Water conservancy project management data processing method and system based on cloud computing
By employing cloud computing-based water conservancy project management data processing methods, utilizing feature extraction networks and time-series data stream splitting technology, and combining environmental noise entropy vectors and water balance differential equations, the accuracy and logical reliability issues of monitoring data in complex environments were resolved, achieving feature consistency in data repair and reducing false alarm rates.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SYMGREEN BEIJING ENVIRONMENTAL PROTECTION TECH CO LTD
- Filing Date
- 2026-04-23
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies struggle to ensure the accuracy and logical reliability of monitoring data in complex environments. Edge devices are susceptible to interference from strong light and ripples, making it difficult to acquire high-precision data and resulting in high bandwidth consumption. Traditional cloud-based restoration methods cannot perceive and retain the microscopic fluctuations of the current hydrological environment, leading to a loss of environmental realism in the supplementary data and resulting in feature fragmentation. Furthermore, the lack of self-consistent verification based on physical mechanisms makes it difficult to distinguish between equipment failures and sudden changes in actual water conditions, resulting in a high false alarm rate.
This paper proposes a cloud-based water conservancy project management data processing method. It utilizes a real-time video frame input feature extraction network to generate weighted feature maps, constructs dynamic mask clipping regions of interest, combines time-series data stream splitting and density clustering operations, uses environmental noise entropy vectors and historical hydrological data to generate and reconstruct time-series data streams, and calculates physical deviation index through water balance differential equations to distinguish between equipment failures and effective mutation data.
It achieves self-consistency verification in both physical and semantic dimensions, accurately distinguishes between equipment fault data and valid mutation data, solves the problem of feature fragmentation in data repair, ensures that the reconstructed time-series data stream is consistent with the historical hydrological dependency features in terms of environmental texture, and reduces the false alarm rate.
Smart Images
Figure CN122090359B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a cloud computing-based method and system for processing water conservancy project management data. Background Technology
[0002] With the widespread application of IoT and cloud computing technologies in water conservancy engineering, real-time monitoring of key hydrological elements using distributed sensing devices has become fundamental to ensuring the safe operation of reservoirs and the optimal allocation of water resources. However, water conservancy projects are typically located in open-air environments, and the collection and processing of monitoring data face complex environmental interferences and challenges.
[0003] Currently, Chinese invention patent application number 202411801794.1 discloses a data analysis and processing method and system based on water conservancy engineering. It locates abnormal data by searching for blank characters or garbled text, determines the missing data type based on data dimensions, and uses algebraic calculations to fill in missing data based on the inherent power conversion law of the equipment, or directly deletes data when the number of missing data exceeds a threshold. However, the existing technology has the following shortcomings: First, source data collection lacks anti-interference capabilities and has a low degree of data structuring. Existing technologies often assume that the data entering the database only has format errors, ignoring the complexity of edge-end data collection. In real-world scenarios, when using visual sensors to read water gauges, they are easily affected by strong light reflection, water surface ripples, and floating objects. Processing only the generated text data at the database level cannot address reading deviations caused by environmental interference at the source. Second, the cloud-based repair logic is simple and rigid, lacking environmental realism. The algebraic interpolation method based on fixed physical formulas is essentially a deterministic linear calculation; however, real water level or flow sequences have extremely strong... The nonlinearity and time-varying nature of the data, coupled with environmental noise such as wind and waves, prevent the generation of smooth theoretical calculations when filling long-term gaps. This makes it impossible to perceive and retain the micro-fluctuation characteristics caused by the current hydrological environment. Third, there is a lack of deep semantic verification based on hydraulic mechanisms. Data screening is based solely on the threshold of the number of missing data, lacking a deep understanding of the physical meaning of the data. In water conservancy projects, data mutations may originate from sensor failures or actual flood discharges or rainstorms. Relying solely on simple statistical rules or algebraic relationships cannot construct a dynamic closed-loop verification system based on the differential equation of water balance. It is difficult to accurately distinguish between equipment failures and effective hydrological mutations from a semantic level, which can easily lead to false alarms or missed alarms. Summary of the Invention
[0004] The technical problem solved by this invention is that existing technologies cannot guarantee the accuracy and logical reliability of monitoring data in complex environments. Edge devices are difficult to obtain high-precision data due to strong light and ripple interference, and have high bandwidth consumption. Traditional cloud-based repair cannot perceive and retain the micro-fluctuation characteristics of the current hydrological environment, resulting in the loss of environmental realism in the supplementary data and the generation of feature fragmentation. Furthermore, it lacks self-consistency verification based on physical mechanisms, making it difficult to distinguish between equipment failure and sudden changes in actual water conditions, resulting in a high false alarm rate.
[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a cloud-based water conservancy project management data processing method, comprising the following steps: Step S1: Input the real-time video frame into the feature extraction network, calculate and generate the feature weight matrix, and weight the initial feature map to obtain the weighted feature map. Calculate the semantic density score based on the weighted feature map and generate a saliency heatmap. Construct a dynamic mask and crop out the region of interest, input it into the feature detection network, regress the output water level pixel coordinates, and map the water level pixel coordinates into structured water level values. Step S2: Based on the structured water level values, the time series data stream is split to obtain non-stationary fluctuation data and time series void data. The non-stationary fluctuation data is split to the first processing channel to obtain the water level process line and the environmental noise entropy vector. The time series void data is split to the second processing channel. Combined with the environmental noise entropy vector and historical hydrological data, the final completion value is obtained, and the reconstructed time series data stream is generated. Step S3: Mark outliers in the reconstructed time-series data stream using density clustering, and logically match the outliers with the historical association rule set. Calculate the theoretical dynamic storage capacity based on the current inflow data and structured water level values, and calculate the physical deviation index between the theoretical dynamic storage capacity and the static storage capacity. Based on the logical matching results and the physical deviation index, classify and mark the outliers as equipment fault data or valid mutation data.
[0006] Preferably, step S1 includes the following sub-steps: Step S101: Obtain the real-time video stream of the water conservancy facility, input the video frames of the real-time video stream into the feature extraction network to obtain the initial feature map, and perform global average pooling on the initial feature map in the horizontal and vertical directions respectively to generate two orthogonal direction-aware feature vectors. A one-dimensional convolutional kernel is used to perform cross-channel interactive calculation on the orientation-aware feature vector to generate a feature weight matrix. The feature weight matrix is then combined with the initial feature map to perform an element-level Hadamard product operation to obtain a weighted feature map. Step S102: Based on the weighted feature map, calculate the semantic density score of each spatial partition of the image, generate a saliency heatmap, calculate the gradient distribution of the saliency heatmap and set a binarization threshold, determine the region in the saliency heatmap whose semantic density score is greater than the binarization threshold as a high response region, construct a dynamic mask covering the high response region, and perform a logical AND operation between the dynamic mask and the video frame to obtain an image of the region of interest containing only the water gauge stripe.
