Tail water tunnel water gushing prediction and early warning method, device, equipment and medium

Abstract

The invention relates to the technical field of data processing and prediction, and provides a tailrace tunnel water gushing prediction and early warning method, device, equipment and medium, and the method comprises the steps: carrying out the flow estimation according to a monitoring video of a tailrace tunnel node, and generating a real-time flow sequence; constructing an adjacency matrix based on the real-time flow sequence, the osmotic pressure data of the osmotic pressure monitoring points and the osmotic path, introducing a multi-scale time attention mechanism for processing based on the adjacency matrix, and outputting a prediction result of the water inflow; and outputting an early warning strategy based on the prediction result and a preset early warning standard. According to the method, full-chain optimization of tailrace tunnel water gushing from monitoring perception and accurate prediction to scientific early warning is realized, the comprehensiveness and the real-time performance of water gushing monitoring and the accuracy and the adaptability of prediction are improved, the timeliness and the effectiveness of response are guaranteed through a clear early warning strategy, the safety and the stability of tailrace tunnel operation are greatly enhanced, and the safety and the reliability of tailrace tunnel water gushing monitoring are improved. Powerful support is provided for reliable operation of the pumped storage power station, and the engineering practical value is remarkable.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and prediction, in particular to a tailrace tunnel gushing water prediction and early warning method, device, equipment and medium. BACKGROUND

[0002] As a key facility for peak shaving and valley filling of power systems and new energy consumption, the safe and stable operation of pumped storage power stations is directly related to the reliable power supply of the power grid. As the core hydraulic structure connecting the underground powerhouse and the inlet / outlet of the lower reservoir of the pumped storage power station, the tailrace tunnel is in a complex operating environment for a long time, affected by multiple factors such as surrounding rock seepage, groundwater level change, reservoir water level amplitude, etc. Gushing water anomalies occur from time to time, which will affect the normal operation of the power station. Therefore, dynamic prediction and early warning of tailrace tunnel gushing water are important links to ensure the safe operation of the power station, and have significant engineering practical value.

[0003] At present, tunnel gushing water monitoring mainly relies on fixed flow meters or manual inspection. In terms of gushing water prediction, traditional methods mostly use empirical formulas or numerical simulation methods. Some technical solutions construct a time-space multi-source database, combine with hydrodynamic constraints to identify water body connectivity, and build a diversion flow dynamic prediction model based on bidirectional long short-term memory network and attention mechanism. In addition, although there has been some progress in the field of tunnel gushing water prediction in recent years, the related models still follow a specific technical path in actual application. However, the monitoring methods of fixed flow meters and manual inspection have the problems of limited coverage of monitoring points, time delay in response to gushing water anomalies, high equipment maintenance and labor costs; traditional empirical formulas simplify the hydrogeological conditions, and numerical simulation methods require a large number of geological parameters for calibration, which are difficult to adapt to the complex conditions of dynamic evolution of seepage channels; the technical solutions for small hydropower station water system connectivity evaluation use water body recognition methods based on remote sensing images and water body existence frequency matrix, which are suitable for open water system scenes, and there are differences in water flow characteristics in the closed environment of tailrace tunnels. Moreover, the related prediction models focus on the correlation between diversion flow and water system connectivity, and the processing effect of complex optical characteristics such as water flow surface ripples is limited in the low-illumination and high-humidity environment of tailrace tunnels, resulting in insufficient stability of flow velocity estimation results. Most of the existing tunnel gushing water prediction models regard each monitoring point as an independent sample, and the spatial topological relationship between monitoring points and the physical constraints of seepage paths are not fully utilized. The prediction model has limited degree of mining of these associated information, which further restricts the prediction accuracy and generalization ability of the model. SUMMARY

[0004] The present application aims to solve at least one of the above problems.

[0005] To solve the above problems, the application provides a tail water tunnel gushing water prediction and early warning method, device, equipment and medium.

[0006] In a first aspect, the application provides a tail water tunnel gushing water prediction and early warning method, comprising: Flow estimation is performed according to the monitoring video of the tail water tunnel node to generate a real-time flow sequence; An adjacency matrix is constructed based on the real-time flow sequence, seepage pressure data of a seepage monitoring point and a seepage path, and a multi-scale time attention mechanism is introduced based on the adjacency matrix for processing to output a prediction result of gushing water volume; An early warning strategy is output based on the prediction result and a preset early warning standard.

[0007] Optionally, the flow estimation according to the monitoring video of the tail water tunnel node to generate a real-time flow sequence comprises: An instance segmentation network is used to extract a water flow area mask in the monitoring video, wherein an improved Mask R-CNN architecture is used to introduce a feature pyramid network to construct an initial segmentation network, and the initial segmentation network is trained based on historical tail water tunnel annotation data to obtain the instance segmentation network; A flow velocity field is estimated based on a time consistency regularization optical flow method according to the mask; The flow velocity field is combined with pre-measured water passage section geometric parameters to calculate instantaneous flow, and the real-time flow sequence is generated based on time sequence.

[0008] Optionally, the adjacency matrix is constructed based on the real-time flow sequence, seepage pressure data of a seepage monitoring point and a seepage path, comprising: An initial matrix is constructed based on the real-time flow sequence, the seepage pressure data and the seepage path; An adaptive adjacency matrix learning mechanism is introduced to obtain an adaptive matrix according to the real-time flow sequence and the seepage pressure data; The initial matrix and the adaptive matrix are combined to obtain a final adjacency matrix.

[0009] Optionally, the multi-scale time attention mechanism is introduced based on the adjacency matrix for processing to output a prediction result of gushing water volume, comprising: Convolution operation is performed on the final adjacency matrix in spatial dimension and time dimension; The final adjacency matrix after the convolution operation is filtered for noise by using a gated time convolution unit; The final adjacency matrix after filtering is processed by introducing a multi-scale time attention mechanism to output the prediction result.

[0010] Optionally, comprising: The future time period meteorological forecast data is encoded into a meteorological feature vector by using a multi-modal feature fusion strategy, and is fused with the final adjacency matrix processed by the multi-scale time attention mechanism; The prediction result is output according to the fused final adjacency matrix.

