Earth-rockfill dam leakage thermal infrared image prediction method and device

Abstract

The invention discloses an earth-rockfill dam leakage thermal infrared image prediction method and device, and belongs to the technical field of earth-rockfill dam leakage monitoring and prediction. The method comprises the following steps: continuously acquiring thermal infrared images of an earth and rockfill dam leakage hidden danger area by using a distributed thermal infrared ground station; and inputting the infrared image with the time sequence into the SA-E3D-LSTM network model for training, and predicting the change condition of the thermal infrared image in a future time period. The device comprises a dual-light thermal infrared imager, a photovoltaic power supply module, a data receiving and transmitting module, a fixing frame, a steering device, a central server and a user terminal. According to the earth-rock dam leakage thermal infrared image prediction method and device provided by the invention, thermal infrared image prediction of earth-rock dam leakage can be realized, the method has the advantages of interference resistance, wide coverage range, good space-time continuity and the like, and the problem that a traditional point type leakage prediction method cannot reflect the distribution and development trend of leakage in space is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of leakage monitoring and prediction of earth-rock dams, and particularly to a method and device for predicting leakage of earth-rock dams by using thermal infrared images. Background Art

[0002] As an important water conservancy infrastructure, the structural safety of earth-rock dams is directly related to the normal operation of infrastructure such as flood control, irrigation, and power generation. Leakage is one of the main reasons for the failure of earth-rock dams. Therefore, how to effectively monitor and predict the leakage of dams has become a key issue in engineering safety management. If leakage is not discovered and disposed of in time, it will lead to damage to the structural integrity of earth-rock dams, a decline in the ability to resist disasters, pose a threat to their daily operation and maintenance, and even trigger serious disasters such as breaches and dam failures, causing irreversible social losses.

[0003] The leakage of earth-rock dams is characterized by large spatio-temporal randomness and strong concealment. Timely discovery of leakage and understanding of the development of leakage are crucial for ensuring the safety of earth-rock dams. At present, the prediction methods for earth-rock dam leakage are mainly based on point-type monitoring data such as water level and water pressure, and the prediction results cannot reflect the distribution and development trend of leakage in space. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and device for predicting leakage of earth-rock dams by using thermal infrared images. Through the extraction of spatio-temporal features of thermal infrared image sequences and time series modeling, the leakage trend of earth-rock dams is predicted, so as to provide technical and equipment support for anti-seepage layout design and leakage emergency disposal.

[0005] To achieve the above purpose, the present invention provides a method and device for predicting leakage of earth-rock dams by using thermal infrared images, which continuously collect thermal infrared images of the leakage hazard area of earth-rock dams by using a distributed thermal infrared ground station;

[0006] The thermal infrared images with time series are input into the SA-E3D-LSTM network model for training, and a future thermal infrared image sequence is output.

[0007] The method includes the following steps:

[0008] S1. Data acquisition: The leakage hazard area of the earth-rock dam is monitored in real time by using a distributed thermal infrared ground station, and continuous thermal infrared images and visible light image sequences are collected. The thermal infrared image sequence reflects the change trend of the surface thermal distribution of the leakage hazard area of the earth-rock dam over time and space, and the image sequence includes the thermal distribution information on the surface of the earth-rock dam; before collecting the thermal infrared images and visible light images at the earth-rock dam site, the scope of the leakage hazard area of the earth-rock dam is divided, and a number of distributed thermal infrared ground stations are arranged;

[0009] S2. Data Input and Preprocessing: The input data is a time-series of thermal infrared images. The thermal infrared image sequence data consists of T frames, each frame having a dimension of H×W×C (height, width, channels). The data is divided into small 3D data blocks, usually divided by a fixed time window, and each window consists of several frames, forming a tensor of H×W×D×C. These 3D data blocks will be the input to the SA-E3D-LSTM network model, where D represents the number of frames in the time window (depth dimension).

