Tailings dam collapse risk early warning system based on multi-source data generative adversarial network

Abstract

This invention belongs to the field of mine safety monitoring technology and discloses a tailings dam failure risk early warning system based on a conditional generative adversarial network (GAN). The system preprocesses multi-source heterogeneous monitoring raw data through spatiotemporal alignment and standardization to construct a standardized spatiotemporal feature tensor. A GAN is then constructed, using the standardized spatiotemporal feature tensor as a conditional variable input to the generator. Combined with random noise, it generates predicted state samples for future moments. These samples are then trained adversarially against real high-risk samples using a discriminator to optimize model parameters. Finally, multiple evolution paths are generated based on real-time data, the probability of falling into a high-risk area is calculated, and an early warning signal is output. This invention solves the problem of extreme imbalance between positive and negative samples, achieves accurate early warning under multi-factor coupling, and significantly improves the ability to identify rare high-risk scenarios and the timeliness of early warnings.

Description

Technical Field

[0001] This invention belongs to the field of safety monitoring technology, specifically relating to a tailings dam failure risk early warning system based on multi-source data generative adversarial networks. Background Technology

[0002] In the field of tailings dam safety management and risk prevention, information technology and intelligent monitoring technology have become the mainstream development trend. Traditional monitoring methods are gradually evolving towards automation and networking, using various sensors to collect key indicators such as dam displacement, seepage line, reservoir water level, and rainfall in real time, aiming to achieve early warning of dam failure accidents.

[0003] In existing technologies, most mainstream early warning schemes are monitoring systems based on threshold discrimination and a single physical model. These systems collect monitoring data in real time, triggering an early warning when a threshold is exceeded. However, fixed thresholds are often set based on static operating conditions and cannot adapt to the dynamic operating environment of tailings dams (such as the difference between flood and dry seasons), leading to high false alarm or missed alarm rates. Secondly, this scheme views various monitoring indicators in isolation, ignoring the nonlinear coupling relationships between multi-source heterogeneous data such as displacement, seepage, and rainfall. Finally, early warning models based on traditional machine learning algorithms train classifiers or regression models using historical data to establish monitoring data... The mapping relationship between tailings dam failure and safe state is important, but since tailings dam failure is an extremely low-probability event, most of the historical data is from normal operation, and data on dam failure precursors is extremely scarce. This makes it impossible to simulate the complex nonlinear dynamic evolution during a real dam failure, resulting in poor generalization ability of the model when faced with real, unprecedented abnormal patterns. Consequently, attempts have been made to use physical numerical simulation methods for dam failure risk prediction. However, the accuracy of these methods is highly dependent on the precision of geological parameters. In actual engineering projects, geological parameters often exhibit significant spatial variability and uncertainty, leading to discrepancies between simulation results and actual monitoring data. Summary of the Invention

[0004] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a tailings dam failure risk early warning system based on multi-source data generative adversarial networks. It can construct a multi-source data fusion framework through deep learning technology, use generative adversarial mechanisms to solve the problem of sample scarcity, and achieve accurate and dynamic early warning of tailings dam failure risk.

[0005] The technical solution adopted in this invention is: a tailings dam failure risk early warning method based on multi-source data generative adversarial networks, comprising the following steps:

[0006] Step 1: Collect multi-source heterogeneous raw monitoring data by deploying a sensor array in the tailings dam, and preprocess the multi-source heterogeneous raw monitoring data, which includes dam displacement data, pore water pressure data, rainfall data and reservoir water level data;

[0007] Step 2: Perform spatiotemporal alignment and standardization on the preprocessed multi-source heterogeneous raw monitoring data to generate a standardized spatiotemporal feature tensor;

[0008] Step 3: Construct a Conditional Generative Adversarial Network (cGAN) model. The cGAN model includes a generator and a discriminator. The generator receives a random noise vector and a standardized spatiotemporal feature tensor as a condition variable, and outputs a predicted state sample for future time. The discriminator receives the predicted state sample and the standardized spatiotemporal feature tensor as a condition variable, and uses historical dam failure case data as real high-risk samples to determine the probability that the predicted state sample is a real high-risk sample.