[0011] Optionally, the prediction result includes a water inflow interval prediction result and an overrun probability exceeding the preset warning standard; and the output of the prediction result according to the fused final adjacency matrix comprises: The water inflow interval prediction result is output by using a quantile regression method according to the fused final adjacency matrix; The overrun probability is approximately calculated by using a piecewise linear interpolation method based on the water inflow interval prediction result and the preset warning standard.

[0012] Optionally, the output of the warning strategy based on the prediction result and the preset warning standard comprises: The warning strategy matching the warning level of the preset warning standard is triggered according to the ratio of the water inflow interval prediction result to a normal benchmark value or the range of the overrun probability.

[0013] In a second aspect, the present application provides a tailrace tunnel water inflow prediction and warning device, comprising: A flow estimation module is configured to estimate the flow according to the monitoring video of the tailrace tunnel node, and generate a real-time flow sequence; A spatio-temporal graph neural network prediction module is configured to construct an adjacency matrix based on the real-time flow sequence, osmotic pressure data of an osmotic pressure monitoring point and a seepage path, introduce a multi-scale time attention mechanism for processing based on the adjacency matrix, and output a prediction result of water inflow; A warning decision module is configured to output a warning strategy based on the prediction result and a preset warning standard.

[0014] In a third aspect, the present application provides an electronic device comprising a memory and a processor; The memory is configured to store a computer program; The processor is configured to implement the tailrace tunnel water inflow prediction and warning method of the first aspect when executing the computer program.

[0015] In a fourth aspect, the present application provides a computer readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the tailrace tunnel water inflow prediction and warning method of the first aspect is implemented.

[0016] The tailrace tunnel gushing water prediction and early warning method has the beneficial effects that: the flow rate is estimated through the monitoring video of the tailrace tunnel node to generate a real-time flow rate sequence, the wide coverage characteristic of the video monitoring is used to break through the point limitation of the fixed flow meter, and the overall perception of the water flow state of the key node of the tailrace tunnel is realized; meanwhile, the complex optical characteristics such as the water flow surface wave can be processed in the flow rate estimation process, the closed environment of the tailrace tunnel is adapted to the low illumination and high humidity, the stability and accuracy of the flow rate estimation are improved, a large amount of manual inspection is not needed, the maintenance and labor costs are reduced, and the response time of the gushing water anomaly is shortened. The adjacent matrix is constructed based on the real-time flow rate sequence, the osmotic pressure monitoring data and the seepage path, the multi-source data related to the gushing water is fully integrated, the coupling relationship between the seepage channel evolution law and the surrounding rock osmotic pressure and the gushing water quantity is accurately described; the multi-scale time attention mechanism is introduced in combination with the adjacent matrix for processing, the spatial topological association between the monitoring points and the physical constraint of the seepage path can be deeply mined, the lag correlation between factors such as the reservoir water level change and the gushing water quantity response can be captured, the complex working conditions of the dynamic evolution of the seepage channel are effectively adapted, and the precision and generalization ability of the gushing water quantity prediction are significantly improved; in addition, the prediction process is simplified without relying on a large number of geological parameter calibration, and the prediction efficiency is further optimized. The warning strategy is output based on the prediction result and the preset warning standard, the accurate gushing water quantity prediction result can be converted into a warning scheme that can directly guide the practice, clear and definite decision-making basis is provided for the operation and management personnel, and when the gushing water risk occurs, the corresponding measures can be taken in time to avoid the risk expansion.

[0017] The present application realizes the whole-chain optimization of the tailrace tunnel gushing water from monitoring and perception to accurate prediction and scientific early warning, not only improves the comprehensiveness and real-time of the gushing water monitoring, the accuracy and adaptability of the prediction, but also guarantees the timeliness and effectiveness of the response through the clear early warning strategy, greatly enhances the safety and stability of the tailrace tunnel operation, provides a strong support for the reliable operation of the pumped storage power station, and has a significant engineering practical value. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 FIG. 1 is a flowchart of the tailrace tunnel gushing water prediction and early warning method of the embodiment of the present application; Figure 2 FIG. 2 is a structural schematic diagram of the tailrace tunnel gushing water prediction and early warning device of the embodiment of the present application; Figure 3 FIG. 3 is a structural schematic diagram of the electronic device of the embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to make the above objectives, characteristics and advantages of the present application more apparent, concrete embodiments of the present application will be described in detail below with reference to the drawings. Although some embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms, and should not be interpreted as being limited to the embodiments set forth herein, but rather, these embodiments are provided in order to more thoroughly and completely understand the present application. It should be understood that the drawings and embodiments of the present application are merely for illustrative purposes, and are not intended to limit the scope of protection of the present application.

[0020] It should be understood that each step described in the method embodiments of the present application can be performed in different orders, and / or in parallel. In addition, the method embodiments can include additional steps and / or omit the steps shown. The scope of the present application is not limited in this respect.

[0021] As used herein, the term "comprises" and its variants are open-ended, meaning "includes but is not limited to"; the term "based on" is "based, at least in part, on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". Related definitions of other terms will be given in the description below. It should be noted that the concepts of "first", "second", etc. mentioned in the present application are only used to distinguish different devices, modules or units, and are not intended to limit the order or interdependence of the functions performed by these devices, modules or units.

[0022] It should be noted that the modification of "one" or "multiple" mentioned in the present application is illustrative rather than limiting, and those skilled in the art should understand that, unless otherwise explicitly indicated in the context, it should be understood as "one or more".

[0023] The names of the messages or information exchanged between the devices in the embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0024] As shown in Figure 1 The tail water tunnel water surge prediction and early warning method provided by the embodiments of the present application comprises: Step S1, estimating the flow according to the monitoring video of the tail water tunnel node to generate a real-time flow sequence.

