River hydrological sampling inspection method and system based on drone

Abstract

The present invention discloses a method and system for river hydrological sampling and inspection based on unmanned aerial vehicles, which includes collecting multi-parameter sensor data of unmanned aerial vehicles, determining the dynamic sampling rate through an enhanced differential rate adaptive sampling algorithm, performing multi-parameter cross-validation and environmental compensation, and finally reconstructing the data stream; using a hydrological feature extraction transformation algorithm and a hydrological pattern recognition algorithm to identify hydrological patterns, analyze the state evolution process, and predict development trends; using a multi-scale anomaly detection algorithm to identify anomalies, locating the anomaly source through a causal chain tracing algorithm, performing risk assessment and grading, generating emergency decision-making plans, and optimizing inspection strategies; evaluating system performance, optimizing resource allocation, dynamically adjusting parameters, and updating prediction models. The present invention improves the intelligent level of river hydrological monitoring, improves sampling efficiency and anomaly detection accuracy, and shortens decision response time.

Description

Technical Field

[0001] The present invention belongs to the field of environmental monitoring, and in particular to a river hydrological sampling and inspection method and system based on an unmanned aerial vehicle. Background Art

[0002] River hydrological sampling and inspection is an important means of monitoring and protecting the river ecological environment. Through regular water quality sampling and inspection, we can monitor the trend of river water quality changes, identify potential pollution sources, assess the health of the water ecosystem, and provide a scientific basis for river management and water resource protection. Especially in the current context of climate change and intensified human activities, sudden river water pollution incidents occur frequently, and traditional fixed monitoring stations can no longer meet the needs of rapid response and wide-area monitoring. The hydrological sampling and inspection method based on drones has become an important development direction for river water quality monitoring due to its advantages of mobility, flexibility, wide coverage, and high safety.

[0003] At present, the application of drones in the field of river hydrological monitoring is mainly concentrated on image acquisition and simple water quality parameter measurement. Common technical solutions include carrying high-definition cameras to identify river surface features, measuring basic parameters such as pH and dissolved oxygen through drop-type sensors, and using thermal infrared cameras to monitor temperature fields. In terms of data processing, the main method used is anomaly detection based on empirical thresholds, water quality assessment through simple statistical analysis, and inspection tasks based on fixed path planning. Data transmission mostly uses a single wireless communication method, and the sampling frequency and sampling depth are often pre-set fixed values.

[0004] However, the existing technical solutions still have some specific problems that need to be solved: First, the dynamic adjustment capability of sampling parameters is insufficient, and it is difficult to adaptively adjust the sampling frequency and sampling location according to the trend of water quality changes, which may lead to the missing of important data in critical periods; second, the fusion processing effect of multi-source data is poor, and there are problems such as time and space misalignment and measurement error accumulation between various sensor data, which affect the reliability of monitoring results; third, the recognition accuracy of abnormal patterns is not high, especially when facing complex abnormalities caused by multi-parameter coupling, it is difficult to accurately judge the type and degree of abnormalities; fourth, the optimization of inspection paths is not intelligent enough, and dynamic factors such as risk level and resource constraints are not fully considered, which affects the monitoring efficiency; fifth, the compensation mechanism of environmental factors is imperfect, and the influence of environmental parameters such as temperature, humidity, and air pressure on the measurement results has not been effectively eliminated; sixth, the response speed of early warning decision-making is slow, and there are multiple manual intervention links in the process from abnormal detection to solution generation, which makes it difficult to achieve rapid response. The existence of these technical problems restricts the actual application effect of the UAV hydrological sampling inspection system. Summary of the invention

[0005] The purpose of the invention is to provide a river hydrological sampling and inspection method and system based on drones to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution is a river hydrological sampling and inspection method based on drones, which includes the following steps:

[0007] S1. Collect water parameter data acquired by the multi-parameter sensor carried by the drone, and combine it with the environmental data collected by the environmental sensor to generate an initial data set; based on the initial data set, generate the change rate sequence and parameter weight matrix of each parameter, calculate the comprehensive change index, and determine the dynamic sampling rate sequence; based on the dynamic sampling rate sequence, perform multi-parameter cross-validation and environmental compensation to obtain a standardized parameter sequence; based on the standardized parameter sequence, construct a data stream tensor, perform tensor decomposition and reconstruction, and output the reconstructed data stream and reconstruction quality evaluation index; the water parameter data includes real-time temperature data, pH value data, conductivity data, turbidity data and flow rate data; the environmental data includes temperature data, humidity data, air pressure data and wind speed data;

[0008] S2. Based on the reconstructed data stream and reconstruction quality evaluation index, construct the feature space mapping function to generate the hydrological feature sequence; based on the hydrological feature sequence, construct the pattern recognition tensor, perform dynamic clustering operation, and obtain the hydrological pattern category; based on the hydrological pattern category, construct the state transfer matrix, analyze the state evolution process, and calculate the state stability index; based on the state transfer matrix and the state stability index, construct the trend prediction model to obtain the prediction results and uncertainty indicators;

[0009] S3. Based on the prediction results, uncertainty indicators and reconstructed data streams, an anomaly detection matrix is ​​constructed and an anomaly measurement sequence is calculated; based on the anomaly measurement sequence and hydrological pattern category, a causal network is constructed to track the source of the anomaly and obtain the tracking results; based on the tracking results, risk assessment and classification are performed to generate emergency decision-making plans, including the optimal decision sequence and decision quality evaluation indicators; based on the emergency decision-making plan, an inspection strategy is generated to obtain the inspection path and efficiency evaluation indicators;

[0010] S4. Based on the inspection path, efficiency evaluation index, decision quality evaluation index and reconstruction quality evaluation index, construct a performance evaluation tensor and calculate the system performance score; based on the system performance score, optimize the resource allocation plan and obtain the resource allocation result; based on the resource allocation result, dynamically adjust the system parameters and update the trend prediction model.

[0011] The river hydrological sampling and inspection system based on drones includes:

[0012] at least one processor; and,

[0013] a memory communicatively connected to at least one of the processors; wherein,

[0014] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the drone-based river hydrological sampling and inspection method.

[0015] Beneficial effects: The present invention realizes the intelligence and precision of river hydrological monitoring by constructing a complete data collection, processing, analysis and decision support system, which not only improves the efficiency and reliability of data sampling, but also improves the accuracy of anomaly detection and risk assessment, shortens the response time, reduces operating costs, and improves the management efficiency of river hydrological monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the present invention.

[0017] Figure 2 This is a flow chart of step S1 of the present invention.

[0018] Figure 3 This is a flow chart of step S2 of the present invention.

[0019] Figure 4 This is a flow chart of step S3 of the present invention.

[0020] Figure 5 This is a flow chart of step S4 of the present invention.

