Joint time-lapse resistivity and hydrologic monitoring for identification of dam leakage

Abstract

The application provides a dam leakage identification method and system combining time-lapse resistivity and hydrological monitoring, and belongs to the technical field of safety monitoring of hydraulic facilities. The application innovatively proposes an optimized interpretation process combined with an artificial intelligence algorithm and associated with multiple important parameters in hydrological monitoring to construct a deep learning model, so as to effectively solve the hidden dangers existing in dam engineering safety monitoring and significantly improve the timeliness and accuracy of the monitoring results. Specifically, a time-lapse resistivity method inversion profile image sequence is constructed into a space-time input tensor, a hydrological monitoring sequence code synchronized with the space-time input tensor is taken as a hydrological feature vector, and a deep learning neural network is constructed. In the feature layer of the deep learning neural network, the hydrological features and the resistivity time sequence space features are fused. A leakage probability heat map corresponding to the profile domain grid is output.

Description

Technical Field

[0001] This application relates to the field of safety monitoring of hydraulic facilities, and specifically proposes a method and system for jointly identifying seepage zones in dams using time-shifted resistivity imaging technology based on multimodal deep learning and hydrological monitoring information. Background Technology

[0002] Embankment engineering is a crucial component of flood control systems, playing a vital role in flood control, water storage, and protection of surrounding areas. It is of paramount importance to safeguarding people's lives and property, as well as protecting the ecological environment. Existing embankments and earth-rock dams in China are characterized by their large number, wide distribution, and long length. Some projects were built relatively early and are prone to defects such as cracks and seepage due to construction conditions and operating environments. In recent years, extreme rainfall events have become more frequent. During high-risk periods such as the flood season, failure to promptly identify and address potential embankment hazards can easily trigger serious disasters such as piping, seepage damage, and even dam failure, causing significant losses. Therefore, non-destructive, accurate, and proactive identification and early warning of embankment hazards are the prerequisite and key to risk prevention and emergency response in embankment engineering.

[0003] The types of potential hazards in current earth-rock dam structures typically include internal defects in the cutoff body or cutoff wall, uneven dam fill quality, structural interface defects, and localized weak zones. Essentially, these hazards manifest as compromised dam integrity while the external appearance may remain stable in the early stages, making it difficult to directly identify the location of the hazard through surface inspection. Seepage failure often initiates at weak points and gradually evolves. Potential piping channels may initially appear as small pores or cracks; when external events such as rainfall or reservoir water level changes cause an increase in water head and alter seepage conditions, water flows into the weak points and, under long-term scouring and erosion, carries out fine particles, causing the pores to gradually expand and develop upstream, potentially forming a concentrated seepage channel connecting upstream and downstream. Because the internal condition of the dam is difficult to observe directly, once visible signs such as muddy water seepage or collapse appear on the dam slope surface, it often indicates that the hazard has progressed to a more serious stage, leaving engineers with a limited window of opportunity for intervention and significantly increasing the difficulty of risk control.

[0004] For monitoring potential hazards in embankments, existing technologies include manual inspection, geophysical exploration, remote sensing and sensor monitoring, hydrogeological monitoring, and data analysis and diagnosis. However, existing technologies still have shortcomings in engineering promotion and pre-flood prevention, mainly in the following aspects: (1) Some methods and equipment deployment and data collection costs are high, and they rely on professional personnel for interpretation and analysis, making it difficult to achieve continuous monitoring of large-scale, long-distance embankments before the flood season; (2) Under complex working conditions, the monitoring response is greatly affected by external factors such as water level fluctuations and rainfall, resulting in fluctuations and deviations in imaging results or characteristic responses, thereby reducing the stability of hazard location and risk assessment; (3) Existing monitoring work often focuses on local verification after the discovery of obvious anomalies, lacking continuous tracking and systematic recording of the hazard evolution process, which limits the ability to identify and quantify hazards in the early stages; (4) Existing technologies generally rely on the experience of professional personnel for interpretation and analysis of monitoring data, lacking unified discrimination rules and standardized output forms, making it difficult to obtain consistent identification conclusions among different engineering scenarios or different personnel, thereby affecting the decision-making efficiency and scalability of hazard disposal. In view of this, this patent application is hereby filed. Summary of the Invention

[0005] The method for identifying dam seepage by combining time-shift resistivity and hydrological monitoring described in this application aims to solve the problems existing in the prior art. It innovatively proposes to combine the optimized interpretation process of artificial intelligence algorithms with multiple important parameters in hydrological monitoring to construct a deep learning model. In order to address the shortcomings in the existing dam hazard monitoring process, it proposes a dam seepage identification solution based on multimodal deep learning that combines time-shift resistivity imaging and hydrological monitoring, in order to effectively solve the hidden dangers in dam engineering safety monitoring and significantly improve the timeliness and accuracy of monitoring results.

[0006] To achieve the above objectives, the method for identifying dam seepage by combining time-shifted resistivity and hydrological monitoring involves constructing a spatiotemporal input tensor from the profile image sequence obtained by the time-shifted resistivity method, encoding the hydrological monitoring sequence synchronized with the spatiotemporal input tensor as a hydrological feature vector, and constructing a deep learning neural network. In the feature layer of the deep learning neural network, the hydrological features and the temporal spatial features of resistivity are fused, and a seepage probability heatmap corresponding one-to-one with the profile domain grid is output.

