Water conservancy project intelligent monitoring system and method

Abstract

The invention relates to the technical field of hydraulic engineering monitoring, and discloses a hydraulic engineering intelligent monitoring system and method. The method comprises the following steps: acquiring multi-source monitoring data such as a water level fluctuation sequence, a flow velocity change sequence and a gate opening change sequence of a water conservancy project monitoring point; spatial-temporal feature extraction processing is carried out on the multi-source monitoring data, and a hydraulic engineering operation state feature matrix containing spatial-temporal correlation data of water level fluctuation features, flow velocity change features and gate opening change features is generated; inputting the feature matrix into an anomaly detection model, and generating a water conservancy project abnormal operation state identifier containing anomaly type, anomaly level and anomaly position information; and finally, based on the abnormal operation state identifier, generating a water conservancy project regulation and control instruction set comprising a water level regulation scheme, a flow speed control scheme and a gate opening degree regulation scheme. According to the method, the engineering operation state can be comprehensively monitored, the abnormity can be accurately identified, the regulation and control scheme can be rapidly generated, and the intelligence and efficiency of hydraulic engineering monitoring can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water conservancy engineering monitoring, in particular to a water conservancy engineering intelligent monitoring system and method. BACKGROUND

[0002] In the process of water conservancy engineering operation management, real-time monitoring and effective regulation of the operation state of the engineering are important work to ensure the safe and stable operation of the engineering. The traditional water conservancy engineering monitoring method relies on manual inspection and single monitoring equipment to collect data, which has obvious limitations. Manual inspection is limited by the experience, physical strength and inspection frequency of the inspectors, and it is difficult to achieve real-time monitoring of the entire water conservancy engineering in all-weather conditions. There are often monitoring blind spots or monitoring delays, which may lead to potential operation abnormalities that cannot be discovered in time. At the same time, data recording in the process of manual inspection relies on paper records or simple electronic form records, and the data sorting and analysis efficiency is low, making it difficult to quickly and accurately determine the operation state of the engineering.

[0003] The single monitoring equipment data collection method can usually only obtain one of the water level, flow rate or gate opening, and cannot achieve the coordinated collection and comprehensive analysis of multi-source data. The operation state of the water conservancy engineering is a complex system problem, and there is a close correlation between the water level, flow rate and gate opening. Single data cannot fully reflect the actual operation of the engineering. For example, when only the abnormal rise of the water level is monitored, without the analysis of the flow rate and gate opening data, it is difficult to determine whether the abnormal water level is caused by the increase of the upstream water inflow, the gate opening fault or the change of the flow rate, so that targeted control measures cannot be developed.

[0004] The traditional monitoring method also has obvious deficiencies in data processing and abnormality detection. The traditional method uses a simple threshold judgment method to identify abnormalities, i.e. setting a fixed threshold range, and determining that the monitoring data is abnormal when it exceeds the threshold range. This method cannot consider the spatio-temporal correlation of the data, ignores the trend of the data changing over time and the spatial correlation between different monitoring points in the operation process of the water conservancy engineering, and is prone to misjudgment or omission. For example, during the flood season, the water level will show a gradual rising trend over time. If a fixed threshold is used for judgment, the normal rise of the water level may be determined as abnormal, or the abnormal rise of the water level may not be identified in time in the early stage because it does not reach the threshold. At the same time, after the traditional method detects the abnormality, it is difficult to quickly generate a scientific and reasonable control instruction, and usually needs to rely on the experience of technical personnel to manually develop a control scheme. The efficiency and scientificity of the development of the control scheme cannot be guaranteed, which may delay the best control opportunity and affect the safe operation of the water conservancy engineering. SUMMARY

[0005] The application aims to provide a water conservancy project intelligent monitoring system and method to solve the problems in the background art.

[0006] To achieve the above-mentioned purpose, the application provides a water conservancy project intelligent monitoring method, which comprises the following steps: obtaining multi-source monitoring data of a water conservancy monitoring point, wherein the multi-source monitoring data comprises a water level fluctuation sequence, a flow rate change sequence and a gate opening change sequence; performing spatio-temporal feature extraction processing on the multi-source monitoring data to generate a water conservancy project operation state feature matrix, wherein the water conservancy project operation state feature matrix contains spatio-temporal correlation data of water level fluctuation features, flow rate change features and gate opening change features; inputting the water conservancy project operation state feature matrix into an anomaly detection model to generate a water conservancy project abnormal operation state identifier, wherein the abnormal operation state identifier contains abnormal type, abnormal level and abnormal position information; generating a water conservancy project control instruction set based on the abnormal operation state identifier, wherein the control instruction set contains a water level adjustment scheme, a flow rate control scheme and a gate opening adjustment scheme.

[0007] Preferably, the spatio-temporal feature extraction processing on the multi-source monitoring data comprises the following steps: establishing a spatio-temporal correlation matrix of the water level fluctuation sequence and the flow rate change sequence, and calculating a transfer coefficient of the influence of water level change on flow rate; constructing a response function of the gate opening change sequence and the water level fluctuation sequence, and determining a lag time parameter of the influence of gate adjustment on water level change; extracting multi-scale features of the water level fluctuation sequence through a spatio-temporal convolutional neural network to generate a water level fluctuation feature vector; fusing local features and global features of the flow rate change sequence through an attention mechanism to generate a flow rate change feature vector; performing spatio-temporal alignment processing on the water level fluctuation feature vector, the flow rate change feature vector and the gate opening change feature to generate a water conservancy project operation state feature matrix.

[0008] Preferably, the establishment of the spatio-temporal correlation matrix of the water level fluctuation sequence and the flow rate change sequence comprises the following steps: dividing a time window of the water level fluctuation sequence, and calculating a water level change rate in each time window; establishing a regression model of the water level change rate and the flow rate change rate in the corresponding time window to determine the transfer coefficient of the influence of water level change on flow rate; constructing a water level-flow rate influence weight matrix according to the spatial position relationship of the monitoring points to calculate the influence weight of an upstream monitoring point on a downstream monitoring point; The transfer coefficient is multiplied by an influence weight matrix to generate a time-space correlation matrix of water level fluctuation sequence and flow rate change sequence.