[0010] The input data is represented as:

[0011] X t ∈R H×W×C for t = 1, 2, …, T

[0012] where: t is the time step, H is the height of the frame, W is the width of the frame, and C is the number of channels (e.g., the number of RGB channels is 3);

[0013] S3. Spatiotemporal Feature Extraction: Spatiotemporal features of the preprocessed thermal infrared image sequence are extracted through 3D convolution, which can extract local spatiotemporal features in both the spatial and temporal dimensions simultaneously;

[0014] S4. Time Series Modeling: The spatiotemporal features extracted from 3D convolution are modeled through a long short-term memory network (LSTM) to capture long-term temporal dependencies in the thermal infrared image sequence;

[0015] S5. Enhancement of the Improved Eidetic Memory Mechanism: During the update process of the LSTM hidden state, 3D convolution operations are used to reconstruct the spatial features of the cell state to enhance the memory ability for spatiotemporal features, and a spatial attention mechanism is introduced to weightedly screen the thermal imaging abnormal features in the leakage area, improving the ability to capture the leakage features of the thermal infrared images of earth-rock dams;

[0016] S6. Thermal Infrared Image Prediction: Based on the output of the LSTM network, the change trend of the thermal infrared images of the potential leakage areas of the earth-rock dam in the future time period is predicted to predict the leakage situation of the earth-rock dam;

[0017] S7. Result Output: The predicted thermal infrared images are output and a field prediction report related to the leakage risk is generated for evaluating the safety status of the earth-rock dam.

[0018] Preferably, in step S1, the distributed thermal infrared ground station has the ability to stably collect thermal infrared images and visible light image sequence data in complex environments (such as mountainous areas, windy areas, rainy areas, etc.). Compared with the drone flight platform, the distributed infrared ground station can obtain more accurate and comprehensive thermal distribution information on the surface of the earth-rock dam.

[0019] Preferably, in step S3, when the 3D convolution extracts the spatio-temporal features of the thermal infrared image, the convolutional kernel slides simultaneously in the spatial and temporal dimensions to capture the features of the surface thermal distribution of the earth-rock dam changing with time;

[0020] At each time step t, the input sequence segment X of the thermal infrared image t is fed into the SA-E3D-LSTM network model. The main function of LSTM is to process the dependencies in the time dimension and capture the dynamic information between the current time step and the past time steps. The structure of LSTM updates and controls the hidden state and memory state through three gating mechanisms:

[0021] Input gate: Determines the input information to be added to the current memory.

[0022] Forget gate: Determines which information to discard from the current memory.

[0023] Output gate: Determines the output of the current time step.

[0024] For each time step, the mathematical expression of LSTM is as follows:

[0025] i t = σ(W i [h t-1 , X t +b i )(Input gate);

[0026] f t = σ(W f [h t-1 , X t +b f )(Forget gate);

[0027] o t = σ(W o [h t-1 , X t +b o )(Output gate);

[0028] (Candidate memory unit);

[0029] (Memory state update);

[0030] h t = o t ⊙ tanh(C t )(Hidden state);

[0031] Where: i t , f t , o tThey are the input gate, forget gate, and output gate respectively. is the candidate memory, C t is the hidden state, σ is the sigmoid activation function, tanh is the hyperbolic tangent function, and ⊙ is the element-wise multiplication.

[0032] LSTM can handle the temporal correlations in the thermal infrared image sequence and capture the relationships between each frame and its preceding and succeeding frames.

[0033] Preferably, in step S4, the input of the LSTM network is the spatio-temporal feature map output by the 3D convolution. The LSTM network models the temporal sequence dependencies of the input spatio-temporal features through a gating mechanism.

[0034] The SA-E3D-LSTM network model introduces a 3D convolutional layer to process the spatial dimension of the thermal infrared image sequence data. The 3D convolution not only performs convolution in the spatial (height, width) dimensions but also in the temporal dimension (depth), thereby enabling the capture of the spatial features that change over time in the thermal infrared image sequence data. The mathematical expression of the 3D convolution operation is:

[0035] X out = W c * X in + b c

[0036] where: * represents the 3D convolution operation, W c is the 3D convolution kernel, usually a convolution kernel of size kH×kW×kD, representing the convolution sizes in the height, width, and depth dimensions respectively. X in is the input data block, X out is the output feature map, and b c is the bias.

[0037] The 3D convolution enables the model to learn the feature changes in the input thermal infrared image sequence not only in the spatial dimensions (H and W) but also in the temporal dimension (D), thereby capturing complex spatio-temporal relationships.