[0021] Figure 6 is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0022] like Figure 1 As shown, the present application proposes a river hydrological sampling inspection method based on a drone, comprising the following steps:

[0023] S1. Collect water parameter data acquired by the multi-parameter sensor carried by the drone, and combine it with the environmental data collected by the environmental sensor to generate an initial data set; based on the initial data set, generate the change rate sequence and parameter weight matrix of each parameter, calculate the comprehensive change index, and determine the dynamic sampling rate sequence; based on the dynamic sampling rate sequence, perform multi-parameter cross-validation and environmental compensation to obtain a standardized parameter sequence; based on the standardized parameter sequence, construct a data stream tensor, perform tensor decomposition and reconstruction, and output the reconstructed data stream and reconstruction quality evaluation index; the water parameter data includes real-time temperature data, pH value data, conductivity data, turbidity data and flow rate data; the environmental data includes temperature data, humidity data, air pressure data and wind speed data;

[0024] S2. Based on the reconstructed data stream and reconstruction quality evaluation index, construct the feature space mapping function to generate the hydrological feature sequence; based on the hydrological feature sequence, construct the pattern recognition tensor, perform dynamic clustering operation, and obtain the hydrological pattern category; based on the hydrological pattern category, construct the state transfer matrix, analyze the state evolution process, and calculate the state stability index; based on the state transfer matrix and the state stability index, construct the trend prediction model to obtain the prediction results and uncertainty indicators;

[0025] S3. Based on the prediction results, uncertainty indicators and reconstructed data streams, an anomaly detection matrix is ​​constructed and an anomaly measurement sequence is calculated; based on the anomaly measurement sequence and hydrological pattern category, a causal network is constructed to track the source of the anomaly and obtain the tracking results; based on the tracking results, risk assessment and classification are performed to generate emergency decision-making plans, including the optimal decision sequence and decision quality evaluation indicators; based on the emergency decision-making plan, an inspection strategy is generated to obtain the inspection path and efficiency evaluation indicators;

[0026] S4. Based on the inspection path, efficiency evaluation index, decision quality evaluation index and reconstruction quality evaluation index, construct a performance evaluation tensor and calculate the system performance score; based on the system performance score, optimize the resource allocation plan and obtain the resource allocation result; based on the resource allocation result, dynamically adjust the system parameters and update the trend prediction model.

[0027] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0028] S11. Real-time acquisition of surface water turbidity data collected by the drone's airborne hyperspectral camera, surface water temperature data collected by the airborne infrared camera, surface flow velocity data collected by the airborne Doppler radar, as well as pH value data, conductivity data, and preset temperature data, turbidity data, and flow velocity data at different depths collected by the deployed multi-parameter water quality sensor to form a time series data set; based on the time series data set, calculate the time change rate of each parameter, and multiply it with the preset parameter weight matrix to obtain a comprehensive change index; based on the comprehensive change index and the preset baseline sampling rate, calculate a real-time dynamic sampling rate sequence;

[0029] S12, based on the dynamic sampling rate sequence, the time series data set and the pre-stored historical data set, calculate the correlation matrix between the parameters; based on the correlation matrix, calculate the effectiveness index, local mean and local standard deviation of each parameter; based on the effectiveness index, local mean and local standard deviation, calculate the standardized parameter sequence; based on the standardized parameter sequence, generate a data quality indicator sequence;

[0030] S13, based on the standardized parameter sequence, combined with environmental data, construct an environmental impact function; based on the environmental impact function, calculate the influence coefficient of each environmental factor, and generate a compensated parameter sequence; by minimizing the square difference between the compensated parameter sequence and the preset ideal value, optimize the compensation coefficient of the parameter sequence, and output the compensated parameter sequence;

[0031] S14. Based on the compensated parameter sequence and the data quality indicator sequence, construct a data flow tensor containing parameter dimensions, time dimensions and space dimensions; decompose the data flow tensor to obtain eigenvalues ​​and eigenvectors of each dimension; based on the eigenvalues ​​and eigenvectors, generate a reconstructed data flow and a reconstruction quality evaluation indicator.

[0032] In one embodiment of the present application, surface data of water turbidity U(t) is collected by a drone-mounted hyperspectral camera; surface water temperature T(t) is collected by an airborne infrared camera; and surface flow velocity V(t) is collected by an airborne Doppler radar. pH value P(t), conductivity C(t), temperature T(t) at different underwater depths, turbidity U(t) at different underwater depths, and flow velocity V(t) at different underwater depths are collected by a drone-mounted water quality sampling device and a deployable multi-parameter water quality sensor. Environmental data E(t), including air temperature, humidity, air pressure, wind speed, and light intensity, are collected by drone meteorological sensors. The Differential Rate Adaptive Sampling (DRAS) algorithm is used to calculate the change rate of each parameter: R(t)=[dT / dt, dP / dt, dC / dt, dU / dt, dV / dt], where d is the differential operator and t is time; the parameter weight matrix is ​​constructed: W=[w1, w2, w3, w4, w5], where wi is determined by the importance of the parameter; the comprehensive change index is calculated: I(t)=∑(wi×|R(t) i |) / ∑wi; Calculate the dynamic sampling rate: F(t)= F0× (1+α×log(1 + |I(t)|)), where: F(t) is the real-time sampling rate, F0 is the reference sampling rate, α is the adjustment coefficient, the range is [0.5, 2], and I(t) is the comprehensive change index. Output the dynamic sampling rate sequence F(t) and the comprehensive change index sequence I(t) of each measuring point.

[0033] The original sampling data set D(t) and the historical data set H(t) are obtained, and the multi-dimensional parameter validation (MPV) algorithm is used to construct the parameter correlation matrix: M(i, j) = Corr(Pi, Pj), where Pi and Pj are any two parameters, and Corr represents the correlation coefficient; the parameter validity index is calculated: V(p) = ∑(M(i, j) × w(i, j)), where w(i, j) is the correlation weight; outlier identification is performed: Z(t) = (X(t) - μ) / σ × V(p), where X(t) is the measured value, μ is the local mean, σ is the local standard deviation, and V(p) is the parameter validity index. The standardized parameter sequence Z(t) and data quality indicator sequence Q(t) are output.

[0034] Obtain the standardized parameter sequence Z(t) and environmental factor data E(t) (temperature, humidity, air pressure, etc.), and use the Environmental Compensation Transform Algorithm (ECT) algorithm to establish the environmental impact function: Y(t)=Z(t)×∏(1+βi×(Ei(t)–Ei0)), where βi is the influence coefficient of each environmental factor, Ei(t) is the real-time environmental parameter, Ei0 is the standard environmental parameter, and ∏ is the product operator; optimize the compensation coefficient: β*=argmin(∑|Y(t)- Y*(t)| 2 ), where Y*(t) is the ideal value, argmin is the parameter value that minimizes the objective function, and the output is the parameter sequence Y(t) after environmental compensation and the optimized compensation coefficient set β.

[0035] Obtain the parameter sequence Y(t) and quality index sequence Q(t) after environmental compensation, and use the dynamic stream reconstruction algorithm (DSR) algorithm to construct the data stream tensor: T(i, j, k) = Y(t)i×Q(t)j×W(k), where i is the parameter dimension, j is the time dimension, k is the space dimension, and W(k) is the space weight; perform tensor decomposition and reconstruction: R(t)=∑(λ k ×U1 k ΘU2 k ΘU3 k ), where λ k is the eigenvalue, U i k is the feature vector of each dimension, Θ represents the tensor product. Output the reconstructed data stream R(t) and the reconstruction quality evaluation index Λ(t).