[0009] Step 4: Train the conditional generative adversarial network (cGAN) model using the loss function based on Wasserstein distance, iteratively update the parameters of the generator and discriminator in the cGAN model until the model converges, and complete the training of the cGAN model.

[0010] Step 5: Input the standardized spatiotemporal feature tensor after real-time acquisition and processing into the trained generator to generate multiple future state evolution paths and calculate the dynamic risk probability value; then input the standardized spatiotemporal feature tensor of the current moment into the trained discriminator to obtain the anomaly score of the current state, and trigger different levels of early warning signals according to the preset graded early warning rules based on the dynamic risk probability value and the anomaly score.

[0011] Preferably, in step 1, the preprocessing includes the following steps:

[0012] A1: Set a data acquisition time threshold to determine the data missing situation of the acquired multi-source heterogeneous raw monitoring data. If the missing time length is less than the threshold, linear interpolation is used to fill it. If it is greater than or equal to the threshold, the data in that time period is marked as invalid and a sensor fault alarm is triggered.

[0013] A2: A sliding window-based Kalman filter algorithm is used to denoise the effective multi-source heterogeneous raw monitoring data.

[0014] Preferably, step 2, the spatiotemporal alignment and standardization process includes:

[0015] In the time dimension, Lagrange interpolation or nearest neighbor interpolation is used to resample sensor data with different sampling frequencies to a unified time step, and statistical characteristic values ​​within the time step are calculated.

[0016] In the spatial dimension, based on the three-dimensional geological model of the tailings dam, the data of each monitoring point is mapped to the preset dam body grid unit. The Min-Max normalization method is used to map the data values ​​of each physical quantity to the [0,1] interval to generate a standardized spatiotemporal feature tensor.

[0017] Preferably, in step 3, the generator adopts an encoder-decoder structure combined with a long short-term memory network, including an input layer, a fully connected layer, a residual network module, and a long short-term memory network LSTM layer.

[0018] The input layer is used to concatenate the random noise vector with the standardized spatiotemporal feature tensor to form a joint representation vector;

[0019] The fully connected layer is used to map the joint representation vector to a high-dimensional feature space;

[0020] The residual network module includes multiple residual blocks, each of which consists of a convolutional layer and skip connections, and is used to extract deep features;

[0021] The Long Short-Term Memory (LSTM) layer is used to capture the temporal dependencies of feature vectors and output the predicted state samples for future time steps.

[0022] The discriminator adopts a structure that combines a one-dimensional convolutional neural network with an attention mechanism, including a one-dimensional convolutional layer, an attention mechanism module, a feature concatenation layer, a fully connected layer, and a sigmoid activation function;

[0023] The one-dimensional convolutional layer is used to extract local features of the input sample;

[0024] The attention mechanism module is used to calculate the weights of each time step and feature dimension in the local features in order to focus on key disaster-causing factors.

[0025] The feature concatenation layer is used to concatenate the features output by the attention mechanism module with the standardized spatiotemporal feature tensor;

[0026] The fully connected layer and the sigmoid activation function are used to output the probability that the input sample is a real high-risk sample.

[0027] Preferably, in step 4, the loss function based on Wasserstein distance includes a discriminator loss function and a generator loss function. The discriminator loss function introduces a gradient penalty term, which is used to constrain the gradient norm of the discriminator to be close to a preset value.

[0028] As a preferred option, step 4, after the adversarial training is completed, also includes high-risk sample augmentation. Using the trained generator, normal or sub-healthy state data are input as conditional variables, and different random noise vectors are used to generate virtual samples that conform to the high-risk distribution. Then, the trained discriminator is used to discriminate and score the virtual samples, and virtual samples with a discriminant score greater than a preset threshold are selected and added to the training set to expand the high-risk sample library.

[0029] As a preferred option, step 5, calculating the dynamic risk probability value includes: for each generated future state evolution path, determining whether its multi-source heterogeneous original monitoring data indicators exceed a preset safety threshold, counting the number of evolution paths that exceed the safety threshold, calculating the ratio of this number to the total number of generated paths, and obtaining the dynamic risk probability.