[0038] Preferably, in step S5, the improved Eidetic memory mechanism enhances the model's ability to capture the spatio-temporal features of the thermal infrared images of earth-rock dams by adding a spatial attention mechanism during the LSTM hidden state update process, adding a spatial attention module after the 3D convolution output, and weighted screening of the thermal imaging abnormal features in the leakage area to suppress background noise (such as solar reflection, vegetation coverage, etc.), enabling the model to focus on the thermal infrared features related to leakage (such as linear temperature anomaly bands).

[0039] The spatial attention mechanism enables the model to dynamically focus on the abnormal temperature areas on the earth-rock dam surface (such as the gradient change area around the leakage point) through a dual-path weighting method. The channel attention suppresses irrelevant feature channels (such as the thermal radiation of background vegetation), and the spatial attention enhances the local high-temperature gradient change area caused by leakage. The two work together to improve the recognition rate of leakage features. The attention weight matrix A ∈ R H×W is expressed mathematically as follows:

[0040] M c = σ(W 2 · ReLU(W 1 · AvgPool(X t )))(channel attention path)

[0041] M s = σ(Conv 5×5 ([MaxPool(X t );AvgPool(X t )]))(spatial attention path)

[0042] A t = λ · M c (X t )+(1 - λ) · M s (X t )(attention weight fusion)

[0043] where W 1 ∈ R C / r×C , W 2 ∈ R C×C / r , r = 4 is the channel compression ratio, σ is the Sigmoid function; Conv 5×5 represents a 5×5 convolution kernel, [;] represents channel concatenation; the learnable parameter λ ∈ [0,1].

[0044] Preferably, similar to the standard LSTM, the SA-E3D-LSTM network model also updates the cell state to capture long-term dependencies. The updated cell state not only considers the gating mechanism of the LSTM but also includes the enhancement of the hidden state by the improved EideticMemory mechanism. The update formula is as follows:

[0045] C′ t = A t (Conv3D([h t-1 ,C t-1 ))⊙C t

[0046] In the formula: Conv3D is a 3×3×3 3D convolution kernel for extracting spatio-temporal correlation features, and ⊙ is element-wise multiplication.

[0047] This update mechanism combines temporal and spatial information, enhancing the effectiveness and spatial relevance of memory through an improved Eidetic Memory.

[0048] Preferably, in the time series processing, the SA-E3D-LSTM network model processes the thermal infrared image sequence data of multiple time steps. The input of each time step undergoes LSTM and 3D convolution processing to update the cell state and hidden state. After the entire sequence ends, the model generates a future sequence of thermal infrared image frames by analyzing the outputs of all time steps, predicts the surface temperature changes that may be caused by seepage in the earth-rock dam, and identifies potential seepage areas.

[0049] Preferably, based on the abnormal changes in the surface temperature of the earth-rock dam in the predicted thermal infrared image sequence, evaluate its seepage risk level and output seepage warning information to help engineering management personnel take preventive measures in advance.

[0050] An earth-rock dam seepage thermal infrared image prediction device includes a number of distributed thermal infrared ground stations, a central server, and a number of user terminals. The distributed thermal infrared ground stations are arranged in the areas with potential seepage hazards of the earth-rock dam, the central server is arranged in the central control room, and the user terminals are carried by the staff. The distributed thermal infrared ground stations are communicatively connected to the central server and the user terminals.

[0051] Preferably, the distributed thermal infrared ground stations are arranged in the areas with potential seepage hazards of the earth-rock dam, specifically including:

[0052] Dual-band infrared imaging module: Continuously acquire thermal infrared images and visible light images of the potential hazard area of the earth-rock dam at a set sampling frequency.

[0053] Photovoltaic power supply module: Includes a photovoltaic charging panel and a storage battery, and is used to provide the required electrical energy for each module and device of the distributed thermal infrared ground station.

[0054] Data transceiver module: Includes a data repeater and a wireless transceiver unit, and uses the wireless transceiver unit to transmit the thermal infrared image time series data temporarily stored in the data repeater to the central server.

[0055] The dual-band infrared thermal imager, the photovoltaic power supply module, and the data transceiver module are all installed on a fixed frame, and the dual-band infrared thermal imager is connected to the fixed frame through a steering device for adjusting the shooting angle of the dual-band infrared thermal imager.

