Hydraulic engineering operation risk monitoring management system based on multi-sensor fusion

Abstract

The invention relates to the technical field of hydraulic engineering monitoring, and discloses a hydraulic engineering operation risk monitoring management system based on multi-sensor fusion. The system comprises a data acquisition module, a data fusion module, an early warning and instruction generation module, and an instruction distribution and verification and correction module. The data acquisition module acquires structural deformation, water flow pressure and environment temperature and humidity data in real time through multiple types of intelligent sensors; the data fusion module inputs the collected data into a dynamic weighted fusion model, and generates a comprehensive monitoring index in combination with a sensor confidence coefficient weight and a historical data deviation rate; the early warning and instruction generation module triggers multi-stage early warning and generates an equipment regulation and control instruction set according to the comprehensive monitoring index and a preset safety threshold value; the instruction distribution module distributes the instruction to an execution terminal through a distributed message queue; and the verification and correction module checks terminal feedback signals by using a back propagation verification mechanism, and corrects the confidence coefficient weight of the model.

Description

Technical Field

[0001] This invention relates to the field of water conservancy project monitoring technology, specifically to a water conservancy project operation risk monitoring and management system based on multi-sensor fusion. Background Technology

[0002] As a vital infrastructure for national economic and social development, water conservancy projects directly impact flood control, water resource allocation, and regional economic stability. With the expansion of construction scale and the increase in operational years, factors such as structural aging, changes in hydrological conditions, and frequent extreme weather events can all trigger operational risks such as structural deformation and abnormal water flow. Failure to monitor and address these risks in a timely manner could lead to project malfunctions, water waste, or even safety accidents.

[0003] Currently, risk monitoring in water conservancy projects largely relies on single-type sensors or decentralized monitoring equipment, which has significant limitations. A single sensor can only collect a specific type of data, such as monitoring only structural deformation or only water flow pressure, and cannot comprehensively capture multi-dimensional status information during project operation. For example, some monitoring systems only deploy displacement sensors to monitor dam deformation, but ignore the impact of environmental temperature and humidity changes on dam material properties, as well as the correlation between water flow pressure fluctuations and deformation. This results in monitoring data that cannot reflect the overall status of project operation, easily leading to misjudgments or omissions in risk assessment.

[0004] Existing monitoring systems often rely on simple aggregation or independent analysis for data processing, failing to consider the performance differences and data reliability of various sensors. Because sensors may be affected by environmental interference and equipment aging during long-term operation, the accuracy of their collected data can fluctuate. Some systems treat all sensor data equally without differentiating them based on the actual confidence level of each sensor, and also fail to correct current data by incorporating historical data deviation rates. This results in low accuracy of the fused monitoring results, making it unreliable for risk assessment.

[0005] The early warning and control mechanisms of existing monitoring systems lack dynamic adaptability. Most systems use fixed safety thresholds to trigger early warnings, failing to consider the dynamic changes in the engineering operation status. For example, the engineering's tolerance threshold to water flow pressure differs between flood season and non-flood season; fixed thresholds can easily lead to overly sensitive or delayed early warnings. Furthermore, some systems, after generating control commands, can only distribute them to the execution terminals via simple communication methods, lacking efficient management of the command transmission process. This can easily lead to command delays and loss, affecting the timeliness of risk response. More importantly, existing systems lack effective feedback correction mechanisms. The operational status feedback signals from the execution terminals are not used to optimize the monitoring model, preventing the system from adjusting monitoring strategies according to actual operating conditions. Over long-term operation, monitoring accuracy and control effectiveness will gradually decline. Summary of the Invention

[0006] The purpose of this invention is to provide a water conservancy project operation risk monitoring and management system based on multi-sensor fusion, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a water conservancy project operation risk monitoring and management system based on multi-sensor fusion, the system comprising: The data acquisition module is used to collect structural deformation data, water flow pressure data, and environmental temperature and humidity data in real time by deploying multiple types of intelligent sensors; The data fusion module is used to input the structural deformation data, water flow pressure data and environmental temperature and humidity data into a dynamic weighted fusion model, and generate a fused comprehensive monitoring index based on the sensor confidence weight and historical data deviation rate. The early warning and instruction generation module is used to trigger a multi-level early warning protocol and generate a corresponding set of equipment control instructions based on the comprehensive monitoring indicators and preset safety thresholds. The instruction distribution module is used to distribute the device control instruction set to each execution terminal through a distributed message queue; The verification and correction module is used to check the changes in the feedback signal parameters of the execution terminal using a backpropagation verification mechanism, and to correct the confidence weights of the dynamic weighted fusion model.

[0008] Preferably, the step of triggering a multi-level early warning protocol and generating a corresponding set of equipment control instructions based on the comprehensive monitoring indicators and preset safety thresholds includes: Based on the dynamic comparison results of comprehensive monitoring indicators and preset safety thresholds, risk level intervals are divided and abnormal data segments are marked. The time series segmentation algorithm is used to decompose the abnormal data segments into trend, periodic and residual components, and the amplitude-frequency characteristics of each component are extracted as risk evolution parameters. A three-dimensional state space model is established based on risk evolution parameters, and the Mahalanobis distance from the current operating state point to the boundary of each risk level interval is calculated through projection mapping. When the Mahalanobis distance exceeds the adaptive warning threshold, a multi-level early warning protocol is triggered and a corresponding set of equipment control instructions is generated.

[0009] Preferably, the construction process of the dynamic weighted fusion model includes: Calculate the inverse variance of the data collected by each type of sensor as the initial confidence weight; A sliding time window is introduced to statistically analyze the deviation rate between historical data of each sensor and the industry standard database; Grey relational analysis was used to quantify the influence coefficient of the deviation rate on the confidence weight. The initial confidence weights and influence coefficients are integrated using the weighted least squares method to generate a dynamically updated confidence weight matrix.

[0010] Preferably, the method for dividing the risk level range includes: The basic safety threshold range shall be determined according to the design specifications of water conservancy projects; The extreme points and inflection points of the integrated monitoring indicators after fusion are extracted as candidate points for the interval boundary; Density clustering algorithm is used to identify the clustering characteristics of candidate points in the time dimension; By combining the basic safety threshold range with clustering characteristics, a three-level risk level range with flexible boundaries is generated.