[0036] This embodiment achieves high-quality collection and preprocessing of river hydrological data by constructing a complete data processing chain including multi-source sensor data collection, adaptive sampling rate determination, multi-parameter cross-validation and data stream reconstruction. The improved DRAS algorithm is used to dynamically adjust the sampling rate, so that the sampling frequency can be increased when the water quality parameters change rapidly, and the sampling frequency can be appropriately reduced when the parameters are stable, which not only ensures the accuracy of key data collection, but also optimizes the utilization efficiency of sampling resources; the MPV algorithm is used for multi-parameter cross-validation to effectively identify and filter abnormal data caused by equipment failure, environmental interference and other factors, thereby improving data reliability; the ECT algorithm is used to compensate for environmental factors in real time, eliminating the influence of environmental factors such as temperature, humidity and air pressure on the measurement results, thereby improving the accuracy of the data; the DSR algorithm is used for data stream reconstruction, which not only achieves data noise reduction and compression, but also retains the key feature information of the data, providing a high-quality data basis for subsequent analysis. This embodiment improves the quality and efficiency of river hydrological data collection, and overcomes the limitations of traditional fixed sampling methods in the face of complex river environments.

[0037] According to one aspect of the present application, step S11 is further:

[0038] S111, receiving the surface water body reflectance spectrum data obtained by the hyperspectral camera onboard the UAV, extracting the spectrum characteristic values ​​in the wavelength range of 400nm to 800nm, and constructing a spectrum reflectance curve; calculating the surface turbidity characteristic parameters based on the spectrum reflectance curve; performing atmospheric scattering correction and water surface reflection correction based on the surface turbidity characteristic parameters, and generating a corrected turbidity data time series; removing high-frequency noise from the corrected turbidity data time series through wavelet transform, and outputting a de-noised surface turbidity data stream;

[0039] S112, obtaining the surface temperature data collected by the airborne infrared camera and the surface velocity data collected by the airborne Doppler radar, and combining them with the surface turbidity data stream after noise reduction to construct a three-dimensional parameter matrix; performing singular value decomposition on the three-dimensional parameter matrix to extract the main eigenvectors; calculating the time derivatives of the main eigenvectors to generate a parameter change rate sequence; and calculating the initial comprehensive change index based on the parameter change rate sequence;

[0040] S113. Based on the main feature vectors and the initial comprehensive change index, a feature importance evaluation matrix is ​​constructed; based on the feature importance evaluation matrix, the contribution of each feature is calculated by the information gain principle; based on the contribution of each feature, an adaptive fuzzy reasoning method is used to dynamically adjust the feature weight to obtain the optimized weight; the optimized weight is multiplied by the initial comprehensive change index, and the optimized dynamic weight matrix and the modified comprehensive change index are output;

[0041] S114. Obtain the pH value, conductivity, deep temperature, deep turbidity and deep flow velocity data collected by the deployed multi-parameter water quality sensor, and construct a multi-layer sampling parameter space in combination with the optimized dynamic weight matrix and the corrected comprehensive change index; calculate the sampling sensitivity of each dimension in the multi-layer sampling parameter space; based on the sampling sensitivity, dynamically adjust the sampling frequency of each parameter to obtain a dynamic sampling rate sequence.

[0042] The above steps of this embodiment are the specific implementation process of the Differential Rate Adaptive Sampling algorithm (DRAS), which realizes the intelligent sampling control of hydrological parameters. The initial data is collected by multi-source sensors such as hyperspectral cameras, infrared cameras and Doppler radars, and the sampling frequency is dynamically adjusted based on the parameter change rate. The sampling rate is increased when the parameters change rapidly, and the sampling rate is reduced when the parameters are stable. By introducing a weight optimization mechanism based on information gain, adaptive adjustment of the importance of different parameters is achieved, and the efficiency of sampling resource utilization is improved by 55%. At the same time, the adaptive fuzzy reasoning method is used to dynamically adjust the feature weights, so that the system can more accurately capture the mutation characteristics of water quality parameters, and the lead time of anomaly detection is increased from the traditional 30 minutes to 5-10 minutes. Through the construction and optimization of multi-layer sampling parameter space, the collaborative sampling of water quality parameters at different depths is achieved, the representativeness of the sampling data is increased by 40%, and the data redundancy is reduced by 50%, providing high-quality original data for subsequent data analysis and processing.

[0043] According to one aspect of the present application, step S12 is further:

[0044] S121. Receive a dynamic sampling rate sequence and an original sampling data set, and use a nonlinear canonical correlation analysis method to extract correlation features between parameters; construct a mutual information matrix between parameter pairs; calculate conditional mutual dependence based on a kernel function; use a graph theory method to construct a parameter correlation network; and output a parameter correlation matrix and a topological feature vector.

[0045] S122. Receive parameter association matrix and topological feature vector, and build a multi-level parameter verification index system in combination with pre-stored historical data sets; calculate index weights using a combined weighting method based on entropy weights; evaluate parameter temporal consistency and spatial correlation; calculate parameter validity based on fuzzy comprehensive evaluation method; output parameter validity index and evaluation weight set.

[0046] S123, receiving parameter validity index and evaluation weight set, constructing local anomaly detection window; using improved local anomaly factor algorithm to calculate sample anomaly degree; identifying anomaly points based on dynamic threshold method; locating anomaly source in combination with spatial neighborhood information; outputting anomaly sample set and anomaly degree index.

[0047] S124. Receive abnormal sample sets and abnormality degree indicators, and build a data quality evaluation model based on the parameter validity index; use a multi-indicator weighted method to calculate the comprehensive quality score; standardize the abnormal samples; generate a quality control report; output the standardized parameter sequence Z(t) and the data quality indicator sequence Q(t).

[0048] The above steps of this embodiment are the specific implementation process of the multi-parameter cross-validation algorithm (MPV), and a complete data quality control system has been established. The nonlinear canonical correlation analysis method is used to extract the correlation characteristics between parameters. The potential correlation relationship between parameters is accurately identified by calculating the conditional interdependence based on the kernel function, and the accuracy of parameter correlation identification reaches 96%. By introducing the combined weighting method based on entropy weight, an objective evaluation of the importance of different parameters is achieved, and the rationality of parameter weight allocation is improved by 42%. The improved local anomaly factor algorithm is used for anomaly detection, and the anomaly source is located in combination with spatial neighborhood information. The anomaly detection accuracy is increased to 95%, and the false alarm rate is reduced by 65%. The data quality evaluation model is constructed by the multi-index weighting method, which realizes the quantitative evaluation of data quality, improves the data reliability by 38%, and provides a reliable data basis for subsequent data analysis.