[0009] Preferably, the water conservancy project operation state feature matrix is input into an anomaly detection model, including: A sample library containing historical anomaly cases is constructed, and feature patterns of the anomaly cases are extracted; A deep belief network is used to perform feature dimension reduction processing on the water conservancy project operation state feature matrix to generate a low-dimensional feature representation; The distance measure value of the low-dimensional feature representation and the normal operation feature is calculated by an anomaly detection algorithm; Based on the comparison result of the distance measure value and the preset threshold value, a water conservancy project abnormal operation state identifier is generated.

[0010] Preferably, the deep belief network is used to perform feature dimension reduction processing on the water conservancy project operation state feature matrix, including: A multi-layer restricted Boltzmann machine network structure is constructed, and the number of hidden layer nodes and the learning rate parameter are set; The water conservancy project operation state feature matrix is input into the first layer of the restricted Boltzmann machine for feature coding processing; The network parameters are optimized by layer-by-layer training to generate a stable deep belief network model; The trained deep belief network model is used to perform nonlinear transformation on the input features to output a low-dimensional feature representation.

[0011] Preferably, the water conservancy project control instruction set is generated based on the abnormal operation state identifier, including: A preset control strategy template is matched according to the abnormal type to obtain a basic control scheme; The parameter intensity of the basic control scheme is adjusted based on the abnormal level to generate an intensity-adapted control scheme; The implementation range of the control scheme is determined in combination with the abnormal position information to generate a spatially positioned control instruction; The water level regulation scheme, the flow rate control scheme and the gate opening adjustment scheme are time-series coordinated to generate a water conservancy project control instruction set.

[0012] Preferably, the parameter intensity of the basic control scheme is adjusted based on the abnormal level, including: A mapping relationship table of abnormal level and control intensity is established to determine the intensity adjustment coefficient corresponding to each level; The product of the basic control scheme parameter value and the intensity adjustment coefficient is calculated to generate an intensity-adapted control parameter; The intensity adjustment coefficient is corrected according to historical control effect data to optimize the calculation accuracy of the control parameter; The output comprises a control scheme containing intensity adaptation parameters.

[0013] Preferably, the method further comprises: Collecting engineering response data after the execution of the control instruction, the engineering response data comprising actual water level changes, actual flow rate changes and actual gate opening changes; Calculating deviation metric values of the engineering response data and the expected control effect; Updating the detection threshold parameters of the anomaly detection model based on the deviation metric values; Optimizing the generation algorithm of the water conservancy engineering control instruction set.

[0014] Preferably, the calculation of the deviation metric values of the engineering response data and the expected control effect comprises: Establishing a difference matrix of the engineering response data and the expected control data, and calculating local deviation values of each monitoring point; Generating a deviation distribution map of the whole region through a spatio-temporal interpolation method, and determining a deviation hot spot region; Calculating an overall deviation index and a maximum local deviation value, and generating a comprehensive deviation metric value; Comparing and analyzing the comprehensive deviation metric value with historical deviation data, and evaluating the control effect.

[0015] Preferably, the present application further comprises a water conservancy engineering intelligent monitoring system, which comprises a memory, a processor and a computer program stored in the memory and running on the processor, and the processor realizes the steps of the above-mentioned water conservancy engineering intelligent monitoring method when executing the computer program.

[0016] Compared with the prior art, the present application has the following beneficial effects: The water conservancy engineering intelligent monitoring method effectively solves the problem that single data acquisition in the traditional monitoring method cannot fully reflect the engineering operation state by acquiring multi-source monitoring data. The multi-source monitoring data comprises water level fluctuation sequences, flow rate change sequences and gate opening change sequences, which can collect water conservancy engineering operation data from multiple dimensions and fully capture various key information in the engineering operation process. These data cover the core parameters of water conservancy engineering operation, so that the monitoring of the engineering operation state is no longer limited to a single index, but can grasp the overall engineering operation situation, providing a comprehensive and rich data basis for subsequent data analysis and state judgment.

[0017] In the process of generating the water conservancy project operation state feature matrix by spatio-temporal feature extraction processing of multi-source monitoring data, the spatio-temporal correlation of the data is fully considered, and the limitation of ignoring the spatio-temporal correlation in the traditional data processing method is broken through. Spatio-temporal feature extraction can deeply mine the spatio-temporal correlation data between water level fluctuation features, flow rate change features and gate opening change features. Not only the change trend of each data in the time dimension is concerned, but also the mutual influence of the data of different monitoring points in the space dimension is analyzed, so that the internal law of the water conservancy project operation state can be more accurately reflected. Through this processing method, the originally dispersed and independent multi-source data can be converted into a feature matrix with internal correlation, so that the project operation information contained in the data can be fully mined and integrated, providing more valuable analysis basis for subsequent anomaly detection.

[0018] The water conservancy project operation state feature matrix is input into the anomaly detection model to generate an abnormal operation state identifier, which significantly improves the accuracy and comprehensiveness of anomaly detection. The anomaly detection model can analyze based on the integrated spatio-temporal correlation feature data, and compared with the traditional fixed threshold judgment method, it can more accurately identify the abnormal situation in the project operation. The generated abnormal operation state identifier contains abnormal type, abnormal level and abnormal position information, so that the technical personnel can clearly and accurately understand the specific situation of the anomaly, and no longer need to judge whether there is an anomaly as in the traditional method, but cannot determine the detailed information of the anomaly. The explicit abnormal type helps to quickly locate the cause of the anomaly, the abnormal level provides a basis for judging the severity of the anomaly, and the abnormal position information facilitates the technical personnel to quickly reach the scene for verification and processing, greatly improving the efficiency of anomaly processing.