[0025] Specifically, the tailrace tunnel node includes a flow monitoring node and a seepage pressure monitoring node, which are usually arranged at the turning part and the cross-section change part of the tailrace tunnel. An infrared enhanced industrial camera with a resolution of not less than 1920*1080 pixels and a frame rate of 25 fps is arranged at the tailrace tunnel node to adapt to the low-illumination and high-humidity environment inside the tailrace tunnel and continuously collect water flow monitoring videos at each node. A fog-proof and waterproof lens can also be used to adapt to the complex environment inside the tailrace tunnel, such as low-illumination, high-humidity and high-dust, and continuously collect water flow dynamic monitoring videos at each node. The YOLOv8 network can be used to process each frame of the monitoring video, output a pixel-level binary mask of the water flow area, and accurately divide the boundary between the water flow area and the non-water flow area. The water flow velocity field can be estimated based on the bidirectional optical flow matching combined with the spatio-temporal constraint regularization method. The high-precision geometric parameters of the water passage cross-section, such as the cross-section profile curve, the cross-section area, the longitudinal slope and the roughness coefficient, are obtained by scanning in advance. The adaptive grid discrete method can be used to calculate the instantaneous flow. That is, according to the distribution characteristics of the water flow in the cross-section, the water passage cross-section is adaptively discretized into grid elements of different sizes. The normal flow velocity component of each element is obtained by interpolation algorithm, the product of the flow velocity and the element area of each element is calculated, and the instantaneous flow is obtained by summing up the velocity correction coefficient based on the dynamic adjustment of the cross-section position and water depth (the value range is 0.82-0.96). Finally, the instantaneous flow data of each node is continuously collected and output at a configurable time resolution, forming a continuous and complete real-time flow sequence, which provides high-fidelity basic data support for subsequent prediction steps.

[0026] In step S2, an adjacency matrix is constructed based on the real-time flow sequence, the seepage pressure data of the seepage pressure monitoring point and the seepage path. Based on the adjacency matrix, a multi-scale time attention mechanism is introduced for processing, and a prediction result of the water inflow is output.

[0027] Specifically, the tailrace tunnel nodes in the tailrace tunnel drainage system, such as the flow monitoring node and the seepage pressure monitoring node, are taken as graph nodes, and the seepage path is taken as a graph edge. The direction of the edge is from the upstream to the downstream of the water flow, and an initial matrix is constructed. Then, a spatio-temporal fusion network is constructed for water inflow prediction based on the initial matrix. The network structure can include 3 layers of spatio-temporal fusion networks, each layer containing a spatial feature aggregation unit and a time feature extraction unit, and the output feature dimensions are 128, 64 and 32, respectively. The convolution kernel size of the time feature extraction unit is set to 5. A multi-scale time attention mechanism based on mutual information is introduced in the spatio-temporal fusion network. Multiple time scales, such as 24 hours, 48 hours and 72 hours, are set for attention windows. The mutual information values of the historical data and the current prediction target at different time scales are calculated to assign attention weights and automatically identify the different scale lag influence periods of various factors on the water inflow. Then, through a stacked fully connected network combined with an interval regression method, the prediction result of the water inflow of each hour in the future time period is output, and the prediction result can be a specific numerical value.

[0028] Step S3, outputting a warning strategy based on the prediction result and a preset warning standard.

[0029] Specifically, the preset warning standard can be multiple warning levels and corresponding preset warning thresholds. The prediction result is compared with the preset warning thresholds, the corresponding warning level is determined, and the corresponding warning prompt is executed, such as generating a complete warning strategy including a warning signal, a treatment process, a responsibility division, and an emergency resource allocation suggestion, prompting an operator.

[0030] The embodiment of the present application estimates the flow rate based on the monitoring video of the tailrace tunnel node and generates a real-time flow rate sequence. With the wide coverage of video monitoring, the point limitation of fixed flow meters is broken through, and the overall perception of the water flow state of the key nodes of the tailrace tunnel is realized. At the same time, complex optical features such as water flow surface ripples can be processed specifically during flow rate estimation, adapting to the closed environment of low illumination and high humidity inside the tailrace tunnel, improving the stability and accuracy of flow rate estimation, and reducing the maintenance and labor costs without the need for a large number of manual inspections, shortening the response time of gushing water anomalies. Based on the real-time flow rate sequence, the seepage pressure monitoring data and the seepage path, an adjacency matrix is constructed, fully integrating multi-source data related to gushing water, accurately depicting the coupling relationship between seepage channel evolution and surrounding rock seepage pressure and gushing water volume. Combined with the adjacency matrix, a multi-scale time attention mechanism is introduced for processing, which can deeply mine the spatial topological correlation between monitoring points and the physical constraints of seepage paths, while capturing the lag correlation between factors such as reservoir water level changes and gushing water volume response, effectively adapting to the complex working conditions of seepage channel dynamic evolution, and significantly improving the accuracy and generalization ability of gushing water volume prediction. In addition, this processing process does not need to rely on a large number of geological parameter calibration, simplifying the prediction process and further optimizing the prediction efficiency. Based on the prediction result and the preset warning standard, a warning strategy is output, which can convert the accurate gushing water volume prediction result into a warning scheme that can directly guide practice, providing clear and explicit decision-making basis for operation and management personnel, ensuring that targeted measures can be taken in time when gushing water risks occur, and avoiding the expansion of risks.

[0031] The embodiment of the present application realizes the whole-chain optimization of tailrace tunnel gushing water from monitoring perception, accurate prediction to scientific warning, not only improves the comprehensiveness and real-time of gushing water monitoring, the accuracy and adaptability of prediction, but also guarantees the timeliness and effectiveness of response through clear warning strategy, greatly enhances the safety and stability of tailrace tunnel operation, provides a strong support for the reliable operation of pumped storage power station, and has significant engineering practical value.

[0032] Optionally, the flow rate is estimated according to the monitoring video of the tailrace tunnel node, and a real-time flow rate sequence is generated, including: The pre-trained instance segmentation network is used to extract the water flow area mask in the monitoring video, wherein an improved MaskR-CNN architecture is used to introduce a feature pyramid network to construct an initial segmentation network, and the initial segmentation network is trained based on historical tail water hole annotation data to obtain the instance segmentation network.

