Hydropower station leakage set water level monitoring method and system based on image time sequence super-division
By using image temporal super-resolution technology and employing ResNet or Transformer models for hydropower station seepage collection water level monitoring, the problems of high computational resources, poor monitoring effect, and low efficiency of manual monitoring in existing technologies are solved, achieving high-precision, low-cost water level monitoring and unmanned inspection.
Patent Information
- Application Number
- CN202511043807.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-28
- Publication Date
- 2025-11-21
AI Technical Summary
Existing water level monitoring technologies for seepage collection wells in hydropower stations suffer from problems such as high computational resource requirements, difficulty in obtaining high-definition images, poor performance of monitoring equipment in complex environments, low efficiency of manual monitoring, and safety hazards.
A temporal super-resolution method based on images is adopted, which uses ResNet or Transformer models for image super-resolution processing, and combines feature extraction and loss function training to achieve high-definition and real-time monitoring of water level images.
It improves the accuracy of water level monitoring and its applicability to complex environments, reduces hardware costs and the lag of manual monitoring, realizes unmanned inspection, and reduces safety risks.
Smart Images
Figure CN120997655A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image processing, in particular to a water level monitoring method and system for a water leakage set of a hydropower station based on image time sequence super-resolution. BACKGROUND
[0002] Real-time monitoring of the water level of the water leakage set of the hydropower station is a key link to ensure the safe operation of the dam. The traditional water level monitoring mainly relies on the immersion liquid pressure type water level sensor, but long-term use may cause measurement drift due to scale, soil adhesion, and even abnormal data, which may cause accidents such as flooding of the plant or burning of the water pump. In addition, the water collection well is usually located in the deep part of the dam, and the environment is narrow, dark, and humid, so manual inspection is inefficient and has safety hazards.
[0003] The existing water level monitoring technology based on security monitoring mainly includes the following steps: (1) completing image data acquisition of the water level of the water leakage set of the hydropower station by installing a camera; (2) building a central control platform; and (3) manually monitoring the central control platform. Through the above steps, the labor and time cost of real-time water level monitoring of the water collection well of the hydropower station is reduced to a certain extent. The disadvantage is that a large amount of human, material and financial resources are consumed, and a more important reason is that the existing monitoring equipment cannot effectively obtain high-definition images of the water level changes of the water leakage collection well in real time due to the narrow and small space, dark light, and serious water vapor in the water leakage collection well, so that the monitoring screen cannot truly reflect the actual situation inside the water leakage collection well, resulting in the failure of manual monitoring.
[0004] In the prior art, in order to improve the monitoring reliability, some hydropower stations introduce a video monitoring method based on a camera to collect images and combine a neural network. However, the existing method has the following defects: 1. The method based on the convolutional neural network can only focus on local features, but ignores global information; 2. Considering the problem of image resolution, the existing resolution network is usually large and requires high computing resources, which limits its actual application in some resource-limited scenarios such as embedded systems and mobile devices; 3. The method involves a convolutional neural network (CNN), which mainly realizes high-resolution reconstruction of images through multi-layer convolution and deconvolution operations; mainly uses convolution operation to extract features, and performs image reconstruction through upsampling and deconvolution operation, which has high computing efficiency, but less involves reconstruction of high-definition images in the water collection well. SUMMARY
[0005] The main purpose of the present application is to provide a water level monitoring method and system for a water leakage set of a hydropower station based on image time sequence super-resolution, which solves at least one of the technical problems involved in the prior art.
[0006] To solve the above technical problems, the technical scheme adopted by the present application is: a water level monitoring method for a leakage set of a hydropower station based on image timing super-resolution, comprising the following steps:
[0007] S1, image acquisition: collect state data of different scenes inside the leakage set water well, and perform data preprocessing; the state data includes water level images;
[0008] S2, image super-resolution processing: using ResNet or Transformer model for image super-resolution processing, high-definition processing of sample images inside the leakage set water well; based on timing control, using resolution timing scaling factor to control the conversion from low-resolution image to high-resolution image;
[0009] S3, image feature extraction: obtaining the image super-resolution processed image, using ResNet or Transformer feature extraction model as the backbone network for feature extraction, super-resolution feature and corresponding original image feature are fused to form double joint features, and after nonlinear processing, the state sample hierarchical features inside the leakage set water well are obtained;
[0010] S4, model training: obtaining each water level image with corresponding label information, setting a corresponding recognition loss function, setting a super-resolution reconstruction loss function for image super-resolution processing, and training the model jointly with the loss function and the super-resolution reconstruction loss function;
[0011] S5, water level monitoring: after completing the model optimization training, input the to-be-tested image data, realize real-time recognition of the water level state through feature matching, and obtain the water level monitoring result.