[0007] Preferably, step S1 includes the following sub-steps: Step S103: Input the region of interest into the feature detection network, use the sparse connection feature aggregation layer in the feature detection network to perform regression operation on the region of interest image, output the vertical pixel coordinates of the water level in the image coordinate system, retrieve the homography matrix pre-stored in the edge computing node, and map the vertical pixel coordinates to the physical Euclidean space through the perspective transformation algorithm to obtain the structured water level value.
[0008] Preferably, step S2 includes the following sub-steps: Step S201: Based on the structured water level values, construct a time series data stream by sorting by timestamps, detect whether there are consecutive missing timestamps in the time series data stream. If there are no consecutive missing timestamps, but the variance of the structured water level values within the preset sliding window is greater than the preset fluctuation threshold, then mark the structured water level value segments within the preset sliding window as non-stationary fluctuation data and divert them to the first processing channel. If there are consecutive missing timestamps, a placeholder data sequence is generated. The placeholder data sequence includes missing timestamps and preset null values. The placeholder data sequence is marked as time-series hole data and diverted to the second processing channel. If there are no consecutive missing timestamps and the variance of the structured water level values within the preset sliding window is less than or equal to the preset fluctuation threshold, then the structured water level value fragments of the corresponding time period are marked as stable and valid data and retained in the reconstructed time series data stream.
[0009] Preferably, step S2 further includes the following sub-steps: Step S202: In the first processing channel, a phase space reconstruction method based on a sliding window is used to map one-dimensional non-stationary fluctuation data into a two-dimensional fluctuation data matrix. Singular value decomposition is performed on the fluctuation data matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. Calculate the energy contribution rate of each singular value in the singular value diagonal matrix and sort them in descending order. The top N singular values whose cumulative energy contribution rate reaches the preset ratio are identified as the main feature singular values, and the remaining singular values are identified as noise singular values. By utilizing the retained main feature singular values, a smooth water level process line is reconstructed. At the same time, the statistical distribution characteristics of noise singular values are extracted to construct an environmental noise entropy vector that characterizes the current water surface fluctuation state.
[0010] Preferably, step S2 further includes the following sub-steps: Step S203: In the second processing channel, a deep temporal reconstruction network is constructed, which includes an encoder, a condition generator, and a discriminator. Retrieve pre-stored historical hydrological sample data, which includes historical water level data, historical inflow data, historical rainfall data, and historical reservoir capacity data; The time-series hole data is input into the encoder for feature compression, and the environmental noise entropy vector is embedded into the condition generator. The condition generator combines the long-term time-series dependency features of historical hydrological sample data with the environmental noise entropy vector to generate a predicted filling value with the characteristics of current environmental fluctuations. The predicted imputation value is used as the generated sample, and the historical hydrological sample data is used as the real sample. They are input into the discriminator together. The generator loss function and the discriminator adversarial loss function are used for iterative optimization game training until the generator loss function and the adversarial loss function reach the preset equilibrium convergence condition, and the final imputation value is output. Step S204: The final completed value, water level process line and stable effective data are spliced and fused in time sequence according to the original timestamp order to generate a continuous and smooth reconstructed time sequence data stream for the entire time period. The reconstructed time sequence data stream is composed of reconstructed water level values arranged in time sequence.
[0011] Preferably, step S3 includes the following sub-steps: Step S301: Set the neighborhood radius threshold and the minimum sample number threshold for the density clustering algorithm, perform a density scan on the reconstructed time series data stream, and identify data points whose number of data points within the neighborhood radius threshold range is less than the minimum sample number threshold as statistical outliers. Based on the historical association rule set pre-mined using the Apriori algorithm, the historical association rule set includes positive correlation rules between rainfall increment and water level rise, and lagged correlation rules between upstream flow and dam front water level. The confidence scores of rainfall data at the time corresponding to the statistical outlier point and the confidence scores of upstream flow data for the lagged correlation rules are calculated, and it is determined whether the confidence scores are greater than the strong association threshold.
[0012] Preferably, step S3 further includes the following sub-steps: Step S302: Obtain the inflow and outflow data at the current moment, construct a dynamic storage capacity calculation model using the water balance differential equation, perform time integration on the difference between the inflow and outflow data between the current moment and the previous moment, calculate the water volume change in the first time period, retrieve the theoretical dynamic storage capacity value of the previous moment stored in the cloud database, and superimpose the water volume change value with the theoretical dynamic storage capacity value of the previous moment to obtain the theoretical dynamic storage capacity value at the current moment. The static reservoir capacity-water level relationship curve is queried to obtain the static reservoir capacity value corresponding to the reconstructed water level value at the current moment. The absolute value of the difference between the theoretical dynamic reservoir capacity value and the static reservoir capacity value is calculated, and the absolute value is divided by the static reservoir capacity value to obtain the physical deviation index.
[0013] Preferably, step S3 further includes the following sub-steps: Step S303: If the confidence level is less than or equal to the strong correlation threshold, or the physical deviation index is greater than the physical self-consistency threshold, then the statistical outlier is determined to violate the hydraulic logic, the statistical outlier is classified and marked as equipment fault data and a data rewrite instruction is triggered. If the confidence level is greater than the strong correlation threshold and the physical deviation index is less than or equal to the physical self-consistency threshold, then the statistical outliers are determined to reflect the true water conditions, and the statistical outliers are classified and marked as valid mutation data and an early warning command is triggered.
[0014] A cloud-based water conservancy project management data processing system includes a mapping module, a reconstruction module, and a verification module. The mapping module is used to input real-time video frames into the feature extraction network, calculate and generate a feature weight matrix, and weight the initial feature map to obtain a weighted feature map. Based on the weighted feature map, the semantic density score is calculated and a saliency heatmap is generated. A dynamic mask is constructed and the region of interest is cropped out. The input is then input into the feature detection network, and the regression output is the water level pixel coordinates. The water level pixel coordinates are then mapped to structured water level values. The reconstruction module is used to split the time-series data stream based on structured water level values to obtain non-stationary fluctuation data and time-series void data. The non-stationary fluctuation data is split to the first processing channel to obtain the water level process line and the environmental noise entropy vector. The time-series void data is split to the second processing channel. Combined with the environmental noise entropy vector and historical hydrological data, the final completion value is obtained to generate the reconstructed time-series data stream. The verification module is used to mark outliers in the reconstructed time-series data stream according to density clustering operations, and to logically match the outliers with the historical association rule set. Based on the current inflow data and structured water level values, it calculates the theoretical dynamic storage capacity and the physical deviation index between the theoretical dynamic storage capacity and the static storage capacity value. Based on the results of logical matching and the physical deviation index, the outliers are classified and marked as equipment failure data or valid mutation data.