[0056] Preferably, the central server includes a central processor, a data preprocessing unit, a data storage unit, and a wireless transceiver unit, and is used for preprocessing and storing the thermal infrared images and visible light images collected by the distributed thermal infrared ground stations, predicting the thermal infrared images of the potential seepage hazard areas at the earth-rock dam site, and pushing the on-site information to the user terminals.

[0057] Among them, the on-site push information includes the push time, the on-site prediction report, the thermal infrared images and visible light images collected by the distributed thermal infrared ground station at the push moment, and the longitude and latitude coordinates of the distributed thermal infrared ground station;

[0058] The user terminal includes a data communication module and a human-computer interaction module, and is used to receive and view the thermal infrared image prediction information of the leakage hidden danger area of the earth-rock dam.

[0059] Therefore, the earth-rock dam leakage thermal infrared image prediction method and device adopting the above structure have the following beneficial effects:

[0060] (1) The present invention uses a distributed thermal infrared ground station to continuously collect thermal infrared images of the leakage hidden danger area of the earth-rock dam and predict the change of thermal infrared images in the future time period, which can effectively capture the change of temperature with time and space during the leakage process of the earth-rock dam, solves the problem that the traditional point-type leakage prediction method cannot reflect the distribution and development trend of leakage in space, and provides strong technical and equipment support for anti-seepage layout design and leakage emergency disposal.

[0061] (2) The data collection of the present invention has spatio-temporal continuity and is less affected by the environment, and can realize all-weather thermal infrared image data collection of the leakage hidden danger area of the earth-rock dam in a complex environment.

[0062] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0063] Figure 1 It is a schematic diagram of the system layout of the present invention;

[0064] Figure 2 It is a schematic diagram of a thermal infrared image of local low-temperature anomaly in the vegetation-free area of the leakage hidden danger area of the earth-rock dam in the first embodiment of the present invention;

[0065] Figure 3 It is a schematic diagram of a thermal infrared image of local low-temperature anomaly in the vegetation-free area of the leakage hidden danger area of the earth-rock dam in the first embodiment of the present invention;

[0066] Figure 4 It is a schematic diagram of a thermal infrared image of local low-temperature anomaly in the vegetation-covered area of the leakage hidden danger area of the earth-rock dam in the first embodiment of the present invention;

[0067] Figure 5 It is a schematic diagram of a thermal infrared image of local high-temperature anomaly in the vegetation-covered area of the leakage hidden danger area of the earth-rock dam in the first embodiment of the present invention;

[0068] Figure 6Schematic diagram of the SA-E3D-LSTM network structure for the thermal infrared image prediction method of soil-rock dams in the present invention;

[0069] Figure 7 Thermal infrared image prediction result of soil-rock dam leakage provided in the first embodiment of the present invention;

[0070] Figure 8 Evaluation of the thermal infrared image prediction result of soil-rock dam leakage provided in the first embodiment of the present invention and comparison result with other models;

[0071] Figure 9 Frame diagram of the thermal infrared image prediction device for soil-rock dam leakage in the present invention;

[0072] Reference numerals

[0073] 1. Dual-band infrared thermal imager; 2. Photovoltaic power supply module; 3. Data transceiver module; 4. Fixed frame; 5. Steering device; 6. Central server; 7. User terminal. Detailed implementation manners

[0074] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0075] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0076] Embodiment 1

[0077] As Figure 1 shown, the present invention provides a thermal infrared image prediction method for soil-rock dam leakage, including the following steps:

[0078] Step S1, divide the hidden danger area of soil-rock dam leakage, and select areas to arrange distributed thermal infrared ground stations;

[0079] Step S2, debug the distributed thermal infrared ground station to make the focal plane of the dual-spectrum thermal imager 1 parallel to the slope surface, and collect the thermal infrared image and visible light image sequences of the leakage hazard area at a frequency of 1 frame / min;

[0080] Step S3, transmit the thermal infrared image and visible light image data collected by the distributed thermal infrared ground station to the central server 6;

[0081] Step S4, the data preprocessing unit in the central server 6 stores the received thermal infrared image and visible light image data in the data storage unit in a one-to-one correspondence form, that is, two corresponding thermal infrared images and visible light images share the same number, such as 0001-A, 0001-B; and preprocess the thermal infrared image sequence, including noise reduction, normalization processing of the thermal infrared image, and frame resampling of the time series;