[0030] An early warning system for tailings dam failure risk based on multi-source data generative adversarial networks.

[0031] It includes a data acquisition module for acquiring multi-source heterogeneous raw monitoring data of the tailings dam, which includes dam displacement data, pore water pressure data, rainfall data, and reservoir water level data.

[0032] The data preprocessing module is used to perform spatiotemporal alignment and standardization on multi-source heterogeneous raw monitoring data to obtain standardized spatiotemporal feature tensors;

[0033] The model building and training module is used to build a conditional generative adversarial network model including a generator and a discriminator. The generator is used to receive a random noise vector and a standardized spatiotemporal feature tensor as a condition variable, and output a predicted state sample for a future time. The discriminator receives the predicted state sample and the standardized spatiotemporal feature tensor as a condition variable, and uses historical dam failure case data as real high-risk samples to determine the probability that the predicted state sample is a real high-risk sample.

[0034] The risk dynamic assessment module, connected to the trained generator and discriminator, is used to input the standardized spatiotemporal feature tensor generated from real-time monitoring data into the generator to generate multiple future states, calculate the dynamic risk probability value, and trigger the corresponding level of early warning signal by combining the anomaly score output by the discriminator.

[0035] Preferably, the ratio of training steps for the generator to the discriminator is 1:5.

[0036] Preferably, the system adopts an edge-cloud collaborative architecture. The data preprocessing module is deployed on the edge computing gateway to clean and package multi-source heterogeneous raw monitoring data. The model building and training module and the risk dynamic assessment module are deployed on the cloud server to perform complex model training and inference calculations.

[0037] The beneficial effects of this invention are as follows:

[0038] (1) A conditional generative adversarial network architecture is adopted, and a generator is used to synthesize high-risk samples, which enables the system to handle the problem of "extreme imbalance of positive and negative samples". Compared with traditional linear interpolation to generate samples, the potential distribution of the tailings pond evolving from the normal state to the high-risk state is learned through adversarial training, generating "virtual high-risk samples" that conform to physical laws but have never occurred in history. This solves the problem of extreme scarcity of positive samples in deep learning model training, improves the recall rate of high-risk samples by more than 35%, and effectively reduces the false negative rate.

[0039] (2) By setting up a multi-source heterogeneous data fusion framework and using the fusion features as conditional variables to input the model, accurate early warning under the coupling effect of multiple factors was achieved; the spatiotemporal correlation of multiple factors such as displacement, seepage, meteorology, and geology was comprehensively considered, and the complex transmission chain of rainfall-water level-seepage-displacement was learned, rather than the threshold judgment of a single indicator, thus avoiding false alarms caused by the erroneous fluctuation of a single indicator, and the early warning accuracy was improved by about 25%;

[0040] (3) By using the Monte Carlo simulation concept, the probability distribution of future states is predicted, realizing the leap from state warning to trend warning, and gaining valuable lead time for emergency response.

[0041] (4) The introduction of Wasserstein distance into the loss function significantly improves training stability. Gradient penalty term is also introduced into the discriminator loss function to constrain the gradient norm of the discriminator, ensuring the dynamic balance between the generator and the discriminator in the game process, so that the model can stably converge to the optimal solution when dealing with complex tailings monitoring data.

[0042] (5) By adopting a long short-term memory network structure in the generator and introducing an attention mechanism in the discriminator, the ability to capture temporal features and the attention to key disaster-causing factors are both improved. The LSTM structure effectively remembers the impact of historical states on the future, while the attention mechanism automatically focuses on high-risk signals such as displacement mutation rate and infiltration line surge, further improving the sensitivity and robustness of the early warning model. Attached Figure Description

[0043] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0045] Example 1

[0046] like Figure 1As shown in the figure, the tailings dam failure risk early warning system based on multi-source data generative adversarial network provided in this embodiment adopts an end-edge-cloud collaborative architecture, including a data acquisition module, a data preprocessing module, a model building and training module, and a risk dynamic assessment module. The data acquisition module is deployed at the tailings dam site, the data preprocessing module is deployed on the edge computing gateway, and is used to clean and package multi-source heterogeneous raw monitoring data. The model building and training module and the risk dynamic assessment module are deployed on the cloud server and are used to perform complex model training and inference calculations.

