Method and system for intelligent recognition of compressed air energy storage power station salt cavity shape

Abstract

The present application relates to the technical field of energy storage, and provides a compressed air energy storage power station salt cavity form intelligent identification method and system, the method comprises the following steps: S1, obtaining seismic data of the compressed air energy storage power station salt cavity area; S2, labeling the salt cavity area of the seismic data, generating a data set with labeled data, and preprocessing the data set; S3, constructing a compressed air energy storage power station salt cavity form identification model based on a deep learning detection algorithm; S4, collecting the seismic data of the compressed air energy storage power station salt cavity to be detected, inputting the compressed air energy storage power station salt cavity form identification model, and obtaining the salt cavity form of the compressed air energy storage power station to be detected. The method has high recognition accuracy and fast recognition speed, can greatly reduce the design cycle of the compressed air energy storage power station salt cavity gas storage in the early stage, can effectively process the case that multiple targets exist in the image and overlap each other, and accurately separates the target from the complex background.

Description

Technical Field

[0001] This invention relates to the field of energy storage technology, and in particular to a method and system for intelligent identification of salt cavity morphology in compressed air energy storage power stations. Background Technology

[0002] Compressed air energy storage, as a long-term energy storage method, features large capacity, high safety, long service life, and low construction cost. It can effectively solve the problem of grid connection difficulties for new energy power generation such as solar and wind power, and significantly improve energy utilization efficiency. Currently, compressed air energy storage has developed from kW and MW to hundreds of MW and 300MW levels. With the increase in scale, gas storage facilities have become the primary consideration in the planning and design of large-scale compressed air energy storage power plants.

[0003] There are various types of compressed air (CA) storage facilities, including above-ground storage facilities, natural salt caverns, and artificial chambers. For large-scale CA power plants, the required storage space is substantial, and above-ground storage facilities are costly, generally suitable for small and medium-sized plants. Given the current trend towards large-scale development of CA power plants, most CA power plants primarily utilize salt caverns and artificial chambers for storage. Salt caverns, as an excellent storage facility, offer advantages such as large volume, good sealing, high storage pressure, low cost, mechanical stability, and small footprint. Furthermore, most salt caverns are currently idle, making the use of salt caverns for CA power plant storage a hot topic.

[0004] Using salt caverns as the gas storage reservoir for compressed air energy storage power plants presents the challenge of identifying the morphology of these caverns. Traditional methods primarily employ seismic exploration or sonar logging, followed by manual analysis of the seismic or sonar logging data to map the morphology of the salt caverns. This method is limited by the analysis process, and manual analysis is slow and has certain accuracy issues; therefore, the results of the salt cavern morphology mapping need further verification. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for intelligent recognition of salt cavity morphology in compressed air energy storage power stations, based on the Faster R-CNN detection algorithm, thereby improving the accuracy of salt cavity morphology characterization.

[0006] The present invention adopts the following technical solution:

[0007] On one hand, the present invention provides a method for intelligent identification of salt cavity morphology in compressed air energy storage power stations, comprising:

[0008] S1. Obtain seismic data for the salt cavity area of ​​the compressed air energy storage power station;

[0009] S2. Label the salt cavity regions of the seismic data to generate a dataset with labeled data, and preprocess the dataset;

[0010] S3. Based on the dataset obtained in step S2, construct a salt cavity morphology recognition model for compressed air energy storage power stations based on deep learning detection algorithms.

[0011] S4. Collect seismic data of the salt cavity of the compressed air energy storage power station to be detected, and input it into the salt cavity morphology recognition model of the compressed air energy storage power station in step S3 to obtain the salt cavity morphology of the compressed air energy storage power station to be detected.

[0012] In addition to any of the possible implementations described above, another implementation is provided in which the seismic data of the salt cavity area of ​​the compressed air energy storage power station in step S1 includes three sources: first, actual three-dimensional seismic data obtained through three-dimensional seismic detection technology; second, forward modeling seismic data obtained based on the theoretical salt cavity model; and third, enhanced seismic data obtained based on actual three-dimensional seismic data through data augmentation technology.

[0013] In addition to any of the possible implementations described above, another implementation is provided in which the actual three-dimensional seismic data is used to acquire different seismic data slices, which are then saved as image data respectively.