[0007] The method includes the following steps:

[0008] Step S1: Obtain the sequence of profile images of the target embankment section within the historical time window, which are obtained by inverting the profile using the time-shift resistivity method, as well as the hydrological environment monitoring sequence during the same period.

[0009] Step S2: Generate binarized mask labels for seepage-sensitive areas using an envelope labeling strategy to construct a multimodal joint training dataset containing inverted profile image sequences, aligned hydrological environment monitoring sequences, and binarized mask labels.

[0010] Step S3: Construct a deep learning neural network based on the U-Net architecture and the Convolutional Long Short-Term Memory (ConvLSTM) network;

[0011] The deep learning neural network adopts a dual-branch coding and physical constraint fusion architecture, including a spatiotemporal coding branch for extracting spatiotemporal features of time-shifted resistivity images, an environmental coding branch for extracting hydrological stress features, and a decoding module that integrates spatial attention.

[0012] Step S4: Supervised training of the deep learning neural network based on the multimodal joint training dataset;

[0013] Based on the leakage prediction results output by the deep learning neural network, the supervised loss is calculated by the binarized mask label and the network model parameters are iteratively updated until the preset termination condition is met, thus obtaining the deep learning neural network model after training.

[0014] In the process of calculating the supervised loss, the non-geological information interference area corresponding to the mask is excluded from the gradient calculation. This forces the deep learning neural network model to focus only on the effective geological features within the dam inversion profile, preventing the model from learning the visual features of color-coded legends or coordinate text.

[0015] Step S5: Obtain the time-shifted resistivity method inversion profile image sequence within the time window to be identified, and its corresponding hydrological environment monitoring sequence, and input them into the trained deep learning neural network to output a leakage probability heat map that corresponds one-to-one with the profile domain grid.

[0016] In step S1, the time-shifted resistivity method inversion profile image sequence is generated by the same set of preset color mark parameters or obtained through color mark unification processing, and the mapping relationship between the pixel values ​​at each time point in the image and the resistivity physical quantity remains consistent.

[0017] In step S2, the leakage sensitive area is marked based on the spatiotemporal evolution characteristics of the inverted profile image sequence, targeting areas where the dynamic change in resistivity exceeds a preset threshold; for inverted profile images containing color mark legends, coordinate axes, borders, or blank filled areas, an ignore mask is generated to mark non-geological information interference areas in the image.

[0018] In step S2, the hydrological environment monitoring sequence is preprocessed as follows:

[0019] Step S2.1: For each frame of inverted profile image, match the hydrological data sampled most recently before its acquisition time as the environmental feature input for that frame, thereby achieving time alignment between the image and the hydrological data;

[0020] Step S2.2: Construct a multidimensional hydrological stress feature vector containing instantaneous boundary factors, dynamic evolution factors, and historical cumulative factors; the instantaneous boundary factors are used to characterize the water head pressure at the current moment, the dynamic evolution factors are used to characterize the hydraulic loading rate, and the historical cumulative factors are used to characterize the soil background saturation and seepage hysteresis effect.

[0021] Step S2.3: Construct a time-series hydrological feature matrix using a time sliding window truncation method;

[0022] Using the acquisition time of each frame in the inversion profile image sequence as the termination anchor point of the sliding window, the historical time step of a preset length is traced back, and the hydrological stress feature vector of multiple consecutive moments within the time window is extracted.

[0023] In step S3, the spatiotemporal coding branch uses the encoder of the U-Net network to extract multi-scale spatial features, and ConvLSTM is embedded in the bottleneck layer of the U-Net network to process the deep feature sequence output by the encoder and capture the temporal evolution of deep features.

[0024] In step S4, the supervision loss is calculated using a composite optimization loss function with the following formula:

[0025]

[0026] The supervision loss value L is composed of the binary cross-entropy loss Li. BCE Dice loss L Dice and focus loss L Focal The weighted composition, where λ1, λ2, and λ3 are weighting coefficients; the L... BCE Used to ensure basic accuracy of pixel-level classification; the L Dice This is used to optimize the overlap between the predicted region and the true label set, solving the problem of imbalanced positive and negative sample ratios; the L FOCAL By introducing a focusing parameter to reduce the weight of easily classified background samples, the model can focus on tiny leaks that are difficult to identify.

[0027] Based on the above-mentioned method for identifying dam seepage using combined time-lapse resistivity and hydrological monitoring, this application proposes a dam seepage identification system using combined time-lapse resistivity and hydrological monitoring, the system comprising:

[0028] The data acquisition module is configured to acquire the time-shifted resistivity method inversion profile image sequence of the target embankment section within a historical time window, as well as the hydrological environment monitoring sequence during the same period.

[0029] The data processing and labeling module is configured to establish a time correlation between the hydrological environment monitoring sequence and the inversion profile image sequence and perform preprocessing, and to determine the seepage sensitive area based on the spatiotemporal evolution characteristics of the inversion profile image sequence and generate corresponding binary mask labels to construct a multimodal joint training dataset.

[0030] The model building module is configured to build a multimodal deep neural network model, which includes a spatiotemporal coding branch for extracting spatiotemporal features of inverted profile image sequences, an environmental coding branch for extracting features of hydrological environmental monitoring sequences, and a decoding module for fusing the two types of features and decoding the output.

[0031] The model training module is configured to train the multimodal deep neural network model based on the multimodal joint training dataset to obtain a trained model.