[0011] Preferably, the execution process of the time series segmentation algorithm includes: Empirical mode decomposition is performed on abnormal data segments to obtain intrinsic mode function components; The instantaneous frequency and amplitude of each component are calculated using the Hilbert transform; Components with amplitudes exceeding the noise threshold are screened for reconstruction, separating the trend term, periodic term, and residual term. Fourier transform is performed on the three reconstructed terms to extract the amplitude-frequency feature vectors.

[0012] Preferably, the method for establishing the three-dimensional state space model includes: The state space is constructed using the trend term amplitude, period term frequency, and residual term entropy as coordinate axes. Calculate the cluster centers of historical normal data in the state space; The polyhedral boundaries of each risk level interval in the state space are generated using the convex hull algorithm. The relative position of the current operating point in the polyhedron is determined by ray casting.

[0013] Preferably, the calculation process of the Mahalanobis distance includes: Calculate the covariance matrix of historical data points within each risk level interval; Singular value decomposition of the covariance matrix yields the eigenvector basis; Project the current running state point onto a new coordinate system formed by the eigenvector basis; The standardized Euclidean distance between the projection point and the centroid of the target interval is calculated as the Mahalanobis distance.

[0014] Preferably, the triggering conditions for the multi-level early warning protocol include: When the Mahal distance exceeds the Level 1 warning line, initiate the data verification process and mark the suspicious sensor; When the distance to the Mahalanobis exceeds the level 2 warning line, a device inspection command is sent to the operation and maintenance terminal. When the distance to the Mach number exceeds the Level 3 warning line, the backup control system is automatically activated and an emergency plan is generated.

[0015] Preferably, the step of using the backpropagation verification mechanism to check the changes in the feedback signal parameters of the execution terminal and correcting the confidence weights of the dynamic weighted fusion model includes: Receive real-time status signals from each execution terminal after executing the device control instruction set; The changes in key parameters are extracted from the real-time status signal and compared with the preset expected parameter range to generate parameter deviation values. Based on the parameter deviation values, the adjustment amount of the confidence weight of each sensor in the dynamic weighted fusion model is calculated by the backpropagation algorithm; The confidence weights of the dynamic weighted fusion model are updated based on the adjustment amount.

[0016] Preferably, the distributed message queue used in the instruction distribution module is a persistent message queue, and its message distribution mechanism includes: Each device control command is assigned a unique serial number and timestamp and stored in a high-availability storage cluster; Instructions are dynamically allocated based on the processing capacity and current load of each execution terminal using a weighted round-robin algorithm. If no confirmation receipt is received from the execution terminal within the preset timeout period, the instruction retransmission mechanism is triggered based on the sequence number and timestamp.

[0017] Compared with the prior art, the beneficial effects of the present invention are: From a data acquisition perspective, the system's data acquisition module employs multiple types of intelligent sensors, enabling simultaneous real-time collection of structural deformation data, water flow pressure data, and environmental temperature and humidity data. This overcomes the limitations of traditional monitoring methods where a single sensor can only acquire information in one dimension. The simultaneous acquisition of multiple data types comprehensively captures various state parameters during the operation of the hydraulic engineering project, covering the deformation of the engineering structure itself, the pressure exerted by water flow on the project, and the impact of external environmental temperature and humidity on the engineering materials and operational status. This allows the system to have a more comprehensive perception of the project's operational status, avoiding biased risk assessments due to missing data dimensions, and ensuring that subsequent data-driven analysis and decision-making are more aligned with the actual operational situation of the project.

[0018] In the data processing stage, the data fusion module introduces a dynamic weighted fusion model. Instead of simply summarizing the collected data, it combines sensor confidence weights with historical data deviation rates to generate comprehensive monitoring indicators. Different sensors exhibit varying data reliability due to factors such as equipment performance, installation location, and environmental interference. By assigning different confidence weights, the data fusion process prioritizes information from high-reliability sensors, reducing the interference of low-confidence data on the fusion results. Simultaneously, correcting current data using historical data deviation rates reduces errors caused by aging and drift during long-term sensor operation. This results in more accurate comprehensive monitoring indicators that better reflect the actual operating status of water conservancy projects, providing a reliable data foundation for subsequent early warning and control.

[0019] In terms of early warning and control, the early warning and command generation module triggers a multi-level early warning protocol based on comprehensive monitoring indicators and preset safety thresholds, and generates corresponding equipment control command sets. Compared with the traditional fixed threshold early warning method, the multi-level early warning protocol can divide different early warning levels according to the degree of deviation between comprehensive monitoring indicators and safety thresholds. Different levels correspond to different response strategies, which can avoid early warning lag or oversensitivity caused by a single threshold, and enable managers and equipment implementers to clearly understand the urgency of the current risk and take targeted measures quickly. At the same time, the generated equipment control command set is directly matched with the early warning level, ensuring the rationality and effectiveness of control measures and reducing resource waste or improper handling caused by mismatch between commands and risks.

[0020] The instruction distribution module employs a distributed message queue for instruction distribution. This distributed architecture effectively improves the efficiency and stability of instruction transmission. Traditional instruction distribution methods are susceptible to network congestion and single points of failure, leading to instruction delays or loss. In contrast, the distributed message queue stores and distributes instructions in an orderly manner. Even if some nodes fail, other nodes can continue to complete instruction transmission, ensuring that instructions are delivered to each execution terminal in a timely and accurate manner. This guarantees that control measures can be implemented quickly and avoids delays in risk management due to instruction transmission problems.