[0014] The forward modeling seismic data is obtained by using a theoretical salt cavity model, designing geological models of spherical salt cavities, columnar salt cavities, or irregularly shaped salt cavities, and then using seismic forward modeling software to save the forward modeling seismic data as image data.

[0015] The enhanced seismic data is obtained by rotating, flipping, cropping, scaling, and adding noise to the actual 3D seismic image data to expand the size of the dataset.

[0016] In addition to any of the possible implementations described above, another implementation is provided in which, in step S2, the salt cavity region of the seismic data is labeled to determine the boundary and outline of the salt cavity and generate a dataset with labeled data; the dataset is preprocessed by normalization, denoising and edge detection.

[0017] In addition to any of the possible implementations described above, another implementation is provided in which the denoising is performed by using median filtering, Gaussian filtering, or bilateral filtering to remove noise from the image data; and the edge detection is performed by using the Canny operator or Sobel operator edge detection algorithm to extract edge information from the image data and highlight the contours of objects in the image.

[0018] In addition to any of the possible implementations described above, another implementation is provided in which, in step S3, the compressed air energy storage power station salt cavity morphology recognition model is constructed using a deep learning detection algorithm. The deep learning detection algorithm adopts the Faster R-CNN algorithm model and includes the following steps:

[0019] S31. Dataset partitioning: Based on the preprocessed seismic image dataset obtained in step S2, it is divided into a training dataset and a test dataset.

[0020] S32. Determine the backbone network: Select the pre-trained VGG16 deep convolutional neural network as the backbone network of the Faster R-CNN algorithm, and extract features from the earthquake image training dataset to obtain feature maps.

[0021] S33. Training the Region Proposal Network (RPN): The feature map obtained in step S32 is input into the RPN. The RPN generates a series of anchor boxes of different scales and aspect ratios at each location on the feature map. Then, the RPN processes each anchor box through convolution operations and outputs the results of two branches: one branch is a classification network that predicts whether each anchor box contains a target, separating the salt cavity from the background; the other branch is a regression network that predicts the offset of the anchor box relative to the real salt cavity target box.

[0022] For classification networks, the cross-entropy loss function is used to calculate the predicted anchor box category and the true label; for regression networks, the smooth L1 loss function is used to calculate the loss between the predicted anchor box offset and the true offset; by calculating the two losses, the total loss of RPN is obtained.

[0023] Based on the calculated total loss, the gradient of the total loss with respect to each parameter in the RPN is calculated using the backpropagation algorithm. The Adam optimization algorithm is then used to update the RPN parameters based on the gradients, gradually bringing the RPN's predictions closer to the true labels.

[0024] RPN training is now complete;

[0025] The parameters of an RPN network include the size and number of convolutional kernels;

[0026] S34. Training the Fast R-CNN detection network:

[0027] Candidate regions are generated on the feature map of the training image using the trained RPN, and the non-maximum suppression (NMS) algorithm is used to remove candidate regions with high overlap.

[0028] The generated candidate regions are mapped back to the feature maps extracted by the backbone network, and then the feature map regions corresponding to each candidate region are adjusted to feature vectors of a fixed size through the RoIPooling layer.

[0029] The feature vectors output by RoIPooling are input into the classification network and the regression network. The classification network predicts the category of the target (including the background category) in each candidate region through fully connected layers and the Softmax function, and the regression network predicts the further offset of the bounding box of each candidate region relative to the real target box through fully connected layers.

[0030] For classification networks, the cross-entropy loss function is used to calculate the loss between the predicted target category and the true label; for regression networks, the smooth L1 loss function is used to calculate the loss between the predicted bounding box offset and the true offset. The two losses are added together to obtain the total loss of the Fast R-CNN detection network.

[0031] Based on the calculated total loss, the gradient of the total loss with respect to each parameter in the Fast R-CNN detection network is calculated using the backpropagation algorithm, and then the parameters of the Fast R-CNN detection network are updated using the optimizer.

[0032] The parameters of the Fast R-CNN detection network include the kernel size, which is obtained through continuous iteration;

[0033] S35. Alternating Training: RPN and Fast R-CNN detection networks are trained alternately;

[0034] First, train the RPN. Then, fix the parameters of the RPN and use the candidate regions generated by the RPN to train the Fast R-CNN detection network. Next, fix the parameters of the Fast R-CNN detection network and train the RPN again. Repeat this process until the model converges and the loss function no longer decreases significantly. At this point, the Fast R-CNN algorithm model training is complete.