[0032] The leakage identification module is configured to input the time-shifted resistivity method inversion profile image sequence and its corresponding hydrological environment monitoring sequence within the time window to be identified into the trained model, output a leakage probability heat map corresponding one-to-one with the profile domain grid, and optionally extract suspected leakage areas based on the leakage probability heat map and generate risk assessment results and graded early warning information.

[0033] This application proposes a computer-readable storage medium storing a computer program and a parameter set, wherein the parameter set is used to configure a preset deep learning neural network, and the computer program, when executed by a processor, implements the above-described method for identifying dam seepage by combining time-shift resistivity and hydrological monitoring.

[0034] This application proposes an electronic device including a processor and a memory, wherein the memory stores a set of parameters, and the processor is used to execute the above-described method for identifying dam leakage by combining time-shifted resistivity and hydrological monitoring.

[0035] In summary, this application has the following advantages and positive effects compared with the prior art:

[0036] 1. This application constructs a deep learning model that correlates the spatiotemporal evolution of hydrological environment and resistivity, which can effectively overcome the shortcomings of existing technologies such as reliance on human experience, susceptibility to environmental interference, and lack of dynamic tracking capabilities. By replacing manual interpretation with a trained neural network, low-cost, all-weather automated monitoring can be achieved, freeing us from excessive reliance on professional personnel.

[0037] 2. By introducing hydrological data as a physical constraint, this application can automatically distinguish between normal background changes caused by water level fluctuations or rainfall and actual leakage anomalies, which greatly reduces the false alarm rate and makes the monitoring data more effective and accurate.

[0038] 3. This application captures the dynamic evolution of resistivity field based on time series analysis technology, realizes the keen identification and continuous tracking of the early form of hidden dangers, and finally outputs standardized objective judgment results, effectively eliminating subjective misjudgment caused by differences in human experience, and significantly improving the stability and engineering practicality of dam hidden danger identification under complex working conditions. Attached Figure Description

[0039] The following drawings constituting this application are used to provide a detailed and targeted description and explanation of the features of the innovative solution. Those skilled in the art can understand the content of the following illustrative embodiments, which do not constitute an improper limitation of this application.

[0040] Figure 1 This is a schematic diagram of the dam leakage identification method based on the combined time-shift resistivity and hydrological monitoring proposed in this application. Detailed Implementation

[0041] The technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described solutions are some embodiments of this application, but not all embodiments. Based on the embodiments proposed in this application, those skilled in the art can make improvements or modifications without creative effort, and all such improvements and modifications should fall within the protection scope of the appended claims.

[0042] Example 1, as Figure 1 As shown, this study aims to comprehensively utilize resistivity time-shift response and hydrological environment information to improve the robustness and accuracy of seepage zone identification, thereby providing a basis for the identification and early warning of dam seepage hazards.

[0043] This application proposes a method for identifying dam seepage by combining time-shifted resistivity and hydrological monitoring. The method constructs a spatiotemporal input tensor by inverting the profile image sequence using the time-shifted resistivity method, and encodes the hydrological monitoring sequence synchronized with the spatiotemporal input tensor as a hydrological feature vector to construct a deep learning neural network. In the feature layer of the deep learning neural network, the hydrological features and the temporal spatial features of resistivity are fused. The output is a seepage probability heatmap that corresponds one-to-one with the profile domain grid.

[0044] Specifically, the method includes the following steps:

[0045] Step S1: Obtain the sequence of profile images of the target embankment section within the historical time window, which are obtained by inverting the profile using the time-shift resistivity method, as well as the hydrological environment monitoring sequence during the same period.

[0046] Specifically, the time-shifted resistivity method inversion profile image sequence is generated by the same set of preset color mark parameters or obtained through color mark unified processing, and the mapping relationship between the pixel values ​​at each time point in the image and the resistivity physical quantity remains consistent.

[0047] One possible implementation method is to match the most recently sampled hydrological data before the acquisition time with each frame of the inverted profile image as the environmental feature input for that frame, and align the image with the hydrological data in time.

[0048] Using a reservoir in a certain region of China as the monitoring target, a resistivity monitoring line was laid out along the dam axis on the back slope of the target embankment section. The burial depth of the line was set to 0.3m, with 112 electrode points laid out along the line, the spacing between adjacent electrodes being 2.6m, and the total length of the line being 291.2m. A Wenner quadrupole device was used as the observation device, and the data acquisition interval was set to 8 hours. This application adds an electrical resistivity monitoring device to the target embankment section to acquire electrical resistivity observation data and generate a time-shifted resistivity inversion profile image sequence.

[0049] Using the above deployment method, resistivity data is continuously collected within the historical monitoring time window. After inversion processing, a time-shifted resistivity inversion profile image sequence of the target embankment section is obtained. The spatial resolution of each inversion profile image is 488×2400 pixels. The single frame image input to the model is in four-channel format, with four channels. The four channels include resistivity physical quantity channels and auxiliary channels (such as coordinate encoding channels, relative time encoding channels, or equivalent feature channels) used to characterize the spatial location and acquisition conditions of the profile domain. Simultaneously, hydrological environment monitoring data is continuously collected through the existing water level stations, rainfall stations, and reservoir capacity management system of the reservoir to obtain hydrological parameters such as water level and rainfall during the same period as resistivity observations, forming a hydrological environment monitoring sequence.

[0050] Step S2: Generate binarized mask labels for seepage-sensitive areas using an envelope labeling strategy to construct a multimodal joint training dataset containing inverted profile image sequences, aligned hydrological environment monitoring sequences, and binarized mask labels.