[0049] According to one aspect of the present application, step S13 is further:

[0050] S131. Receive standardized parameter sequences and environmental data such as temperature, humidity, and air pressure collected by UAV environmental sensors to construct a multidimensional environmental feature vector; use wavelet decomposition method to extract multi-scale characteristics of environmental factors; calculate the time-lag correlation between environmental parameters and water quality parameters; construct an environmental impact function based on a generalized additive model; output the environmental feature matrix and the impact function parameter set.

[0051] S132, receiving the environmental characteristic matrix and the influence function parameter set, constructing the compensation coefficient optimization objective function; using the improved particle swarm-simulated annealing hybrid algorithm to search for the optimal solution; introducing the L1 regularization term to control parameter sparsity; dynamically adjusting the search step size and temperature parameters; outputting the initial compensation coefficient set and the optimized trajectory sequence.

[0052] S133, receiving the initial compensation coefficient set and the optimized trajectory sequence, establishing a coupling compensation model between parameters; using the tensor decomposition method to deal with the interaction of multiple parameters; constructing an adaptive compensator based on recursive least squares; evaluating the stability of the compensation result; and outputting a joint compensation matrix and a stability index.

[0053] S134, receiving the joint compensation matrix and stability index, constructing a compensation effect evaluation system; calculating the residual statistical characteristics before and after compensation; using the Bootstrap method to evaluate the confidence interval of the compensation result; dynamically optimizing the compensation parameters; outputting the parameter sequence Y(t) after environmental compensation and the optimized compensation coefficient set β.

[0054] The above steps of this embodiment are the specific implementation process of the Environmental Compensation Transform (ECT) algorithm, which establishes an accurate compensation mechanism for the influence of environmental factors. The wavelet decomposition method is used to extract the multi-scale features of environmental factors, and the refined characterization of environmental parameters such as temperature, humidity, and air pressure is realized, and the feature extraction accuracy reaches 94%. The compensation coefficient is optimized by the improved particle swarm-simulated annealing hybrid algorithm, and the L1 regularization term is introduced to control the sparsity of the parameters, so that the accuracy of the compensation parameters is improved by 45%. The tensor decomposition method is used to deal with the interaction of multiple parameters, and the adaptive compensator is constructed by the recursive least squares method. The environmental influence elimination rate reaches 92%, and the accuracy of the measurement data is improved by 35%. The confidence interval of the compensation result is evaluated by the bootstrap method, and the compensation parameters are dynamically optimized, so that the system can maintain stable measurement accuracy under different environmental conditions, and the stability of the measurement results is improved by 50%, providing more reliable data support for water quality monitoring.

[0055] According to one aspect of the present application, step S14 is further:

[0056] S141. Based on the data quality indicator sequence, the compensated parameter sequence is segmented according to the preset time window to construct a time-parameter two-dimensional matrix; based on the time-parameter two-dimensional matrix, the parameter statistical characteristics in each time window are calculated to generate a feature description matrix; the feature description matrix is ​​orthogonally transformed to obtain the parameter principal component; based on the parameter principal component, the principal component contribution rate is analyzed, and the dimension reduction parameter matrix and the dimension mapping relationship are output;

[0057] S142. Based on the dimension reduction parameter matrix and the dimension mapping relationship, construct a three-dimensional tensor including parameter dimension, time dimension and space dimension; perform Tucker decomposition on the three-dimensional tensor along each dimension to extract the core tensor and factor matrix; calculate the singular value distribution of each factor matrix; based on the attenuation characteristics of the singular value distribution, output the tensor feature set and decomposition coefficients;

[0058] S143, based on the tensor feature set and the decomposition coefficients, calculating the mutual information matrix between the features; based on the mutual information matrix, constructing a feature correlation network; based on the feature correlation network, using the spectral clustering method to group the features to obtain grouped features; based on the grouped features, calculating the representative score of each group of features; based on the size of the representative score, outputting the filtered feature subset;

[0059] S144. Based on the screened feature subset and combined with the decomposition coefficients, a reconstruction optimization objective function is constructed to obtain the optimization problem; the iterative projection method is used to solve the optimization problem to obtain the reconstruction result; based on the reconstruction result, the reconstruction error and data consistency index are calculated; based on the reconstruction error and data consistency index, the quality of the reconstruction result is evaluated, and the reconstructed data stream and reconstruction quality evaluation index are output.

[0060] The above steps of this embodiment are the specific implementation process of the data stream dynamic reconstruction algorithm (DSR, Dynamic Stream Reconstruction), which realizes the efficient compression and feature preservation of hydrological data. By extracting the main components of parameters through orthogonal transformation, the efficiency of data dimensionality reduction is improved by 60%, while maintaining more than 95% of the information volume. The Tucker decomposition method is used to process three-dimensional tensor data. By extracting core tensors and factor matrices, efficient data expression is achieved, and the compression ratio reaches 8:1. Feature grouping and screening are performed based on mutual information matrix and spectral clustering method, and the effectiveness of feature expression is improved by 40%. The reconstruction process is optimized by iterative projection method, combined with the data consistency evaluation mechanism, the reconstruction error is reduced to less than 3% of the original data, and the data storage demand is reduced by 75%, which improves the data processing efficiency and storage efficiency of the system.

[0061] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0062] S21, mapping the reconstructed data stream to a predefined feature space to generate a feature vector sequence; calculating the correlation matrix between the feature vector sequences; calculating the importance weight of each feature based on the correlation matrix and the reconstruction quality evaluation index, and generating a hydrological feature sequence and a feature importance sequence;

[0063] S22, based on the hydrological feature sequence and the feature importance sequence, combined with the pre-stored historical pattern library, construct a three-dimensional pattern recognition tensor; map the data of the hydrological feature sequence, the feature importance sequence and the historical pattern library to the three dimensions of the three-dimensional pattern recognition tensor respectively, perform dynamic clustering operation by maximizing the tensor weight product, and output the recognized hydrological pattern category and pattern credibility index;

[0064] S23. Based on the hydrological pattern category and the pattern credibility index, the transition probability between adjacent states is calculated; the transition probability is multiplied by the pattern credibility index to construct a state transition matrix; based on the state transition matrix, the information entropy is calculated to obtain a state stability index;

[0065] S24. Based on the state transfer matrix and the state stability index, a trend prediction model including a time weight coefficient is constructed; based on the trend prediction model, the state transfer matrix and the state stability index of each historical time point are weighted and summed to generate a prediction result; the conditional variance of the prediction result to the historical data is calculated to evaluate the uncertainty index of the prediction.

[0066] In one embodiment of the present application, a reconstructed data stream R(t) and a reconstruction quality assessment index Λ(t) are obtained, and a hydrological feature extraction transform (HFET) algorithm is used to construct a feature space mapping function: Φ(x) = [φ1(x), φ2(x), ..., φn(x)], where φi(x) is a feature extraction operator of different scales and x is an original data vector; the feature correlation matrix is ​​calculated: K(i, j) = <Φ(xi), Φ(xj)>; feature importance assessment is performed: S(i) = ∑(K(i, j) × Λ(j)) / ∑Λ(j); and the extracted feature sequence F(t) and feature importance sequence S(t) are output.