[0035] S36. Verify the dataset test;

[0036] The validation dataset is input into the Fast R-CNN detection network obtained from the training dataset. The accuracy of object detection is calculated. If the recognition accuracy is greater than 90%, the Fast R-CNN detection network is used for subsequent recognition processes. Otherwise, the model parameters are adjusted to further optimize the model until the set accuracy is achieved.

[0037] In addition to any of the possible implementations described above, another implementation is provided in which, in step S4, the seismic data of the salt cavity of the compressed air energy storage power station to be detected is input into the Fast R-CNN detection network through a normalization operation to obtain the salt cavity morphology recognition result.

[0038] In addition to any of the possible implementations described above, another implementation is provided in which, in step S4, the salt cavity morphology recognition result is further characterized by a threshold segmentation method; the threshold segmentation is based on the difference in grayscale between the target and the background, and multiple thresholds are set to achieve clear separation between the target and the background; the threshold segmentation adopts multi-threshold segmentation.

[0039] On the other hand, the present invention also provides an intelligent recognition system for the salt cavity morphology of a compressed air energy storage power station, the system being used to implement the above-mentioned method, the system comprising:

[0040] The seismic data acquisition unit is used to acquire seismic data of the salt cavity area of ​​the compressed air energy storage power station.

[0041] The seismic data annotation and preprocessing unit is used to annotate the salt cavity regions of seismic data, generate a dataset with labeled data, and preprocess the dataset.

[0042] The model building and training unit is used to build and train a salt cavity morphology recognition model for compressed air energy storage power stations.

[0043] The salt cavity morphology recognition unit inputs the seismic data of the salt cavity of the compressed air energy storage power station to be detected into the compressed air energy storage power station salt cavity morphology recognition model to obtain the salt cavity morphology of the compressed air energy storage power station to be detected.

[0044] In addition to any of the possible implementations described above, another implementation is provided in which the seismic data of the salt cavity area of ​​the compressed air energy storage power station includes three sources: first, actual three-dimensional seismic data obtained through three-dimensional seismic detection technology; second, forward-modeled seismic data obtained based on the theoretical salt cavity model; and third, enhanced seismic data obtained based on actual three-dimensional seismic data through data augmentation technology.

[0045] The beneficial effects of this invention are as follows:

[0046] 1. The salt cavity morphology recognition method provided by this invention has high recognition accuracy and fast recognition speed, which can greatly reduce the early design cycle of compressed air energy storage power station salt cavern gas storage and reduce construction costs, thereby increasing the profitability of compressed air energy storage power station.

[0047] 2. This invention also provides multiple methods for obtaining seismic data, reducing the cost and time required for acquiring 3D seismic data. Furthermore, with the increasing number of salt cavern compressed air energy storage projects, the amount of 3D seismic data will increase, leading to a greater training dataset with salt cavern morphology labels, which can significantly improve the accuracy of the salt cavern morphology recognition model.

[0048] 3. The Faster R-CNN algorithm provided by this invention can effectively handle situations where there are multiple targets in an image and the targets overlap with each other, accurately separating the targets from complex backgrounds, and is suitable for complex scenes. Attached Figure Description

[0049] Figure 1 The diagram shown is a flowchart illustrating an intelligent identification method for salt cavity morphology in a compressed air energy storage power station according to an embodiment of the present invention.

[0050] Figure 2 The diagram shown illustrates the training process of the Fast R-CNN algorithm in this embodiment. Detailed Implementation

[0051] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered in isolation, but can be combined with each other to achieve better technical effects.

[0052] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for intelligent identification of salt cavity morphology in a compressed air energy storage power station, comprising:

[0053] S1. Obtain seismic data for the salt cavity area of ​​the compressed air energy storage power station;

[0054] S2. Label the salt cavity regions of the seismic data to generate a dataset with labeled data, and preprocess the dataset;

[0055] S3. Based on the dataset obtained in step S2, construct a salt cavity morphology recognition model for compressed air energy storage power stations based on deep learning detection algorithms.