[0047] The acquisition module is used to acquire multi-source heterogeneous raw monitoring data of the tailings dam. In this embodiment, the acquisition module uses various types of sensor arrays. The different types of sensors are distributed in different locations on the tailings dam site according to the different monitoring objects. The sensors mainly include GNSS displacement sensors, piezometers, rain gauges, water level gauges, etc.

[0048] GNSS displacement sensors are deployed on the dam surface and at key monitoring points to collect dam displacement data, specifically X / Y / Z three-dimensional displacement and displacement rate.

[0049] Piezometers are buried at different depths inside the dam body to monitor pore water pressure data inside the dam body, specifically transient pore water pressure. The pore water pressure data can reflect changes in the height of the phreatic line.

[0050] Rain gauges are placed in open areas around the reservoir to collect rainfall data, specifically 24-hour rainfall.

[0051] Water level gauges are installed in drainage wells or on the banks of reservoirs to monitor reservoir water level data, specifically the reservoir water level, and can reflect changes in the length of dry beach through the reservoir water level data;

[0052] Due to the harsh on-site environment, sensor data often contains noise or is missing. For noise, a sliding window-based Kalman filter algorithm is used to denoise the data. For missing data, if the missing time is less than the set acquisition time threshold, linear interpolation is used to fill it in. If the missing time is greater than or equal to the set acquisition time threshold, the data for that time period is marked as invalid and a sensor fault alarm is triggered.

[0053] The data preprocessing module is used to perform spatiotemporal alignment and standardization on multi-source heterogeneous raw monitoring data to obtain a standardized spatiotemporal feature tensor.

[0054] In terms of time, since different sensors have different sampling frequencies, directly inputting the data into the model will result in a mismatch due to insufficient time. In this embodiment, the Lagrange interpolation method is used for time alignment. Specifically, the system reference clock is selected and a uniform time step is set. In this embodiment, it is set to 1 hour. For high-frequency data, the average value within the set time step is calculated as the feature value. For low-frequency data, the nearest neighbor interpolation method is used to fill in the missing data. The data collected at each time step and the filled data are used as the feature values.

[0055] In terms of space, the data preprocessing module has a built-in three-dimensional geological model of the tailings dam. Based on the coordinates of each sensor, the data of each monitoring point is mapped to the preset dam grid unit, thereby constructing the three-dimensional state field of the dam.

[0056] After time and space alignment, standardization is performed. The Min-Max normalization method is used to map the data values ​​of each physical quantity to the [0,1] interval to generate a standardized spatiotemporal feature tensor. In this embodiment, the preferred time window size is 72 hours, and the preferred total feature dimension is 12, including features such as X / Y / Z three-dimensional displacement and displacement rate, wetting line height, dry beach length, reservoir water level, 24-hour rainfall, and transient pore water pressure.

[0057] The model building and training module is used to build and train a conditional generative adversarial network (cGAN) model; the model includes two sub-networks: a generator and a discriminator.

[0058] The generator adopts an encoder-decoder structure and combines a Long Short-Term Memory (LSTM) network layer, including an input layer, a fully connected layer, a residual network module, and a Long Short-Term Memory (LSTM) network layer.

[0059] The input layer receives two parts of input: one is a random noise vector sampled from a standard normal distribution, preferably with a dimension of 100, and the other is a condition variable, namely a standardized spatiotemporal feature tensor.

[0060] The input layer concatenates the random noise vector with the standardized spatiotemporal feature tensor to form a joint representation vector;

[0061] The fully connected layer is used to map the joint representation vector to a high-dimensional feature space. In this embodiment, two fully connected layers are set. The joint representation vector is mapped to the high-dimensional feature space through the two fully connected layers.

[0062] The residual network module includes multiple residual blocks, each residual block consisting of a convolutional layer and a skip connection. In this embodiment, each residual block consists of two convolutional layers and one skip connection. After the feature vector enters the residual network module, deep features are extracted and the gradient vanishing problem of deep networks is solved.