[0051] Specifically, the leakage sensitive area is marked based on the spatiotemporal evolution characteristics of the inverted profile image sequence, targeting areas where the dynamic change in resistivity exceeds a preset threshold; for inverted profile images containing color mark legends, coordinate axes, borders, or blank filled areas, an ignore mask is generated to mark non-geological information interference areas in the image.

[0052] Furthermore, the above-mentioned hydrological environment monitoring sequences can be preprocessed. The preprocessing process includes the following steps:

[0053] Step S2.1: For each frame of inverted profile image, match the hydrological data sampled most recently before its acquisition time as the environmental feature input for that frame, thereby achieving time alignment between the image and the hydrological data;

[0054] Step S2.2: Construct a multidimensional hydrological stress feature vector containing instantaneous boundary factors, dynamic evolution factors, and historical cumulative factors; the instantaneous boundary factors are used to characterize the water head pressure at the current moment, the dynamic evolution factors are used to characterize the hydraulic loading rate, and the historical cumulative factors are used to characterize the soil background saturation and seepage hysteresis effect.

[0055] Step S2.3: Construct a time-series hydrological feature matrix using a time sliding window truncation method;

[0056] Using the acquisition time of each frame in the inversion profile image sequence as the termination anchor point of the sliding window, the historical time step of a preset length is traced back, and the hydrological stress feature vector of multiple consecutive moments within the time window is extracted.

[0057] Taking a reservoir in a certain region of China as the monitoring target, the preprocessing procedure of the above-mentioned hydrological environment monitoring sequence is performed as follows: First, the acquired raw data is preprocessed. The inversion profile image maintains a preset resolution of 488×2400. All inversion profile images are generated by the same set of preset color scale parameters or obtained through unified color scale processing to maintain the consistency of the mapping relationship between the pixel values ​​of images at different time frames and the physical quantity of resistivity. The color scale parameters include the resistivity value range and the color mapping function. In a preferred embodiment, the resistivity value range can be set to a preset interval (e.g., 10~1000 Ω·m) and the color scale mapping is unified within this preset interval to ensure the comparability of inversion profile images at different times.

[0058] Hydrological environmental monitoring data is feature-constructed and matrix-organized to form hydrological stress feature inputs that can be learned by the environmental coding branch. The hydrological data can retain the original physical quantity units or, after normalization, be organized into a hydrological feature matrix according to a preset sliding window; preferably, the dimension of the hydrological feature matrix is ​​T. h ×F T T h For the preset time step, F TThe hydrological stress characteristics are defined as follows: In this embodiment, the hydrological stress characteristics can be divided into three categories according to the physical driving mechanism: instantaneous boundary factors, dynamic evolution factors, and historical cumulative factors. These three categories of factors are used to characterize the seepage boundary conditions, loading changes, and infiltration lag effects of the dam. Specifically, they include: (1) Instantaneous boundary factors, which include hydrological quantity characteristics used to characterize instantaneous boundary conditions, such as the current water level L(t) and the current rainfall-related quantity R(t); (2) Dynamic evolution factors, which include the change characteristics used to characterize short-term disturbances and transient loading, such as the water level change or rate of change ΔH, the rainfall change or rate of change ΔR, and the water level statistics within a preset time window (such as the maximum and minimum values); (3) Historical cumulative factors, which include the cumulative characteristics used to characterize the infiltration and saturation lag effects, such as the previous impact rainfall index, the attenuated cumulative rainfall at a preset time scale (such as 3 days, 7 days, 15 days), the water level sliding average at a preset time scale (such as 5 days), and the water level potential energy index used to characterize the cumulative effect of water level loading. The attenuated cumulative rainfall can be used to quantify the background wetting degree or saturation level of the soil, and the water level sliding average can be used to characterize the cumulative hysteresis effect of continuous high water level soaking on pore water pressure. The attenuation coefficient, statistical window length and feature dimension can be set according to the monitoring frequency and engineering requirements, and this embodiment is not limited thereto.

[0059] Hydrological data retains its original physical quantity units and is organized into a hydrological feature matrix using a sliding window. Due to the significant physical lag in pore water pressure transmission within the soil medium, hydrological boundary conditions at a single moment are often insufficient to explain current resistivity inversion anomalies. This embodiment constructs a feature matrix of dimension T×D, forcing the model to jointly analyze the instantaneous stress at the current moment and the cumulative stress over the past T moments. This accurately captures the dynamic driving mechanism of rainfall infiltration and water level fluctuations on the seepage field inside the dam, significantly improving the model's physical interpretability and prediction accuracy under complex hydrological conditions. In this embodiment, the hydrological feature matrix has a dimension of 5×12, where 5 represents the preset time step and 12 represents the feature dimensions, including water level, rainfall, reservoir capacity, and their statistical derivatives (such as water level change rate, cumulative rainfall, water level extremes, etc.). Due to the inconsistency in sampling periods or time resolution between electrical resistivity monitoring and hydrological monitoring, the time-shifted inversion profile image sequence and the hydrological environmental monitoring sequence are difficult to correspond one-to-one in the time dimension. Therefore, time correlation processing is performed on the hydrological environment monitoring sequence to align it with the inverted profile image sequence, resulting in a hydrological environment feature sequence that corresponds one-to-one with each image frame.