[0056] S4. Collect seismic data of the salt cavity of the compressed air energy storage power station to be detected, and input it into the salt cavity morphology recognition model of the compressed air energy storage power station in step S3 to obtain the salt cavity morphology of the compressed air energy storage power station to be detected.

[0057] In one specific embodiment, in step S1, the seismic data of the salt cavity area of ​​the compressed air energy storage power station includes three sources: first, actual three-dimensional seismic data obtained through three-dimensional seismic detection technology; second, forward-modeled seismic data obtained based on the theoretical salt cavity model; and third, enhanced seismic data obtained based on the actual three-dimensional seismic data through data augmentation technology. In one specific embodiment, the actual three-dimensional seismic data is used to acquire different seismic data slices (a seismic slice refers to seismic data at a certain depth), which are then saved as image data.

[0058] The forward modeling seismic data is obtained by using a theoretical salt cavity model, designing geological models of spherical salt cavities, columnar salt cavities, or irregularly shaped salt cavities, and then using seismic forward modeling software to save the forward modeling seismic data as image data.

[0059] The enhanced seismic data is obtained by rotating, flipping, cropping, scaling, and adding noise to the actual 3D seismic image data to expand the size of the dataset.

[0060] In one specific embodiment, in step S2, the salt cavity region of the seismic data is labeled to determine the boundary and outline of the salt cavity and generate a dataset with labeled data; the dataset is preprocessed by normalization, denoising and edge detection.

[0061] In one specific embodiment, the image is normalized to ensure that the image data input into the network has a uniform format and scale, which facilitates subsequent calculation and feature extraction.

[0062] The denoising process employs median filtering, Gaussian filtering, or bilateral filtering to remove noise from image data and improve image quality.

[0063] The edge detection uses the Canny or Sobel edge detection algorithm to extract edge information from image data and highlight the outlines of objects in the image.

[0064] In one specific embodiment, in step S3, the compressed air energy storage power station salt cavity morphology recognition model is constructed using a deep learning detection algorithm. The deep learning detection algorithm uses the Faster R-CNN (Faster Region-based Convolutional Neural Networks) algorithm, which is a deep learning-based target detection algorithm that can accurately detect the position and category of the target in the input image, thereby achieving an efficient target detection task.

[0065] Model building and training, such as Figure 2 As shown, the specific steps include:

[0066] S31. Dataset partitioning: Based on the preprocessed seismic image dataset obtained in step S2, it is divided into a training dataset and a test dataset; preferably, the proportion of the training dataset is 70% and the proportion of the test dataset is 30%.

[0067] S32. Determine the backbone network: Select the pre-trained VGG16 deep convolutional neural network as the backbone network of the Faster R-CNN algorithm, and extract features from the earthquake image training dataset to obtain the feature map.

[0068] S33. Training the Region Proposal Network (RPN): The feature map obtained in step S32 is input into the RPN. The RPN generates a series of anchor boxes of different scales and aspect ratios at each location on the feature map. Then, the RPN processes each anchor box through convolution operations and outputs the results of two branches: one branch is a classification network that predicts whether each anchor box contains a target, separating the salt cavity from the background; the other branch is a regression network that predicts the offset of the anchor box relative to the real salt cavity target box.

[0069] For classification networks, the cross-entropy loss function is used to calculate the predicted anchor box category and the true label; for regression networks, the smooth L1 loss function is used to calculate the loss between the predicted anchor box offset and the true offset; by calculating the two losses, the total loss of RPN is obtained.

[0070] Based on the calculated total loss, the gradient of the total loss with respect to each parameter in the RPN network is calculated using the backpropagation algorithm; the parameters of the RPN network are updated using the Adam optimization algorithm based on the gradient, so that the prediction results of the RPN gradually approach the true labels, and the RPN training is completed.

[0071] The parameters of an RPN network include the size and number of convolutional kernels;

[0072] S34. Training the Fast R-CNN detection network:

[0073] The trained RPN is used to generate candidate regions on the feature map of the training image, and the non-maximum suppression (NMS) algorithm is used to remove candidate regions with high overlap.

[0074] The generated candidate regions are mapped back to the feature maps extracted by the backbone network, and then the feature map regions corresponding to each candidate region are adjusted to feature vectors of a fixed size through the RoIPooling layer.