[0019] Based on the abnormal operation state identifier, a set of water conservancy project control instructions is generated, realizing the quick connection from anomaly detection to control scheme formulation, and solving the problem of low efficiency and insufficient scientificity of control scheme formulation after anomaly detection in the traditional method. The control instruction set includes water level regulation scheme, flow rate control scheme and gate opening adjustment scheme. These schemes are formulated based on accurate abnormal information and can provide targeted control measures for different types and levels of abnormal situations. Instead of relying on manual experience to formulate control schemes, the system generates scientific and reasonable control instructions through systematic analysis and calculation, ensuring the scientificity and effectiveness of the control scheme. At the same time, the rapid generation of control instructions can timely respond to abnormal situations in project operation, avoid further deterioration of abnormal situations due to untimely control, and ensure that the water conservancy project can quickly recover to normal operation state when an abnormal situation occurs, maintaining the safety and stability of the project.

[0020] The whole intelligent monitoring method forms a complete process from data collection, data processing, abnormality detection to regulation instruction generation, realizes the automation and intelligence of water conservancy project monitoring. Compared with the traditional manual monitoring and simple data processing method, the method greatly reduces the manual intervention, reduces the error and low efficiency problems caused by manual operation. The automatic data collection and processing can realize the real-time monitoring of water conservancy project, avoid the hysteresis of manual inspection; the intelligent abnormality detection and regulation instruction generation can quickly respond to various situations in the project operation, improve the overall efficiency and level of water conservancy project operation management, and ensure that the water conservancy project can long-term stably play its functions of flood control, irrigation, water supply and the like. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A working principle diagram of the intelligent monitoring method of the water conservancy project is described. Figure 2 A working principle diagram of the spatiotemporal feature extraction and processing of the multi-source monitoring data is described. Figure 3 A working principle diagram of the water conservancy project running state feature matrix inputting the abnormality detection model is described. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0023] Please refer to Figure 1The application provides a water conservancy project intelligent monitoring system and method, which realizes real-time monitoring and regulation of project operation by integrating multi-source monitoring data and intelligent algorithms. The method obtains multi-source monitoring data of water conservancy project monitoring points, including water level fluctuation sequence, flow rate change sequence and gate opening change sequence. These sequences are continuously collected by sensors arranged at key positions to form time series data. Then, spatio-temporal feature extraction processing is performed on the multi-source monitoring data to capture the temporal dynamics and spatial correlation in the data, generating a water conservancy project operation state feature matrix that integrates spatio-temporal correlation information of water level fluctuation features, flow rate change features and gate opening features, providing structured input for subsequent analysis. Then, the water conservancy project operation state feature matrix is input into an anomaly detection model that identifies deviations by comparing with historical normal operation patterns, and outputs water conservancy project abnormal operation state identification, which includes abnormal types such as water level abnormality or flow rate abnormality, abnormal levels such as mild or severe, and abnormal position information such as specific monitoring point number. Finally, based on the abnormal operation state identification, a set of water conservancy project regulation instructions is generated, including water level regulation schemes such as adjusting gate opening to control water level, flow rate control schemes such as adjusting flow rate through pump station, and gate opening adjustment schemes such as accurately setting gate position. These schemes are implemented through control actuators to realize automatic regulation of the project.

[0024] Embodiment 1: refer to Figure 2 The obtained water level fluctuation sequence, flow rate change sequence and gate opening change sequence are first sent to a preprocessing module for data cleaning and standardization to eliminate the influence of sensor errors and dimension differences. The preprocessed data form a regular time series data set, preparing for subsequent deep feature extraction. The first step of feature extraction is to establish a spatio-temporal correlation matrix of water level fluctuation sequence and flow rate change sequence. This process captures the phased features of hydrological element changes by dividing reasonable time windows, for example, dividing continuous monitoring data into segments of several hours or days according to the regulation period of the water conservancy project. The average change rate of water level is calculated within each time window, which reflects the overall trend of water level change in that period. The change rate data of flow rate in the corresponding period is also extracted. On this basis, a statistical regression model between water level change rate and flow rate change rate is constructed, and the least squares method is used to fit the mathematical relationship between them to determine the quantitative transfer coefficient of water level change on flow rate. This coefficient represents the amount of flow rate change caused by unit water level change. Further consideration needs to be given to the influence of spatial layout of monitoring points. According to the relationship between flow direction and distance, different positions of monitoring points are assigned influence weights. Changes in upstream stations usually have a stronger influence on downstream. A water level-flow rate influence weight matrix is generated through spatial interpolation algorithm. Finally, the calculated transfer coefficient and spatial influence weight matrix are combined and operated to generate a matrix structure that fully expresses the spatio-temporal correlation characteristics of water level and flow rate.

[0025] The response function of the gate opening change sequence and the water level fluctuation sequence needs to analyze the dynamic relationship between the gate operation and the water level response. The opening and closing operation of the gate will directly cause the change of the upstream and downstream water level, but this change has a certain time lag effect. By analyzing the corresponding relationship between the gate opening adjustment event and the subsequent water level change in the historical data, a dynamic response model between the two is established by using system identification method. The model can accurately describe the time delay parameter required from the gate action to the significant change of the water level. The establishment of the response function considers various factors such as gate type and flow conditions. The influence range and duration of water level adjustment by gates of different scales are significantly different, and the appropriate function form needs to be determined according to the specific engineering characteristics. The multi-scale features of the water level fluctuation sequence are extracted by using a spatio-temporal convolutional neural network, which uses a deep learning method. The network structure includes multiple convolutional layers and pooling layers to capture fluctuation patterns at different time scales. The shallow convolutional kernel focuses on identifying small water level fluctuations in the short term, while the deep network can capture the long-term trends and periodic patterns of water level changes. The network scans the input sequence step by step to extract representative local features through a sliding window method. In the process of generating multi-scale features, the network automatically learns important features at different time dimensions, and converts the original water level sequence into a feature vector rich in information, which contains both the instantaneous change characteristics and the long-term evolution law of the water level.