[0021] The verification and correction module uses a backpropagation verification mechanism to check the changes in the feedback signal parameters from the execution terminal and corrects the confidence weights of the dynamically weighted fusion model, enabling the system to continuously self-optimize. The feedback signals from the execution terminal reflect the actual execution effect of control commands and changes in the engineering operation status. By analyzing the changes in the parameters of the feedback signals, it can be determined whether the current model's confidence weights are reasonable. If a large deviation is found between the data from a certain type of sensor and the actual operating status, its confidence weight can be adjusted through backpropagation, making the subsequent data fusion process more closely match the current engineering operation characteristics. Simultaneously, as the engineering operation time increases and the external environment changes, the model can continuously adapt to new operating conditions through continuous correction, avoiding a decrease in monitoring accuracy due to model rigidity and ensuring the reliability and adaptability of the system during long-term operation. Attached Figure Description

[0022] Figure 1 This is a timing diagram of the water conservancy project operation risk monitoring and management system based on multi-sensor fusion as described in this invention; Figure 2 A flowchart illustrating the process of building a dynamic weighted fusion model; Figure 3 A flowchart illustrating the process of establishing a three-dimensional state-space model. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1This invention provides a water conservancy project operation risk monitoring and management system based on multi-sensor fusion. The system includes: a system that integrates multiple functional modules to achieve real-time monitoring and management of water conservancy project operation risks. A data acquisition module deploys multiple types of intelligent sensors, including structural deformation sensors, water flow pressure sensors, and environmental temperature and humidity sensors. These sensors continuously collect raw data at a fixed sampling frequency. A data fusion module receives structural deformation data, water flow pressure data, and environmental temperature and humidity data transmitted from the data acquisition module. It processes the data using a dynamic weighted fusion model, which calculates the fused comprehensive monitoring index based on sensor confidence weights and historical data deviation rates. An early warning and instruction generation module compares the comprehensive monitoring index with preset safety thresholds. Based on the comparison results, it triggers a multi-level early warning protocol and generates a set of equipment control instructions. An instruction distribution module sends the equipment control instruction set to each execution terminal through a distributed message queue, ensuring reliable instruction transmission. A verification and correction module monitors the feedback signals from the execution terminals, analyzes parameter changes using a backpropagation verification mechanism, and dynamically adjusts the confidence weights of the dynamic weighted fusion model.

[0025] Example 1: See Figure 2The dynamic weighted fusion model construction process in the data fusion module involves multiple steps, including sensor data variance calculation, historical data deviation rate statistics, grey relational analysis, and weighted least squares fusion. The inverse variance of the data collected by each type of sensor is calculated as the initial confidence weight. Variance calculation is based on the data sequence within a fixed-length sliding time window. The size of this window is set according to the specific characteristics of the water conservancy project and monitoring requirements. For example, a larger window can be used for slowly changing data such as structural deformation to smooth short-term fluctuations, while a smaller window is used for potentially rapidly changing data such as water flow pressure to preserve details. Another sliding time window is introduced to statistically analyze the deviation rate between the historical data of each sensor and the industry standard database. The industry standard database stores historical benchmark data sequences of verified water conservancy projects under safe operating conditions. The deviation rate is calculated by comparing the Euclidean distance between the current sensor data sequence and the benchmark data sequence of the industry standard database at the corresponding time point. The magnitude of the Euclidean distance directly reflects the degree to which the current data deviates from the standard. Grey relational analysis is used to quantify the influence coefficient of the deviation rate on the initial confidence weight. Grey relational analysis constructs an ideal data sequence as a reference and calculates the correlation coefficient between the actual data sequence of each sensor and this ideal sequence at different time points. The average of these correlation coefficients is ultimately quantified as an influence coefficient between zero and one. The higher the correlation, the closer the influence coefficient is to zero, indicating a smaller negative impact of the deviation rate on the weight. The initial confidence weight and the deviation rate influence coefficient are then fused using weighted least squares. Weighted least squares aims to minimize the sum of squared errors between the fused comprehensive monitoring index and a theoretical true value estimate, iteratively solving for the optimal dynamic weight combination. This generates a confidence weight matrix that can be dynamically updated over time. This matrix serves as the core parameter of the dynamic weighted fusion model and continuously affects the real-time incoming sensor data.

[0026] The early warning and instruction generation module dynamically compares the comprehensive monitoring indicators output by the dynamic weighted fusion model with preset safety thresholds. The comparison results are used to divide risk level intervals and mark abnormal data segments exceeding the thresholds. The division of risk level intervals is a dynamic process. The data stream of the comprehensive monitoring indicators is scanned in real time, and its extreme points and inflection points are extracted using mathematical algorithms as candidate points for interval boundaries. Identification of extreme points relies on finding points where the first derivative of the data is zero and the sign of the second derivative changes. Identification of inflection points focuses on moments when the data curvature changes significantly. A time series segmentation algorithm is used to decompose the marked abnormal data segments into three components: a trend term, a periodic term, and a residual term. The time series segmentation algorithm uses adaptive methods such as Empirical Mode Decomposition (EMD) to handle non-stationary monitoring data. EMD decomposes complex signals into a finite number of intrinsic mode function (IMF) components, which contain information from different scales of the original signal, from high frequency to low frequency. The instantaneous frequency and amplitude of each intrinsic mode function component are calculated using Hilbert transform. The Hilbert transform converts each component into an analytic signal, allowing for precise extraction of its instantaneous characteristics. Components with amplitudes exceeding a preset noise threshold are then selected for reconstruction. The noise threshold is dynamically set based on the background noise level of the sensor under normal conditions. The reconstruction process merges the selected effective components according to specific rules, ultimately separating a trend term representing the long-term direction of change, a periodic term reflecting periodic patterns, and a residual term containing random fluctuations and noise. Fourier transforms are then performed on the reconstructed trend, periodic, and residual terms to extract their amplitude-frequency characteristic vectors. The Fourier transform converts the time-domain signal to the frequency domain, revealing the frequency composition and corresponding energy distribution of each component. These amplitude-frequency characteristic vectors, such as the average amplitude of the trend term, the dominant frequency of the periodic term, and the spectral entropy of the residual term, are collectively referred to as risk evolution parameters.