[0033] Specifically, first, an initial segmentation network is constructed, taking MaskR-CNN as the basic architecture, and selecting ResNet-101 as the backbone network to extract image bottom texture, middle contour and high-level semantic features, and introducing a feature pyramid network to integrate different scale feature maps through top-down feature transmission and horizontal connection, to ensure that both small water flow areas and large water flow areas can be effectively identified, forming the initial segmentation network. Collect image data of the tail water hole under three typical working conditions of normal flow, high flow and low flow, covering different light intensities, different water mist concentrations and other scenes, and the cumulative number of images is not less than 5000. Using manual annotation method, the water flow area, water surface reflection area and other categories in each image are clearly labeled to form a labeled data set, which is divided into training set, validation set and test set in the ratio of 7:2:1. The training set is input into the initial segmentation network for training. The test set is used to verify the performance of the trained instance segmentation network, which meets the segmentation accuracy requirements in the tail water hole scene, and finally the usable instance segmentation network is obtained. The monitoring video of the key node of the tail water hole is disassembled into static images by frame, and the trained instance segmentation network is input, and the network outputs the pixel-level binary mask of the water flow area in each image, wherein the water flow area pixels are marked as valid pixels, and the non-water flow area pixels are marked as invalid pixels, to realize the accurate separation of the water flow area and the interference area.

[0034] According to the mask, the flow velocity field is estimated based on the time consistency regularization optical flow method.

[0035] Specifically, the traditional optical flow method is prone to be disturbed by factors such as ripples and reflections when processing the water flow surface, resulting in large fluctuations in the flow velocity estimation result. To solve this problem, the time consistency regularization optical flow method is designed to force the flow velocity change of adjacent frames to meet the hydraulic continuity constraint. The loss function of the time consistency regularization optical flow method includes a luminosity consistency loss term, a spatial smoothness loss term, a time consistency loss term and a hydraulic constraint loss term, which is represented by the following formula: , Wherein, L total is the loss function of the time consistency regularization optical flow method, L photo is the luminosity consistency loss term, which is used to measure the brightness difference between corresponding pixels in adjacent frames, and is calculated by the inverse mapping difference value of the image gray value. In the specific calculation, Charbonnier robust function is used to reduce the influence of abnormal values, and L smoothLspatialis the spatial smoothness loss term for constraining the continuity of the optical flow field in space, which is obtained by calculating the gradient norm of the optical flow vector in horizontal and vertical directions, and adopts an edge-sensitive weighting strategy to impose stronger smoothness constraints in uniform areas where the image gradient is small, and the edge sensitivity parameter is set to 50, L temporal Ltemporalis the temporal consistency loss term, L hydro Lhydraulicis the hydraulic constraint loss term, λ s λspatialis the spatial smoothness weight coefficient, λ t λtemporalis the temporal consistency weight coefficient, λ h λhydraulicis the hydraulic constraint weight coefficient. λ s λspatialis set to 0.1, λ t λtemporalis set to 0.5, λ h λhydraulicis set to 0.3, and this parameter configuration is verified by a large number of experiments to be able to effectively suppress the noise interference caused by ripples while ensuring the accuracy of the flow velocity estimation. At the same time, the weight coefficients are dynamically and adaptively adjusted, when the water flow is strongly turbulent, the temporal consistency weight coefficient is automatically increased according to the optical flow change between adjacent iterations, and the adjustment factor is 0.2; when the water depth changes significantly, the hydraulic constraint weight coefficient is automatically increased according to the water depth change rate, and the adjustment factor is 0.15.

[0036] The formula of the temporal consistency loss term is: , wherein, L temporal Ltemporalis the temporal consistency loss term, N is the total number of pixels in the water flow area, v i （t） vitis the optical flow vector at the i-th pixel in the t-th frame, v i （t-1） vitis the optical flow vector at the i-th pixel in the t-1-th frame, ▽I i （t） ▽Iitis the image gradient at the i-th pixel in the t-th frame, σ g is the gradient sensitivity parameter, || ||2denotes the norm operation. Wherein, N, v i （t） , v i （t-1） , ▽I i （t） is obtained according to the mask. The physical meaning of this formula is that in the flat area where the image gradient is small, the optical flow change between adjacent frames should be small; while in the edge area where the image gradient is large, the optical flow is allowed to have a larger change to adapt to the dynamic characteristics of the water flow.

[0037] The formula of the hydraulic constraint loss term is: , wherein, Lhydro M is the total number of control volume in the velocity field, ▽•v j is the velocity divergence in the jth control volume, h j is the water depth in the jth control volume, is the partial derivative of water depth with respect to time. This formula forces the optical flow field to satisfy the mass conservation equation, so that the velocity estimation result conforms to the actual hydraulic law. By minimizing the total loss function through gradient descent method, the initial optical flow vector is iteratively optimized until the loss function converges, and finally a stable and accurate water flow velocity field is obtained.

[0038] The instantaneous flow is calculated in combination with the velocity field and the pre-measured geometric parameters of the water cross section, and the real-time flow sequence is generated based on time series.

[0039] Specifically, the geometric parameters of the water cross section of the tailrace tunnel key nodes are measured by combining on-site three-dimensional laser scanning with manual review. The cross section contour curve data is obtained, and then the cross section area is calculated; the cross section longitudinal slope is measured by a level instrument; the roughness coefficient is determined in combination with the drainage ditch material to form a complete geometric parameter data set, which is stored in the system database. Based on the measured geometric parameters of the water cross section, the water cross section is uniformly discretized into K small units, the area of each unit is calculated by integral of the cross section contour curve, ensuring that the discrete units can completely cover the entire water cross section, and the unit size meets the flow calculation accuracy requirement. The normal velocity component of each discrete unit is extracted from the velocity field, and the flow contribution value of each unit is calculated in combination with the unit area. According to the roughness coefficient and water depth of the drainage ditch, the velocity correction coefficient of each unit is determined, and the unit flow contribution value is corrected. The sum of the corrected flow contribution values of all units is summed up to obtain the instantaneous flow Q(t) at time t. It is expressed by the formula: , Wherein, Q(t) is the instantaneous flow at time t, A is the water cross section area, v n (x, y, t) is the normal velocity component at the cross section position (x, y) at time t, K is the total number of units after discretization of the water cross section, v n，k (t) is the normal velocity component of the kth unit at time t, △A k is the area of the kth unit, α k is the velocity correction coefficient of the kth unit. The velocity correction coefficient is used to correct the difference between the surface velocity and the average velocity of the cross section, and its value is determined according to the roughness and water depth of the drainage ditch. Preferably, the value of α k ranges from 0.80 to 0.95.