[0026] The attention mechanism is used to fuse the local features and global features of the flow rate change sequence to balance the importance of recent changes and long-term rules. The attention mechanism calculates the weight of each time point in the sequence to contribute to the current state, achieving adaptive weighting of features. Local features reflect the instantaneous fluctuations of flow rate, which may indicate sudden conditions. Global features represent the overall trend and periodicity of flow rate changes. The attention mechanism dynamically adjusts the contribution of different time point features by evaluating their importance. This weighted fusion method strengthens the features of key time points and appropriately suppresses less important features. The generated flow rate change feature vector focuses more on the flow rate change patterns that have a significant impact on the operation of the project. The spatio-temporal alignment of the water level fluctuation feature vector, the flow rate change feature vector, and the gate opening change feature is a key step in generating a unified feature matrix. Since there are differences in the collection frequency and transmission delay of different monitoring data, time synchronization processing is necessary. The timestamp alignment method is used to unify all features to the same time reference. Missing data points are filled using appropriate interpolation methods. Spatial alignment integrates discrete point features into continuous spatial distribution according to the geographical location information of the monitoring points. The aligned feature data is organized into a regular three-dimensional tensor structure, where the time dimension, spatial dimension, and feature dimension together form a complete water conservancy project operation state feature matrix. This matrix provides high-quality input data for subsequent anomaly detection.

[0027] The implementation of the spatio-temporal feature extraction process relies on efficient computing architecture and algorithm optimization. Given the large amount of data and high real-time requirements of water conservancy monitoring, a distributed computing framework is used for parallel processing. The feature extraction module is designed as a configurable pipeline structure, allowing for adjustments to the feature extraction strategy and parameters based on specific project needs. This flexibility enables the method to adapt to different types and scales of water conservancy monitoring scenarios. The entire processing process emphasizes the interpretability and physical meaning of the features. The extracted features not only have statistical significance but also maintain consistency with the actual operation mechanism of water conservancy projects. Special attention is paid to handling common noise and outliers in the monitoring data during the feature extraction process. Robust statistical methods and filtering techniques are used to eliminate data anomalies caused by occasional sensor failures or transmission interference. Additionally, environmental factors are introduced as auxiliary information for feature extraction to improve the adaptability of feature representation. This allows the generated feature matrix to accurately reflect the engineering operation state under various working conditions. The results of spatio-temporal feature extraction not only serve real-time monitoring tasks but also can be used for historical operation pattern analysis and summary.

[0028] Example 2: Refer to Figure 3 , a sample library containing rich historical abnormal cases is constructed. The establishment of the sample library requires the system to collect and organize various abnormal working condition data accumulated during the long-term operation of water conservancy projects. These data come from historical records of multiple monitoring parameters such as water level, flow rate, and gate opening. The selection of abnormal cases covers different types of operation anomalies, including abnormal rise or fall of water level, sudden change of flow rate, and abnormal operation of gates. Each case is carefully labeled and confirmed by domain experts to ensure the accuracy and representativeness of the sample data. During the construction of the sample library, feature engineering methods are used to extract meaningful feature patterns from the original monitoring data. These patterns can capture the typical characteristics and rules of abnormal occurrences, providing reliable reference for subsequent anomaly detection. Deep belief network is used for feature dimensionality reduction processing of water conservancy operation state feature matrix, which is an important step based on deep learning. Deep belief network is stacked by multiple restricted Boltzmann machines and can automatically learn the distributed representation of input data through unsupervised learning. The network structure design needs to consider the dimension and complexity of input features. The number of nodes in the hidden layer is determined according to the size of the feature matrix. The optimal configuration is usually selected by comparing the performance of different structures through experiments. During the training process, the contrastive divergence algorithm is used to update the network parameters layer by layer. Each layer of restricted Boltzmann machine learns different abstract representations of input data. Finally, the high-dimensional feature matrix is converted into low-dimensional but information-rich feature representation. This dimensionality reduction processing not only reduces the computational complexity but also eliminates the redundancy in the original features, improving the efficiency and accuracy of subsequent anomaly detection.

[0029] The distance metric value of the low-dimensional feature representation and the normal operation feature is calculated by an anomaly detection algorithm. A variety of statistical and machine learning methods are adopted. The normal operation feature is obtained by analyzing a large amount of historical normal operation data, and is usually represented in the form of a feature vector to represent the typical mode of normal operation. The distance metric can be calculated by selecting the Euclidean distance, Mahalanobis distance or other suitable distance indicators. These indicators can quantify the deviation between the current operation state and the normal mode. The distance value calculation takes into account the weights and correlations of different feature dimensions. Important features are given higher weights, so that the distance metric can better reflect the actual abnormal situation. A time sliding window mechanism is also introduced in the calculation process, which not only considers the current feature value, but also takes into account the trend of recent historical data, thereby more comprehensively evaluating the operation state.

[0030] The abnormal operation state identifier of the water conservancy project is generated based on the comparison result of the distance metric value and the preset threshold value. A scientific threshold setting mechanism needs to be established. The selection of the threshold value directly affects the sensitivity and specificity of the anomaly detection. It needs to be optimized and adjusted according to historical data and engineering actual situation. When the distance metric value exceeds the preset threshold value, the system determines that the current operation state is abnormal, and determines the severity of the anomaly according to the distance value. The abnormality level is usually divided into several levels, from slight abnormality to severe abnormality. The identification of abnormal type is achieved by analyzing the contribution of each dimension in the feature vector, finding out the main feature dimension that causes the distance to increase, and determining whether the abnormality is in water level, flow rate or gate opening. The determination of abnormal position information combines the spatial information in the feature matrix, and accurately locates the specific area or equipment where the anomaly occurs by analyzing the abnormality degree distribution of different monitoring points.