[0015] The beneficial effects of this invention are as follows: By performing phase space reconstruction and singular value decomposition on non-stationary fluctuation data, the noise singular values, which are usually discarded information, are inversely transformed into environmental noise entropy vectors that characterize the current water surface fluctuation state. The environmental noise entropy vectors are then embedded as prior conditions into the condition generator of the second processing channel, strongly linking the denoising process and the generation process. This ensures that the generated final completed value satisfies the long-term time-series dependence of historical hydrological sample data while inheriting the micro-fluctuation characteristics of the current environment, thus solving the problem of feature fragmentation in data repair. Combined with the physical deviation index calculated based on the water balance differential equation for secondary verification, the reconstructed time-series data stream is ensured to be self-consistent in both the physical and semantic dimensions, thereby accurately distinguishing between equipment failure data and effective mutation data. Attached Figure Description
[0016] Figure 1 A flowchart illustrating the steps of a cloud computing-based water conservancy project management data processing method according to an embodiment of the present invention; Figure 2 A basic flowchart of a cloud-based water conservancy project management data processing system provided in one embodiment of the present invention. Detailed Implementation
[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0018] Example 1, referring to Figure 1 This paper provides a cloud-based data processing method for water conservancy project management, including the following steps: Step S1: Input the real-time video frame into the feature extraction network, calculate and generate the feature weight matrix, and weight the initial feature map to obtain the weighted feature map. Calculate the semantic density score based on the weighted feature map and generate a saliency heatmap. Construct a dynamic mask and crop out the region of interest, input it into the feature detection network, regress the output water level pixel coordinates, and map the water level pixel coordinates into structured water level values.
[0019] Step S2: Based on the structured water level values, the time-series data stream is split to obtain non-stationary fluctuation data and time-series void data. The non-stationary fluctuation data is split to the first processing channel to obtain the water level process line and the environmental noise entropy vector. The time-series void data is split to the second processing channel. Combined with the environmental noise entropy vector and historical hydrological data, the final completion value is obtained, and the reconstructed time-series data stream is generated.
[0020] Step S3: Mark outliers in the reconstructed time-series data stream using density clustering, and logically match the outliers with the historical association rule set. Calculate the theoretical dynamic storage capacity based on the current inflow data and structured water level values, and calculate the physical deviation index between the theoretical dynamic storage capacity and the static storage capacity. Based on the logical matching results and the physical deviation index, classify and mark the outliers as equipment fault data or valid mutation data.
[0021] This invention performs phase space reconstruction and singular value decomposition on non-stationary fluctuation data, inversely transforming noise singular values, which are usually discarded information, into environmental noise entropy vectors representing the current water surface fluctuation state. These environmental noise entropy vectors are then embedded as prior conditions into the condition generator of the second processing channel, strongly linking the denoising and generation processes. This ensures that the generated final completed value satisfies the long-term temporal dependence of historical hydrological sample data while inheriting the micro-fluctuation characteristics of the current environment, solving the problem of feature fragmentation in data restoration. A secondary verification is performed using a physical deviation index calculated based on the water balance differential equation, ensuring the reconstructed time-series data stream is self-consistent in both the physical and semantic dimensions, thereby accurately distinguishing between equipment failure data and valid mutation data.
[0022] In a specific embodiment, step S1 includes the following sub-steps: Step S101: Obtain the real-time video stream of the water conservancy facility, input the video frames of the real-time video stream into the feature extraction network to obtain the initial feature map, and perform global average pooling on the initial feature map in the horizontal and vertical directions to generate two orthogonal direction-aware feature vectors.
[0023] A one-dimensional convolutional kernel is used to perform cross-channel interactive computation on the orientation-aware feature vectors, generating a feature weight matrix containing long-range dependencies. The feature weight matrix is then subjected to element-wise Hadamard product operation with the initial feature map to obtain a weighted feature map. The weights of water surface reflection regions with high-frequency random texture attributes are suppressed nonlinearly, while the weights of water level scale regions with regular geometric attributes are enhanced.
[0024] Step S102: Based on the weighted feature map, calculate the semantic density score of each spatial partition of the image to generate a saliency heatmap. Calculate the gradient distribution of the saliency heatmap and set a binarization threshold. Identify regions in the saliency heatmap with semantic density scores greater than the binarization threshold as high-response regions. Construct a dynamic mask covering the high-response regions. Perform a logical AND operation between the dynamic mask and the video frame to obtain a region of interest image containing only the water gauge stripe. Isolate shoreline vegetation and floating debris interference.
[0025] In this embodiment, the binarization threshold is used as the critical value to segment the saliency heatmap into high-response regions (water level) and background regions. This embodiment adopts an adaptive threshold strategy instead of a fixed value, specifically using the maximum inter-class variance method to automatically calculate the binarization threshold. To enhance recall in areas with weak texture, the binarization threshold was set to [value missing]. Outdoor lighting conditions vary drastically, such as large differences between day and night light. A fixed threshold can easily lead to missed detections in dim environments or over-detections in bright light. The maximum inter-class variance method dynamically finds the optimal segmentation point based on the image histogram, ensuring that the dynamic mask accurately covers the water gauge strip under different lighting conditions, effectively eliminating interference from shoreline vegetation and floating debris. The typical range of binarization threshold values is usually within... Within the range.
[0026] Semantic density scores are used to quantify the confidence level of each pixel region in an image as a water level target. Let the weighted feature map be... ,in, For the number of channels, For spatial dimensions. Calculate spatial coordinates. The semantic density score at a given location is expressed mathematically as follows: ; in, For semantic density score, For the first Each channel in spatial coordinates Activation value at that location, For the first The feature importance weights of each channel are determined by the feature weight matrix in step S101.
[0027] To eliminate the influence of numerical scaling, the score matrix of the entire image is... After performing Min-Max normalization, the resulting significance heatmap is the normalized score matrix. .
[0028] Step S103: Input the region of interest into the feature detection network, use the sparse connection feature aggregation layer in the feature detection network to perform regression operation on the region of interest image, output the vertical pixel coordinates of the water level in the image coordinate system, retrieve the homography matrix pre-stored in the edge computing node, and use the perspective transformation algorithm to map the vertical pixel coordinates to the physical Euclidean space to obtain the structured water level value with physical units.