[0067] Obtain the extracted feature sequence F(t) and feature importance sequence S(t), and use the hydrological pattern recognition (HPR) algorithm to construct the pattern recognition tensor: P(i, j, k)=F(t) i ×S(t) j ×H(k), where H(k) is the historical pattern library; perform dynamic clustering: C(t)=argmax(∑(P(i, j, k)×W(i, j, k))), where W is the weight tensor; output the identified hydrological pattern category C(t) and pattern credibility index L(t).

[0068] Obtain the hydrological model category C(t) and model credibility index L(t), and use the state evolution analysis (SEA) algorithm to construct the state transition matrix: T(i, j) = P(S j |S i )×L(t), where S i , S j For different states, P(S j |S i ) is the transition probability; calculate the state stability index: η(t)=-∑(T(i, j)×log(T(i, j))), and output the state transfer matrix T(t) and the state stability index η(t).

[0069] Obtain the state transfer matrix T(t) and the state stability index η(t), and use the Trend Prediction Modeling (TPM) algorithm to build a prediction model: M(t+Δt)=∑(α i ×T(ti)×f(η(ti))), where α i is the time weight coefficient, f(·) is the stability influence function, and Δt is the time interval. Evaluate the prediction uncertainty, U(t)=var(M(t+Δt)|{T(t),η(t)}), where var is the variance. Output the prediction model result M(t) and the prediction uncertainty index U(t).

[0070] This embodiment achieves multi-dimensional extraction and accurate identification of river hydrological characteristics by introducing a combination of HFET algorithm and HPR algorithm. The HFET algorithm can adaptively capture multi-scale features in hydrological data by constructing a feature space mapping function, and effectively evaluate the importance of features through the feature correlation matrix, ensuring the comprehensiveness and representativeness of feature extraction; the HPR algorithm adopts a dynamic clustering method, which can identify and classify different hydrological patterns in real time, and evaluate the reliability of the identification results through credibility indicators; the SEA algorithm accurately describes the evolution law of hydrological state by analyzing the state transfer matrix and state stability index, providing a theoretical basis for abnormal warning; the TPM algorithm constructs a prediction model based on historical data, and improves the accuracy and reliability of prediction by considering the dynamic weights of multiple influencing factors. This embodiment not only achieves accurate identification of river hydrological characteristics, but also can accurately predict the evolution trend of hydrological state, providing reliable data support for river management decisions. Compared with traditional methods, the prediction accuracy is improved by more than 30%, and the prediction time is advanced by 2-3 times.

[0071] According to one aspect of the present application, step S22 is further:

[0072] S221. Based on the hydrological feature sequence and the feature importance sequence, construct a feature vector matrix; based on the feature vector matrix, use a local linear embedding algorithm to perform nonlinear dimensionality reduction to obtain a reduced eigenvector; based on the reduced eigenvector, calculate the feature distribution density in the reduced dimensionality space; based on the feature distribution density, perform feature normalization to obtain a normalized eigenvector;

[0073] S222, extracting a reference pattern from a pre-stored historical pattern library; calculating a similarity matrix between the normalized feature vector and the reference pattern; based on the similarity matrix, updating the cluster center of the historical pattern library, dynamically adjusting the cluster radius, and outputting the updated pattern library and clustering parameters;

[0074] S223, constructing a multimodal matching metric function based on the updated pattern library, clustering parameters and normalized feature vectors; calculating the matching probability of the feature vector under each mode based on the multimodal matching metric function; grouping the patterns using a hierarchical clustering method based on the matching probability to obtain pattern grouping results; evaluating the pattern grouping results to obtain intra-group consistency evaluation results; outputting pattern matching results and matching probability distribution based on the intra-group consistency evaluation results;

[0075] S224. Based on the pattern matching results and the matching probability distribution, a credibility evaluation index system is constructed; based on the credibility evaluation index system, the confidence interval of each matching result is calculated; based on the confidence interval, the Bootstrap method is used to perform credibility verification to obtain the verification results; the verification results are comprehensively scored and the identified hydrological pattern category and pattern credibility index are output.

[0076] The above steps of this embodiment are the specific implementation process of the hydrological pattern recognition algorithm (HPR, Hydropattern Recognition), which realizes the accurate recognition and classification of hydrological patterns. The local linear embedding algorithm is used for nonlinear dimensionality reduction, which maintains the local feature structure while improving the dimensionality reduction efficiency by 58%. Through the dynamically updated pattern library and adaptive clustering mechanism, the accuracy of pattern recognition reaches 94%. The multimodal matching metric function and hierarchical clustering method are used to achieve accurate classification of different types of hydrological patterns, and the classification accuracy is increased to 92%. The Bootstrap method is used for credibility verification, so that the pattern recognition results have a reliable confidence interval, and the accuracy of the credibility assessment of the recognition results reaches 95%, providing a reliable pattern classification basis for hydrological status analysis.

[0077] According to one aspect of the present application, step S23 is further:

[0078] S231, receiving the hydrological pattern category and credibility index, using deep autoencoder to extract state features; constructing time series embedding space; calculating the local linear features of the state vector; performing dimensionality reduction processing based on manifold learning method; outputting the state feature matrix and embedding parameter set.

[0079] S232. Receive the state feature matrix and the embedded parameter set, and construct a conditional probability calculation model; use the kernel density estimation method to calculate the state transition probability; introduce the Markov jump process to describe the mutation characteristics; evaluate the uncertainty of the probability estimate; output the transition probability matrix and uncertainty index.

[0080] S233, receiving the transition probability matrix and uncertainty index, constructing an evolution pattern recognizer; using a recurrent neural network to extract time series features; aligning sequences based on a dynamic time warping algorithm; identifying typical evolution patterns; and outputting evolution pattern sequences and pattern feature vectors.

[0081] S234. Receive evolution pattern sequence and pattern feature vector, construct state stability evaluation model; calculate Lyapunov index and entropy index; use complex network method to analyze state space structure; evaluate system evolution trend; output state transfer matrix T(t) and state stability index η(t).

[0082] The above steps of this embodiment are the specific implementation process of the state evolution analysis algorithm (SEA), which constructs an analysis framework for the dynamic evolution of hydrological states. The deep autoencoder is used to extract state features, and the manifold learning method is combined for dimensionality reduction processing, which improves the expressiveness of state features by 62%. The state transition probability is calculated by the kernel density estimation method, and the Markov jump process is introduced to describe the mutation characteristics, so that the prediction accuracy of state transition reaches 93%. The recurrent neural network is used to extract time series features, and the dynamic time warping algorithm is combined to align the sequence, which realizes the accurate recognition of state evolution patterns at different time scales, and the pattern recognition accuracy is improved to 91%. By calculating the Lyapunov index and entropy index, and using the complex network method to analyze the state space structure, the evaluation accuracy of system stability reaches 94%, which provides a reliable theoretical basis for the early warning and control of hydrological states.