[0031] The implementation of the anomaly detection model focuses on real-time and adaptability. The model can process continuous input monitoring data online and timely issue an abnormal alarm to provide decision support for engineering operation personnel. The model also has self-learning ability, which can continuously update and optimize detection parameters according to new abnormal cases, and improve the identification ability of future abnormal situations. The detection process takes into account the seasonal and periodic characteristics of water conservancy project operation, and adopts appropriate detection strategies for different working conditions to avoid misjudgment of normal working conditions as abnormal. The abnormal detection result is output in a structured form, including detailed information such as abnormal type, level and position, which provides accurate and reliable input for subsequent control instruction generation. The entire anomaly detection process emphasizes reliability and robustness, and adopts various technical means to reduce the risk of false positives and false negatives, including setting multiple detection thresholds, introducing a voting mechanism, etc. The detection system also has a fault self-diagnosis function, which can identify sensor faults or data transmission interruptions and other external problems, avoiding misjudgment of these technical problems as operation abnormalities. The anomaly detection model is closely integrated with the upstream feature extraction module and the downstream control instruction generation module, forming a complete monitoring closed loop to ensure the safe and stable operation of the water conservancy project.

[0032] In the embodiment 3, the network structure is constructed according to the dimension of the input feature matrix, the number of nodes in the hidden layer is generally set to decrease layer by layer, and the number of nodes in each layer is determined through multiple experiments to ensure that the network has sufficient expression ability and does not have overfitting phenomenon. The learning rate parameter is set by using an adaptive adjustment strategy, that is, a larger value is used in the initial training to accelerate convergence, and the value is gradually reduced in the later period to improve the accuracy of parameter optimization. The introduction of the momentum parameter helps to smooth the training process and avoid falling into a local optimal solution. When the feature matrix of the water conservancy project operation state is input into the first layer of the restricted Boltzmann machine for feature coding processing, the machine realizes nonlinear transformation of data through the bidirectional connection between the visible layer and the hidden layer. The visible layer node receives the input feature vector, calculates the activation probability of the hidden layer node through the weight matrix and the bias parameter, and then the hidden layer node reconstructs the visible layer output. The whole process learns the probability distribution of the input data by minimizing the energy function. The activation function used in the feature coding process is usually the sigmoid function, which can convert continuous values into probability representation, facilitating subsequent sampling and reconstruction calculation.

[0033] The network parameters are optimized by using a greedy hierarchical pre-training strategy through layer-by-layer training. Each layer of the restricted Boltzmann machine is independently pre-trained unsupervised, and the distribution characteristics of the input data are learned and output as the input of the next layer. During the training process, the contrast divergence algorithm is used to quickly approximate the maximum likelihood estimation, and the network parameters are updated through limited-step Gibbs sampling, which greatly improves the training efficiency. The update of the network parameters follows the gradient descent principle, and the adjustment amount of the weight matrix and the bias vector is determined by the reconstruction error and the learning rate. After multiple iterations, the network gradually converges to a stable state. After generating a stable deep belief network model, overall fine-tuning training is still needed. The back propagation algorithm is used to supervise the fine adjustment of all layer parameters to further improve the quality of feature representation. When the trained deep belief network model is used to perform nonlinear transformation on the input features, the original high-dimensional feature vector is transmitted through multiple layers of the network, each layer performs linear transformation and nonlinear activation on it, and more abstract and essential feature representations are gradually extracted. The final output of the low-dimensional feature representation retains the most distinctive information in the original data and eliminates redundancy and noise. The feature dimension is usually reduced to one-tenth to one-fifth of the original dimension. The low-dimensional feature representation not only reduces the computational overhead of subsequent calculations, but also improves the accuracy of anomaly detection because the irrelevant feature dimensions are removed, reducing the complexity of data distribution.

[0034] The training quality of DBN directly affects the effect of feature dimension reduction, so various measures are needed to ensure the stability of the training process. The preprocessing of training data includes standardization and denoising to ensure that the input data meets the requirements of the network. The changes in reconstruction error and likelihood function value are monitored during the training process, and the hyperparameters are adjusted in time to prevent overfitting. Regularization techniques such as Dropout are used to improve the generalization ability of the model. The choice of network depth needs to balance the representation ability and training difficulty. Too deep a network may cause gradient vanishing problems, while too shallow a network may not be able to fully extract the hierarchical features of the data.

[0035] The form of the energy function in the DBN model is:

[0036] Where: represents the state vector of the visible layer unit, is the th unit of the visible layer, represents the state vector of the hidden layer unit, is the th unit of the hidden layer, is the bias parameter of the visible layer unit , is the bias parameter of the hidden layer unit , is the connection weight between the visible layer unit and the hidden layer unit , represents the total number of visible layer units, represents the total number of hidden layer units.

[0037] The implementation of feature dimension reduction process takes into account the special nature of water conservancy data. The monitoring data has obvious time correlation and spatial correlation, so the network design introduces corresponding mechanisms to maintain these characteristics. Recurrent connections are used in the time dimension to capture long-term dependencies, and convolution operations are used in the spatial dimension to maintain local feature structures, so that the reduced features still contain rich spatio-temporal information. The dimension reduction result not only serves the current anomaly detection task, but also can be used for other data analysis purposes such as running state clustering, trend prediction, etc., providing multi-faceted support for the intelligent management of water conservancy projects. The entire feature dimension reduction module adopts modular design, allowing flexible adjustment of network structure and parameters according to different engineering needs, while providing interfaces for real-time processing and historical data analysis. By continuously collecting new running data and periodically retraining the network, the feature dimension reduction model can adapt to changes in engineering operating conditions, maintaining long-term effectiveness and accuracy. The reduced feature representation maintains an interpretable correspondence with the original features, making it easy for engineers to understand and analyze the results and make correct decisions.