[0012] In the preferred scheme, the sample data set in S1 is divided by batch, and each batch contains images under different viewing angles, lighting or water vapor conditions.
[0013] In the preferred scheme, after the data is collected by the camera, sensor or mobile device, the data is de-duplicated and diversified;
[0014] Random cropping, scaling, twisting and other operations are performed to facilitate subsequent super-resolution enhancement processing and enhance the diversity of the data and the generalization ability of the model.
[0015] In the preferred scheme, in S2, the image super-resolution processing is performed, that is, the high-resolution image is restored from the low-resolution image, and the Gaussian noise is gradually added through the forward timing diffusion model, which specifically includes;
[0016] An image degradation model is used, and the formula is:
[0017] I LR =φ(I HR ,θ)(1);
[0018] where θ represents the blur kernel, down-sampling operation, noise, and other model parameters, I LR is the observed LR image, I HR is the original HR image.
[0019] Image super-resolution uses the inverse process of the degradation model to recover the super-resolution image I LR from the low-resolution image I SR , which is expressed as:
[0020] I SR = φ -1 (I LR , δ) = S(I LR , δ) (2).
[0021] where S is the super-resolution model (inverse degradation model), δ is the set of parameters, I SR is the image super-resolution result.
[0022] In the preferred embodiment, the image super-resolution processing based on time sequence control in S2 uses the resolution time sequence scaling factor to control the different degrees of super-resolution image, which is specifically:
[0023] Given a pair of high and low resolution images (Y LR , Y HR ), where Y LR , Y HR represent the low and high resolution image states, respectively, and the resolution output time sequence control factor is λ, which is used to control the different degrees of image resolution, a forward time sequence diffusion model S t is defined, which is expressed as:
[0024]
[0025] where β t ∈ (0, 1) is the Gaussian noise, T is the number of time sequence iterations, y0 is the image of the given resolution, y t is the resolution image at time t, and I is the identity matrix.
[0026] In the preferred embodiment, the degradation model is used to denoise the image (low-resolution image) to recover the target image and reconstruct the high-definition image, which is expressed as:
[0027]
[0028] wherein In order to obtain continuous super-resolution image output, the actual noise model is expressed as x is the given input image, and the entire time sequence super-resolution optimization loss function L isl□ The formula is:
[0029]
[0030] wherein t ~ (1,..., T), and lambda is subject to a uniform distribution.
[0031] In a preferred solution, the loss function in S4 includes a super-resolution reconstruction loss and a water level classification loss function, that is, after obtaining the double joint features, a classification loss function is used for training to realize water level monitoring and training recognition, and the formulas are respectively:
[0032]
[0033] wherein, represents the predicted label, and k represents the true label, that is, the real class label.
[0034] In a preferred solution, the joint loss function is a combination of recognition classification loss and super-resolution reconstruction loss, and the hyperparameter tau dynamically balances the weights of super-resolution reconstruction and water level classification, and the formula is:
[0035] L = tau * L ID + (1-tau) * L lh (9);
[0036] wherein L is the overall loss function, and tau is a hyperparameter for controlling recognition and water level reconstruction.
[0037] In a preferred solution, S3 is specifically:
[0038] After the S2 time series super-resolution operation, the reconstructed super-resolution sample obtained is again subjected to a ResNet50 or Transformer backbone network to obtain the corresponding super-resolution feature map, and then a fully connected layer is used to obtain the final super-resolution reconstruction feature.
[0039] The state image data after preprocessing in step S1 is extracted by forward propagation to form the final double joint features.