[0063] The Long Short-Term Memory (LSTM) layer is used to capture the temporal dependencies of feature vectors and output predicted state samples for future time. Since the state evolution of the tailings dam has a significant time accumulation effect, the displacement at the current time is closely related to the rainfall and seepage over a period of time in the past. The LSTM layer can effectively remember historical information and forget irrelevant noise.

[0064] The discriminator is designed to distinguish whether the input sample is a real high-risk sample from historical data or a predicted state sample synthesized by the generator.

[0065] The discriminator employs a structure combining a one-dimensional convolutional neural network and an attention mechanism, including one-dimensional convolutional layers, an attention mechanism module, a feature concatenation layer, a fully connected layer, and a sigmoid activation function; the discriminator's input is paired data. ,in For the sample to be judged, For the corresponding condition variables, i.e., the standardized spatiotemporal feature tensor;

[0066] The input data is processed by a one-dimensional convolutional layer to extract local features from the input samples. An attention mechanism module is then introduced, which calculates the weights of each time step and feature dimension within the local features. It automatically focuses on key disaster-causing factors sensitive to dam failure risk (such as displacement mutation rate and seepage line surge), while ignoring irrelevant background noise. The flattened features extracted by the attention mechanism module are then concatenated with a standardized spatiotemporal feature tensor through a feature concatenation layer. Finally, a scalar value is output through a fully connected layer and a sigmoid activation function. , indicating that under given conditions Below, sample The probability that it is a genuine high-risk sample;

[0067] The training process of the cGAN model adopts a minimax game strategy, and the loss function is defined to include the generator loss function and the discriminator loss function.

[0068] The generator loss function is defined as follows:

[0069] The discriminant loss function is defined as:

[0070]

[0071] in, The weights of the gradient penalty term are... This is a gradient penalty term used to force the L2 norm of the discriminator's gradient with respect to the input sample to approach 1, thereby satisfying the 1-Lipschitz constraint.

[0072] The generator and discriminator are trained using a loss function based on Wasserstein distance, and a gradient penalty term is introduced in the discriminator training to constrain the gradient norm of the discriminator to be close to a preset value. In this embodiment, in order to ensure the balance of capabilities between the generator and discriminator during training, a training strategy of training the discriminator 5 times and the generator 1 time is adopted until the loss function converges and the model training is completed.

[0073] Since the training dataset includes a large number of normal state samples and a small number of real dam failure precursor samples, through adversarial training, the generator gradually learns how to generate realistic high-risk state evolution paths under given normal state conditions, while the discriminator gradually improves its ability to distinguish between real and fake samples.

[0074] After the adversarial training is completed, the trained generator is used to input normal or sub-healthy state data as condition variables and, together with different random noise vectors, generate a large number of physically meaningful high-risk state evolution samples, that is, generate virtual samples that conform to the high-risk distribution. Then, the trained discriminator is used to discriminate and score the virtual samples, and virtual samples with discriminant scores greater than a preset threshold are selected and added to the training set, thereby expanding the high-risk sample library.

[0075] The dynamic risk assessment module is used for real-time risk warning during actual operation. It is connected to the trained generator and discriminator. The acquired real-time monitoring stream data is preprocessed to obtain a standardized spatiotemporal feature tensor, which is the condition variable at the current moment. The condition variable at the current moment is input into the trained generator to perform Monte Carlo simulation. Specifically, N different noise vectors are sampled from the noise distribution, preferably N=100, to generate N possible future state evolution paths. For each evolution path, its key indicators (indicators of multi-source heterogeneous original monitoring data) are calculated to see if they exceed a preset safety threshold. The number of paths exceeding the safety threshold is counted, and the ratio of this number to the total number of generated paths is calculated to obtain the dynamic risk probability. The condition variable is then input into the discriminator to obtain the anomaly score value under the current conditions. Based on the dynamic risk probability value and combined with the anomaly score, different levels of warning signals are triggered according to the preset graded warning rules.