[0060] The time correlation processing includes at least one of time resampling, interpolation mapping, or backtracking matching based on time anchors. In this embodiment, data alignment is achieved by matching the most recent reference time no later than the acquisition time: a fixed time each day (8:00 AM) is set as the hydrological data sampling reference time. For each frame of inverted profile image, the nearest sampling reference time before its acquisition time is retrieved, and the hydrological data at that time is used as the environmental feature input for that frame of image. For example, if the acquisition time of a frame of image is 6:00 PM on the same day, then the hydrological data at 8:00 AM on the same day is matched; if the acquisition time of the image is 2:00 AM the next day, then the hydrological data at 8:00 AM the previous day is matched.

[0061] An envelope-based annotation strategy is employed to generate binary mask labels for leakage-sensitive areas. Specifically, resistivity anomalies in the inverted profile image are labeled. Resistivity anomalies typically manifest as low-resistivity anomalies, i.e., areas with resistivity values ​​significantly lower than the surrounding background. During annotation, an envelope line is drawn along the boundary of the anomaly region, forming a closed anomaly region outline. A binary mask with the same size as the inverted profile image is generated based on the envelope line. The mask labels are consistent with the inverted profile image on the profile domain raster, with a mask size of 488×2400. In the mask, the pixel values ​​of leakage-sensitive areas are set to 1, and the pixel values ​​of non-leakage areas are set to 0. The labeled inverted profile image sequence, the aligned hydrological feature sequence, and the binary mask labels are combined to construct a multimodal joint training dataset. The dataset can be divided into training, validation, and test sets according to a preset ratio.

[0062] Step S3: Construct a deep learning neural network based on the U-Net architecture and the Convolutional Long Short-Term Memory (ConvLSTM) network;

[0063] The deep learning neural network adopts a dual-branch coding and physical constraint fusion architecture, including a spatiotemporal coding branch for extracting spatiotemporal features of time-shifted resistivity images, an environmental coding branch for extracting hydrological stress features, and a decoding module that integrates spatial attention.

[0064] Specifically, the spatiotemporal coding branch uses the encoder of the U-Net network to extract multi-scale spatial features, and ConvLSTM is embedded in the bottleneck layer of the U-Net network to process the deep feature sequence output by the encoder and capture the temporal evolution of deep features.

[0065] In this embodiment, the spatiotemporal coding branch is used to extract multi-scale spatial features of the resistivity inversion profile image sequence. It is set with a four-level coding structure, each level of which includes a convolution module and a max pooling downsampling module. The input data is a resistivity inversion image sequence tensor containing temporal information, with dimensions of B×T×C×H×W. B is the batch size, T is set to 5 (representing 5 consecutive time steps), C is set to 4 (containing resistivity values, X coordinates, Z coordinates, and relative time coding), H is 488, and W is 2400.

[0066] The spatiotemporal coding branch processes as follows: The input image sequence first undergoes convolution processing in the first-level coding structure, maintaining a feature map resolution of 488×2400 and expanding the number of channels to 64. Then, it undergoes a max-pooling layer for a 2x downsampling, reducing the feature map resolution to 244×1200. After processing in the second-level coding structure, the feature map resolution is downsampled to 122×600, and the number of channels expands to 128. After processing in the third-level coding structure, the feature map resolution is downsampled to 61×300, and the number of channels expands to 256. After processing in the fourth-level coding structure, the feature map resolution is downsampled to 30×150, and the number of channels expands to 512. A bottleneck layer is placed at the very end of the encoder, where the feature map undergoes convolution processing again, further expanding the number of channels to 1024 while maintaining a spatial resolution of 30×150. At this point, the data dimension output by the bottleneck layer is B×T×1024×30×150.

[0067] A convolutional long short-term memory network (ConvLSTM) is introduced into the bottleneck layer of the U-Net network. The ConvLSTM receives the T-frame feature sequence output by the bottleneck layer and uses the convolutional gating mechanism to model the temporal dependency of the feature sequence.

[0068] In this embodiment, ConvLSTM aggregates temporal features and outputs a deep static feature tensor containing spatiotemporal context information, with dimensions of B×1024×30×150.

[0069] Meanwhile, the environmental coding branch is used to extract semantic features from external environmental data. Its input data is a multi-dimensional hydrological vector sequence with a time window length of T, containing features of water level, rainfall, and water level variability, with a dimension of B×5×12. This branch uses a gated recurrent unit (GRU) as an encoder to map time-series hydrological data into a global environmental feature vector, with an output dimension of B×1×128.

[0070] The decoding module employs a feature modulation mechanism, which uses the environmental feature vector output by the environmental coding branch to perform weighted fusion on the feature tensor output by the spatiotemporal coding branch to introduce hydrophysical constraints.

[0071] In this embodiment, the decoding module employs a spatial attention mechanism for feature fusion, comprising a four-level upsampling structure. Each level includes a bilinear interpolation upsampling layer, a skip connection layer, and a convolutional decoding layer. The fused multimodal deep features are first upsampled to a resolution of 61×300, and then concatenated with the features output from the fourth level of the image spatiotemporal coding branch via skip connections, reducing the number of channels to 512. Subsequently, the feature map is restored level by level, passing through 122×600 and 244×1200 resolution levels, and the shallow features of the corresponding coding levels are fused sequentially. Finally, the feature map is restored to the original resolution of 488×2400, and the number of channels is reduced to 64. The final output layer outputs a single-channel leakage probability heatmap with a size of 488×2400×1, using a 1×1 convolution and a sigmoid activation function. The pixel value in the heatmap represents the probability of leakage at that location.