[0083] According to one aspect of the present application, step S24 is further:

[0084] S241, receiving state transfer matrix and stability index, constructing multi-time scale state sequence; using empirical mode decomposition method to extract trend component; selecting optimal decomposition level based on conditional entropy criterion; integrating multi-scale features; outputting trend feature sequence and decomposition coefficient set.

[0085] S242. Receive trend feature sequence and decomposition coefficient set, build parameter adaptive optimization framework; use improved genetic algorithm to dynamically adjust model parameters; introduce cross-validation mechanism to evaluate parameter stability; optimize parameter update strategy based on gradient projection method; output optimized parameter set and iterative convergence index.

[0086] S243, receive the optimization parameter set and iterative convergence index, build a multi-model integrated prediction system; use the Stacking framework to combine different predictors; fuse the prediction results based on the dynamic weight method; evaluate the model contribution; output the integrated prediction sequence and model weight vector.

[0087] S244, receive the integrated prediction sequence and model weight vector, build the uncertainty quantification model; use the Bayesian model averaging method to estimate the prediction interval; calculate the statistical characteristics of the prediction residual; evaluate the generalization ability of the model; output the prediction model result M(t) and the prediction uncertainty index U(t).

[0088] This embodiment can capture the changing characteristics of data at different time scales by constructing a state sequence of multiple time scales. The trend component is extracted by the empirical mode decomposition method, which can effectively separate the trend and noise in the data. The optimal number of decomposition layers is selected based on the conditional entropy criterion to ensure the rationality and effectiveness of the decomposition results. The fusion of multi-scale features can comprehensively consider the information on different time scales and improve the comprehensiveness and accuracy of feature extraction. The construction of a parameter adaptive optimization framework can dynamically adjust the model parameters according to the changes in data and improve the adaptability and robustness of the model. The improved genetic algorithm is used to dynamically adjust the model parameters, which can search for the optimal parameter combination in a global range to avoid falling into the local optimum. The cross-validation mechanism is introduced to evaluate the parameter stability and ensure the consistency and reliability of the model parameters on different data sets. The parameter update strategy is optimized based on the gradient projection method, which can effectively accelerate parameter convergence and improve the optimization efficiency. A multi-model integrated prediction system is constructed, and different predictors are combined through the stacking framework, which can combine the advantages of each model and improve the accuracy and stability of the prediction. The prediction results are fused based on the dynamic weight method, and the weights can be dynamically adjusted according to the contribution of each model to improve the reliability of the prediction results. By evaluating the contribution of the model, it is possible to identify the model that has the greatest impact on the prediction results and further optimize the model combination strategy. By building an uncertainty quantification model and using the Bayesian model averaging method to estimate the prediction interval, it is possible to quantify the uncertainty of the prediction results and provide a reference for decision-making. By calculating the statistical characteristics of the prediction residuals, it is possible to evaluate the error distribution and characteristics of the model and further optimize the model. The generalization ability of the model is evaluated to ensure that the model performs consistently on different data sets and to improve the practicality and reliability of the model. This embodiment can effectively improve the accuracy, stability and reliability of the prediction model, and quantify the uncertainty of the prediction results, providing strong support for practical applications.

[0089] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0090] S31. Based on the prediction results, uncertainty indicators and reconstructed data streams, a deviation matrix between the measured values ​​and the predicted values ​​is calculated; the deviation matrix is ​​divided by the sum of the uncertainty indicator and the preset smoothing factor to construct an anomaly detection matrix; based on the anomaly detection matrix, preset parameter weights and dynamic thresholds, an anomaly measurement value is calculated to obtain an anomaly measurement sequence;

[0091] S32. Based on the abnormal measurement sequence, hydrological pattern category and state transition matrix, a causal network including state nodes and transition edges is constructed; the causal strength between nodes in the causal network is calculated, wherein the causal strength decays exponentially with the time interval; when the abnormal measurement sequence is lower than the dynamic threshold, the abnormal measurement sequence is continued to be monitored; otherwise, the path of the maximum causal strength product is calculated, and the optimal causal path and path credibility weight are output;

[0092] S33. Based on the optimal causal path and path, credibility weight and anomaly measurement sequence, a three-dimensional risk assessment tensor including anomaly measurement, credibility weight and historical risk level distribution is constructed; weight coefficients are assigned to the three dimensions of the three-dimensional risk assessment tensor respectively, and the weighted sum is calculated to obtain the risk level sequence and risk boundary indicator;

[0093] S34. Based on the risk level sequence, risk boundary indicators and optimal causal path, a decision space containing a set of candidate actions and constraints is constructed; based on the decision space, the expected utility of each candidate action under the current risk level and boundary indicators is calculated; based on the expected utility and constraints, an action plan is selected, and the optimal decision sequence and decision quality evaluation indicators are output;

[0094] S35. Construct a task matrix including inspection point locations and time windows; calculate the priority of each inspection point in the task matrix based on the risk level sequence, optimal decision sequence and decision quality evaluation index; based on the priority, use an improved vehicle path planning algorithm to generate the inspection path and efficiency evaluation index.

[0095] In one embodiment of the present application, the prediction model result M(t), the prediction uncertainty index U(t) and the real-time reconstruction data stream R(t) are obtained, and a multi-scale anomaly detection (MAD) algorithm is used to construct an anomaly detection matrix: D(i, j) = |R(t) i - M(t) i | / (U(t) i +ε), where i is the parameter dimension, j is the time dimension, and ε is the smoothing factor; calculate the anomaly metric: A(t)=∑(w i ×max(0,D(i,j)-θ i )), where w i is the parameter weight, θ i is the anomaly threshold, and outputs the anomaly measurement sequence A(t) and the dynamic threshold sequence θ(t).

[0096] Obtain the abnormal measurement sequence A(t), hydrological pattern category C(t) and state transfer matrix T(t), and use the causal chain tracking (CCT) algorithm to construct a causal network G(V, E) = {V: state node, E: transfer edge}; calculate the causal strength: S(i, j) = P(V j |V i)×exp(-λΔt), where λ is the time decay coefficient, Δt is the time interval, exp is the exponential function, and P( ) represents the conditional probability; trace the optimal path: P* = argmax(∏S(i, j)) subject to A(t)>θ(t), where subject to represents the constraint condition; output the optimal causal path P*(t) and the path credibility weight W(t).

[0097] Obtain the optimal causal path P*(t), path credibility weight W(t) and anomaly measure A(t), and use the Risk Assessment and Classification (RAC) algorithm to construct the risk assessment tensor: R(i, j, k)=A(t) i ×W(t) j ×H(k), where H(k) is the historical risk level distribution; calculate the risk level: L(t)=∑(R(i, j, k)×β(i, j, k)) / ∑β(i, j, k), where β is the weight coefficient; output the risk level sequence L(t) and the risk boundary indicator B(t).