[0075] The feature vectors output by RoIPooling are input into the classification and regression networks. The classification network predicts the category of the target (including the background category) in each candidate region through fully connected layers and a softmax function, and the regression network predicts the further offset of the bounding box of each candidate region relative to the ground truth target box through fully connected layers.

[0076] For classification networks, the cross-entropy loss function is used to calculate the loss between the predicted target category and the true label; for regression networks, the smooth L1 loss function is used to calculate the loss between the predicted bounding box offset and the true offset. The two losses are added together to obtain the total loss of the Fast R-CNN detection network.

[0077] Based on the calculated total loss, the gradient of the total loss with respect to each parameter in the Fast R-CNN detection network is calculated using the backpropagation algorithm, and then the parameters of the Fast R-CNN detection network are updated using the optimizer.

[0078] The parameters of the Fast R-CNN detection network include the kernel size, which are obtained through continuous iteration.

[0079] S35. Alternating Training: RPN and Fast R-CNN detection networks are trained alternately;

[0080] First, train the RPN. Then, fix the parameters of the RPN and use the candidate regions generated by the RPN to train the Fast R-CNN detection network. Next, fix the parameters of the Fast R-CNN detection network and train the RPN again. Repeat this process until the model converges and the loss function no longer decreases significantly. At this point, the Fast R-CNN algorithm model training is complete.

[0081] S36. Verify the dataset test;

[0082] The validation dataset is input into the Fast R-CNN detection network obtained from the training dataset (the input is step seismic image data, and the output is the salt cavity morphology). The accuracy of target detection is calculated. If the recognition accuracy is greater than 90%, the Fast R-CNN model is used for subsequent recognition processes. Otherwise, the model parameters are adjusted to further optimize the model until the set accuracy is achieved.

[0083] In one specific embodiment, in step S4, the seismic data of the salt cavity of the compressed air energy storage power station to be detected is normalized and input into the Fast R-CNN detection network to obtain the salt cavity morphology recognition result.

[0084] In one specific embodiment, in step S4, the salt cavity morphology recognition result is further characterized by a threshold segmentation method; the threshold segmentation is based on the difference in grayscale between the target and the background, and by setting multiple thresholds, the target and the background are clearly separated; the threshold segmentation adopts multi-threshold segmentation, and the morphology of the salt cavity can be better characterized by threshold segmentation.

[0085] This invention provides an intelligent identification system for the salt cavity morphology of a compressed air energy storage power station. The system is used to implement the above-described method and includes:

[0086] The seismic data acquisition unit is used to acquire seismic data of the salt cavity area of ​​the compressed air energy storage power station.

[0087] The seismic data annotation and preprocessing unit is used to annotate the salt cavity regions of seismic data, generate a dataset with labeled data, and preprocess the dataset.

[0088] The model building and training unit is used to build and train a salt cavity morphology recognition model for compressed air energy storage power stations.

[0089] The salt cavity morphology recognition unit inputs the seismic data of the salt cavity of the compressed air energy storage power station to be detected into the compressed air energy storage power station salt cavity morphology recognition model to obtain the salt cavity morphology of the compressed air energy storage power station to be detected.

[0090] In one specific embodiment, the seismic data of the salt cavity area of ​​the compressed air energy storage power station includes three sources: first, actual three-dimensional seismic data obtained through three-dimensional seismic detection technology; second, forward-modeled seismic data obtained based on the theoretical salt cavity model; and third, enhanced seismic data obtained based on actual three-dimensional seismic data through data augmentation technology.

[0091] The identification method of this invention has high identification accuracy and fast identification speed, which can greatly reduce the design cycle and construction cost of compressed air energy storage power station salt cavern gas storage, and improve the profitability of compressed air energy storage power station; it can effectively handle the situation where there are multiple targets in the image and the targets overlap with each other, and accurately separate the target from the complex background.

[0092] While several embodiments of the present invention have been provided herein, those skilled in the art should understand that modifications can be made to these embodiments without departing from the spirit of the invention. The above embodiments are merely exemplary and should not be construed as limiting the scope of the invention.