[0072] The decoding module performs element-wise multiplication of the spatial weight map with the deep static feature tensor output by ConvLSTM, and dynamically enhances the model's response weights to high-risk areas using hydrological prior knowledge, resulting in fused multimodal deep features. Compared with direct splicing or simple additive fusion, this multiplicative attention mechanism can more effectively suppress background noise during non-leakage periods, significantly reducing the false alarm rate of the model during dry seasons or low water levels.

[0073] Step S4: Supervised training of the deep learning neural network based on the multimodal joint training dataset;

[0074] Based on the leakage prediction results output by the deep learning neural network, the supervised loss is calculated by the binarized mask label and the network model parameters are iteratively updated until the preset termination condition is met, thus obtaining the deep learning neural network model after training.

[0075] In the process of calculating the supervised loss, the non-geological information interference area corresponding to the mask is excluded from the gradient calculation. This forces the deep learning neural network model to focus only on the effective geological features within the dam inversion profile, preventing the model from learning the visual features of color-coded legends or coordinate text.

[0076] Specifically, the supervision loss is calculated using a composite optimization loss function with the following formula:

[0077]

[0078] The supervision loss value L is composed of the binary cross-entropy loss Li. BCE Dice loss L Dice and focus loss L Focal The weighted composition, where λ1, λ2, and λ3 are weighting coefficients; the L... BCE Used to ensure basic accuracy of pixel-level classification; the LDice This is used to optimize the overlap between the predicted region and the true label set, solving the problem of imbalanced positive and negative sample ratios; the L FOCAL By introducing a focusing parameter to reduce the weight of easily classified background samples, the model can focus on small, hard-to-identify leaks.

[0079] In this embodiment, the main network parameters and hardware conditions for supervised training are as follows: the network is built based on the PyTorch deep learning framework, the computing environment uses an NVIDIA GeForce RTX 4090 graphics card with 24GB of VRAM (or an equivalent high-performance computing card), and parallel computing is performed using the CUDA acceleration library. The main network parameters are: AdamW optimizer, momentum parameter (betas) set to (0.9, 0.999), and batch size set to 8. The initial learning rate is 1e-3, dynamically adjusted using a cosine annealing strategy, with a minimum learning rate of 1e-6; weight decay is 1e-2, the maximum number of training epochs is set to 100, and an early stopping mechanism is set to terminate training when the validation set loss does not decrease for 10 consecutive iterations. The supervised training may employ a pixel-level binary classification loss function or a combination thereof (e.g., binary cross-entropy loss, Dice loss, or focus loss), and update the network parameters through gradient-based iterative optimization until a preset termination condition is met; the training platform, optimizer type, learning rate strategy, and hyperparameters may be set according to computing power and data scale, and this application is not limited thereto.

[0080] The above supervised training process adopts an end-to-end backpropagation optimization strategy, which includes four steps: forward prediction generation, loss calculation under mask constraints, gradient backpropagation, and parameter iterative update. The detailed implementation process is as follows:

[0081] The time-shifted resistivity inversion image sequence and the aligned hydrological feature sequence are input into a deep neural network model;

[0082] The data stream undergoes feature extraction and fusion processing through a spatiotemporal coding branch, an environment coding branch, and a spatial attention fusion unit. Finally, the decoding module outputs a leakage probability prediction map P with the same spatial resolution as the input image. Each pixel value P in the prediction map P... i,j This characterizes the confidence level that leakage exists at this location. The leakage probability prediction map P output by the model is compared with the true binary label Y.

[0083] Based on the aforementioned composite optimization loss function formula, the initial error is calculated across the entire map. At this point, the generated original loss matrix L... raw Defined as:

[0084]

[0085] The L raw The elements in the matrix represent the mathematical difference between the model's predicted value and the actual value at the corresponding pixel location, but do not yet distinguish whether the pixel belongs to an effective geological region. To eliminate interference from non-geological information, a binarization neglect mask M with the same spatial size as the input image is introduced. ignore The mask is pre-generated based on the inverted image, where the pixel value of the effective monitoring area is set to 1, and the pixel value of non-geological interference areas (such as color marks and borders) is set to 0. Based on this, error filtering and scalar calculation are performed:

[0086]

[0087] In the above formula, L vaild This is the effective loss matrix;

[0088] The original loss matrix L obtained in the first step raw With binarization ignoring the mask M ignore Element-wise multiplication is performed to force the original error in the interference region to zero, thereby generating an effective loss matrix L that retains only the effective qualitative information. vaild For the effective loss matrix L vaild Accumulate all elements and divide by the effective pixels in the mask. M vaild The total number of values ​​is used to obtain the final scalar total loss used to drive model optimization. L total :

[0089]

[0090] In the above formula, i , j Represents the position of any pixel; H, W These represent the height and width of the resistivity inversion image, respectively.

[0091] Based on the scalar total loss obtained above L total The gradient of L with respect to the model weight parameters is calculated using the backpropagation algorithm. Since L is embedded with M during the calculation process... ignore According to the chain rule, for interference regions where the mask value is zero, the corresponding gradient calculation term is always equal to zero because it contains zero weight coefficients. This mathematical property constitutes the physical truncation mechanism of the gradient, ensuring that the error only propagates backward within the effective monitoring region. Finally, the optimizer iteratively updates the model parameters based on this physically truncated effective gradient until the model converges.

[0092] Step S5: Obtain the time-shifted resistivity method inversion profile image sequence within the time window to be identified, and its corresponding hydrological environment monitoring sequence, and input them into the trained deep learning neural network to output a leakage probability heat map that corresponds one-to-one with the profile domain grid.