[0040] A water level monitoring system for a hydropower station leakage set based on image time series super-resolution, comprising:
[0041] An image acquisition module for acquiring state data under different scenes inside the leakage set water well after data preprocessing; the state data includes water level images.
[0042] An image super-resolution processing module is configured to perform image super-resolution processing by using a ResNet or a Transformer model, and to perform high-definition processing on sample images inside the leakage collection well; based on time sequence control, a resolution time sequence scaling factor lambda is used to control the conversion of a low-resolution image into a high-resolution image;
[0043] An image feature extraction module is configured to obtain the image processed by the image super-resolution, and to use a feature extraction model of the ResNet or the Transformer as a backbone network to perform feature extraction, to fuse the super-resolution features and the corresponding original image features to form double combined features, and to obtain the sample level features of the internal state of the leakage collection well after nonlinear processing.
[0044] A model training module is configured to obtain corresponding label information for each water level image, to set a corresponding recognition loss function, to set a super-resolution reconstruction loss function for the image super-resolution processing, and to jointly train the model by using the recognition loss function and the super-resolution reconstruction loss function.
[0045] A water level monitoring module is configured to input image data to be measured after the model optimization training is completed, to realize real-time recognition of the water level state by feature matching, and to obtain a water level monitoring result.
[0046] The application provides a water level monitoring method and system for a leakage collection well of a hydropower station based on image time sequence super-resolution, which comprises the following steps: S1, image acquisition: collecting state data of different scenes inside the leakage collection well and performing data preprocessing; S2, image super-resolution processing: performing image super-resolution processing by using a ResNet or a Transformer model, and performing high-definition processing on sample images inside the leakage collection well; based on time sequence control, a resolution time sequence scaling factor lambda is used to control the conversion of a low-resolution image into a high-resolution image; S3, image feature extraction: obtaining the image processed by the image super-resolution, and using a feature extraction model of the ResNet or the Transformer as a backbone network to perform feature extraction, to fuse the super-resolution features and the corresponding original image features to form double combined features; S4, model training: obtaining corresponding label information for each water level image, setting a corresponding recognition loss function, setting a super-resolution reconstruction loss function for the image super-resolution processing, and jointly training the model by using the recognition loss function and the super-resolution reconstruction loss function; S5, water level monitoring: inputting image data to be measured after the model optimization training is completed, realizing real-time recognition of the water level state by feature matching, and obtaining a water level monitoring result; the low-quality monitoring image is processed by using the time sequence super-resolution algorithm, the blur interference caused by water vapor and insufficient light is eliminated, the accuracy of data processing is improved, the applicability of the complex environment is improved, the hardware deployment and maintenance cost is reduced, the compatibility of the existing monitoring system is improved, unmanned inspection is realized, and the lag and safety risk of manual monitoring are reduced. BRIEF DESCRIPTION OF DRAWINGS
[0047] The application will be further described below in conjunction with the accompanying drawings and embodiments:
[0048] Fig. 1 is a flow chart of the water level monitoring method of the application;
[0049] Fig. 2 is a schematic diagram of the overall structure of the water level monitoring model of the application;
[0050] Fig. 3 is a comparison diagram of the original picture and the classified image of the water level monitoring method of the application. DETAILED DESCRIPTION
[0051] Embodiment 1
[0052] As shown in the figure, a water level monitoring method for a leakage set of a hydropower station based on image time sequence super-resolution includes the following steps: Figs. 1-3
[0053] S1, image acquisition: collect state data under different scenes inside the leakage set water well, and perform data preprocessing; the state data includes water level images.
[0054] S2, image super-resolution processing: using a ResNet or Transformer model f(θ) to perform image super-resolution processing, and performing high-definition processing on the sample images inside the leakage set water well: based on time sequence control, using a resolution time sequence scaling factor λ to control the conversion of a low-resolution (LR) image to a high-resolution (HR) image.
[0055] S3, image feature extraction: obtaining the image after super-resolution processing, using a feature extraction model of ResNet or Transformer as a backbone network to perform feature extraction, fusing the super-resolution features and the corresponding original image features to form double joint features, and after nonlinear processing, obtaining the hierarchical features of the state samples inside the leakage set water well.