[0076] Example 2

[0077] This embodiment is basically the same as Embodiment 1, except that a physical constraint term is introduced into the loss function, that is, a physical consistency loss is added to the loss function. The physical consistency loss includes fluid dynamics constraints and soil mechanics constraints. The fluid dynamics constraints are based on Darcy's law, requiring that the changes in the seepage field in the generated sample must conform to the direction of the hydraulic gradient. For example, if the height of the seepage line in the generated sample fluctuates drastically in a short period of time, and this fluctuation does not match the rainfall data (such as a sudden increase in the seepage line in the absence of rainfall), then the physical constraint term will generate a large loss value. The soil mechanics constraints are based on the principle of slope stability analysis. In this embodiment, a simplified limit equilibrium method calculation module is embedded in the model training process. For the future state samples generated by the generator, the corresponding slope stability safety factor is calculated in real time. If the generated sample shows a sharp increase in displacement, but the calculated safety factor is still greater than 1.2 (i.e., it is in a stable state), it indicates that the sample violates the laws of soil mechanics, and the physical constraint term will penalize it.

[0078] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications and substitutions based on the technical solutions and inventive concepts provided by the present invention should be covered within the scope of protection of the present invention.

Claims

1. A tailings dam failure risk early warning method based on multi-source data generative adversarial networks, characterized in that: Includes the following steps: Step 1: Collect multi-source heterogeneous raw monitoring data by deploying a sensor array in the tailings dam, and preprocess the multi-source heterogeneous raw monitoring data, which includes dam displacement data, pore water pressure data, rainfall data and reservoir water level data; Step 2: Perform spatiotemporal alignment and standardization on the preprocessed multi-source heterogeneous raw monitoring data to generate a standardized spatiotemporal feature tensor; Step 3: Construct a Conditional Generative Adversarial Network (cGAN) model. The cGAN model includes a generator and a discriminator. The generator receives a random noise vector and a standardized spatiotemporal feature tensor as a condition variable, and outputs a predicted state sample for future time. The discriminator receives the predicted state sample and the standardized spatiotemporal feature tensor as a condition variable, and uses historical dam failure case data as real high-risk samples to determine the probability that the predicted state sample is a real high-risk sample. Step 4: Train the conditional generative adversarial network (cGAN) model using the loss function based on Wasserstein distance, iteratively update the parameters of the generator and discriminator in the cGAN model until the model converges, and complete the training of the cGAN model. Step 5: Input the standardized spatiotemporal feature tensor after real-time acquisition and processing into the trained generator to generate multiple future state evolution paths and calculate the dynamic risk probability value; then input the standardized spatiotemporal feature tensor of the current moment into the trained discriminator to obtain the anomaly score of the current state, and trigger different levels of early warning signals according to the preset graded early warning rules based on the dynamic risk probability value and the anomaly score.

2. The tailings dam failure risk early warning method based on multi-source data generative adversarial networks according to claim 1, characterized in that: In step 1, the preprocessing includes the following steps: A1: Set a data acquisition time threshold to determine the data missing situation of the acquired multi-source heterogeneous raw monitoring data. If the missing time length is less than the threshold, linear interpolation is used to fill it. If it is greater than or equal to the threshold, the data in that time period is marked as invalid and a sensor fault alarm is triggered. A2: A sliding window-based Kalman filter algorithm is used to denoise the effective multi-source heterogeneous raw monitoring data.

3. The tailings dam failure risk early warning method based on multi-source data generative adversarial networks according to claim 1, characterized in that: Step 2, spatiotemporal alignment and normalization, includes: In the time dimension, Lagrange interpolation or nearest neighbor interpolation is used to resample sensor data with different sampling frequencies to a unified time step, and statistical characteristic values ​​within the time step are calculated. In the spatial dimension, based on the three-dimensional geological model of the tailings dam, the data of each monitoring point is mapped to the preset dam body grid unit. The Min-Max normalization method is used to map the data values ​​of each physical quantity to the [0,1] interval to generate a standardized spatiotemporal feature tensor.