[0040] Set the flow data sampling time resolution, calculate the instantaneous flow at each time according to the time interval, and arrange all the instantaneous flow data in time sequence to form a real-time flow sequence.

[0041] Optionally, the adjacency matrix is constructed based on the real-time flow sequence, the osmotic pressure data of the osmotic pressure monitoring point, and the seepage path, comprising: An initial matrix is constructed based on the real-time flow sequence, the osmotic pressure data, and the seepage path.

[0042] An adaptive adjacency matrix learning mechanism is introduced to obtain an adaptive matrix according to the real-time flow sequence and the osmotic pressure data. Specifically, the adaptive adjacency matrix learning mechanism is represented as: , wherein A adapt is the adaptive adjacency matrix, E1∈R Nv×de and E2∈R Nv×de are learnable node embedding matrices corresponding to the real-time flow sequence and the osmotic pressure data, N v is the total number of graph nodes, d e is the embedding dimension, ReLU is the rectified linear unit activation function, softmax is the normalized exponential function, and T represents the transpose operation of the matrix. Through back propagation in the training process, the node embedding matrix can automatically capture the spatial dependency relationship implied in the data, making up for the possible incompleteness of the initial adjacency matrix. The adaptive adjacency matrix learning mechanism breaks through the limitations of traditional reliance on physical surveys in a data-driven manner, and can automatically capture the potential spatial dependency relationship between monitoring points, making up for the omission of hidden seepage channels or dynamic associations by the initial adjacency matrix. Learning based on the real-time flow sequence and the osmotic pressure data enables the matrix to reflect the dynamic association changes in the actual operation process, improving the adaptability of the adjacency matrix to complex working conditions and providing a more comprehensive association basis for subsequent accurate mining of spatial and temporal dependency relationships.

[0043] The initial matrix and the adaptive matrix are combined to obtain a final adjacency matrix.

[0044] Specifically, based on the tail water tunnel operation characteristics and data reliability analysis, the weight coefficients of the initial adjacency matrix and the weight coefficients of the adaptive adjacency matrix are set. The initial adjacency matrix is multiplied by the corresponding weight coefficient to obtain a weighted matrix of physical prior association; the adaptive adjacency matrix is multiplied by the corresponding weight coefficient to obtain a weighted matrix of data implied association; the two weighted matrices are added element by element to obtain a final adjacency matrix, represented as: , wherein A represents the final adjacency matrix, β is the combination weight coefficient, and preferably, β is 0.6.init represents an initial adjacency matrix.

[0045] The final adjacency matrix is normalized to unify the value range of each element in the matrix to 0-1, avoid the influence of too large value difference on the calculation stability of the subsequent spatio-temporal graph convolution network, and ensure that the matrix can be directly input into the network for spatial feature aggregation operation. The physical prior advantage of the initial adjacency matrix and the data-driven advantage of the adaptive adjacency matrix are combined through weighted combination, so that the final adjacency matrix not only conforms to the actual physical law of the tailwater tunnel drainage system, but also can cover potential associations that have not been discovered, realizing comprehensive representation of physical laws and data implied relationships. The normalization processing ensures the calculation adaptability of the matrix, provides high-quality input for the spatio-temporal graph convolution network in aggregating upstream and downstream node seepage pressure information, mining monitoring point spatial topological relationship and seepage path physical constraints, and significantly improves the accuracy and generalization ability of subsequent water inflow prediction.

[0046] Optionally, based on the adjacency matrix, a multi-scale time attention mechanism is introduced for processing, and a prediction result of water inflow is output, including: Perform convolution operation on the final adjacency matrix in spatial dimension and time dimension.

[0047] The final adjacency matrix after the convolution operation is filtered by a gated time convolution unit.

[0048] Specifically, the spatial graph convolution operation is represented as: , wherein, is the final adjacency matrix of the l+1th layer, is the final adjacency matrix of the lth layer, N v is the total number of graph nodes, T 长 is the length of time series, d l is the feature dimension of the lth layer, is the final adjacency matrix after adding a self-loop, I is an identity matrix, is a degree matrix of, is a learnable weight matrix of the lth layer, and sigma is an activation function. The formula indicates that the feature update of each node not only considers the historical information of itself, but also aggregates the information of neighbor nodes, thereby realizing the spatial transmission of upstream and downstream node seepage pressure information.

[0049] The time convolution operation adopts a gated time convolution unit and is represented as: , , , Wherein, P is the output feature of time convolution, G is the gating signal, the value range is between 0 and 1, and is used to control the passing ratio of time feature, when the gating signal is close to 1, it indicates that the feature at this moment is highly related to the water inrush prediction and should be fully reserved, when the gating signal is close to 0, it indicates that the feature at this moment is noise and should be filtered out, W p and W g is the time convolution kernel, b p and b g is the bias term, * represents the causal convolution operation in the time dimension, ⊙ represents element-wise multiplication, tanh is the hyperbolic tangent activation function, and sigma is the activation function such as the sigmoid activation function, is the feature matrix after time convolution. The gating mechanism can selectively retain the time features related to the water inrush prediction and filter out the noise information.

[0050] The filtered final adjacency matrix is processed by introducing a multi-scale time attention mechanism, and the prediction result is output.

[0051] Specifically, in order to capture the lag correlation between reservoir water level change and water inrush response, the application introduces a multi-scale time attention mechanism in the spatio-temporal graph convolution network. Three time scale windows are set to cover short, medium and long period time dependence; the feature matrix after noise filtering is converted into a query vector, a key vector and a value vector, the similarity between the query vector and the key vector under different time scales is calculated to obtain the initial attention weight; the initial attention weight is normalized by softmax to obtain the normalized attention weight under each time scale, and the greater the weight value, the more significant the influence of the corresponding time step feature on the prediction result. The normalized attention weight under each time scale is weighted and summed with the corresponding value vector to fuse the key features of different time scales and output the final adjacency matrix after multi-scale attention weighting. The final adjacency matrix is input into a multilayer perceptron to calculate and output the water inrush prediction result in the future preset period through a quantile regression model.