[0029] In this embodiment, the feature extraction network uses an improved ResNet-50 residual network as its backbone architecture. Specifically, the fully connected layers and the last two pooling layers of the original ResNet-50 residual network are removed, while the convolutional layers from Stage 1 to Stage 4 are retained to extract multi-scale spatial features. The input real-time video frames are adjusted to... A three-channel RGB image of pixels. The orientation-aware feature vector is generated by analyzing the feature map output from Stage 4 (dimension 1). ) respectively execute the horizontal direction ( ) and vertical direction ( Global average pooling is used to compress the two-dimensional features into two orthogonal one-dimensional directional sensing feature vectors, which capture the scale distribution features of the water gauge in the vertical direction and the water surface ripple features in the horizontal direction, respectively.
[0030] The sparse connection feature aggregation layer in the feature detection network specifically employs a sparse convolutional layer improved from a dilated spatial pyramid pooling structure. This sparse connection feature aggregation layer contains four parallel convolutional branches, specifically one... Convolution and three Dilated convolutions with dilation rates set to 6, 12, and 18 are used. This sparse connection method allows the feature detection network to significantly expand its receptive field without increasing the number of parameters, capturing long-range contextual information of the watermark bands. The feature maps output from each branch are concatenated along the channel dimension and then... The convolutional layers are fused together and the final input is fed into the regression head, which consists of two fully connected layers with 256 and 1 neurons respectively. The output is the vertical pixel coordinates of the waterline in the image coordinate system. .
[0031] This invention addresses the problem that drastic changes in outdoor lighting conditions can easily lead to missed detections in dim environments or over-detections under strong light, and the problem of being easily disturbed by water reflections, shore vegetation, and floating objects. Considering the need to decouple the target from complex environmental noise and achieve dynamic perception segmentation during the feature extraction stage, global average pooling is performed on the initial feature map in the horizontal and vertical directions to generate two orthogonal direction-aware feature vectors. Cross-channel interactive calculation is used to generate a feature weight matrix containing long-range dependencies, and element-wise Hadamard product operation is performed with the initial feature map to obtain a weighted feature map. Based on the weighted feature map, semantic density scores are calculated to generate a saliency heatmap, and an adaptive thresholding strategy using the maximum inter-class variance method is adopted to automatically calculate the binarization threshold and construct a dynamic mask. Finally, regression operations and homography matrix mapping are performed through a sparsely connected feature aggregation layer. The weight of the water surface reflection area with high-frequency random texture attributes was suppressed by nonlinearity, while the weight of the water gauge scale area with regular geometric attributes was enhanced. This not only ensured that the dynamic mask could accurately cover the water gauge strip under different lighting conditions, effectively isolated and eliminated the interference of shore vegetation and floating objects on the water surface, but also ultimately accurately transformed the pixel coordinate perspective to physical Euclidean space, outputting high-precision structured water level values with physical units.
[0032] In a specific embodiment, step S2 includes the following sub-steps: Step S201: Based on the structured water level values, construct a time-series data stream by sorting by timestamps, detect whether there are consecutive missing timestamps in the time-series data stream. If there are no consecutive missing timestamps, but the variance of the structured water level values within a preset sliding window is greater than a preset fluctuation threshold, then mark the structured water level value segments within the preset sliding window as non-stationary fluctuation data and divert them to the first processing channel.
[0033] If there are consecutive missing timestamps, a placeholder data sequence is generated. The placeholder data sequence includes missing timestamps and preset null values. The placeholder data sequence is marked as time-series hole data and diverted to the second processing channel.
[0034] If there are no consecutive missing timestamps and the variance of the structured water level values within the preset sliding window is less than or equal to the preset fluctuation threshold, then the structured water level value fragments of the corresponding time period are marked as stable and valid data and directly retained in the reconstructed time series data stream.
[0035] In step S202, in the first processing channel, a phase space reconstruction method based on a sliding window is used to map one-dimensional non-stationary fluctuation data into a two-dimensional fluctuation data matrix. Singular value decomposition is performed on the fluctuation data matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix.
[0036] Calculate the energy contribution rate of each singular value in the singular value diagonal matrix and sort them in descending order. The top N singular values whose cumulative energy contribution rate reaches the preset ratio are identified as the main feature singular values, and the remaining singular values are identified as noise singular values.
[0037] By utilizing the retained main feature singular values, a smooth water level process line is reconstructed. At the same time, the statistical distribution characteristics of noise singular values are extracted to construct an environmental noise entropy vector that characterizes the current water surface fluctuation state.
[0038] In this embodiment, the preset sliding window is set to a time span of 30 to 60 seconds. If the video sampling rate is 1fps, the window size is 30 to 60 data points. The preset fluctuation threshold is set to 3 times the average variance of water level during historical stable periods; for example, if the historical variance of still water level is... The preset fluctuation threshold is set to It can accurately distinguish between continuous high-frequency fluctuations (non-stationary data) caused by wind and waves and instantaneous jumps (outliers) caused by equipment failure, thus achieving precise data splitting.
[0039] The preset ratio is used to determine the cumulative energy percentage standard of the number N singular values of the main feature. The preset ratio is set to 90%. In specific implementation, the cumulative energy contribution rate curve is calculated, and the N value corresponding to the abrupt change point of the curve slope is selected. It is required that the cumulative contribution rate is greater than 90%. The first 90% of the energy contains the main trend of the signal, namely the water level change, while the last 10% of the low energy part contains environmental noise, namely the wind and wave texture. This not only preserves the real physical change trend of the water level (reconstructed from the first N singular values), but also successfully separates the pure environmental noise component (extracted from the remaining singular values), providing a clean noise source for the subsequent construction of the environmental noise entropy vector.
[0040] Constructing the environmental noise entropy vector specifically includes: After performing singular value decomposition on the fluctuation data matrix, a diagonal singular value sequence is obtained. Select the top performers based on the cumulative energy contribution rate threshold (e.g., 90%). The main feature is one, and the rest are... The singular values constitute the noise singular value set. .
[0041] To quantify the degree of disorder in the current water surface fluctuations, i.e., the environmental entropy, the noise singular values are normalized into a probability distribution. The mathematical expression for the probability distribution is: ; Calculating environmental entropy using the Shannon entropy formula The mathematical expression for environmental entropy is: ; To more comprehensively characterize the environmental state, in addition to the environmental entropy value, the mean of the noise singular values is also calculated. and variance .
[0042] Constructed environmental noise entropy vector Defined as a three-dimensional feature vector ,Will It is transmitted to the second processing channel as a priori input to the condition generator.
[0043] Step S203: In the second processing channel, a deep temporal reconstruction network is constructed, which includes an encoder, a condition generator, and a discriminator.
[0044] Retrieve pre-stored historical hydrological sample data, which includes historical water level data, historical inflow data, historical rainfall data, and historical reservoir capacity data.