[0038] In Example 4, assume that the monitoring system of a water conservancy project detects an abnormal rise in water level at monitoring point S12, the abnormal type is determined to be water level overrun, the abnormal level is secondary (medium severity), and the abnormal position is accurate to the S12 monitoring point and the 500-meter area downstream. The system first matches the preset control strategy template according to the abnormal type, retrieves the basic control scheme corresponding to the water level overrun type from the knowledge base, and the scheme includes basic operation parameters in three dimensions of water level regulation, flow rate control, and gate opening adjustment. The matching process of the control strategy template uses a rule-based retrieval mechanism, the water level abnormal type corresponds to the control logic, the gate opening adjustment value in the basic scheme is an increase of 5%, the flow rate control target is a decrease of 0.2 m / s, and the water level regulation target is a decrease of 15 cm. These basic parameters are derived from the historical operation experience and expert knowledge of the project, forming a standardized response measure library. The retrieval process also considers the current overall operation state of the project, including upstream inflow, downstream water demand, and other background factors, to ensure that the extracted basic scheme is consistent with the actual situation.

[0039] The parameter intensity of the basic control scheme based on the abnormal level needs to refer to the preset level-intensity mapping relationship. The project establishes a five-level abnormal level system, and each level corresponds to an intensity adjustment coefficient. The intensity adjustment coefficient of the secondary abnormal level is 1.2, indicating an increase of 20% in the control intensity based on the basic scheme parameter value. In the calculation process, the basic gate opening adjustment value of 5% is adjusted to 6% after intensity adaptation, the flow rate control target is adjusted from 0.2 m / s to 0.24 m / s, and the water level regulation target is adjusted from 15 cm to 18 cm. The determination of the intensity adjustment coefficient considers the nonlinear relationship between the severity of the abnormal level and the control effect. Higher-level abnormalities require a larger proportion of intensity increase. The mapping relationship between the abnormal level and the control intensity is continuously corrected based on historical control effect data. The system records the actual effect of each control action, analyzes the correlation between control intensity and abnormal elimination speed, and optimizes the coefficient values in the mapping relationship table through machine learning algorithms to make the control intensity setting more accurate and effective. The intensity adjustment calculation uses a piecewise linear interpolation method to maintain the smooth transition of the adjustment coefficient within different level intervals, avoiding the sharp jump of control parameters.

[0040] The implementation range of the regulation scheme determined in combination with the abnormal position information needs to analyze the spatial structure characteristics of the water conservancy project. The S12 monitoring point is located in the middle section of the main channel, and within a range of 500 meters downstream of the S12 point, there are three water diversion gates and two drainage outlets. The system calculates the abnormal influence range according to the water flow propagation model, determines that the gates that need to be adjusted include the main control gate upstream of the S12 point and the two water diversion gates downstream of the S12 point. The spatial positioning process utilizes the geographic information system data of the project to construct the correspondence between the regulation operation and the spatial position, and generates targeted regulation instructions. The spatial positioning of the regulation instructions takes into account the time and space characteristics of water flow movement. The adjustment of the upstream gate will affect the downstream point after a certain time delay, so the instruction includes the operation time sequence arrangement. For the water level anomaly of the S12 point, first, adjust the main control gate 200 meters upstream of the S12 point and increase the opening degree by 6%; 10 minutes later, adjust the first water diversion gate downstream of the S12 point and reduce the water diversion amount by 15%; 20 minutes later, adjust the second water diversion gate and adjust the water diversion proportion accordingly. This spatio-temporal coordinated regulation scheme ensures the stable recovery of the water level in the entire affected area.

[0041] The time sequence coordination processing of the water level regulation scheme, the flow velocity control scheme and the gate opening adjustment scheme needs to establish a multi-objective optimization model, considering the mutual influence and time sequence dependence relationship between the regulation measures. Water level regulation is achieved through gate operation, and gate opening change will affect flow velocity, while flow velocity change will in turn affect water level fluctuation. This coupling relationship needs to be properly handled in time sequence arrangement. The coordination algorithm takes the overall stability as the goal to avoid mutual conflict between regulation measures and ensure smooth transition of the project to normal state. The time sequence coordination processing adopts model predictive control method, which predicts the implementation effect of different regulation schemes based on the hydraulic model of the water conservancy project, selects the optimal operation time sequence, and for the detected water level anomaly, the system generates an instruction set containing specific operation time, operation object and operation parameter. The instructions maintain a reasonable time interval, allowing each operation to produce an effect before subsequent adjustments are made. The entire instruction set constitutes a complete regulation process, guiding project personnel or automatic control system to perform corresponding operations.

[0042] Table 1: Abnormal regulation parameter adaptation table

[0043] The generation process of the control instruction set also considers the constraint conditions of engineering operation, including physical constraints such as gate mechanical limit, maximum allowable flow rate, and minimum operating water level. The system checks whether the generated control parameters are within the allowed range, automatically corrects the instructions that exceed the limit, and ensures the safety and feasibility of the control scheme. The output format of the instruction set uses a standardized structure, including fields such as operation device number, operation type, operation parameter, and execution time, which facilitates the parsing and execution of the automatic control system. The entire control instruction generation process emphasizes practicality and reliability, and the generated instructions are not only based on theoretical calculations but also incorporate practical experience from engineering operation. The system provides a manual review and modification interface, allowing experienced operators to fine-tune the automatically generated instructions, fully leveraging the advantages of human-machine collaboration. Through continuous accumulation of control cases and effect data, the instruction generation algorithm is continuously optimized, improving the ability to handle complex abnormal situations.