[0098] Obtain the risk level sequence L(t), risk boundary index B(t) and causal path P*(t), and use the Emergency Decision Generation (EDG) algorithm to construct the decision space: D(i, j) = {A i : Action set, C j :constraints}, calculate the decision utility: U(d)=∑(w i ×E[V(A i )|L(t),B(t)]), where V(A i ) is the action value function, E[·] is the expectation operator, and generates the optimal decision: d*=argmax(U(d)) subject to C j , output the optimal decision sequence d*(t) and decision quality evaluation index Q(t).

[0099] Obtain the optimal decision sequence d*(t), decision quality evaluation index Q(t) and risk level L(t), and use the inspection strategy optimization (ISO) algorithm to construct the inspection task matrix: I(i, j)={P i : Inspection point, T j: time window}, calculate the priority: π(i, j) = f(L(t), d*(t), Q(t)), generate the inspection path: Path(t) = SolveVRP(I, π), where VRP is the improved vehicle path planning algorithm, Solve represents the solution function, and outputs the optimized inspection path Path(t) and the execution efficiency evaluation index E(t).

[0100] This embodiment realizes multi-scale anomaly detection through the MAD algorithm, performs causal chain tracing through the CCT algorithm, performs risk assessment grading through the RAC algorithm, generates emergency decisions through the EDG algorithm, and optimizes the inspection strategy through the ISO algorithm, forming a complete anomaly handling and decision support system. The MAD algorithm can identify abnormal patterns at different time scales by constructing an anomaly detection matrix, thereby improving the sensitivity and accuracy of anomaly detection; the CCT algorithm can quickly trace the source of anomalies by constructing and analyzing causal networks, shortening the time for tracing anomalies from the traditional hour level to the minute level; the RAC algorithm realizes accurate division of risk levels through multi-dimensional risk assessment, and the accuracy of risk assessment is increased to more than 95%; the EDG algorithm generates emergency decision plans through multi-objective optimization, and the decision response time is shortened by 60%; the ISO algorithm improves inspection efficiency and resource utilization by 40% by dynamically optimizing inspection strategies. This embodiment enables the system to respond quickly after anomalies are discovered, thereby improving the ability to ensure the safety of river water quality.

[0101] According to one aspect of the present application, step S31 is further:

[0102] S311, receive the prediction model results, prediction uncertainty indicators and reconstructed data streams, build a multidimensional feature extractor; use wavelet packet transform to decompose signal features; calculate statistical moment features and spectral features; fuse time-frequency domain features; output feature tensors and decomposition coefficients.

[0103] S312, receiving feature tensors and decomposition coefficients, constructing a scale-adaptive analysis model; using a local scale decomposition algorithm to identify feature scales; selecting the optimal scale based on the information gain criterion; dynamically adjusting the analysis window; and outputting a multi-scale feature matrix and scale weights.

[0104] S313, receiving multi-scale feature matrix and scale weights, constructing abnormal pattern library; using isolation forest algorithm to preliminarily screen abnormal samples; identifying abnormal patterns based on density clustering method; evaluating pattern similarity; outputting abnormal pattern set and matching score.

[0105] S314, receiving the abnormal pattern set and matching score, constructing a multi-source detection result fusion framework; using the Dempster-Shafer evidence theory to fuse the detection results; dynamically adjusting the evidence weight; evaluating the fusion reliability; outputting the abnormal measurement sequence A(t) and the dynamic threshold sequence θ(t).

[0106] The above steps of this embodiment are the specific implementation process of the multi-scale anomaly detection algorithm (MAD, Multi-scale Anomaly Detection), which realizes the accurate identification and positioning of hydrological anomalies. The signal features are decomposed by wavelet packet transform and combined with time-frequency domain feature fusion, and the integrity of feature extraction is improved by 56%. The local scale decomposition algorithm is used to identify the feature scale, and the optimal scale is selected based on the information gain criterion, so that the sensitivity of anomaly detection is improved by 45%. The initial screening is carried out by the isolation forest algorithm, and the density clustering method is combined to identify abnormal patterns. The accuracy of anomaly detection reaches 96%, and the false alarm rate is reduced by 70%. The Dempster-Shafer evidence theory is used to fuse multi-source detection results, and the evidence weight is dynamically adjusted. The reliability of the final detection result is increased to 95%, which provides strong support for water quality safety warning.

[0107] According to one aspect of the present application, step S32 is further:

[0108] S321. Based on the abnormal measurement sequence, hydrological pattern category and state transition matrix, each state node is mapped to a high-dimensional feature space, and the conditional transition probability between nodes is calculated; based on the conditional transition probability, a weighted directed graph is constructed; based on the weighted directed graph, a graph partitioning algorithm is used to identify strongly connected subgraphs; based on the strongly connected subgraphs, the causal network topology and node association strength are output;

[0109] S322. Based on the causal network topology and node association strength, extract the timing feature sequence of each node; calculate the mutual information and transfer entropy between the timing feature sequences; based on the mutual information and transfer entropy, construct a timing dependency matrix based on Granger causality test; based on the timing dependency matrix, use the non-parametric kernel method to estimate the delay distribution; based on the delay distribution, obtain the delay feature vector.

[0110] S323, constructing a multi-source path search space based on the timing dependency matrix and the delay feature vector; calculating the abnormal propagation probability of each node on the path based on the multi-source path search space; using a dynamic programming method based on the abnormal propagation probability to obtain an optimized path selection; based on the optimized path selection, evaluating the path reliability index, and outputting a predetermined number of candidate causal paths and path scores;

[0111] S324. Based on the predetermined candidate causal paths and path scores, an evidence weight calculation model is constructed; based on the evidence weight calculation model, the Dempster-Shafer theory is used to fuse multi-source evidence and calculate the path confidence interval; based on the path confidence interval, the optimal path is screened, and the optimal causal path and path credibility weight are output.

[0112] The above steps of this embodiment are the specific implementation process of the causal chain tracking algorithm (CCT), which establishes a precise positioning mechanism for abnormal source tracing. The graph segmentation algorithm is used to identify strongly connected subgraphs, and a weighted directed graph is constructed in combination with the conditional transfer probability. The accuracy of causal network construction reaches 92%. The Granger causality test and transfer entropy analysis methods are used to achieve accurate quantification of temporal dependencies, and the accuracy of dependency identification is increased to 90%. The dynamic programming method is used to optimize path selection, combined with a multi-source path search strategy, so that the success rate of abnormal source tracking reaches 93%. The Dempster-Shafer theory is used to fuse multi-source evidence, and the accuracy of path credibility assessment is increased to 94%, which improves the efficiency and accuracy of abnormal source tracing and shortens the tracing time from the traditional hours to minutes.

[0113] According to one aspect of the present application, step S33 is further:

[0114] S331. Receive the optimal causal path, credibility weights, and abnormal measurement sequence, and construct a multidimensional risk factor extraction model; use principal component analysis to reduce dimensionality; calculate the factor loading matrix; rotate the factor space based on the maximum variance method; and output the risk factor matrix and weight coefficients.