[0056] S4, model training: obtaining each water level image with corresponding label information, setting a corresponding recognition loss function, setting a super-resolution reconstruction loss function for image super-resolution processing, and jointly training the model using the identification loss function and the super-resolution reconstruction loss function.
[0057] S5, water level monitoring: after completing the model optimization training, inputting the to-be-tested image data, realizing real-time identification of the water level state through feature matching, and obtaining the water level monitoring result.
[0058] The embodiment realizes training and optimization of the model by collecting and diversifying and preprocessing scene data inside the leakage water collecting well to obtain diversified state feature sample data, realizes high-definition processing of the leakage water collecting well internal scene sample image features, realizes scaling operation of low-quality images to high-definition denoising images through time sequence super-resolution control, performs high-definition decoding reconstruction on low-quality images, realizes joint learning of two modal features by combining original input samples and high-definition reconstruction samples, and realizes real-time monitoring and identification of the water level of the leakage water collecting well through the water level monitoring model.
[0059] The embodiment is implemented as follows.
[0060] 1. Leakage water collecting well internal scene sample data collection.
[0061] 1.1 Obtain the leakage water collecting well internal scene sample data (images or videos) to be processed. The data is collected by a camera, a sensor or the like. The collected data is screened and processed, and the data with high similarity is removed to avoid excessive similar defect samples and increase the diversity of data categories.
[0062] 1.2 Construct a data set, randomly select a sequence of data samples as a batch from the data set, and as much as possible, select samples with different perspectives, different scales and the like under different scene types for the image samples in a batch, so that as many categories as possible are presented.
[0063] In the embodiment, a camera of Hikvision / Tian Di Weiye is installed, and a searchlight is arranged beside the camera to obtain real-time water level monitoring data in the water collecting well.
[0064] In the preferred solution, the sample data set in step S1 is divided by batch (batch), and each batch contains images under different perspectives, illuminations or water vapor conditions.
[0065] In the preferred solution, after the data is collected by the camera, the sensor or the mobile device, the data is removed and screened for diversity.
[0066] Random cropping, scaling, twisting and the like are further performed to facilitate subsequent super-resolution enhancement processing and enhance the diversity of the data and the generalization ability of the model.
[0067] The embodiment preprocesses the obtained image data, mainly including data checking, data cleaning, data sample enhancement and the like, specifically:
[0068] A1 data inspection: check whether the number of state sample data (record number) meets the minimum requirement of super-resolution algorithm, and whether the internal scene of leakage collection well meets the standard and requirement required by the method. Specifically, the obtained state image is normalized and the mean value is subtracted, so that the image has the same size and proportion, and the image is adjusted to a fixed size.
[0069] A2 data cleaning: for the obvious defects, missing values and abnormal samples found in the data inspection process, appropriate methods are selected for cleaning, which is conducive to the subsequent statistical analysis to obtain reliable feature expression.
[0070] A3 data sample conversion enhancement: data analysis emphasizes the comparability of objects, and some statistical indicators need to be analyzed before data transformation, dimensionless processing, linear transformation, summary and aggregation, moderate generalization, normalization and attribute construction; It is also a preliminary assessment and judgment of whether the data meets the needs of statistical analysis, and whether the data amount needs to be increased or decreased. In order to increase the diversity of data and the robustness of the model, the data samples are enhanced by rotation, clipping and scaling, etc. to expand the data set scale and improve the model generalization ability. And use simple linear model, scatter plot, histogram and other image transformations for exploratory analysis, use correlation analysis, consistency check and other methods to verify the accuracy of the data, ensure that no error and bias data is brought into the actual data analysis, and improve the accuracy of the data.
[0071] In step S2, sample high-definition processing based on image super-resolution processing is performed. Image super-resolution, simply referred to as image super-resolution, aims to recover a high-resolution image from a low-resolution image. SR focuses on recovering the corresponding HR image from the observed LR image.
[0072] In a preferred scheme, image super-resolution processing is performed in step 2, and Gaussian noise is gradually added through a forward time diffusion model, specifically including:
[0073] An image degradation model is used, and the formula is:
[0074] I LR = φ(I HR , θ) (1).