[0044] Example 5: Suppose that after a water level abnormality control action, the system starts collecting actual operation data after executing the control instructions. These data are obtained in real time through a sensor network arranged at key positions in the project, including water level meters, flow rate meters, and gate opening sensors, etc. The collection frequency of engineering response data is synchronized with the monitoring system, usually set to collect once every minute, and the data collection interval of important positions may be shortened to seconds to ensure that the dynamic response process of the control measures can be captured. The collected data include actual water level change value, actual flow rate change value, and actual gate opening change value. These data form a corresponding relationship with the expected target values in the control instructions. Calculating the deviation measure value of engineering response data and expected control effect requires establishing a scientific and reasonable comparison method. The expected control effect is derived from the theoretical prediction value of the water power calculation model, including expected water level change, expected flow rate change, and expected gate opening change. The deviation calculation is first performed independently at each monitoring point, and the absolute difference and relative difference between the actual value and the expected value are calculated for each monitoring parameter to form a preliminary local deviation evaluation. Taking the water level parameter as an example, if the expected water level at a monitoring point decreases by 15 centimeters and the actual decrease is 13 centimeters, the absolute deviation is 2 centimeters, and the relative deviation is about 13.3%. Similar calculation methods are applied to flow rate and gate opening parameters.

[0045] The calculation of the local deviation value takes into account the influence of measurement errors and random fluctuations, and uses a moving average filtering technique to preprocess the original data, eliminating the interference caused by short-term fluctuations. The difference matrix between the engineering response data and the expected control data needs to take into account the spatial and temporal dimensions. The rows of the difference matrix represent different monitoring point positions, and the columns represent consecutive time points. The matrix elements record the deviation values of each point at each time. The construction of this matrix is based on the geographic information system platform, combining the deviation data of discrete monitoring points with the spatial layout of the project to form a deviation distribution map with spatial continuity. The blank areas between monitoring points are processed through a spatio-temporal interpolation method. The Kriging interpolation algorithm is used to estimate the deviation of unmonitored areas based on the deviation values of adjacent points, generating a deviation distribution map covering the entire project. The deviation distribution map visually displays the spatial variation characteristics of the control effect in the form of a heat map, with color depth representing the deviation size, making it easy to identify hotspots with larger deviations.

[0046] The overall deviation index and the maximum local deviation value are calculated using a multi-index comprehensive evaluation method. The overall deviation index is obtained by weighted averaging of the deviation values of all monitoring points at all time points. The weight coefficients are determined according to the importance and reliability of the monitoring points. The maximum local deviation value is selected from all monitoring points with the largest absolute deviation value, reflecting the most undesirable situation of the control effect. The overall deviation index represents the overall implementation effect of the control measures, and the smaller the value, the closer the actual effect is to the expected target. The maximum local deviation value reveals the uniformity of the control effect, and a large value indicates that there is a significant imbalance in the project response. These two indicators combined can comprehensively evaluate the quality of the control action and provide directional guidance for subsequent optimization. Comparing and analyzing the comprehensive deviation measurement value with historical deviation data requires querying the historical database to retrieve historical deviation records under similar control scenarios, including control types, engineering conditions, hydrological conditions, and other background information. The comparison and analysis focuses on the relative level of this control effect in the historical sequence, determining whether the expected improvement target has been achieved. Historical data provide a benchmark reference to help distinguish whether the deviation is caused by systematic factors or accidental factors. If the deviation value is consistently higher than the historical average, it indicates that there are problems in the control strategy or execution that need to be improved. The analysis results are displayed in the form of a trend chart, which visually presents the long-term trend of the control effect.

[0047] The detection threshold parameters of the anomaly detection model are updated based on the deviation metric value using an adaptive adjustment mechanism. The detection threshold parameters include key parameters such as distance threshold for anomaly determination and confidence level. When the actual regulation effect deviates from the expected target by a small amount, it indicates that the current anomaly detection model has good accuracy, and the detection threshold can be appropriately reduced to improve sensitivity. Conversely, when the deviation is large, the threshold may need to be adjusted to avoid false positives caused by excessive sensitivity. The threshold adjustment follows a gradual principle, with the adjustment amplitude controlled within a reasonable range to avoid drastic fluctuations in the threshold due to a single abnormal situation. The adjusted threshold parameters need to be verified through simulation testing to ensure that the model performance is improved rather than deteriorated under the new threshold setting. The optimization of the generation algorithm for the water conservancy project regulation instruction set focuses on improving parameter calculation logic and timing arrangement strategy. The algorithm optimization identifies weak links based on deviation analysis results. If it is found that the actual water level change after gate regulation is generally less than the expected value, the corresponding relationship coefficient between gate opening and water level change is adjusted, and the expected water level adjustment amount corresponding to unit opening change is appropriately increased. If the flow control effect shows obvious spatial unevenness, the spatial distribution algorithm of the flow control scheme is optimized to enhance the coordination between different parts. The timing arrangement optimization adjusts the execution time and interval of the regulation instructions, and resets the time delay between instructions based on the actual response speed to form a better synergistic effect for multi-step regulation actions.

[0048] The optimization process uses a data-driven method to establish a mapping relationship between regulation parameters and implementation effects through training on a large number of historical cases, continuously correcting model parameters and improving prediction accuracy. The optimization algorithm introduces machine learning technology, allowing the system to learn from successful and failed cases and gradually improve the generation quality of regulation instructions. After each optimization update, the system records the performance comparison data of the new and old algorithms, establishes a version management mechanism, and ensures the traceability and reversibility of the optimization process. Through continuous iterative optimization, the regulation instruction generation algorithm gradually adapts to changes in engineering characteristics, maintaining long-term effectiveness and reliability. The system regularly generates optimization effect evaluation reports to summarize the effects and experiences of recent improvement measures, providing references for subsequent optimization. The optimization process focuses on robustness, with any adjustment of algorithm parameters thoroughly verified to avoid performance fluctuations caused by over-optimization. Through this continuous improvement mechanism, the water conservancy project monitoring system can continuously adapt to changes in operating conditions, improve the accuracy and efficiency of anomaly handling, and provide reliable protection for the safe and stable operation of the project.