[0082] Step S5, input the preprocessed thermal infrared image sequence into the SA-E3D-LSTM network model for model training, output the future thermal infrared image frame sequence, predict the surface temperature change that may be caused by the leakage of the earth-rock dam, and identify potential leakage areas;

[0083] Step S6, evaluate the risk level according to the abnormal change of the surface temperature of the earth-rock dam in the predicted thermal infrared image sequence, and output the leakage warning information;

[0084] Step S7, the information push unit of the central server 6 pushes the on-site information of the leakage hazard area of the earth-rock dam to the user terminal 7;

[0085] The on-site information includes the push time, on-site report, and longitude and latitude coordinates of the distributed thermal infrared ground station;

[0086] The on-site report evaluates the potential risk level of the dam leakage based on the change of the temperature distribution in the predicted thermal infrared image sequence;

[0087] Step S8, engineering managers carry out anti-seepage layout design and organize leakage emergency disposal work according to the on-site information of the leakage hazard area of the earth-rock dam on the user terminal 7.

[0088] Such as Figures 2 - 5As shown in the figure, the thermal infrared images of the leakage risk area of ​​the earth-rock dam show normal, local low temperature anomaly, and local high temperature anomaly in terms of temperature. The low temperature anomaly and the high temperature anomaly are relative. The low temperature anomaly indicates that the leakage outlet temperature is lower than the background temperature (i.e., the non-leakage area), and the high temperature anomaly indicates that the leakage outlet temperature is higher than the background temperature. Thermal infrared images are mapped to color spaces (such as grayscale color space, RGB color space, HSV color space, etc.) according to their temperature values, and the color scale values ​​in the image represent their corresponding temperature values. The contour and texture features of the thermal infrared images of earth-rock dam leakage are the key information for the SA-E3D-LSTM network model to predict thermal infrared images.

[0089] The thermal infrared image dataset of leakage risk areas of earth-rock dams contains a total of 400 thermal infrared images with time series. Before training, the training set and the validation set are divided into a ratio of 8:2, the Timestep is set to 10, and each group of Timestep outputs 1 frame of thermal infrared image. The initial learning rate is set to 0.0001, the Batchsize is set to 32, the maximum number of iterations is set to 500, and the validation frequency is set to 50.

[0090] like Figure 6 , Figure 7 As shown in the figure, the thermal infrared image with time series is input into the SA-E3D-LSTM network model, and the model is trained and outputs the future thermal infrared image frame sequence. It can be seen that the predicted frame is very similar to the ground truth in the appearance of thermal image feature details and color relationship, with only slight differences in image details. The future development of leakage can be intuitively predicted based on the predicted frame.

[0091] The MSE and SSIM indicators are used to quantitatively evaluate the prediction effect of the model, measuring the similarity between the predicted frame and the groundtruth from the error value level and the perception level respectively, and the prediction results of the E3D-LSTM model are calculated and summarized in Figure 8 It can be seen that compared with the E3D-LSTM model, the SA-E3D-LSTM (Spatial-Attention Eidetic 3D-LSTM) model provided by the present invention has a higher SSIM value and a lower and more stable MSE value, indicating that the model has significantly improved the prediction accuracy of thermal infrared images of earth-rock dam leakage and is relatively less affected by background noise.

[0092] like Figure 9 As shown, a thermal infrared image prediction device for earth-rock dam leakage is used to execute any of the methods described above, including a plurality of distributed thermal infrared ground stations, a central server 6 and a plurality of user terminals 7, wherein the distributed thermal infrared ground stations are in communication with the central server 6 and the user terminals 7;

[0093] The distributed thermal infrared ground station is arranged in the hidden danger area of the earth-rock dam leakage, including a dual-spectrum infrared thermal imager 1, a photovoltaic power supply module 2, and a data transceiver module 3. The dual-spectrum infrared thermal imager 1 is electrically connected to the photovoltaic power supply module 2 and the data transceiver module 3, and is used to continuously acquire the time series data of the thermal infrared images and visible light images in the hidden danger area of the earth-rock dam, and send them to the central server 6;

[0094] The dual-spectrum infrared thermal imager 1, the photovoltaic power supply module 2, and the data transceiver module 3 are all installed on the fixing frame 4; the dual-spectrum infrared thermal imager 1 is connected to the fixing frame 4 through a steering device 5, and is used to adjust the shooting angle of the dual-spectrum infrared thermal imager 1.