[0093] Specifically, threshold segmentation and connected component analysis can be performed on the leakage probability heatmap to extract suspected leakage areas, and hierarchical early warning information can be generated based on the probability value and spatial scale of the suspected leakage areas.

[0094] In this embodiment, the time window to be identified can be a real-time monitoring period or a historical backtracking period;

[0095] The acquisition method for the time-shifted resistivity inversion profile image sequence is the same as in step S1. Electrical resistivity monitoring data is continuously collected at a preset acquisition frequency using the deployed electrical resistivity monitoring device, and the inversion profile image sequence is obtained after inversion processing. The inversion profile images are processed using the same color scale parameters as in the training phase to ensure consistent resistivity value range. The hydrological environment monitoring sequence is continuously collected through the existing hydrological monitoring system of the reservoir, including parameters such as water level and rainfall.

[0096] The aforementioned leakage probability heatmap can be presented as a heatmap on the display interface to help locate potential leakage areas and provide quantitative data. The heatmap uses color gradients to represent the probability, with colors closer to red indicating a higher probability of leakage and colors closer to blue indicating a lower probability of leakage.

[0097] Furthermore, a threshold is set for the leakage probability heatmap, and areas with probability values ​​greater than the threshold are identified as high-risk leakage areas. The threshold can be set according to actual engineering needs; in this embodiment, it is set to 0.5. Connectivity analysis is performed on the identified leakage areas to mark independent leakage anomaly areas, and the area, location coordinates, and maximum probability value of each anomaly area are statistically analyzed.

[0098] The temporal evolution trend of the seepage area can also be analyzed by using seepage probability heatmaps over multiple consecutive time steps. The area change rate and location shift of the seepage area in adjacent time steps are calculated to determine whether the seepage is expanding, weakening, or migrating. When the seepage area continues to increase or the probability value continues to rise, an early warning mechanism is triggered, providing timely risk warnings for dam safety management.

[0099] Based on the aforementioned method for identifying dam seepage using combined time-lapse resistivity and hydrological monitoring, this application also proposes a dam seepage identification system using combined time-lapse resistivity and hydrological monitoring, which includes:

[0100] The data acquisition module is configured to acquire the time-shifted resistivity method inversion profile image sequence of the target embankment section within a historical time window, as well as the hydrological environment monitoring sequence during the same period.

[0101] The data processing and labeling module is configured to establish a time correlation between the hydrological environment monitoring sequence and the inversion profile image sequence and perform preprocessing, and to determine the seepage sensitive area based on the spatiotemporal evolution characteristics of the inversion profile image sequence and generate corresponding binary mask labels to construct a multimodal joint training dataset.

[0102] The model building module is configured to build a multimodal deep neural network model, which includes a spatiotemporal coding branch for extracting spatiotemporal features of inverted profile image sequences, an environmental coding branch for extracting features of hydrological environmental monitoring sequences, and a decoding module for fusing the two types of features and decoding the output.

[0103] The model training module is configured to train the multimodal deep neural network model based on the multimodal joint training dataset to obtain a trained model.

[0104] The leakage identification module is configured to input the time-shifted resistivity method inversion profile image sequence and its corresponding hydrological environment monitoring sequence within the time window to be identified into the trained model, output a leakage probability heat map corresponding one-to-one with the profile domain grid, and optionally extract suspected leakage areas based on the leakage probability heat map and generate risk assessment results and graded early warning information.

[0105] This application also proposes a computer-readable storage medium storing a computer program and a parameter set, wherein the parameter set is used to configure a preset deep learning neural network, and the computer program, when executed by a processor, implements the dam leakage identification method based on joint time-shift resistivity and hydrological monitoring as described above.

[0106] This application also proposes an electronic device comprising a processor and a memory, wherein the memory stores a set of parameters, and the processor is used to execute the dam leakage identification method based on joint time-shift resistivity and hydrological monitoring as described above.

[0107] Although the present application has been presented and described with the above embodiments, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying dam seepage by combining time-shift resistivity and hydrological monitoring, characterized in that: The time-shifted resistivity method is used to invert the profile image sequence and construct it as a spatiotemporal input tensor. The hydrological monitoring sequence synchronized with the spatiotemporal input tensor is encoded as a hydrological feature vector and a deep learning neural network is constructed. In the feature layer of the deep learning neural network, the hydrological features are fused with the temporal spatial features of resistivity. The output is a seepage probability heatmap that corresponds one-to-one with the profile domain grid. Includes the following steps, Step S1: Obtain the sequence of profile images of the target embankment section within the historical time window, which are obtained by inverting the profile using the time-shift resistivity method, as well as the hydrological environment monitoring sequence during the same period. Step S2: Generate binarized mask labels for seepage-sensitive areas using an envelope labeling strategy to construct a multimodal joint training dataset containing inverted profile image sequences, aligned hydrological environment monitoring sequences, and binarized mask labels; generate an ignore mask for inverted profile images containing color-coded legends, coordinate axes, borders, or blank-filled areas to mark non-geological information interference areas in the images. Step S3: Construct a deep learning neural network based on the U-Net architecture and the Convolutional Long Short-Term Memory (ConvLSTM) network; The deep learning neural network adopts a dual-branch coding and physical constraint fusion architecture, including a spatiotemporal coding branch for extracting spatiotemporal features of time-shifted resistivity images, an environmental coding branch for extracting hydrological stress features, and a decoding module that integrates spatial attention. Step S4: Supervised training of the deep learning neural network based on the multimodal joint training dataset; Based on the leakage prediction results output by the deep learning neural network, the supervised loss is calculated by the binarized mask label and the network model parameters are iteratively updated until the preset termination condition is met, thus obtaining the deep learning neural network model after training. In the process of calculating the supervised loss, the non-geological information interference area corresponding to the mask is excluded from the gradient calculation. This forces the deep learning neural network model to focus only on the effective geological features within the dam inversion profile, preventing the model from learning the visual features of color-coded legends or coordinate text. Step S5: Obtain the time-shifted resistivity method inversion profile image sequence within the time window to be identified, and its corresponding hydrological environment monitoring sequence, and input them into the trained deep learning neural network to output a leakage probability heat map that corresponds one-to-one with the profile domain grid.