[0052] Optionally, after the filtered final adjacency matrix is processed by introducing a multi-scale time attention mechanism, the following is included: A multi-modal feature fusion strategy is adopted to encode the future time period meteorological forecast data into a meteorological feature vector, and the final adjacency matrix processed by the multi-scale time attention mechanism is fused.

[0053] Specifically, the future time period such as 24 hours is determined as the core meteorological forecast period, three key meteorological parameters such as rainfall, air temperature and air pressure are screened, and the corresponding hourly forecast data of the time period is obtained in real time from the meteorological service platform through a standardized interface to ensure the timeliness and integrity of the data. The obtained meteorological forecast data is preprocessed, for example, the Z-score standardization method is used to eliminate the dimensional difference between different meteorological parameters, and the numerical values of each parameter are converted to a unified numerical interval; for the possible missing data, the linear interpolation method is used for completion to avoid the influence of data missing on subsequent processing. A single-layer fully connected neural network is constructed as a feature encoding network, and the preprocessed meteorological forecast data is input into the network. The input dimension of the network is the number of meteorological parameter types, and the output dimension is set to be consistent with the dimension of the features processed by the multi-scale time attention mechanism; the weight parameters are optimized through network training, and the multi-dimensional meteorological forecast data is mapped to a fixed-dimensional one-dimensional meteorological feature vector, realizing feature condensation and format adaptation of meteorological data.

[0054] The meteorological feature fusion is represented as: , wherein, H fused is the final adjacency matrix after fusion, H stgcn is the final adjacency matrix of the spatio-temporal graph convolution network, H weather is the meteorological feature vector, [, ] represents a feature splicing operation, W fuse is a fusion weight matrix, and b fuse is a bias term.

[0055] By screening key meteorological parameters and standardizing preprocessing, the dimensional interference and data missing problems of meteorological data are eliminated, and the data quality is ensured. The feature encoding network converts the original meteorological forecast data into a unified-dimensional feature vector, realizes the format compatibility of meteorological data and the features processed by the multi-scale time attention mechanism, lays a foundation for subsequent multi-source feature fusion, and at the same time, the condensed meteorological feature vector can accurately retain the influence law of meteorological factors on water inflow, improving the representation efficiency of the features.

[0056] Optionally, the prediction result includes a water inflow interval prediction result and an over-limit probability exceeding the preset warning standard; and the output of the prediction result according to the final adjacency matrix after fusion includes: According to the final adjacency matrix after fusion, the water inflow interval prediction result is output by using a quantile regression method.

[0057] Specifically, a multilayer perceptron is constructed as a quantile regression prediction network, the input layer dimension of the network is consistent with the feature dimension of the fused final adjacency matrix, the output layer dimension corresponds to the preset number of quantiles, the ReLU activation function is used in the hidden layer to enhance the non-linear expression ability of the network and ensure that it can fit the complex water inrush prediction mapping relationship. A set of preset quantiles is defined, where the 10% quantile corresponds to a water inrush value that has a 10% probability of being lower than the value, the 50% quantile corresponds to a water inrush value that has a 50% probability of being lower or higher than the value, the corresponding quantile, the 90% quantile corresponds to a water inrush value that has a 90% probability of being lower than the value, and the three quantiles comprehensively cover the probable value interval of the water inrush. The feature data corresponding to the fused final adjacency matrix and the actual water inrush data are used as training samples, a quantile loss function is used to measure the prediction error, an optimizer such as Adam is selected for network training, the network weight parameters are continuously optimized through back propagation until the quantile loss function converges, and the network can accurately output the water inrush prediction value corresponding to each quantile. The fused final adjacency matrix of the to-be-predicted period is input into the trained quantile regression network, and the network outputs the water inrush prediction values corresponding to the 10%, 50%, and 90% quantiles, respectively, and the three values together constitute the water inrush interval prediction result, which fully presents the probability distribution range of future water inrush.

[0058] The loss function of quantile prediction is represented as: , where L quantile is the quantile loss, N s is the number of training samples, q is the quantile, Q is the set of preset quantiles, preferably Q = {0.1, 0.5, 0.9}, y i is the true water inrush of the i-th sample, is the prediction value of the i-th sample at quantile q, and p q is the quantile loss function.

[0059] The definition of the quantile loss function is: , where p q (u) is the quantile loss function of u, u is the prediction residual, and l u＜0 is the indicator function, which takes the value 1 when u < 0, and otherwise takes the value 0.

[0060] The quantile regression method breaks through the limitation of traditional point prediction that can only output a single numerical value, forms interval results through multi-quantile prediction, effectively quantifies the uncertainty of the water inflow prediction, and provides a more comprehensive risk reference for the operation and management personnel. The combination of multi-layer perceptron and quantile loss function enables the model to accurately fit the complex nonlinear relationship between water inflow and multi-source features, and the output interval prediction result not only contains the probable value range of water inflow, but also clearly shows the core trend through the median prediction value. Compared with traditional prediction methods, the practicality and reference value of the prediction result are significantly improved, laying a solid foundation for subsequent over-limit probability calculation and warning decision-making.

[0061] Based on the water inflow interval prediction result and the preset warning standard, the over-limit probability is approximately calculated using a piecewise linear interpolation method.

[0062] Specifically, the preset warning standard includes warning levels and preset warning thresholds corresponding to each warning level, the quantile prediction value in the water inflow interval prediction result, and the preset warning threshold corresponding to the warning level, which verifies the data integrity and ensures that there are no missing or abnormal values affecting the calculation result. Assuming that the water inflow is uniformly distributed within the 10% quantile to 90% quantile interval, the over-limit probability is calculated through a linear mapping relationship. Based on the relative position of the preset warning threshold in the water inflow interval prediction result, a linear correspondence is established between the corresponding quantile interval and the probability interval (10%-90%), and the over-limit probability is calculated using the interpolation formula. When the 90% quantile prediction value is less than the preset warning threshold, the over-limit probability is 0, and when the 10% quantile prediction value is greater than the preset warning threshold, the over-limit probability is 1, ensuring that the result can reflect the relationship between the position of the preset warning threshold in the interval and the over-limit risk.