[0027] A three-dimensional state-space model is established based on the extracted risk evolution parameters. The three coordinate axes of the three-dimensional state-space model are composed of the trend term amplitude, the periodic term frequency, and the residual term entropy value, respectively. The position of each data point in the state space intuitively reflects the operational characteristics of the water conservancy project system at a specific moment. The Mahalanobis distance from the current operational state point to the predefined boundaries of each risk level interval is calculated using the projection mapping method. Projection mapping first requires determining a projection plane or direction of the state space. The calculation of Mahalanobis distance depends on the covariance matrix of historical data points in each risk level interval. Unlike Euclidean distance, it considers the correlation between different dimensions of the data and the dispersion of the data itself, making the distance metric scale-independent and more consistent with the actual distribution of the data. When the calculated Mahalanobis distance exceeds the system's adaptive warning threshold, the warning and command generation module will trigger the corresponding multi-level warning protocol. The adaptive warning threshold is not fixed but dynamically adjusted based on the accuracy of historical warnings and the system's false alarm rate to balance sensitivity and reliability. The multi-level warning protocol is divided into different levels according to the severity of the Mahalanobis distance exceeding the limit. Each warning level is associated with a set of preset equipment control commands. The command set includes the identifier of the execution terminal to be operated, the specific control command parameters, and the execution priority order.

[0028] The confidence weight update process of the dynamic weighted fusion model is closely linked to and works collaboratively with the early warning triggering process. Sensor data streams from the data acquisition module are continuously input into the data fusion module. The dynamic weighted fusion model uses the latest confidence weight matrix to fuse these multi-source data and outputs comprehensive monitoring indicators to the early warning and command generation module. The early warning and command generation module continuously monitors and analyzes the incoming comprehensive monitoring indicators. The risk level interval boundaries within the module are not static but are flexibly adjusted based on the statistical characteristics of recent data. This flexible mechanism allows the system to adapt to the long-term, slow changes in the operating status of water conservancy projects and seasonal impacts. The digital filtering technology used in the time series segmentation algorithm to process abnormal data segments can effectively separate high-frequency details and low-frequency contours in the signal. The combination of empirical mode decomposition and Hilbert transform ensures the adaptability of the signal decomposition process and the accuracy of feature extraction, providing a high-quality data foundation for subsequent risk evolution parameter calculations. The construction of a three-dimensional state space model provides a visual and quantifiable analytical framework for multi-dimensional monitoring data. Each point in the state space uniquely corresponds to a system operating mode. The convex hull algorithm is used to generate the minimum convex polyhedron boundary that encloses each risk level data point, thus clearly defining different areas such as safe zones, warning zones, and danger zones in the state space. The projection mapping process uses a linear transformation to map the current state point to a new coordinate system composed of the principal component eigenvectors of historical data. The Mahalanobis distance is calculated in this new coordinate system, and its value directly reflects the statistical significance of the current state deviating from the normal or safe zone. The introduction of an adaptive warning line enables the system to self-optimize based on actual operating performance, avoiding excessive false alarms or missed alarms due to unreasonable threshold settings. The entire process, from data fusion to risk warning, forms a closed loop. The dynamic weighted fusion model provides reliable input for state assessment, while the warning results indirectly verify and provide feedback to adjust the parameters of the fusion model. Through this cyclical and progressive approach, the system continuously improves the accuracy and timeliness of risk monitoring for water conservancy projects.

[0029] Example 2: The method for dividing risk level intervals originates from the design specifications of water conservancy projects. These specifications clearly define the safe threshold ranges for parameters such as structural deformation, water flow pressure, and environmental temperature and humidity. This basic safe threshold range forms the benchmark framework for risk interval division. Extreme points and inflection points of the integrated monitoring indicators in a continuous time series are extracted as candidate interval boundaries. Extreme points are identified by scanning the data sequence to find local maximum and minimum values, while inflection points are located by analyzing the zero-crossing points of the second derivative of the data sequence. These candidate points mark potential locations where significant changes in data behavior occur. A density clustering algorithm is used to identify the clustering characteristics of a large number of candidate points in the time dimension. The density clustering algorithm discovers high-density areas by calculating the spatial distance between candidate points and setting a neighborhood radius. Candidate points located at the core of high-density areas are identified as effective cluster centers, while points distributed in low-density areas are considered noise points and excluded. Combining the basic safety threshold range with the clustering characteristics identified by the density clustering algorithm, the system generates a three-level risk level range with flexible boundaries. Flexible boundaries mean that the upper and lower limits of each risk range are not fixed values, but rather a dynamically changing range formed according to the center position and distribution density of the candidate point clusters. The low-risk range corresponds to the area where the cluster center is close to the historical normal data range, the medium-risk range corresponds to the area where the data begins to deviate but is still within a controllable range, and the high-risk range corresponds to the area where the data deviates severely and the cluster center is close to the edge of the safety threshold.

[0030] The execution process of the time series segmentation algorithm focuses on the abnormal data segments marked by the early warning module. Empirical Mode Decomposition (EMD) is performed on these segments to obtain a set of intrinsic mode function (EMF) components. EMD is an adaptive signal processing method that iteratively decomposes complex, non-stationary original signals into a finite number of EMF components arranged from high to low frequencies. Each EMF component contains local feature information of the original signal at different time scales. The instantaneous frequency and amplitude of each EMF component are calculated using Hilbert Transform (HFT). Hilbert Transform converts each component of a real signal into a complex signal, thus accurately extracting the instantaneous amplitude and instantaneous frequency—two physical quantities that change over time. The instantaneous frequency reflects the rate of oscillation of the signal component, while the instantaneous amplitude reflects the energy strength of that oscillation. EMF components with instantaneous amplitudes exceeding a preset noise threshold are selected for reconstruction. The noise threshold is set based on statistical analysis of the background noise level of historical data under normal operating conditions; only those signal components with significant amplitudes and carrying real physical information are retained for subsequent analysis. The selected effective components are weighted and reconstructed to separate the trend term (representing the long-term direction of data change), the periodic term (reflecting periodic fluctuations), and the residual term (containing random noise and unmodeled dynamics). The trend term reflects the macroscopic trend of the monitoring indicators, the periodic term reveals regular changes related to day / night, tides, or seasonality, and the residual term contains possible sudden anomalies or noise. Fourier transforms are performed on the reconstructed trend, periodic, and residual terms to extract their amplitude-frequency characteristic vectors. The Fourier transform converts each signal in the time domain to the frequency domain, obtaining its amplitude spectrum. From the amplitude spectrum, characteristic parameters such as the amplitude of the dominant frequency, the spectral centroid, and the bandwidth can be extracted. These amplitude-frequency characteristic vectors quantitatively describe the slow change intensity of the trend term, the main oscillation frequency of the periodic term, and the degree of randomness of the residual term.