4. The tailings dam failure risk early warning method based on multi-source data generative adversarial networks according to claim 1, characterized in that: In step 3, the generator adopts an encoder-decoder structure combined with a long short-term memory network, including an input layer, a fully connected layer, a residual network module, and a long short-term memory network LSTM layer; The input layer is used to concatenate the random noise vector with the standardized spatiotemporal feature tensor to form a joint representation vector; The fully connected layer is used to map the joint representation vector to a high-dimensional feature space; The residual network module includes multiple residual blocks, each of which consists of a convolutional layer and skip connections, and is used to extract deep features; The Long Short-Term Memory (LSTM) layer is used to capture the temporal dependencies of feature vectors and output the predicted state samples for future time steps. The discriminator adopts a structure that combines a one-dimensional convolutional neural network with an attention mechanism, including a one-dimensional convolutional layer, an attention mechanism module, a feature concatenation layer, a fully connected layer, and a sigmoid activation function; The one-dimensional convolutional layer is used to extract local features of the input sample; the attention mechanism module is used to calculate the weights of each time step and feature dimension in the local features to focus on key disaster-causing factors. The feature concatenation layer is used to concatenate the features output by the attention mechanism module with the standardized spatiotemporal feature tensor; The fully connected layer and the sigmoid activation function are used to output the probability that the input sample is a real high-risk sample.

5. The tailings dam failure risk early warning method based on multi-source data generative adversarial networks according to claim 4, characterized in that: In step 4, the loss function based on Wasserstein distance includes a discriminator loss function and a generator loss function. The discriminator loss function introduces a gradient penalty term, which is used to constrain the gradient norm of the discriminator to be close to a preset value.

6. The tailings dam failure risk early warning method based on multi-source data generative adversarial networks according to claim 5, characterized in that: In step 4, after the adversarial training is completed, high-risk sample augmentation is also included. Using the trained generator, normal or sub-healthy state data are input as conditional variables, and different random noise vectors are used to generate virtual samples that conform to the high-risk distribution. Then, the trained discriminator is used to discriminate and score the virtual samples, and virtual samples with discriminant scores greater than a preset threshold are selected and added to the training set to expand the high-risk sample library.

7. The tailings dam failure risk early warning system based on multi-source data generative adversarial networks according to claim 1, characterized in that: In step 5, calculating the dynamic risk probability value includes: for each generated future state evolution path, determining whether its multi-source heterogeneous original monitoring data indicators exceed the preset safety threshold, counting the number of evolution paths that exceed the safety threshold, calculating the ratio of this number to the total number of generated paths, and obtaining the dynamic risk probability.

8. The early warning system for the tailings dam failure risk early warning method based on multi-source data generative adversarial networks according to any one of claims 1-7, characterized in that: It includes a data acquisition module for acquiring multi-source heterogeneous raw monitoring data of the tailings dam, which includes dam displacement data, pore water pressure data, rainfall data, and reservoir water level data. The data preprocessing module is used to perform spatiotemporal alignment and standardization on multi-source heterogeneous raw monitoring data to obtain standardized spatiotemporal feature tensors; The model building and training module is used to build a conditional generative adversarial network model including a generator and a discriminator. The generator is used to receive a random noise vector and a standardized spatiotemporal feature tensor as a condition variable, and output a predicted state sample for a future time. The discriminator receives the predicted state sample and the standardized spatiotemporal feature tensor as a condition variable, and uses historical dam failure case data as real high-risk samples to determine the probability that the predicted state sample is a real high-risk sample. The risk dynamic assessment module, connected to the trained generator and discriminator, is used to input the standardized spatiotemporal feature tensor generated from real-time monitoring data into the generator to generate multiple future states, calculate the dynamic risk probability value, and trigger the corresponding level of early warning signal by combining the anomaly score output by the discriminator.

9. The early warning system according to claim 8, characterized in that: The ratio of training steps for the generator to those for the discriminator is 1:

5.

10. The early warning system according to claim 8, characterized in that: The system adopts an edge-cloud collaborative architecture. The data preprocessing module is deployed on the edge computing gateway to clean and package multi-source heterogeneous raw monitoring data. The model building and training module and the risk dynamic assessment module are deployed on the cloud server to perform complex model training and inference calculations.

Images