[0063] The piecewise linear interpolation method realizes accurate estimation of the over-limit probability based on the quantile prediction result, has simple logic and efficient calculation, and can quickly convert the interval prediction result into an intuitive risk probability indicator. Through scene-specific judgment and interpolation calculation, the rationality of the over-limit probability in extreme cases (threshold exceeding the 10%-90% quantile interval) is ensured, and the risk gradient change when the threshold is within the high-probability interval is accurately described, avoiding the limitations of single threshold determination. The output over-limit probability provides a quantitative risk basis for warning decision-making, making the determination of warning levels more scientific and objective, effectively reducing the false alarm and missed alarm probabilities, and improving the reliability of warning decision-making.

[0064] Optionally, the output of the warning strategy based on the prediction result and the preset warning standard includes: Triggering a warning strategy matching the warning level of the preset warning standard according to the ratio of the water inflow interval prediction result to the normal reference value or the range of the over-limit probability.

[0065] Specifically, during the normal reference value setting, the historical operation data of the tailrace tunnel without gushing water anomaly for nearly 3 years can be collected, the extreme values are removed, and the arithmetic mean value is taken, combined with the design maximum allowable safety flow, such as 60%, for calibration, and finally the normal reference value is determined to ensure that it fits the actual safety operation threshold of the tailrace tunnel.

[0066] The preset early warning standard includes four early warning levels, and the trigger condition setting includes: Blue early warning: when the 50% quantile (median) of the gushing water interval prediction result exceeds 120% but does not exceed 150% of the normal reference value, or the over-limit probability is greater than 30% but does not exceed 50%, the blue early warning is triggered; Yellow early warning: when the 50% quantile of the gushing water interval prediction result exceeds 150% but does not exceed 200% of the normal reference value, or the over-limit probability is greater than 50% but does not exceed 70%, the yellow early warning is triggered; Orange early warning: when the 50% quantile of the gushing water interval prediction result exceeds 200% but does not exceed 300% of the normal reference value, or the over-limit probability is greater than 70% but does not exceed 90%, the orange early warning is triggered; Red early warning: when the 50% quantile of the gushing water interval prediction result exceeds 300% of the normal reference value, or the over-limit probability is greater than 90%, the red early warning is triggered.

[0067] An early warning standard self-adaptive adjustment mechanism is established, with 1 month as the adjustment period, combined with historical gushing water data, seasonal precipitation characteristics and surrounding rock seepage changes, the trigger thresholds of various early warning levels are dynamically corrected, the adjustment range is not more than 20% of the reference threshold, and the early warning standard is adapted to different operation conditions.

[0068] The normal reference value is determined by combining historical data calibration and design parameters, which ensures the rationality and scientificity of early warning determination. The gradient division of four early warning levels realizes the fine control of gushing water risk, and avoids the problem of excessive or insufficient risk response caused by single threshold determination. The dynamic adjustment mechanism makes the early warning standard adapt to the changes of the tailrace tunnel operation environment, improves the adaptability and flexibility of the early warning standard, and lays a foundation for the subsequent precise triggering of early warning strategies.

[0069] From the results output by the spatio-temporal graph neural network prediction module, the ratio of the 50% quantile prediction value of the gushing water interval prediction to the corresponding early warning threshold and the over-limit probability specific value are extracted. The "maximum principle" is used for early warning level determination, that is, the corresponding candidate early warning level is determined according to the gushing water ratio and the over-limit probability respectively, and the higher level is selected as the final early warning level, which ensures that any index anomaly can trigger the matching high-level early warning, avoiding the risk of missing judgment. According to the final early warning level, the preset response measure library is called to generate a targeted early warning strategy: Blue early warning strategy: increase monitoring frequency, inform on-duty personnel to pay attention to water inflow trend; Yellow early warning strategy: start emergency duty, check the operation state of drainage facilities, and prepare to pump drainage equipment; Orange early warning strategy: start emergency pumping, restrict personnel from entering dangerous areas, and notify the superior administrative department; Red early warning strategy: start emergency pumping, restrict personnel from entering dangerous areas, and notify the superior administrative department.

[0070] After the generation of the early warning strategy, the early warning signal and disposal suggestion are pushed through multiple channels such as a visual platform, a mobile phone APP, a short message, and an industrial control terminal, so as to ensure that the relevant personnel receive and respond in time.

[0071] The generation of the early warning strategy comprehensively considers the water inflow interval prediction result and the over-limit probability, and the decision rule is represented as: , Wherein, W level is the early warning level, is the water inflow interval prediction result, P exceed is the over-limit probability, when the threshold value is between the 10% and 90% quantile prediction, the corresponding over-limit probability is calculated by linear interpolation, f Q and f p are grade mapping functions based on water inflow and over-limit probability respectively.

[0072] As shown in Figure 2 , the tailrace water inflow prediction and early warning device 200 provided by the embodiment of the application comprises: A flow estimation module 210 is configured to estimate flow according to the monitoring video of the tailrace node, and generate a real-time flow sequence; A spatio-temporal graph neural network prediction module 220 is configured to construct an adjacency matrix based on the real-time flow sequence, osmotic pressure data of an osmotic pressure monitoring point, and a seepage path, introduce a multi-scale time attention mechanism for processing based on the adjacency matrix, and output a prediction result of water inflow; An early warning decision module 230 is configured to output an early warning strategy based on the prediction result and a preset early warning standard.

[0073] As shown in Figure 3 , the electronic device 300 provided by the embodiment of the application comprises a memory 310 and a processor 320; the memory 310 is configured to store a computer program; and the processor 320 is configured to implement the tailrace water inflow prediction and early warning method as described above when executing the computer program.