[0031] The application of risk level interval division and time series segmentation algorithms forms a coherent data processing chain. Comprehensive monitoring indicators are continuously input into the risk assessment module, which internally runs algorithms for detecting extreme points and inflection points in real time. New data points are constantly generated, and the candidate point set is dynamically updated accordingly. Density clustering algorithms periodically re-cluster the updated candidate point set. The clustering results directly drive adjustments to the elastic boundaries of the three-level risk level intervals. This dynamic adjustment mechanism enables the system to adapt to the slow performance drift caused by long-term service, environmental changes, or changes in operational strategies in water conservancy projects. When processing anomalous data segments, the adaptive nature of empirical mode decomposition (EMD) allows it to effectively handle nonlinear and non-stationary engineering data, while Hilbert transform provides accurate time-frequency analysis tools. The component reconstruction process eliminates noise interference, making the separated trend, periodic, and residual terms more physically meaningful. Fourier transform converts time-domain features into frequency-domain features that are easier to quantify and compare. The extracted amplitude-frequency feature vectors, used as risk evolution parameters, provide a solid data foundation for accurately characterizing the system's operating state in three-dimensional state space, calculating Mahalanobis distance, and conducting risk warnings. The entire implementation method embodies a systematic approach from raw data to feature extraction and dynamic risk interval division. The flexible boundary design of the risk level intervals enhances the system's adaptability to complex operating conditions, while the fine decomposition of the time series segmentation algorithm ensures the accuracy and comprehensiveness of risk feature capture.

[0032] Example 3: See Figure 3The method for establishing a three-dimensional state-space model uses three feature quantities output from the time series segmentation algorithm as the basic coordinate axes: the trend term amplitude, the periodic term frequency, and the residual term entropy. The trend term amplitude characterizes the average intensity of long-term changes in the monitored data and is obtained by calculating the root mean square value of the trend term signal over a complete cycle. The periodic term frequency reflects the main rate of periodic fluctuations in the data and is determined by identifying the frequency corresponding to the main peak in the power spectrum of the periodic term. The residual term entropy quantifies the degree of uncertainty or randomness of the residual sequence and is calculated using the Shannon entropy formula to calculate the amplitude distribution of the residual term. A three-dimensional Euclidean space is constructed using these three physically meaningful parameters, and the operating state of the water conservancy project at any given time can be uniquely represented by a point in this space. The cluster centers of historical normal data in the state space are calculated. Historical normal data refers to the long-term accumulated data records of the system under known safe operating conditions. The calculation of the cluster centers is completed using the k-means algorithm, with k set to 1 to find a single center representing the overall normal state distribution. This center point constitutes a reference anchor point in the state space. The convex hull algorithm generates the polyhedral boundaries of each risk level interval in the state space. The algorithm iterates through all historical data points belonging to a specific risk level, finding the smallest convex polygon that can enclose all these points. In three-dimensional space, this smallest convex polygon is a convex polyhedron, whose faces, edges, and vertices collectively define the geometric boundary of that risk level. A ray casting method is used to determine the relative position of the current running state point within the polyhedron. Ray casting emits a ray from the current state point in any direction in space, and the number of intersections between this ray and each face of the target polyhedron is calculated. The parity of the number of intersections determines whether the current point is inside or outside the polyhedron.

[0033] Mahalanobis distance is a multivariate statistical metric. It calculates the covariance matrix of historical data points within each risk level interval. The covariance matrix describes the changes in the trend term magnitude, periodic term frequency, and residual term entropy of data points within the risk level interval, as well as their respective dispersion. Singular value decomposition (SVD) is performed on the covariance matrix to obtain eigenvector bases. SVD decomposes the covariance matrix into the product of three matrices, one of which contains a set of column vectors that form a new set of orthogonal basis vectors. These eigenvector bases indicate the main directions of the data distribution. The current running state point is projected onto a new coordinate system formed by the eigenvector bases. This projection operation is achieved by multiplying the coordinate vector of the current state point with the transformation matrix formed by the eigenvector bases. The projected coordinates represent the components of the original state point in the main data direction. The standardized Euclidean distance between the projected point and the centroid of the target interval is calculated as the Mahalanobis distance. The centroid of the target interval is the projection of the average of all data points within that risk level interval onto the feature space. The standardized Euclidean distance takes into account the variance differences in each main direction during calculation. The formula for Mahalanobis distance is expressed as:

[0034] Where: symbol This represents the scalar value of the Mahalanobis distance to be calculated. (Symbol) It is a three-dimensional column vector representing the coordinates of the current running state point in the original state space, i.e. , of which components Represents the magnitude of the trend term, components Represents the frequency of periodic terms and components. Represents the entropy value of the residual term. Symbol Similarly, it is a three-dimensional column vector representing the centroid coordinates of all historical data points within the target risk level range, i.e. .symbol It is a 3x3 covariance matrix of historical data points within the target risk level range. (Superscript) The superscript represents the matrix transpose operation. This represents the matrix inversion operation. The physical meaning of the formula is to calculate the points... Compared to points Centered on, with The statistical distance describing the distribution pattern, the distance value It is a dimensionless number whose magnitude reflects the degree to which the current state deviates significantly from the typical state of the target interval.

[0035] The three-dimensional state-space model and Mahalanobis distance calculation together constitute the geometric and statistical framework for risk assessment. The state-space model transforms multidimensional data into intuitive spatial points, giving complex operational states a visual representation. Cluster centers provide a baseline location for the normal state, while the polyhedral boundaries generated by the convex hull algorithm transform abstract risk levels into concrete spatial regions. Ray projection can efficiently locate and determine points. The Mahalanobis distance calculation introduces a statistical perspective; the inverse of the covariance matrix acts as a normalization, eliminating the influence of different dimensional units and variance differences, allowing the distance metric to truly reflect the deviation of data from the probability distribution. The projection transformation of the eigenvector basis aligns the data to its eigendirection. The Mahalanobis distance formula ultimately integrates spatial location and distribution characteristics; a larger value indicates that the current operational state deviates more from the normal or expected risk level range, providing a precise and quantitative decision-making basis for triggering multi-level early warning protocols. The entire implementation process embodies a systematic approach from geometric modeling to statistical analysis. The three-dimensional state space model provides physical intuition, while Mahalanobis distance provides statistical rigor. The combination of the two enables an in-depth and objective assessment of the operational risks of water conservancy projects.