2. The method for identifying dam seepage by combining time-shift resistivity and hydrological monitoring according to claim 1, characterized in that: In step S1, the time-shifted resistivity method inversion profile image sequence is generated by the same set of preset color mark parameters or obtained through color mark unification processing, and the mapping relationship between the pixel values ​​at each time point in the image and the resistivity physical quantity remains consistent.

3. The method for identifying dam seepage by combining time-shift resistivity and hydrological monitoring according to claim 1, characterized in that: In step S2, the leakage sensitive area is marked based on the spatiotemporal evolution characteristics of the inverted profile image sequence, targeting areas where the dynamic change in resistivity exceeds a preset threshold.

4. The method for identifying dam seepage by combining time-shift resistivity and hydrological monitoring according to claim 1, characterized in that: In step S2, the hydrological environment monitoring sequence is preprocessed as follows; Step S2.1: For each frame of inverted profile image, match the hydrological data sampled most recently before its acquisition time as the environmental feature input for that frame, thereby achieving time alignment between the image and the hydrological data; Step S2.2: Construct a multidimensional hydrological stress feature vector that includes instantaneous boundary factors, dynamic evolution factors, and historical cumulative factors; The instantaneous boundary factor is used to characterize the water head pressure at the current moment, the dynamic evolution factor is used to characterize the hydraulic loading rate, and the historical cumulative factor is used to characterize the soil background saturation and seepage hysteresis effect. Step S2.3: Construct a time-series hydrological feature matrix using a time sliding window truncation method; Using the acquisition time of each frame in the inversion profile image sequence as the termination anchor point of the sliding window, the historical time step of a preset length is traced back, and the hydrological stress feature vector of multiple consecutive moments within the time window is extracted.

5. The method for identifying dam seepage by combining time-shift resistivity and hydrological monitoring according to claim 1, characterized in that: In step S3, the spatiotemporal coding branch uses the encoder of the U-Net network to extract multi-scale spatial features, and ConvLSTM is embedded in the bottleneck layer of the U-Net network to process the deep feature sequence output by the encoder and capture the temporal evolution of deep features.

6. The method for identifying dam seepage by combining time-shift resistivity and hydrological monitoring according to claim 1, characterized in that: In step S4, the supervision loss is calculated using a composite optimization loss function with the following formula: Among them, the value of supervision loss L By binary cross-entropy loss L BCE , Dice loss L Dice and focus loss L Focal The weighted composition, where λ1, λ2, and λ3 are weighting coefficients; L BCE Used to ensure basic accuracy in pixel-level classification; the aforementioned L Dice This is used to optimize the overlap between the predicted region and the true label set, addressing the problem of imbalanced positive and negative sample ratios; L Focal By introducing a focusing parameter to reduce the weight of easily classified background samples, the model can focus on tiny leaks that are difficult to identify.

7. A dam leakage identification system based on the combined time-lapse resistivity and hydrological monitoring method as described in any one of claims 1 to 6, characterized in that: The system includes, The data acquisition module is configured to acquire the time-shifted resistivity method inversion profile image sequence of the target embankment section within a historical time window, as well as the hydrological environment monitoring sequence during the same period. The data processing and labeling module is configured to establish a time correlation between the hydrological environment monitoring sequence and the inversion profile image sequence and perform preprocessing, and to determine the seepage sensitive area based on the spatiotemporal evolution characteristics of the inversion profile image sequence and generate corresponding binary mask labels to construct a multimodal joint training dataset. The model building module is configured to build a multimodal deep neural network model, which includes a spatiotemporal coding branch for extracting spatiotemporal features of inverted profile image sequences, an environmental coding branch for extracting features of hydrological environmental monitoring sequences, and a decoding module for fusing the two types of features and decoding the output. The model training module is configured to train the multimodal deep neural network model based on the multimodal joint training dataset to obtain a trained model. The leakage identification module is configured to input the time-shifted resistivity method inversion profile image sequence and its corresponding hydrological environment monitoring sequence within the time window to be identified into the trained model, output a leakage probability heat map corresponding one-to-one with the profile domain grid, and optionally extract suspected leakage areas based on the leakage probability heat map and generate risk assessment results and graded early warning information.

8. A computer-readable storage medium storing a computer program and a parameter set, characterized in that: The parameter set is used to configure a preset deep learning neural network, and the computer program, when executed by the processor, implements the dam leakage identification method based on the combined time-shift resistivity and hydrological monitoring as described in any one of claims 1 to 6.

9. An electronic device comprising a processor and a memory, characterized in that: The memory stores a set of parameters, and the processor is used to execute the dam leakage identification method based on combined time-shift resistivity and hydrological monitoring as described in any one of claims 1 to 6.

Images