[0074] Or, an electronic device 300, comprising a memory 310 and a processor 320 coupled to the memory 310;The memory 310 is configured to store a computer program;The processor 320 is configured to execute the following operations when executing the computer program: According to the monitoring video of the tail water tunnel node, the flow estimation is carried out, and the real-time flow sequence is generated; Based on the real-time flow sequence, the osmotic pressure data of the osmotic pressure monitoring point and the seepage path, an adjacency matrix is constructed, based on the adjacency matrix, a multi-scale time attention mechanism is introduced for processing, and a prediction result of the water inflow is output; Based on the prediction result and the preset warning standard, a warning strategy is output.

[0075] The computer readable storage medium provided by the embodiment of the application, the storage medium stores a computer program, when the computer program is executed by the processor, the tail water tunnel water inflow prediction and warning method is realized.

[0076] Or, a non-volatile computer readable storage medium, the storage medium stores a computer program, when the computer program is executed by the processor, the processor executes the following operations: According to the monitoring video of the tail water tunnel node, the flow estimation is carried out, and the real-time flow sequence is generated; Based on the real-time flow sequence, the osmotic pressure data of the osmotic pressure monitoring point and the seepage path, an adjacency matrix is constructed, based on the adjacency matrix, a multi-scale time attention mechanism is introduced for processing, and a prediction result of the water inflow is output; Based on the prediction result and the preset warning standard, a warning strategy is output.

[0077] Now it will be described that the electronic device 300 can be as the server or the client of the application, it is the example of the hardware device that can be applied to each aspect of the application. Electronic device 300 is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 300 can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are merely examples and are not intended to limit the implementations of the application described and / or claimed herein.

[0078] The electronic device 300 includes a computing unit that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The computing unit, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0079] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, the processes of the above-mentioned embodiments can be included. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM), a random access memory (RAM), or the like. In this application, the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiments of the present application according to actual needs. In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0080] Although the present application is disclosed as above, the protection scope of the present application is not limited to this. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application, and these changes and modifications will fall within the protection scope of the present application.

Claims

1. A method for predicting and warning of tailrace tunnel inrush, characterized in that, include: Flow is estimated based on monitoring video of the tailrace tunnel node, and a real-time flow sequence is generated; Based on the real-time flow sequence, the seepage pressure data from the seepage pressure monitoring points, and the seepage path, an adjacency matrix is ​​constructed. Based on the adjacency matrix, a multi-scale time attention mechanism is introduced for processing, and the predicted result of the inflow is output. Based on the prediction results and preset early warning standards, an early warning strategy is output.

2. The tailrace tunnel inrush prediction and early warning method according to claim 1, characterized in that, The step of estimating flow rate and generating a real-time flow sequence based on monitoring video of the tailrace tunnel node includes: A pre-trained instance segmentation network is used to extract water flow region masks in surveillance videos. Specifically, an improved Mask R-CNN architecture is adopted, a feature pyramid network is introduced to construct an initial segmentation network, and the instance segmentation network is trained based on historical tailrace hole annotation data. Based on the mask, the velocity field is estimated using the time-consistent regularized optical flow method. The instantaneous flow rate is calculated by combining the velocity field and the pre-measured geometric parameters of the cross-section, and the real-time flow sequence is generated based on the time series.

3. The tailrace tunnel inrush prediction and early warning method according to claim 1, characterized in that, The construction of an adjacency matrix based on the real-time flow sequence, seepage pressure data from seepage monitoring points, and seepage paths includes: An initial matrix is ​​constructed based on the real-time flow sequence, the seepage pressure data, and the seepage path. An adaptive adjacency matrix learning mechanism is introduced to obtain an adaptive matrix based on the real-time flow sequence and the seepage pressure data; The final adjacency matrix is ​​obtained by weighting and combining the initial matrix and the adaptive matrix.

4. The tailrace tunnel inrush prediction and early warning method according to claim 3, characterized in that, The process, based on the adjacency matrix, incorporates a multi-scale time attention mechanism to process the data and outputs the predicted inflow rate, including: Perform convolution operations on the final adjacency matrix in both spatial and temporal dimensions; A gated temporal convolutional unit is used to filter noise from the final adjacency matrix after the convolution operation. A multi-scale temporal attention mechanism is introduced to process the filtered final adjacency matrix and output the prediction result.

5. The tailrace tunnel inrush prediction and early warning method according to claim 4, characterized in that, After the multi-scale temporal attention mechanism is introduced to process the filtered final adjacency matrix, the following steps are included: A multimodal feature fusion strategy is used to encode meteorological forecast data for future time periods into meteorological feature vectors, and then fuse them with the final adjacency matrix processed by the multi-scale time attention mechanism. The prediction result is output based on the final adjacency matrix after fusion.

6. The tailrace tunnel inrush prediction and early warning method according to claim 5, characterized in that, The prediction results include the predicted inflow range and the probability of exceeding the preset warning standard; the step of outputting the prediction results based on the fused final adjacency matrix includes: Based on the final adjacency matrix after fusion, the quantile regression method is used to output the predicted water inflow interval. Based on the predicted water inflow range and the preset early warning standard, the probability of exceeding the limit is approximately calculated using a piecewise linear interpolation method.

7. The tailrace tunnel inrush prediction and early warning method according to claim 6, characterized in that, The early warning strategy based on the prediction results and preset early warning standards includes: The warning strategy is triggered based on the ratio of the predicted water inflow range to the normal baseline value or the range of the probability of exceeding the limit, and the warning level is matched with the preset warning standard.

8. A tailrace tunnel inrush prediction and early warning device, characterized in that, include: The flow estimation module is used to estimate the flow based on the monitoring video of the tailrace tunnel node and generate a real-time flow sequence; The spatiotemporal graph neural network prediction module is used to construct an adjacency matrix based on the real-time flow sequence, the seepage pressure data of the seepage pressure monitoring point, and the seepage path. Based on the adjacency matrix, a multi-scale time attention mechanism is introduced for processing, and the predicted result of the inflow is output. The early warning decision module is used to output an early warning strategy based on the prediction results and preset early warning standards.

9. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the tailrace tunnel inrush prediction and early warning method as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the tailrace tunnel inrush prediction and early warning method as described in any one of claims 1 to 7.

Images