[0036] Example 4: The triggering conditions of the multi-level early warning protocol are based on the real-time calculation of Mahalanobis distance and the comparison of dynamic thresholds. Mahalanobis distance is calculated by a three-dimensional state space model and reflects the statistical deviation of the current operating state from the centroid of each risk level interval. The system presets three levels of adaptive warning lines, each corresponding to a Mahalanobis distance threshold. These thresholds are not fixed but are periodically adjusted based on the early warning accuracy and false alarm rate in the system's operating history. When the real-time calculated Mahalanobis distance exceeds the first-level warning line, the system automatically initiates a data verification process and marks the sensors that contributed abnormal data within that time window. The data verification process includes reading the sensor's self-diagnostic status information, comparing redundant sensor readings, and checking the communication quality of the data acquisition link. The confidence weight of the marked suspicious sensors will be temporarily reduced in subsequent data fusion. When the Mahalanobis distance further increases and exceeds the second-level warning line, the early warning and instruction generation module generates specific equipment inspection instructions and immediately sends them to the terminal devices of maintenance personnel through the instruction distribution module. The equipment inspection instructions detail the list of equipment that needs to be verified on-site, the suggested inspection items, and the operation priority. When the Mahalanobis distance continues to rise and exceeds the Level 3 warning line, it indicates that the system considers the risk to be imminent. The warning and instruction generation module will automatically activate the pre-configured backup control system and generate a structured emergency plan based on the current comprehensive monitoring indicators, equipment status database, and emergency plan database. The emergency plan includes specific operating steps, resource allocation suggestions, and personnel division of labor.

[0037] The verification and correction module utilizes a backpropagation verification mechanism to perform a closed-loop evaluation of the execution effect of control commands issued by the early warning and command generation module. The module continuously receives real-time status signals from each execution terminal after executing the equipment control command set. These real-time status signals include the execution terminal's own status parameters and the local environmental parameters it perceives. Key parameter changes are extracted from the real-time status signals. The selection of key parameters is directly related to the target of the previously issued equipment control commands. For example, if the command is to open a floodgate, the key parameters include the gate opening degree, downstream flow rate, and upstream water level. The extracted changes are the differences or rates of change of these parameters after the command execution relative to before execution. The extracted key parameter changes are compared with a preset expected parameter range. This expected parameter range is stored in the system knowledge base and defines the reasonable range of key parameter changes that should exist when the command is correctly executed and the system responds normally. The comparison result generates a quantified parameter deviation value. Based on the parameter deviation values, the adjustment amount of the confidence weights of each sensor in the dynamic weighted fusion model is calculated using the backpropagation algorithm. The backpropagation algorithm treats the parameter deviation values ​​as a loss function, propagating the error in the reverse direction of data processing. It calculates the gradient of the loss function with respect to the confidence weights of each sensor in the dynamic weighted fusion model, and determines the magnitude and direction of the weight adjustment based on the gradient direction and the learning rate. The confidence weights of the dynamic weighted fusion model are updated according to the calculated adjustment amounts. Weight updates are performed incrementally to avoid making overly drastic judgments about sensor reliability due to a single deviation. The updated confidence weights directly affect the contribution of each sensor's data in subsequent data fusion processes, thereby achieving adaptive optimization of the model.

[0038] The triggering of the multi-level early warning protocol and the operation of the backpropagation verification mechanism constitute a complete closed loop of perception-decision-execution-feedback. Mahalanobis distance, as the core indicator of risk quantification, drives the leap in early warning levels through continuous numerical changes. The adaptive warning line design allows the system to balance the sensitivity and stability of early warnings. The data verification process is the first line of defense against potential data source problems, using technical means to eliminate false alarms caused by sensor malfunctions or data interference. The issuance of equipment inspection commands indicates that the risk may have shifted from data anomalies to potential equipment hazards, requiring manual intervention for confirmation and initial handling. The activation of the backup control system and the generation of emergency plans are the system's final automated defense measures, aiming to control the situation to the greatest extent possible. The backpropagation verification mechanism runs through the period after command execution. Like a supervisor, it analyzes the execution effect to infer the accuracy of the previous data perception and fusion stages. The magnitude of the parameter deviation directly reflects the effectiveness of command execution and the accuracy of the initial state judgment. The gradient calculation in the backpropagation algorithm clarifies the responsibility allocation of each sensor's data for the final decision deviation, and the learning rate setting controls the pace of model updates. The entire implementation method embodies the complete intelligent management logic of the system, from risk identification and early warning escalation to control execution, and then to effect evaluation and model correction. The multi-level early warning protocol enables tiered response, and the backpropagation verification mechanism ensures the system's continuous self-learning and improvement capabilities. Refer to Table 1, which shows the correspondence between different Mahalanobis distance intervals and early warning levels, response actions, and verification focus.