Claims

1. A method for intelligent recognition of salt cavity morphology in a compressed air energy storage power station, characterized in that, The method includes: S1. Obtain seismic data for the salt cavity area of ​​the compressed air energy storage power station; S2. Label the salt cavity regions in the seismic data, determine the boundaries and outlines of the salt cavities, generate a dataset with labeled data, and preprocess the dataset. S3. Based on the dataset obtained in step S2, construct a salt cavity morphology recognition model for compressed air energy storage power stations based on a deep learning detection algorithm; the deep learning detection algorithm adopts the Faster R-CNN algorithm model, including: A pre-trained VGG16 deep convolutional neural network was selected as the backbone network of the Faster R-CNN algorithm to extract features from the earthquake image training dataset and obtain feature maps. The feature map is input into the RPN, which generates a series of anchor boxes of different scales and aspect ratios at each location on the feature map. The RPN processes each anchor box through convolution operations and outputs the results of two branches: one branch is a classification network that predicts whether each anchor box contains a target, separating the salt cavity from the background; the other branch is a regression network that predicts the offset of the anchor box relative to the real salt cavity target box. The generated candidate regions are mapped back onto the feature maps extracted by the backbone network, and the feature map regions corresponding to each candidate region are adjusted to feature vectors of a fixed size through the RoIPooling layer. The feature vectors output by RoIPooling are input into the classification network and the regression network; S4. Collect seismic data of the salt cavity of the compressed air energy storage power station to be detected, and input it into the salt cavity morphology recognition model of the compressed air energy storage power station in step S3 to obtain the salt cavity morphology of the compressed air energy storage power station to be detected.

2. The intelligent identification method for salt cavity morphology in a compressed air energy storage power station as described in claim 1, characterized in that, In step S1, the seismic data of the salt cavity area of ​​the compressed air energy storage power station includes three sources: first, actual three-dimensional seismic data obtained through three-dimensional seismic exploration technology; second, forward-modeled seismic data obtained based on the theoretical salt cavity model. Third, enhanced seismic data is obtained by using data augmentation techniques based on actual 3D seismic data.

3. The intelligent recognition method for salt cavity morphology in a compressed air energy storage power station as described in claim 2, characterized in that, The actual three-dimensional seismic data is used to obtain different seismic data slices, which are then saved as image data. The forward modeling seismic data is obtained by using a theoretical salt cavity model, designing geological models of spherical salt cavities, columnar salt cavities, or irregularly shaped salt cavities, and then using seismic forward modeling software to save the forward modeling seismic data as image data. The enhanced seismic data is obtained by rotating, flipping, cropping, scaling, and adding noise to the actual 3D seismic image data to expand the size of the dataset.

4. The intelligent identification method for salt cavity morphology in a compressed air energy storage power station as described in claim 1, characterized in that, In step S2, the salt cavity regions of the seismic data are labeled to determine the boundaries and outlines of the salt cavities, generating a dataset with labeled data; the dataset is preprocessed by performing normalization, denoising, and edge detection in sequence.

5. The intelligent identification method for salt cavity morphology in a compressed air energy storage power station as described in claim 4, characterized in that, The denoising process employs median filtering, Gaussian filtering, or bilateral filtering to remove noise from the image data; the edge detection uses the Canny or Sobel operator edge detection algorithm to extract edge information from the image data, highlighting the contours of objects in the image.