[0045] The time-series void data is input into the encoder for feature compression. Based on the missing timestamps in the time-series void data, the nearest non-stationary fluctuation data segment is searched on the time axis, and the corresponding environmental noise entropy vector is extracted. If no non-stationary fluctuation data segment exists within a preset time neighborhood, the environmental noise entropy vector is set to the mean of the environmental noise entropy vectors for the same historical period and embedded into the condition generator. If the time-series void data does not lack timestamps, the environmental noise entropy vector is directly embedded into the condition generator. The condition generator combines the long-term temporal dependency features of historical hydrological sample data with the environmental noise entropy vector to generate predicted imputation values that reflect the current environmental fluctuation characteristics.
[0046] The predicted imputation value is used as the generated sample, and historical hydrological sample data is used as the real sample. They are input into the discriminator together. The generator loss function and the discriminator adversarial loss function are used for iterative optimization game training until the generator loss function and the adversarial loss function reach the preset equilibrium convergence condition, and the final imputation value is output.
[0047] In this embodiment, the deep temporal reconstruction network adopts a sequence-to-sequence conditional generative adversarial network architecture, specifically including: The encoder employs a bidirectional gated cyclic unit, with input consisting of a split, time-series, holed data segment of dimension [dimension number missing]. The encoder maps time-series data fragments with holes into hidden state vectors. This is to capture bidirectional time dependencies. The hidden layer dimension of the bidirectional gated loop unit is set to 128.
[0048] The condition generator employs a decoder structure with an environmental feature injection mechanism. To effectively embed the environmental noise entropy vector into the generation process, the condition generator introduces an adaptive instance normalization layer. The specific logic of the embedding operation is as follows: Let the environmental noise entropy vector be... , dimension The feature map of the condition generator at the current time step is The network first utilizes two independent multilayer perceptrons to... Mapped to scaling factor and bias factor Then, the environment texture is injected into the generator's features: ; in, This is the feature map after environmental feature injection. and Features The mean and standard deviation of the environmental noise entropy vector are no longer simply input splicing, but rather control the statistical distribution characteristics of the generated data, so that the generated fill value is consistent with the current wind and wave environment in terms of texture.
[0049] The discriminator employs a PatchGAN structure based on a one-dimensional convolutional neural network. It receives the complete time series and the corresponding environmental noise entropy vector. The complete time series includes both real and generated samples. The output is the probability that the complete time series is not only numerically real but also matches the environmental texture.
[0050] The training process employs the WGAN-GP loss function to address the mode collapse problem during training and ensure the stability of the generated gradients. The equilibrium convergence condition is that the fluctuation range of the generator loss and discriminator loss is less than 5% over 50 consecutive epochs.
[0051] Step S204 involves splicing and merging the final completed values, water level process lines, and stable and valid data in the order of the original timestamps to generate a continuous and smooth reconstructed time-series data stream. The reconstructed time-series data stream consists of reconstructed water level values arranged in time sequence.
[0052] This invention addresses the problem that traditional interpolation methods often generate smooth theoretical values when filling time-series gaps, leading to the loss of realistic water surface fluctuation characteristics. Considering that the low-energy portion (the bottom 10% of the signal) is not useless data but contains pure environmental noise components such as wind and wave textures, this invention proposes a reconstruction mechanism based on diversion and cross-channel injection. It performs precise diversion processing based on the variance and timestamp characteristics of structured water level values. In the first processing channel, it performs phase space reconstruction and singular value decomposition based on a sliding window on non-stationary fluctuation data. Using the noise singular values, it calculates statistical distribution characteristics such as Shannon entropy, and inversely constructs an environmental noise entropy vector representing the current water surface fluctuation state. In the second processing channel, a sequence-to-sequence conditional generative adversarial network is constructed. By introducing an adaptive instance normalization layer into the conditional generator, the environmental noise entropy vector is mapped to scaling and bias factors using a multilayer perceptron and injected into the features of the generator. This makes the environmental noise entropy vector no longer a simple input concatenation, but directly controls the statistical distribution features of the generated data. This ensures that the generated final completion value, while combining the long-term time-series dependency features of historical hydrological sample data, is consistent with the current wind and wave environment in texture. This makes the final output reconstructed time-series data stream not only numerically realistic, but also matching in environmental texture, solving the problem of environmental feature fragmentation during data repair.
[0053] In a specific embodiment, step S3 includes the following sub-steps: Step S301: Set the neighborhood radius threshold and the minimum sample number threshold for the density clustering algorithm, perform a density scan on the reconstructed time series data stream, and identify data points whose number of data points within the neighborhood radius threshold range is less than the minimum sample number threshold as statistical outliers.
[0054] Based on the historical association rule set pre-mined by the Apriori algorithm, the historical association rule set includes positive correlation rules between rainfall increment and water level rise, as well as lagged correlation rules between upstream flow and dam front water level. The confidence scores of rainfall data at the time corresponding to outliers and the confidence scores of upstream flow data for lagged correlation rules are calculated and statistically analyzed to determine whether the confidence scores are greater than the strong association threshold.
[0055] In this embodiment, the specific components of the historical association rule set pre-mined based on the Apriori algorithm include: Retrieve pre-stored long-term historical hydrological synchronous observation data from the cloud database. This data includes historical rainfall sequences, historical upstream water level sequences, and historical upstream inflow sequences. Based on a pre-defined numerical interval division standard, the continuous numerical hydrological sequences are discretized and mapped. Specifically, this includes: The incremental rainfall is mapped to a first discrete state item set. Based on the 24-hour cumulative rainfall, the first discrete state item set is divided into four categories: no rain, light rain, moderate rain, and heavy rain. No rain is defined as rainfall less than 10 mm, light rain as rainfall between 10 mm and 24.9 mm, moderate rain as rainfall between 25 mm and 49.9 mm, and heavy rain as rainfall greater than or equal to 50 mm. The rate of change of upstream inflow is mapped to a second discrete state item. Based on the percentage change in flow over two hours, the second discrete state item is divided into three categories: stable flow, slow increase in flow, and surge in flow. Stable flow is defined as a rate of change of upstream inflow less than 5%, slow increase in flow is defined as a rate of change of upstream inflow between 5% and 20%, and surge in flow is defined as a rate of change of upstream inflow greater than 20%. The corresponding water level rise is mapped to a third discrete state term. Based on hourly water level changes, the third discrete state term is divided into three categories: stable water level, slight water level rise, and steep water level rise. Stable water level is defined as a rise less than 0.05m, slight water level rise is defined as a rise between 0.05m and 0.2m, and steep water level rise is defined as a rise greater than 0.2m. The first, second, and third discrete state terms within the matching time window are concatenated to construct a single historical transaction data entry. Based on the constructed historical transaction data set, the Apriori algorithm is used for iterative mining layer by layer to obtain frequent itemsets, and the support and confidence of each association rule are calculated. The mathematical expression for the confidence is: ; in, For confidence level, Indicates precondition item Support in the entire historical data set, i.e., the frequency of occurrence of the preceding hydrological conditions alone. Indicates precondition item With post-result items The joint support that occurs within the same historical time window, i.e., the frequency with which preceding hydrological conditions and subsequent hydrological results occur concurrently, and the confidence level. Characterization in terms of preconditions In a subset of historical data, subsequent result items also appear. The conditional probability and confidence level are used to accurately quantify the reliability of causal transmission relationships in hydraulics.