[0075] Where θ represents model parameters such as blur kernel, downsampling operation, noise, etc. I LR is the observed LR image, and I HR is the original HR image.
[0076] Image super-resolution uses the inverse process of the degradation model to recover the super-resolution image I LR from the low-resolution image I SR , and the expression is:
[0077] I SR =φ -1 (I LR ,δ)=S(I LR ,δ) (2);
[0078] where S is a super-resolution model (inverse degradation model), δ is a set of parameters, I SR is the image super-resolution result.
[0079] The image super-resolution processing is further controlled in time sequence.
[0080] In the preferred embodiment, the image super-resolution processing based on time sequence control in step S2 is performed in the manner of resolution time sequence scaling factor to control different degrees of super-resolution image, specifically:
[0081] Given a pair of high and low resolution images (Y LR , Y HR ), where Y LR , Y HR represent low and high resolution image states respectively, the resolution output time sequence control factor is λ, which is used to control different degrees of image resolution, and a forward time sequence diffusion model S t is defined, and the formula is:
[0082]
[0083] where β t ∈(0,1) is Gaussian noise, T is the number of time sequence iterations, y0 is the image of given resolution, y t is the resolution image at time t, and I is the unit matrix.
[0084] In the preferred embodiment, the degradation model is used to denoise the image (low resolution image) to recover the target image and reconstruct the high-definition image, and the formula is:
[0085]
[0086] wherein, In order to obtain continuous super-resolution image output, the actual noise model is represented as x is the given input image, and the entire time sequence super-resolution optimization loss function L lh is:
[0087]
[0088] where t ~ (1, …, T), and λ is uniformly distributed.
[0089] In a preferred solution, the double combined features in step S3 are fused through a full connection layer.
[0090] Step S3 extracts features from the super-resolution image, and the specific steps include:
[0091] 1) Super-resolution sample feature acquisition: after the time sequence super-resolution operation in step S2, the reconstructed super-resolution sample Again, using ResNet50 or Transformer as the backbone network, the corresponding super-resolution feature map is obtained, and then the final super-resolution reconstruction feature is obtained through a full connection layer.
[0092] 2) Original feature extraction: considering the error information generated by the super-resolution reconstruction feature, the original image data is supplemented and improved as necessary, so the feature extraction is completed through forward propagation for the preprocessed image data in step S1. Combined with the super-resolution reconstruction feature, the final double combined feature is formed, and the internal state sample hierarchical feature of the leakage collection well after nonlinear processing is completed.
[0093] After obtaining the double combined features (super-resolution features and original features) through the water level monitoring model based on ResNet50 or Transformer, the corresponding super-resolution reconstruction loss and water level classification loss function are selected to train the water level monitoring model, and the water level monitoring and recognition are completed.
[0094] In a preferred solution, the loss function in step S4 includes the super-resolution reconstruction loss and the water level classification loss function, that is, after obtaining the double combined features, the classification loss function is used for training, and the formula is:
[0095]
[0096] wherein, represents the predicted label, and k represents the true label, that is, the real class label.
[0097] In a preferred solution, the combined loss function is a combination of the recognition classification loss and the super-resolution reconstruction loss, and the super parameter τ dynamically balances the weights of super-resolution reconstruction and water level classification, and the formula is:
[0098] L = τ * L ID + (1-τ) * L lh (9);
[0099] wherein, L is the overall loss function, and τ is the super parameter for controlling recognition and water level reconstruction.
[0100] Further, system development: after completing the corresponding water level image super-resolution and water level monitoring algorithm development, the corresponding model is encapsulated and connected to the software platform developed, realizing real-time water level monitoring of leakage collection well: specifically:
[0101] 1) Development of water level online monitoring software system for collection well, using micro-service architecture, combining software development process, according to demand analysis, function system design, implementation development, test application, etc. Real-time water level online monitoring system is developed.
[0102] 2) Model migration. The water level monitoring model trained by this method is transplanted to the corresponding developed software platform to realize interface calling and complete real-time monitoring of the water level of the leakage collection well of the hydropower station.