[0039] Table 1: Correspondence between Mahalanobis distance and early warning response Mahalanobis distance interval Warning level Main system response actions Key Considerations for the Verification and Correction Module Below the Level 1 Alert Line normal Continuous monitoring, no unusual activity detected. Monitor parameter change trends and accumulate normal state data. Exceeding the Level 1 warning line, but below the Level 2 warning line Level 1 warning Initiate the data review process and flag suspicious sensors. Analyze sensor data consistency and assess historical biases of labeled sensors. Exceeding the Level II warning line, but below the Level III warning line Level II Warning Send device check command to maintenance terminal Extract the equipment status feedback signal and compare the parameters of the equipment under test with the expected range. Exceeding the Level 3 Alert Line Level III Warning Activate the backup control system and generate and execute the emergency plan. Comprehensive monitoring of the overall system response, extraction of key safety parameter changes, and calculation of comprehensive parameter deviation values. Table 1 clearly outlines the different strategies adopted by the system as risks escalate. The Mahalanobis distance interval discretizes the continuous risk level into different warning levels, each corresponding to a specific operating procedure. Under normal conditions, the system maintains silent monitoring, and the verification and correction module uses this stable data to calibrate the model's baseline. At Level 1 warning, the response focuses on verifying the reliability of the data source; the evaluation results of the verification and correction module are directly used to temporarily adjust the input weights of the dynamic weighted fusion model. Level 2 warning signifies an escalation of action; the system begins to collaborate with external maintenance personnel. At this point, the verification and correction module focuses on whether instructions are correctly received and the expected changes in equipment status after manual intervention. Level 3 warning is the highest level of response; the system initiates automated emergency procedures, and the verification and correction module's work shifts to evaluating the effectiveness of the entire emergency response process. The calculated parameter deviation value at this stage will be used to significantly correct the model weights based on the overall reliability of the sensor data in this major event.

[0040] Example 5: The distributed message queue used in the instruction distribution module is a persistent message queue. The core characteristic of a persistent message queue is that all messages are written to the disk of the high-availability storage cluster, ensuring message content is not lost even in the event of system power interruption or node failure. Each device control instruction is appended with a unique sequence number and timestamp. The unique sequence number is assigned by the central sequence number generator of the instruction distribution module, ensuring that the sequence number of each instruction is unique throughout the system's lifecycle. The timestamp is taken from a high-precision time server, accurately recording the UTC time when the instruction enters the instruction distribution module. The complete message with the appended sequence number and timestamp is synchronously written to the high-availability storage cluster. The high-availability storage cluster employs a multi-replica mechanism, where the same message data is stored as copies on multiple physically isolated server nodes. Only after successfully writing a specified number of copies does the instruction distribution module consider message persistence complete. Instructions are dynamically allocated based on the processing capacity and current load of each execution terminal using a weighted round-robin algorithm. This algorithm requires maintaining a list containing information about all execution terminals, with each terminal having a preset static weight value based on its hardware configuration and historical performance benchmark test results. The instruction distribution module periodically sends lightweight heartbeat packets to each execution terminal. Based on the latency of the execution terminal's response to the heartbeat packet and its self-reported current CPU and memory usage, a dynamic load factor is calculated. The static weight value is combined with the dynamic load factor to determine the actual weight of each execution terminal in processing messages during the current polling cycle. Execution terminals with higher weights will obtain more device control instructions from the persistent message queue for processing.

[0041] If no confirmation is received from the execution terminal within the preset timeout period, a command retransmission mechanism is triggered based on the sequence number and timestamp. The preset timeout period is not a fixed value but is dynamically adjusted based on the statistical average of historical communication delays between the command distribution module and the execution terminal. After the command retransmission mechanism is triggered, the command distribution module first checks the heartbeat status of the target execution terminal. After confirming that the execution terminal is online and the network connection is basically smooth, the command distribution module rereads the original command message from the high-availability storage cluster based on the command's sequence number and timestamp. The retransmission process records the number of retries. When the number of retries exceeds the set upper limit threshold and no confirmation is received, the command distribution module marks the execution terminal as "suspected faulty," temporarily sets its weight to zero, suspends the distribution of new commands to it, and sends a device offline alarm message to the operation and maintenance alarm system.

[0042] A concrete example clearly illustrates the workflow of the instruction distribution module. Suppose the early warning and instruction generation module determines that the No. 3 floodgate needs to be opened to lower the reservoir water level and generates an equipment control instruction with the content "Open the No. 3 floodgate to a five-meter opening." This instruction is sent to the instruction distribution module, which attaches a unique serial number to it, in the format "CMD-20241027-085632-001," and a timestamp accurate to milliseconds, "2024-10-27 08:56:32.123UTC." The complete instruction, with these metadata appended, is constructed into a standard message packet. The instruction distribution module synchronously writes this message packet to a high-availability storage cluster consisting of three nodes. After receiving successful write confirmations from at least two nodes, the instruction distribution module considers the message to be securely persisted. At this point, the instruction distribution module needs to decide which execution terminal should execute this instruction. Currently online execution terminals include gate controller A (high static weight, strong processing capacity), gate controller B (medium static weight), and a standby controller C (low static weight). A weighted round-robin algorithm checks their real-time status: controller A reports a CPU utilization of 75%, a heartbeat response latency of 150 milliseconds, and a high dynamic load factor; controller B reports a CPU utilization of 40%, a heartbeat response latency of 50 milliseconds, and a low dynamic load factor; controller C is in idle standby mode. The weighted round-robin algorithm calculates that controller B has the highest actual weight in this cycle. Therefore, the instruction distribution module retrieves the instruction message to open floodgate No. 3 from the persistent message queue and sends it to gate controller B via the network.

[0043] The command distribution module starts a timer, waiting for acknowledgment from controller B. The preset timeout is three seconds, based on the average round-trip time of previous communications with controller B. If the command distribution module does not receive an acknowledgment from controller B within the three-second timeout period, the command retransmission mechanism is activated. The command distribution module checks controller B's heartbeat status and finds that it is still responding normally to heartbeat packets, indicating that the network connection is basically normal. Using the sequence number "CMD-20241027-085632-001" and the timestamp "2024-10-27 08:56:32.123UTC" as an index, the command distribution module accurately retrieves the original command message from the high-availability storage cluster. The command distribution module resends this command message to the gate controller B and increments the retry counter. After the second transmission, the command distribution module still does not receive acknowledgment within the new timeout period. When the number of retries reaches a preset upper limit (e.g., three times), the instruction distribution module marks gate controller B as "suspected fault," resets its weight to zero, and suspends the allocation of any new instructions to gate controller B. Simultaneously, the instruction distribution module generates an alarm message to notify maintenance personnel. Next, the weighted round-robin algorithm recalculates the weights among the remaining available execution terminals (gate controller A and standby controller C). Due to the higher load on controller A, standby controller C may receive a higher actual weight this time. The instruction distribution module successfully distributes the gate-opening instruction to standby controller C and successfully receives confirmation from controller C, thus completing the reliable distribution of this critical instruction.