[0095] The main parameters of the dual-spectrum infrared thermal imager 1 used in the embodiment of the present invention are shown in Table 1.

[0096] Table 1 Main parameters of the dual-spectrum infrared thermal imager 1

[0097]

[0098]

[0099] The central server 6 is arranged in the central control room, including a central processor, a data preprocessing unit, a data storage unit, and a wireless transceiver unit, and is used to preprocess and store the thermal infrared time series data uploaded by the distributed thermal infrared ground station, predict the thermal infrared images in the future time period, and push the on-site information of the hidden danger area of the earth-rock dam leakage to the user terminal 7;

[0100] The user terminal 7 is equipped for relevant project management personnel, including a data communication module and a human-computer interaction module, and is used to receive and view the on-site information of the hidden danger area of the earth-rock dam leakage;

[0101] The on-site push information includes the push time, the on-site prediction report, the thermal infrared images and visible light images collected by the distributed thermal infrared ground station at the push moment, and the longitude and latitude coordinates of the distributed thermal infrared ground station.

[0102] The equipment components involved in the present invention include a dual-spectrum infrared thermal imager, a photovoltaic charging panel, a storage battery, a data repeater, a wireless transceiver unit, a central processor, a data preprocessing unit, a data storage unit, an information push unit, and a user terminal. There are many product models that can be used for the above components in the prior art, and those skilled in the art can select appropriate models according to actual needs. The specific models are not listed one by one in this embodiment.

[0103] Therefore, by adopting the above-mentioned method and device for predicting the thermal infrared image of the earth-rock dam leakage, the present invention can realize the prediction of the thermal infrared image of the earth-rock dam leakage, and has the advantages of anti-interference, wide coverage, good spatio-temporal continuity, etc., and solves the problem that the traditional point-type leakage prediction method cannot reflect the distribution and development trend of leakage in space.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions of the present invention or make equivalent replacements, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A thermal infrared image prediction method for earth-rock dam leakage, characterized by: The following steps are involved: S1. Data collection: Real-time monitoring of leakage potential areas of earth-rock dams is carried out through distributed thermal infrared ground stations, and continuous thermal infrared images and visible light image sequences are collected. The image sequences include thermal distribution information on the surface of the earth-rock dams. Before collecting thermal infrared images and visible light images on the site of the earth-rock dams, the scope of the leakage potential areas of the earth-rock dams is divided and several distributed thermal infrared ground stations are arranged; S2. Data preprocessing: preprocess the thermal infrared image sequence, including noise removal, brightness correction and time frame resampling of the thermal infrared image sequence to ensure the consistency and accuracy of the input data; S3, spatiotemporal feature extraction: The spatiotemporal features of the preprocessed thermal infrared image sequence are extracted through 3D convolution. 3D convolution can extract local spatiotemporal features in both spatial and temporal dimensions. S4. Time series modeling: The spatiotemporal features extracted from 3D convolution are modeled using a long short-term memory network (LSTM) to capture the long-term temporal dependencies in thermal infrared image sequences. S5. Improved EideticMemory mechanism enhancements: In the process of updating the LSTM hidden state, 3D convolution operation is used to reconstruct the spatial features of the cell state to enhance the memory ability of spatiotemporal features. A spatial attention mechanism is introduced to weightedly screen the thermal imaging abnormal features of the leakage area, thereby improving the ability to capture the leakage features of thermal infrared images of earth-rock dams. S6. Thermal infrared image prediction: Based on the output of the LSTM network, the change trend of thermal infrared images in the leakage risk area of ​​the earth-rock dam in the future time period is predicted, and the leakage situation of the earth-rock dam is predicted; S7. Result output: The predicted thermal infrared image is output and a field prediction report related to leakage risk is generated to evaluate the safety status of the earth-rock dam.

2. The method for predicting leakage of earth-rock dams by thermal infrared images according to claim 1, characterized in that: In step S1, the distributed thermal infrared ground station has the ability to stably collect thermal infrared images and visible light image sequence data in complex environments. Compared with the UAV flight platform, the distributed infrared ground station can obtain the thermal distribution information on the surface of the earth-rock dam more accurately and comprehensively.