[0103] As Fig. 3 shown, the contrast chart obtained after the method of this embodiment is run shows that the clarity of the image after super-resolution processing is better than that of the original picture. Through the time series super-resolution algorithm, the low-quality monitoring image is processed to high definition, eliminating the blur interference caused by water vapor and insufficient light, improving the accuracy of data processing, improving the applicability of complex environments, and realizing real-time and efficient monitoring of the water level of the leakage collection well of the hydropower station.
[0104] Example 2
[0105] In combination with Example 1, a water level monitoring system for leakage collection of a hydropower station based on image time series super-resolution includes:
[0106] An image acquisition module is used to acquire state data under different scenes inside the leakage collection well, and the state data includes water level images after data preprocessing.
[0107] An image super-resolution processing module is used to perform image super-resolution processing using a ResNet or Transformer model to process the sample images inside the leakage collection well to high definition: based on time series control, a resolution time series scaling factor λ is used to control the conversion from a LR image to a HR image.
[0108] An image feature extraction module is used to obtain the image super-resolution processed image, use a ResNet or Transformer feature extraction model as a backbone network to perform feature extraction, fuse the super-resolution features with the corresponding original image features to form double joint features, and then perform nonlinear processing to obtain the state sample hierarchical features inside the leakage collection well.
[0109] A model training module is used to obtain each water level image with corresponding label information, set a corresponding recognition loss function, set a super-resolution reconstruction loss function for image super-resolution processing, and train the model jointly with the loss function and the super-resolution reconstruction loss function.
[0110] The water level monitoring module is used for inputting image data to be measured after model optimization training, realizing real-time identification of the water level state through feature matching, and obtaining a water level monitoring result.
[0111] The above embodiments are only preferred technical solutions of the present application, and should not be regarded as limitations of the present application. The protection scope of the present application should be based on the technical solutions recited in the claims, including equivalent replacement solutions of the technical features recited in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present application.
Claims
1. A water level monitoring method for a hydropower station leakage set based on image timing super-resolution, characterized in that, Comprising the following steps: S1, image acquisition: collect the state data of different scenes inside the leakage water collecting well, and perform data preprocessing; the state data includes water level images; S2, image super-resolution processing: using ResNet or Transformer model for image super-resolution processing, high-definition processing of sample images inside the leakage water collecting well; based on time sequence control, using resolution time sequence scaling factor to control the conversion from low resolution image to high resolution image; S3, image feature extraction: obtaining the image super-resolution processed image, using ResNet or Transformer feature extraction model as the backbone network for feature extraction, super-resolution feature and corresponding original image feature fusion to form double combined features, and after nonlinear processing, obtaining the leakage water collecting well internal state sample hierarchical features; S4, model training: obtaining each water level image with corresponding label information, setting the corresponding recognition loss function, setting the super-resolution reconstruction loss function for image super-resolution processing, and jointly training the model with the loss function and the super-resolution reconstruction loss function; S5, water level monitoring: after completing the model optimization training, inputting the test image data, realizing real-time recognition of water level state through feature matching, and obtaining water level monitoring result.
2. The water level monitoring method of hydropower station leakage set based on image timing super-resolution according to claim 1, characterized in that, The sample data set in S1 is divided by batch, and each batch contains images under different viewing angles, light or water vapor conditions.
3. The water level monitoring method of hydropower station leakage set based on image timing super-resolution according to claim 1, characterized in that, In S1, the data is collected by camera, sensor or mobile device, and then de-duplication and diversity screening is performed on the data; Random cropping, scaling, twisting and other operations are performed to facilitate subsequent super-resolution enhancement processing and enhance the diversity of data and the generalization ability of the model.