[0056] Invalid rules with support or confidence levels less than the preset minimum support or confidence level are removed. The remaining valid association rules are aggregated to generate a historical association rule set, which is a set of mapping relationships between preceding hydrological conditions and subsequent hydrological results.
[0057] The minimum support is preset to 0.03 because, in the objective laws of hydrology and climate, extreme weather (such as heavy rain causing a sharp rise in water levels) is a low-frequency, high-risk event. In the total historical data sample for the whole year, the absolute frequency of extreme situations is low, meaning the support is low. If the minimum support is set too high (e.g., 20%), these most critical flood warning rules will be filtered out by the algorithm as rare noise. Setting the minimum support to 0.03 ensures that low-frequency but important strong correlation rules of extreme hydrological events are preserved. The minimum confidence level is preset to 0.80 because hydraulic systems have extremely strong physical causal transmission; for example, continuous heavy rain upstream will inevitably lead to rising water levels downstream. Setting a strict confidence lower limit greater than 80% is to fully filter out weakly correlated rules caused by accidental factors in nature, such as localized showers but extremely dry soil causing runoff not to flow into the reservoir, ensuring that the extracted historical correlation rule set possesses a near-certain physical causal law. This parameter combination strategy, which combines low support and high confidence, is tailor-made for water conservancy scenarios. It not only ensures the ability to capture rare and extreme water conditions, but also ensures the absolute rigor of the extracted rules in logical derivation, thereby significantly reducing the false alarm rate in real-time detection.
[0058] In this embodiment, the neighborhood radius threshold is set to the physical measurement accuracy or allowable error range of the water level gauge sensor, for example, 2cm to 5cm. The minimum sample size threshold is set to 3 to 5 points. Dramatic changes in water level within a very short time (such as three consecutive sampling points) are generally physically impossible (unless a dam breaks). If no at least three neighboring points are found within a 5cm radius of a point, it indicates that the point is isolated noise or a malfunction. Compared to simple threshold filtering, this method effectively identifies clusters of false abrupt changes, such as continuous anomalies caused by the sensor being entangled in aquatic plants, significantly reducing the false alarm rate.
[0059] Because hydraulic logic has strong causal relationships—for example, if water is released upstream, the water level downstream is likely to rise—an 80% confidence level can filter out weak correlations caused by accidental factors, retaining only rules with strong causal relationships. Therefore, the strong correlation threshold is set to 0.8 to ensure the rigor of logical matching and prevent misjudgments caused by weak correlation rules.
[0060] Step S302: Obtain the inflow and outflow data at the current moment, construct a dynamic storage capacity calculation model using the water balance differential equation, perform time integration on the difference between the inflow and outflow data between the current moment and the previous moment to calculate the water volume change in the first time period, retrieve the theoretical dynamic storage capacity value of the previous moment stored in the cloud database, and superimpose the water volume change value with the theoretical dynamic storage capacity value of the previous moment to obtain the theoretical dynamic storage capacity value at the current moment.
[0061] The static reservoir capacity-water level relationship curve is queried to obtain the static reservoir capacity value corresponding to the reconstructed water level value at the current moment. The absolute value of the difference between the theoretical dynamic reservoir capacity value and the static reservoir capacity value is calculated, and the absolute value is divided by the static reservoir capacity value to obtain the physical deviation index.
[0062] In this embodiment, the physical deviation index is used to quantify the reliability of the current monitored water level in terms of hydraulic physical mechanisms. The mathematical expression for the physical deviation index is: ; in, For the current moment The theoretical dynamic storage capacity, in units of The theoretical dynamic storage capacity value is based on the theoretical storage capacity at the previous moment. The net water inflow during the current period is calculated as the inflow rate into the reservoir. Subtract outbound flow The time integral satisfies the water balance differential equation: ; For the current moment The static storage capacity value, in units of The static reservoir capacity value is obtained by querying the preset reservoir capacity-water level relationship curve and reconstructing the water level value output in step S204. The reservoir capacity value is obtained directly from a table. The physical deviation index reflects the degree of deviation between the theoretical water volume calculated based on flow rate and the water volume obtained from a table based on water level observation.
[0063] If the physical deviation index is greater than the preset physical self-consistency threshold, it indicates that the current flow rate and water level relationship violates the law of conservation of mass, suggesting that the system has water level sensor drift or flow meter failure.
[0064] Step S303: If the confidence level is less than or equal to the strong correlation threshold, or the physical deviation index is greater than the physical self-consistency threshold, then the statistical outlier is determined to violate the hydraulic logic, the statistical outlier is classified and marked as equipment fault data and a data rewrite instruction is triggered.
[0065] If the confidence level is greater than the strong correlation threshold and the physical deviation index is less than or equal to the physical self-consistency threshold, then the statistical outliers are determined to reflect the true water conditions, and the statistical outliers are classified and marked as valid mutation data and an early warning instruction is triggered.
[0066] In this embodiment, the physical self-consistency threshold is a preset tolerance ratio, defining the maximum error range allowed between the theoretical reservoir capacity calculated based on flow rate and the static reservoir capacity based on water level observation. This aims to filter out normal instrument measurement errors and only identify severe deviations exceeding the physical self-consistency threshold as equipment failures violating the water balance principle. Considering the measurement error of the flow meter (typically between 2% and 5%) and the fitting error of the reservoir capacity curve, setting a tolerance range of 5% to 10% is consistent with engineering practice. Therefore, the physical self-consistency threshold is set to 5% to 10%, which, while tolerating normal instrument errors, keenly detects severe anomalies that violate the water balance law, such as sensor drift causing a deviation greater than 20%, thereby accurately marking equipment failures.