[0049] It has to be noted that, in the present document, the terms "first", "second", etc. merely serve to identify different entities or actions from each other in a relative way and do not necessarily require or imply any actual relationship or order between these entities or actions, and the term "comprises", "comprising", or any other variant is intended to cover their non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0050] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. The scope of the application is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring of hydraulic structures, characterized in that, The method comprises the following steps: obtaining multi-source monitoring data of a water conservancy monitoring point, the multi-source monitoring data comprising a water level fluctuation sequence, a flow rate change sequence and a gate opening change sequence; performing spatio-temporal feature extraction processing on the multi-source monitoring data to generate a water conservancy operation state feature matrix, the water conservancy operation state feature matrix containing spatio-temporal correlation data of water level fluctuation features, flow rate change features and gate opening change features; inputting the water conservancy operation state feature matrix into an anomaly detection model to generate a water conservancy abnormal operation state identifier, the abnormal operation state identifier containing abnormal type, abnormal level and abnormal location information; generating a water conservancy regulation instruction set based on the abnormal operation state identifier, the regulation instruction set containing a water level regulation scheme, a flow rate control scheme and a gate opening adjustment scheme.

2. The method for intelligent monitoring of hydraulic engineering according to claim 1, characterized in that, The spatio-temporal feature extraction processing on the multi-source monitoring data comprises: establishing a spatio-temporal correlation matrix of the water level fluctuation sequence and the flow rate change sequence, and calculating a transfer coefficient of the influence of water level change on flow rate; constructing a response function of the gate opening change sequence and the water level fluctuation sequence, and determining a lag time parameter of the influence of gate regulation on water level change; extracting multi-scale features of the water level fluctuation sequence through a spatio-temporal convolutional neural network to generate a water level fluctuation feature vector; fusing local features and global features of the flow rate change sequence through an attention mechanism to generate a flow rate change feature vector; performing spatio-temporal alignment processing on the water level fluctuation feature vector, the flow rate change feature vector and the gate opening change feature to generate a water conservancy operation state feature matrix.

3. The method for intelligent monitoring of hydraulic engineering according to claim 2, characterized in that, The establishment of the spatio-temporal correlation matrix of the water level fluctuation sequence and the flow rate change sequence comprises: dividing a time window of the water level fluctuation sequence, and calculating a water level change rate in each time window; establishing a regression model of the water level change rate and the flow rate change rate in the corresponding time window to determine the transfer coefficient of the influence of water level change on flow rate; constructing a water level-flow rate influence weight matrix according to the spatial position relationship of the monitoring points to calculate the influence weight of the upstream monitoring points on the downstream monitoring points; performing product operation on the transfer coefficient and the influence weight matrix to generate the spatio-temporal correlation matrix of the water level fluctuation sequence and the flow rate change sequence.

4. The method for intelligent monitoring of hydraulic engineering according to claim 1, characterized in that, The inputting of the water conservancy operation state feature matrix into the anomaly detection model comprises: constructing a sample library containing historical abnormal cases, and extracting feature patterns of the abnormal cases; performing feature dimension reduction processing on the water conservancy operation state feature matrix through a deep belief network to generate a low-dimensional feature representation; calculating distance measurement values of the low-dimensional feature representation and normal operation features through an anomaly detection algorithm; generating a water conservancy abnormal operation state identifier based on the comparison result of the distance measurement values and a preset threshold.

5. The method for intelligent monitoring of hydraulic engineering according to claim 4, characterized in that, The feature dimension reduction processing on the water conservancy operation state feature matrix through the deep belief network comprises: constructing a multi-layer restricted Boltzmann machine network structure, and setting the number of hidden layer nodes and the learning rate parameter; inputting the water conservancy operation state feature matrix into a first layer of the restricted Boltzmann machine for feature encoding processing; optimizing network parameters through layer-by-layer training to generate a stable deep belief network model; The trained deep belief network model is used to perform nonlinear transformation on the input features to output low-dimensional feature representations.

6. The method for intelligent monitoring of hydraulic engineering according to claim 1, characterized in that, The water conservancy project control instruction set is generated based on the abnormal operation state identification, including: According to the abnormal type, a preset control strategy template is matched to obtain a basic control scheme; Based on the abnormal level, the parameter intensity of the basic control scheme is adjusted to generate an intensity-adapted control scheme; Combined with the abnormal position information, the implementation range of the control scheme is determined to generate a spatially oriented control instruction; The water level regulation scheme, flow rate control scheme and gate opening adjustment scheme are time-coordinated to generate a water conservancy project control instruction set.

7. The method for intelligent monitoring of hydraulic engineering according to claim 6, characterized in that, The parameter intensity of the basic control scheme is adjusted based on the abnormal level, including: A mapping relationship table between abnormal level and control intensity is established to determine the intensity adjustment coefficient corresponding to each level; The product of the basic control scheme parameter value and the intensity adjustment coefficient is calculated to generate an intensity-adapted control parameter; According to the historical control effect data, the intensity adjustment coefficient is corrected to optimize the calculation accuracy of the control parameter; The control scheme containing the intensity-adapted parameter is output. 8.The water conservancy project intelligent monitoring method according to claim 1, characterized in that, Also includes: Collecting project response data after executing the control instruction, the project response data including actual water level changes, actual flow rate changes and actual gate opening changes; Calculating the deviation measure value of the project response data and the expected control effect; Updating the detection threshold parameter of the abnormal detection model based on the deviation measure value; Optimizing the generation algorithm of the water conservancy project control instruction set.

9. The method for intelligent monitoring of hydraulic engineering according to claim 8, characterized in that, The deviation measure value of the project response data and the expected control effect is calculated, including: A difference matrix of the project response data and the expected control data is established to calculate the local deviation value of each monitoring point; A deviation distribution map of the whole region is generated by a spatio-temporal interpolation method to determine the deviation hot spot area; The overall deviation index and the maximum local deviation value are calculated to generate a comprehensive deviation measure value; The comprehensive deviation measure value is compared and analyzed with historical deviation data to evaluate the control effect.

10. A hydraulic engineering intelligent monitoring system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, realizes the steps of the water conservancy project intelligent monitoring method according to any one of claims 1 to 9.

Images