4. The water level monitoring method of hydropower station leakage set based on image timing super-resolution according to claim 1, characterized in that, In S2, the image super-resolution processing is performed, that is, the high-resolution image is restored from the low-resolution image, and the Gaussian noise is gradually added through the forward time sequence diffusion model, which specifically includes: An image degradation model is used, and the formula is: I LR = φ(I HR , θ) (1); where θ denotes model parameters such as blur kernel, down-sampling operation, noise, etc., I LR is the observed LR image, HR is the original HR image; Image super-resolution uses the inverse process of the simulation degradation model to recover a super-resolution image I LR from a low-resolution image I SR , expressed as: I sR = φ -1 (I LR , δ) = S(I LR , δ) (2); Wherein, S is a super-resolution model (inverse degradation model), δ is a set of parameters, I SR is the image super-resolution result.
5. The water level monitoring method of hydropower station leakage set based on image timing super-resolution according to claim 4, characterized in that, In S2, the image super-resolution processing based on time sequence control uses the resolution time sequence scaling factor to control the super-resolution image of different degrees, which is specifically: Given a pair of high and low resolution image pairs (Y LR ,Y HR ), where Y LR ,Y HR represent low and high resolution image states respectively, with a resolution output temporal control factor λ for controlling different levels of image resolution, a forward temporal diffusion model S t is defined, with the formula: where β t ∈(0,1) is a Gaussian noise, T is the number of time iterations, y0is an image of a given resolution, y t is a resolution image at time t, and I is an identity matrix.
6. The water level monitoring method of hydropower station leakage set based on image timing super-resolution according to claim 4, characterized in that, Based on a degradation model Denoised image (low resolution image) restores target image Reconstructing a high definition image, the formula is: wherein, In order to obtain a continuous super-resolution image output, the actual noise model is represented as x is a given input image, so that the entire temporal super-resolution optimization loss function L lh The formula is: Wherein, t~(1,…,T), and lambda obeys uniform distribution.
7. The water level monitoring method of hydropower station leakage set based on image timing super-resolution according to claim 6, characterized in that, In S4, the loss function includes super-resolution reconstruction loss and water level classification loss function, that is, after obtaining the double combined features, the classification loss function is used for training to realize water level monitoring and training recognition, and the formulas are respectively: wherein, denotes the predicted label, k denotes the true label, i.e., the true class label.
8. The water level monitoring method of hydropower station leakage set based on image timing super-resolution according to claim 7, characterized in that, The joint loss function is the combination of recognition classification loss and super-resolution reconstruction loss, and the super parameter tau dynamically balances the weight of super-resolution reconstruction and water level classification, and the formula is: L = τ * L ID + (1 - τ) * L lh (9) ; Wherein, T is the overall loss function, and tau is the super parameter for controlling recognition and water level reconstruction.
9. The water level monitoring method of hydropower station leakage set based on image timing super-resolution according to claim 1, characterized in that, In S3, it is specifically: After the S2 timing super-resolution operation, the reconstructed super-resolution samples obtained Again, after taking ResNet50 or Transformer as the backbone network, the corresponding super-resolution feature map is obtained, and then the last super-resolution reconstruction feature is obtained through the fully connected layer. The state image data after preprocessing in S1 is extracted by forward propagation to form the final double combined features.
10. A water level monitoring system for a hydropower station leakage set based on image timing super-resolution, characterized in that, It includes: An image acquisition module for acquiring state data of different scenes inside the leakage water collecting well, and performing data preprocessing; the state data includes water level images; An image super-resolution processing module is configured to perform image super-resolution processing by using a ResNet or a Transformer model, and to perform high-definition processing on the sample image inside the infiltration catch basin; based on time sequence control, a resolution time sequence scaling factor λ is used to control the conversion of a low-resolution image to a high-resolution image; An image feature extraction module is configured to obtain the image processed by the image super-resolution processing, to use a feature extraction model of the ResNet or the Transformer as a backbone network to perform feature extraction, to fuse the super-resolution features with the corresponding original image features to form double combined features, and to obtain the sample hierarchical features of the internal state of the infiltration catch basin after nonlinear processing. A model training module is configured to obtain each water level image with corresponding label information, to set a corresponding recognition loss function, to set a super-resolution reconstruction loss function for the image super-resolution processing, and to perform model training in combination with the recognition loss function and the super-resolution reconstruction loss function. A water level monitoring module is configured to input the to-be-tested image data after the model optimization training, to realize real-time recognition of the water level state through feature matching, and to obtain a water level monitoring result.