[0067] This invention addresses the problem that existing systems, relying solely on simple threshold filtering, struggle to identify clusters of false mutations and suffer from high false alarm rates due to a lack of deep semantic verification based on hydraulic mechanisms. Considering that drastic water level jumps within a very short time are generally physically impossible, and that hydraulic logic exhibits strong causality, the relationship between flow rate and water level must not violate the law of conservation of mass, this invention innovatively constructs a verification model that is both physically and semantically self-consistent. First, a density clustering algorithm is used to set a neighborhood radius threshold and a minimum sample number threshold for density scanning to identify statistical outliers. Then, confidence scores are calculated based on the historical association rule set pre-mined by the Apriori algorithm for logical matching. Finally, the flow rate difference is integrated over time using the water balance differential equation to calculate the theoretical dynamic reservoir capacity. The theoretical dynamic reservoir capacity is then compared with the static reservoir capacity obtained by querying the static reservoir capacity-water level relationship curve to calculate a physical deviation index reflecting the degree of deviation between the theoretical water volume and the table-lookup water volume. By comparing the confidence level with the strong correlation threshold and the physical deviation index with the physical consistency threshold, it is possible to accurately determine whether statistical outliers are equipment failure data that violates hydraulic logic or effective abrupt changes that reflect the true water conditions from the dual dimensions of physical mechanism and semantic logic. This not only effectively identifies continuous anomalies caused by sensors being entangled in aquatic plants, but also eliminates misjudgments caused by simply relying on numerical jumps, greatly improving the accuracy of triggering data rewrite commands or early warning commands.
[0068] Example 2, refer to Figure 2 It provides a cloud-based water conservancy project management data processing system, including a mapping module, a reconstruction module, and a verification module.
[0069] The mapping module is used to input real-time video frames into the feature extraction network, calculate and generate a feature weight matrix, and weight the initial feature map to obtain a weighted feature map. Based on the weighted feature map, the semantic density score is calculated and a saliency heatmap is generated. A dynamic mask is constructed and the region of interest is cropped out. The input is then fed into the feature detection network, and the regression output is the pixel coordinates of the water level line. The pixel coordinates of the water level line are then mapped to structured water level values.
[0070] The reconstruction module is used to split the time-series data stream based on structured water level values to obtain non-stationary fluctuation data and time-series void data. The non-stationary fluctuation data is split to the first processing channel to obtain the water level process line and environmental noise entropy vector. The time-series void data is split to the second processing channel. Combined with the environmental noise entropy vector and historical hydrological data, the final completion value is obtained, and the reconstructed time-series data stream is generated.
[0071] The verification module is used to mark outliers in the reconstructed time-series data stream based on density clustering operations, and to logically match the outliers with the historical association rule set. Based on the current inflow data and structured water level values, it calculates the theoretical dynamic storage capacity and the physical deviation index between the theoretical dynamic storage capacity and the static storage capacity. Based on the results of logical matching and the physical deviation index, the outliers are classified and marked as equipment failure data or valid mutation data.
[0072] This invention addresses the problems of existing water monitoring systems, such as susceptibility to visual interference during source extraction in complex field environments, loss of environmental realism in cloud data restoration, and high false alarm rates due to a lack of physical consistency verification. It proposes an end-to-end systematic integration of source visual anti-interference, temporal feature fidelity preservation, and hydraulic logic verification. Through a mapping module that generates salient heatmaps based on semantic density scores and constructs dynamic masks, it achieves high-precision mapping of structured water level values resistant to environmental interference. Recognizing that interference noise in non-stationary fluctuation data can be reversibly transformed into features characterizing water surface status, a reconstruction module performs temporal data stream splitting, combining the environmental noise entropy vector extracted from the first processing channel with historical hydrological data across channels. According to the method, the time-series gap data of the second processing channel is filled to obtain the final complete value that inherits the texture of the real environment to generate the reconstructed time-series data stream. Considering that a single numerical jump cannot distinguish between true and false, the verification module performs logical matching based on historical association rule set and physical deviation index based on theoretical dynamic capacity and static capacity value on the statistical outliers marked by density clustering operation. This makes the final reconstructed time-series data stream perfectly retain the micro-texture features of the current wind and wave environment. Moreover, under the dual control of physical constraints and logical rules, it can accurately classify and mark statistical outliers as equipment fault data or effective mutation data, which solves the technical bottleneck of functional fragmentation and frequent false alarms in the existing system.
[0073] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.
Claims
1. A cloud computing-based water conservancy project management data processing method, characterized in that, Includes the following steps: Step S1: Input the real-time video frame into the feature extraction network, calculate and generate the feature weight matrix, and weight the initial feature map to obtain the weighted feature map. Calculate the semantic density score based on the weighted feature map and generate a saliency heatmap. Construct a dynamic mask and crop out the region of interest, input it into the feature detection network, regress the output water level pixel coordinates, and map the water level pixel coordinates into structured water level values. Step S2: Based on the structured water level values, the time series data stream is split to obtain non-stationary fluctuation data and time series void data. The non-stationary fluctuation data is split to the first processing channel to obtain the water level process line and the environmental noise entropy vector. The time series void data is split to the second processing channel. Combined with the environmental noise entropy vector and historical hydrological data, the final completion value is obtained, and the reconstructed time series data stream is generated. Step S2 includes the following sub-steps: Step S201: Based on the structured water level values, construct a time series data stream by sorting by timestamps, detect whether there are consecutive missing timestamps in the time series data stream. If there are no consecutive missing timestamps, but the variance of the structured water level values within the preset sliding window is greater than the preset fluctuation threshold, then mark the structured water level value segments within the preset sliding window as non-stationary fluctuation data and divert them to the first processing channel. If there are consecutive missing timestamps, a placeholder data sequence is generated. The placeholder data sequence includes missing timestamps and preset null values. The placeholder data sequence is marked as time-series hole data and diverted to the second processing channel. If there are no consecutive missing timestamps and the variance of the structured water level values within the preset sliding window is less than or equal to the preset fluctuation threshold, then the structured water level value fragments of the corresponding time period are marked as stable and valid data and retained in the reconstructed time series data stream. Step S202: In the first processing channel, a phase space reconstruction method based on a sliding window is used to map one-dimensional non-stationary fluctuation data into a two-dimensional fluctuation data matrix. Singular value decomposition is performed on the fluctuation data matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. Calculate the energy contribution rate of each singular value in the singular value diagonal matrix and sort them in descending order. The top N singular values whose cumulative energy contribution rate reaches the preset ratio are identified as the main feature singular values, and the remaining singular values are identified as noise singular values. The smooth water level process line is reconstructed by utilizing the retained main feature singular values, and the statistical distribution characteristics of noise singular values are extracted to construct an environmental noise entropy vector characterizing the current water surface fluctuation state. Step S203: In the second processing channel, a deep temporal reconstruction network is constructed, which includes an encoder, a condition generator, and a discriminator. Retrieve pre-stored historical hydrological sample data, which includes historical water level data, historical inflow data, historical rainfall data, and historical reservoir capacity data; The time-series hole data is input into the encoder for feature compression, and the environmental noise entropy vector is embedded into the condition generator. The condition generator combines the long-term time-series dependency features of historical hydrological sample data with the environmental noise entropy vector to generate a predicted filling value with the characteristics of current environmental fluctuations. The predicted imputation value is used as the generated sample, and the historical hydrological sample data is used as the real sample. They are input into the discriminator together. The generator loss function and the discriminator adversarial loss function are used for iterative optimization game training until the generator loss function and the adversarial loss function reach the preset equilibrium convergence condition, and the final imputation value is output. Step S204: The final completed value, water level process line and stable effective data are spliced and fused in time sequence according to the original timestamp order to generate a continuous and smooth reconstructed time sequence data stream for the entire time period. The reconstructed time sequence data stream is composed of reconstructed water level values arranged in time sequence. Step S3: Mark outliers in the reconstructed time-series data stream using density clustering, and logically match the outliers with the historical association rule set. Calculate the theoretical dynamic storage capacity based on the current inflow data and structured water level values, and calculate the physical deviation index between the theoretical dynamic storage capacity and the static storage capacity. Based on the logical matching results and the physical deviation index, classify and mark the outliers as equipment fault data or valid mutation data.