[0044] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A water conservancy project operation risk monitoring and management system based on multi-sensor fusion, characterized in that, include: The data acquisition module is used to collect structural deformation data, water flow pressure data, and environmental temperature and humidity data in real time by deploying multiple types of intelligent sensors; The data fusion module is used to input the structural deformation data, water flow pressure data and environmental temperature and humidity data into a dynamic weighted fusion model, and generate a fused comprehensive monitoring index based on the sensor confidence weight and historical data deviation rate. The early warning and instruction generation module is used to trigger a multi-level early warning protocol and generate a corresponding set of equipment control instructions based on the comprehensive monitoring indicators and preset safety thresholds. The instruction distribution module is used to distribute the device control instruction set to each execution terminal through a distributed message queue; The verification and correction module is used to check the changes in the feedback signal parameters of the execution terminal using a backpropagation verification mechanism, and to correct the confidence weights of the dynamic weighted fusion model.

2. The water conservancy project operation risk monitoring and management system based on multi-sensor fusion according to claim 1, characterized in that, The step of triggering a multi-level early warning protocol and generating a corresponding set of equipment control instructions based on the comprehensive monitoring indicators and preset safety thresholds includes: Based on the dynamic comparison results of comprehensive monitoring indicators and preset safety thresholds, risk level intervals are divided and abnormal data segments are marked. The time series segmentation algorithm is used to decompose the abnormal data segments into trend, periodic and residual components, and the amplitude-frequency characteristics of each component are extracted as risk evolution parameters. A three-dimensional state space model is established based on risk evolution parameters, and the Mahalanobis distance from the current operating state point to the boundary of each risk level interval is calculated through projection mapping. When the Mahalanobis distance exceeds the adaptive warning threshold, a multi-level early warning protocol is triggered and a corresponding set of equipment control instructions is generated.

3. The water conservancy project operation risk monitoring and management system based on multi-sensor fusion according to claim 1, characterized in that, The construction process of the dynamic weighted fusion model includes: Calculate the inverse variance of the data collected by each type of sensor as the initial confidence weight; A sliding time window is introduced to statistically analyze the deviation rate between historical data of each sensor and the industry standard database; Grey relational analysis was used to quantify the influence coefficient of the deviation rate on the confidence weight. The initial confidence weights and influence coefficients are integrated using the weighted least squares method to generate a dynamically updated confidence weight matrix.

4. The water conservancy project operation risk monitoring and management system based on multi-sensor fusion according to claim 1, characterized in that, The methods for dividing the risk level ranges include: The basic safety threshold range shall be determined according to the design specifications of water conservancy projects; The extreme points and inflection points of the integrated monitoring indicators after fusion are extracted as candidate points for the interval boundary; Density clustering algorithm is used to identify the clustering characteristics of candidate points in the time dimension; By combining the basic safety threshold range with clustering characteristics, a three-level risk level range with flexible boundaries is generated.

5. The water conservancy project operation risk monitoring and management system based on multi-sensor fusion according to claim 1, characterized in that, The execution process of the time series segmentation algorithm includes: Empirical mode decomposition is performed on abnormal data segments to obtain intrinsic mode function components; The instantaneous frequency and amplitude of each component are calculated using the Hilbert transform; Components with amplitudes exceeding the noise threshold are screened for reconstruction, separating the trend term, periodic term, and residual term. Fourier transform is performed on the three reconstructed terms to extract the amplitude-frequency feature vectors.

6. The water conservancy project operation risk monitoring and management system based on multi-sensor fusion according to claim 2, characterized in that, The method for establishing the three-dimensional state-space model includes: The state space is constructed using the trend term amplitude, period term frequency, and residual term entropy as coordinate axes. Calculate the cluster centers of historical normal data in the state space; The polyhedral boundaries of each risk level interval in the state space are generated using the convex hull algorithm. The relative position of the current operating point in the polyhedron is determined by ray casting.

7. The water conservancy project operation risk monitoring and management system based on multi-sensor fusion according to claim 2, characterized in that, The calculation process of the Mahalanobis distance includes: Calculate the covariance matrix of historical data points within each risk level interval; Singular value decomposition of the covariance matrix yields the eigenvector basis; Project the current running state point onto a new coordinate system formed by the eigenvector basis; The standardized Euclidean distance between the projection point and the centroid of the target interval is calculated as the Mahalanobis distance.

8. The water conservancy project operation risk monitoring and management system based on multi-sensor fusion according to claim 7, characterized in that, The triggering conditions for the multi-level early warning protocol include: When the Mahal distance exceeds the Level 1 warning line, initiate the data verification process and mark the suspicious sensor; When the distance to the Mahalanobis exceeds the level 2 warning line, a device inspection command is sent to the operation and maintenance terminal. When the distance to the Mach number exceeds the Level 3 warning line, the backup control system is automatically activated and an emergency plan is generated.

9. The water conservancy project operation risk monitoring and management system based on multi-sensor fusion as described in claim 1, characterized in that, The step of using a backpropagation verification mechanism to check the changes in the feedback signal parameters of the execution terminal and correcting the confidence weights of the dynamic weighted fusion model includes: Receive real-time status signals from each execution terminal after executing the device control instruction set; The changes in key parameters are extracted from the real-time status signal and compared with the preset expected parameter range to generate parameter deviation values. Based on the parameter deviation values, the adjustment amount of the confidence weight of each sensor in the dynamic weighted fusion model is calculated by the backpropagation algorithm; The confidence weights of the dynamic weighted fusion model are updated based on the adjustment amount.

10. The water conservancy project operation risk monitoring and management system based on multi-sensor fusion according to claim 1, characterized in that, The distributed message queue used in the instruction distribution module is a persistent message queue, and its message distribution mechanism includes: Each device control command is assigned a unique serial number and timestamp and stored in a high-availability storage cluster; Instructions are dynamically allocated based on the processing capacity and current load of each execution terminal using a weighted round-robin algorithm. If no confirmation receipt is received from the execution terminal within the preset timeout period, the instruction retransmission mechanism is triggered based on the sequence number and timestamp.

Images