storage power station according to claim 2, characterized in that: Construct a regional dataset, including: Collect images of the area where the pumped storage power station is located; The collected images are labeled, including the dam body, the water surface of the reservoir area and the mountainous area, to obtain the original labeled data set 1; The data enhancement technology is used to expand the original labeled data set 1 to obtain the regional data set.

4. The method for detecting key operation targets of a pumped storage power station according to claim 2, characterized in that: Configure the training parameters of the first-stage Mask2Former model, including: configuring the operating environment parameters, initial training parameters, and training control parameters of the Mask2Former model; The operating environment parameters include GPU servers and deep learning frameworks; the initial training parameters include the number of GPUs, the number of images processed simultaneously by each GPU, and the recognition categories; the training control parameters include the number of batches, the total number of iterations, the learning rate, and the weight decay coefficient.

5. The method for detecting key operation targets of a pumped storage power station according to claim 4, characterized in that: Configure the training parameters of the one-stage Mask2Former model, including: Determine the loss function: ， in, is the total loss, is the cross entropy loss, is the dice loss, is the bipartite matching loss.

6. The method for detecting key operation targets of a pumped storage power station according to claim 1, characterized in that: The two-stage key target detection model is a Mask2Former model, and the training method of the two-stage key target detection model includes: Build key target datasets; Configure the training parameters of the two-stage Mask2Former model; The Mask2Former model is trained using the constructed key target dataset and the configured training parameters to obtain a trained two-stage key target detection model.

7. The method for detecting key operation targets of a pumped storage power station according to claim 6, characterized in that: Construct key target datasets, including: Collect images of the area where the pumped storage power station is located; The collected images are labeled with key targets, including structural cracks, geological landslides, and floating objects on the water surface, to obtain the original labeled dataset 2; The data augmentation technology is used to expand the original labeled data set 2 to obtain the key target data set.

8. The method for detecting key operation targets of a pumped storage power station according to claim 6, characterized in that: Configure the training parameters of the two-stage Mask2Former model, including: configure the operating environment parameters, initial training parameters, and training control parameters of the two-stage Mask2Former model; The operating environment parameters include GPU servers and deep learning frameworks; the initial training parameters include the number of GPUs, the number of images processed simultaneously by each GPU, and the recognition categories; the training control parameters include the number of batches, the total number of iterations, the learning rate, and the weight decay coefficient.

9. The method for detecting key operation targets of a pumped storage power station according to claim 8, characterized in that: Configure the training parameters of the two-stage Mask2Former model, including: Determine the loss function: ， in, is the total loss, is the cross entropy loss, is the dice loss, is the bipartite matching loss.

10. A pumped storage power station operation key target detection device, characterized in that: include: A data acquisition module, used to collect images of the area where the pumped storage power station is located; The one-stage detection module is used to input the collected image into the trained one-stage region detection model and output the region detection result image; The two-stage detection module is used to input the regional detection result image into the trained two-stage key target detection model and output the key target detection result of the pumped storage power station operation; Among them, the first-stage regional detection model is used to divide the image of the area where the pumped-storage power station is located to obtain several sub-images of set categories; the second-stage key target detection model is used to detect key targets in each sub-image using windows of different scales.

Images