2. The cloud computing-based water conservancy project management data processing method as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: Obtain the real-time video stream of the water conservancy facility, input the video frames of the real-time video stream into the feature extraction network to obtain the initial feature map, and perform global average pooling on the initial feature map in the horizontal and vertical directions respectively to generate two orthogonal direction-aware feature vectors. A one-dimensional convolutional kernel is used to perform cross-channel interactive calculation on the orientation-aware feature vector to generate a feature weight matrix. The feature weight matrix is then combined with the initial feature map to perform an element-level Hadamard product operation to obtain a weighted feature map. Step S102: Based on the weighted feature map, calculate the semantic density score of each spatial partition of the image, generate a saliency heatmap, calculate the gradient distribution of the saliency heatmap and set a binarization threshold, determine the region in the saliency heatmap whose semantic density score is greater than the binarization threshold as a high response region, construct a dynamic mask covering the high response region, and perform a logical AND operation between the dynamic mask and the video frame to obtain an image of the region of interest containing only the water gauge stripe.
3. The cloud computing-based water conservancy project management data processing method as described in claim 2, characterized in that, Step S1 includes the following sub-steps: Step S103: Input the region of interest into the feature detection network, use the sparse connection feature aggregation layer in the feature detection network to perform regression operation on the region of interest image, output the vertical pixel coordinates of the water level in the image coordinate system, retrieve the homography matrix pre-stored in the edge computing node, and map the vertical pixel coordinates to the physical Euclidean space through the perspective transformation algorithm to obtain the structured water level value.
4. The cloud computing-based water conservancy project management data processing method as described in claim 3, characterized in that, Step S3 includes the following sub-steps: Step S301: Set the neighborhood radius threshold and the minimum sample number threshold for the density clustering algorithm, perform a density scan on the reconstructed time series data stream, and identify data points whose number of data points within the neighborhood radius threshold range is less than the minimum sample number threshold as statistical outliers. Based on the historical association rule set pre-mined using the Apriori algorithm, the historical association rule set includes positive correlation rules between rainfall increment and water level rise, and lagged correlation rules between upstream flow and dam front water level. The confidence scores of rainfall data at the time corresponding to the statistical outlier point and the confidence scores of upstream flow data for the lagged correlation rules are calculated, and it is determined whether the confidence scores are greater than the strong association threshold.
5. The cloud computing-based water conservancy project management data processing method as described in claim 4, characterized in that, Step S3 further includes the following sub-steps: Step S302: Obtain the inflow and outflow data at the current moment, construct a dynamic storage capacity calculation model using the water balance differential equation, perform time integration on the difference between the inflow and outflow data between the current moment and the previous moment, calculate the water volume change in the first time period, retrieve the theoretical dynamic storage capacity value of the previous moment stored in the cloud database, and superimpose the water volume change value with the theoretical dynamic storage capacity value of the previous moment to obtain the theoretical dynamic storage capacity value at the current moment. The static reservoir capacity-water level relationship curve is queried to obtain the static reservoir capacity value corresponding to the reconstructed water level value at the current moment. The absolute value of the difference between the theoretical dynamic reservoir capacity value and the static reservoir capacity value is calculated, and the absolute value is divided by the static reservoir capacity value to obtain the physical deviation index.
6. The cloud computing-based water conservancy project management data processing method as described in claim 5, characterized in that, Step S3 further includes the following sub-steps: Step S303: If the confidence level is less than or equal to the strong correlation threshold, or the physical deviation index is greater than the physical self-consistency threshold, then the statistical outlier is determined to violate the hydraulic logic, the statistical outlier is classified and marked as equipment fault data and a data rewrite instruction is triggered. If the confidence level is greater than the strong correlation threshold and the physical deviation index is less than or equal to the physical self-consistency threshold, then the statistical outliers are determined to reflect the true water conditions, and the statistical outliers are classified and marked as valid mutation data and an early warning command is triggered.
7. A cloud-based water conservancy project management data processing system, applied in any one of the cloud-based water conservancy project management data processing methods as described in claims 1-6, characterized in that, It includes a mapping module, a reconstruction module, and a verification module; The mapping module is used to input real-time video frames into the feature extraction network, calculate and generate a feature weight matrix, and weight the initial feature map to obtain a weighted feature map. Based on the weighted feature map, the semantic density score is calculated and a saliency heatmap is generated. A dynamic mask is constructed and the region of interest is cropped out. The input is then input into the feature detection network, and the regression output is the water level pixel coordinates. The water level pixel coordinates are then mapped to structured water level values. The reconstruction module is used to split the time-series data stream based on structured water level values to obtain non-stationary fluctuation data and time-series void data. The non-stationary fluctuation data is split to the first processing channel to obtain the water level process line and the environmental noise entropy vector. The time-series void data is split to the second processing channel. Combined with the environmental noise entropy vector and historical hydrological data, the final completion value is obtained to generate the reconstructed time-series data stream. The verification module is used to mark outliers in the reconstructed time-series data stream according to density clustering operations, and to logically match the outliers with the historical association rule set. Based on the current inflow data and structured water level values, it calculates the theoretical dynamic storage capacity and the physical deviation index between the theoretical dynamic storage capacity and the static storage capacity value. Based on the results of logical matching and the physical deviation index, the outliers are classified and marked as equipment failure data or valid mutation data.