6. The intelligent identification method for salt cavity morphology in a compressed air energy storage power station as described in claim 1, characterized in that, In step S3, the compressed air energy storage power station salt cavity morphology recognition model is constructed using a deep learning detection algorithm. The deep learning detection algorithm adopts the Faster R-CNN algorithm model and includes the following steps: S31. Dataset partitioning: Based on the preprocessed seismic image dataset obtained in step S2, it is divided into a training dataset and a test dataset. S32. Determine the backbone network: Using the backbone network described in claim 1, extract features from the earthquake image training dataset to obtain a feature map; S33. Training the Region Proposal Network (RPN): The feature map obtained in step S32 is input into the RPN. The RPN generates a series of anchor boxes of different scales and aspect ratios at each location on the feature map. Then, the RPN processes each anchor box through convolution operations and outputs the results of two branches: one branch is a classification network that predicts whether each anchor box contains a target, separating the salt cavity from the background; the other branch is a regression network that predicts the offset of the anchor box relative to the real salt cavity target box. For classification networks, the cross-entropy loss function is used to calculate the predicted anchor box category and the true label; for regression networks, the smooth L1 loss function is used to calculate the loss between the predicted anchor box offset and the true offset; by calculating the two losses, the total loss of RPN is obtained. Based on the calculated total loss, the gradient of the total loss with respect to each parameter in the RPN is calculated using the backpropagation algorithm; the parameters of the RPN are updated using the Adam optimization algorithm based on the gradient, so that the prediction results of the RPN gradually approach the true labels, and the training of the RPN is completed. The parameters of an RPN network include the size and number of convolutional kernels; S34. Training the Fast R-CNN detection network: The trained RPN is used to generate candidate regions on the feature map of the training image, and the non-maximum suppression algorithm is used to remove candidate regions with high overlap. The generated candidate regions are mapped back to the feature maps extracted by the backbone network, and then the feature map regions corresponding to each candidate region are adjusted to feature vectors of a fixed size through the RoI Pooling layer. The feature vectors output by RoI Pooling are input into the classification network and the regression network. The classification network predicts the category of the target in each candidate region through fully connected layers and the Softmax function, and the regression network predicts the further offset of the bounding box of each candidate region relative to the real target box through fully connected layers. For classification networks, the cross-entropy loss function is used to calculate the loss between the predicted target category and the true label; for regression networks, the smooth L1 loss function is used to calculate the loss between the predicted bounding box offset and the true offset. The two losses are added together to obtain the total loss of the Fast R-CNN detection network. Based on the calculated total loss, the gradient of the total loss with respect to each parameter in the Fast R-CNN detection network is calculated using the backpropagation algorithm, and then the parameters of the Fast R-CNN detection network are updated using the optimizer. The parameters of the Fast R-CNN detection network include the kernel size, which is obtained through continuous iteration; S35. Alternating Training: RPN and Fast R-CNN detection networks are trained alternately; First, train the RPN. Then, fix the parameters of the RPN and use the candidate regions generated by the RPN to train the Fast R-CNN detection network. Next, fix the parameters of the Fast R-CNN detection network and train the RPN again. Repeat this process until the model converges and the loss function no longer decreases significantly. At this point, the Fast R-CNN algorithm model training is complete. S36. Verify the dataset test; The validation dataset is input into the Fast R-CNN detection network obtained from the training dataset. The accuracy of object detection is calculated. If the recognition accuracy is greater than 90%, the Fast R-CNN detection network is used for subsequent recognition processes. Otherwise, the model parameters are adjusted to further optimize the model until the set accuracy is achieved.

7. The intelligent recognition method for salt cavity morphology in a compressed air energy storage power station as described in claim 6, characterized in that, In step S4, the seismic data of the salt cavity of the compressed air energy storage power station to be detected is normalized and input into the Fast R-CNN detection network to obtain the salt cavity morphology recognition result.

8. The intelligent identification method for salt cavity morphology in a compressed air energy storage power station as described in claim 1, characterized in that, In step S4, the salt cavity morphology recognition result is further characterized by a threshold segmentation method. The threshold segmentation is based on the difference in grayscale between the target and the background. By setting multiple thresholds, the target and the background can be clearly separated. The threshold segmentation adopts multi-threshold segmentation.

9. A smart recognition system for the salt cavity morphology of a compressed air energy storage power station, characterized in that, The system is used to implement the method as described in any one of claims 1-8, the system comprising: The seismic data acquisition unit is used to acquire seismic data of the salt cavity area of ​​the compressed air energy storage power station. The seismic data annotation and preprocessing unit is used to annotate the salt cavity regions of seismic data, generate a dataset with labeled data, and preprocess the dataset. The model building and training unit is used to build and train a salt cavity morphology recognition model for compressed air energy storage power stations. The salt cavity morphology recognition unit inputs the seismic data of the salt cavity of the compressed air energy storage power station to be detected into the compressed air energy storage power station salt cavity morphology recognition model to obtain the salt cavity morphology of the compressed air energy storage power station to be detected.

10. The intelligent recognition system for salt cavity morphology in a compressed air energy storage power station as described in claim 9, characterized in that, The seismic data for the salt cavity area of ​​the compressed air energy storage power station includes three sources: first, actual three-dimensional seismic data obtained through three-dimensional seismic detection technology; second, forward-modeled seismic data obtained based on the theoretical salt cavity model. Third, enhanced seismic data is obtained by using data augmentation techniques based on actual 3D seismic data.

Images