3. The method for predicting leakage of earth-rock dam by thermal infrared image according to claim 1, characterized in that: In step S3, when 3D convolution extracts the spatiotemporal features of the thermal infrared image, the convolution kernel slides simultaneously in the spatial and temporal dimensions to capture the characteristics of the thermal distribution on the surface of the earth-rock dam changing with time.

4. The method for predicting leakage of earth-rock dam by thermal infrared image according to claim 1, characterized in that: In step S4, the input of the LSTM network is the spatiotemporal feature map output by the 3D convolution, and the LSTM network performs time series dependency modeling on the input spatiotemporal features through a gating mechanism.

5. The method for predicting leakage of earth-rock dam by thermal infrared image according to claim 1, characterized in that: In step S5, the improved Eidetic memory mechanism adds a spatial attention mechanism to the LSTM hidden state update process. It focuses on the temperature anomaly area dynamically through the attention weight according to the non-uniform characteristics of heat conduction at the leakage outlet during the seepage process of the earth-rock dam, so as to enhance the network's ability to capture the spatial and temporal dependencies in the thermal infrared image sequence. Specifically, it includes: a) dividing the input thermal infrared image sequence into a tensor of H×W×D×C according to a fixed time window; b) performing a 5×5×5 3D convolution operation on the input tensor to extract spatiotemporal features; c) performing dual-path spatial attention weighting on the extracted spatiotemporal features; d) fusing the weighted features with the original cell state to update the cell state.

6. The method for predicting leakage of earth-rock dam by thermal infrared image according to claim 1, characterized in that: In step S6, the potential leakage areas and temperature anomalies on the surface of the earth-rock dam are detected by generating the changes of the thermal infrared images in the future period of time.

7. The method for predicting leakage of earth-rock dams by thermal infrared images according to claim 1, characterized in that: The input of the thermal infrared image sequence is a plurality of continuous frames, and the time intervals of the plurality of continuous frames are adjusted according to actual monitoring requirements to adapt to the monitoring frequency requirements in different leakage detection scenarios.

8. The method for predicting leakage of earth-rock dams by thermal infrared images according to claim 1, characterized in that: In step S7, the on-site prediction report includes the thermal infrared image prediction results, historical thermal infrared images and visible light images, and evaluates the potential risk level of dam leakage based on the changes in temperature distribution in the predicted thermal infrared image sequence.

9. A thermal infrared image prediction device for earth-rock dam leakage, characterized in that: A thermal infrared image prediction method for earth-rock dam leakage is used to execute any one of claims 1-8, comprising a plurality of distributed thermal infrared ground stations, a central server and a plurality of user terminals, wherein the distributed thermal infrared ground stations are arranged in the leakage risk areas of the earth-rock dam, the central server is arranged in the central control room, the user terminals are carried by the staff, and the distributed thermal infrared ground stations are communicatively connected with the central server and the user terminals.

10. The thermal infrared image prediction device for earth-rock dam leakage according to claim 9, characterized in that: The distributed thermal infrared ground station includes a dual-light infrared thermal imager, a photovoltaic power supply module and a data transceiver module. The dual-light infrared thermal imager is electrically connected to the photovoltaic power supply module and the data transceiver module to continuously obtain thermal infrared images and visible light images of the potential hazard areas of the earth-rock dam and send them to the central server. The dual-light infrared thermal imager, the photovoltaic power supply module and the data transceiver module are all installed on the fixing frame, and the dual-light infrared thermal imager is connected to the fixing frame through a steering device for adjusting the shooting angle of the dual-light infrared thermal imager; The central server includes a central processor, a data preprocessing unit, a data storage unit and a wireless transceiver unit, which are used to preprocess and store thermal infrared images and visible light images collected by the distributed thermal infrared ground station, predict thermal infrared images of leakage risk areas on the earth-rock dam site, and push site information to user terminals; The on-site push information includes the push time, on-site forecast report, thermal infrared images and visible light images collected by the distributed thermal infrared ground station at the push time, and the longitude and latitude coordinates of the distributed thermal infrared ground station; The user terminal includes a data communication module and a human-computer interaction module, which are used to receive and view thermal infrared image prediction information of leakage risk areas of earth-rock dams.

Images