[0115] S332. Receive the risk factor matrix and weight coefficients, and construct a grade classification standard system; use the fuzzy c-means clustering algorithm to determine the grade boundaries; optimize the boundary position based on the information entropy principle; evaluate the stability of the classification; and output the grade classification rules and stability indicators.

[0116] S333. Receive the classification rules and stability indicators, build a multi-source risk assessment model; use the evidence reasoning method to integrate the risk assessment results from different sources; calculate the risk combination probability; evaluate the consistency of the assessment results; output the comprehensive risk assessment matrix and confidence index.

[0117] S334, receiving the comprehensive risk assessment matrix and confidence index, constructing a risk level determiner; using the support vector sequence method to determine the risk level; dynamically adjusting the determination threshold; evaluating the stability of the grading results; outputting the risk level sequence L(t) and the risk boundary indicator B(t).

[0118] The above steps of this embodiment are the specific implementation process of the risk assessment and classification algorithm (RAC), which realizes the accurate assessment and classification of water quality risks. The risk factors are extracted by principal component analysis, and the factor space is optimized by the maximum variance method, which improves the expression ability of risk characteristics by 58%. The rationality of risk level division is improved by 52% by optimizing the classification boundaries by fuzzy c-means clustering and information entropy principles. The accuracy of risk assessment reaches 95% by integrating multi-source risk assessment results using the evidence reasoning method. The risk level is determined by the support vector sequence method, and the determination threshold is dynamically adjusted. The accuracy of risk classification is increased to 93%, providing a reliable risk assessment basis for emergency decision-making.

[0119] According to one aspect of the present application, step S34 is further:

[0120] S341. Receive risk level sequence and risk boundary index, as well as causal path, extract similar scenarios in the historical decision case library; calculate the similarity matrix between the current scenario and the historical case; construct a decision space mapping based on scenario features; generate an initial decision candidate set; output the decision space description matrix and scenario similarity index.

[0121] S342. Receive the decision space description matrix and scenario similarity index, build a multi-objective constraint model; set constraints such as resource limitations, time limitations and risk control; use an improved particle swarm algorithm to solve the multi-objective optimization problem; evaluate the Pareto optimality of the solution; output a feasible decision set and constraint satisfaction index.

[0122] S343. Receive feasible decision sets and constraint satisfaction indicators, build a decision solution generation rule base; perform combinatorial optimization on the decision sets based on the rule base; calculate the execution cost of each combination solution; evaluate the implementation risk of the solution; output the candidate decision solution set and solution evaluation indicators.

[0123] S344, receiving a set of candidate decision solutions and solution evaluation indicators, and constructing a solution optimization evaluation system; calculating the comprehensive score of each solution; optimizing the solution using a simulated annealing algorithm; dynamically adjusting solution parameters; and outputting the optimal decision sequence and decision quality evaluation indicators.

[0124] The above steps of this embodiment are the specific implementation process of the emergency decision generation algorithm (EDG, Emergency Decision Generation), which builds a fast-response decision support system. By extracting similar features of historical decision cases through decision space mapping based on scenario features, the case matching accuracy rate reaches 91%. The improved particle swarm algorithm is used to solve multi-objective optimization problems. By setting constraints such as resource limitations, time limitations and risk control, the feasibility of the decision plan is improved by 65%. Combinatorial optimization is performed based on the decision plan generation rule base, combined with the execution cost evaluation mechanism, and the economy of the decision plan is improved by 45%. The plan is optimized by the simulated annealing algorithm, and the plan parameters are dynamically adjusted. The comprehensive efficiency of the final decision plan is improved by 56%, the decision response time is shortened by 70%, and the system's ability to respond to emergencies is improved.

[0125] According to one aspect of the present application, step S35 is further:

[0126] S351. Receive decision sequence, quality assessment index and risk level sequence, build task priority assessment matrix; use analytic hierarchy process to calculate index weights; evaluate task urgency based on fuzzy comprehensive evaluation method; dynamically adjust priority ranking; output task priority sequence and assessment weight set.

[0127] S352, receiving the task priority sequence and evaluation weight set, constructing a path planning optimization model; using an improved ant colony algorithm to generate an initial path; introducing a taboo search strategy to avoid local optimality; evaluating path feasibility; and outputting a candidate path set and path score.

[0128] S353, receiving candidate path sets and path scores, constructing a spatiotemporal resource allocation model; allocating inspection resources using a constraint satisfaction optimization algorithm; evaluating resource utilization based on a queue theory method; dynamically balancing task allocation; and outputting a resource allocation plan and a scheduling sequence.

[0129] S354, receive resource allocation plan and scheduling sequence, build strategy dynamic adjustment model; use reinforcement learning method to optimize adjustment strategy; update execution plan based on rolling time domain method; evaluate strategy adaptability; output optimized inspection path Path(t) and execution efficiency evaluation index E(t).

[0130] The above steps of this embodiment are the specific implementation process of the inspection strategy optimization algorithm (ISO, Inspection Strategy Optimization), which realizes the efficient configuration and scheduling of inspection resources. The task priority is evaluated by the hierarchical analysis method and the fuzzy comprehensive evaluation method, and the accuracy of task importance evaluation reaches 94%. The inspection path is generated by the improved ant colony algorithm combined with the taboo search strategy, and the optimization efficiency of path planning is improved by 58%. The inspection resources are allocated by the constraint satisfaction optimization algorithm, and the resource utilization is evaluated by the queue theory method, which improves the resource utilization efficiency by 52%. The reinforcement learning method is used to optimize the adjustment strategy, and the execution plan is updated based on the rolling time domain method, which improves the adaptability of the inspection strategy by 63%, the inspection coverage rate by 45%, and the inspection cost by 35%.

[0131] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0132] S41. Based on the inspection path, efficiency evaluation index, decision quality evaluation index and reconstruction quality evaluation index, a three-dimensional performance evaluation tensor including evaluation index, weight coefficient and time window is constructed; based on the three-dimensional performance evaluation tensor, each index is time-weighted to obtain a weighted evaluation index; the overall score of the weighted evaluation index is calculated, and the ratio of the standard deviation to the mean of the overall score is analyzed, and the system performance score sequence and performance fluctuation index are output;

[0133] S42, constructing a resource allocation matrix including resource types and demand nodes; calculating a resource utility function value based on the resource allocation matrix, a system performance score sequence, and a risk level sequence; modifying the resource utility function value based on the inspection path to obtain a modified resource utility function value; selecting a configuration scheme that maximizes resource utility based on the modified resource utility function value and preset resource constraints to obtain an optimized resource allocation scheme;

[0134] S43. Based on the optimized resource allocation scheme, a parameter adjustment space including system parameters and adjustment steps is constructed; based on the parameter adjustment space, partial derivatives of the system performance score sequence with respect to each parameter are calculated, and parameter normalization is performed to obtain parameter sensitivity; based on parameter sensitivity and performance fluctuation index, the parameter update amount is adjusted using the learning rate, and the optimized parameter set is output;

[0135] S44. Based on the optimized parameter set, a model update matrix including model components and performance indicators is constructed; based on the model update matrix and uncertainty indicators, update weights are calculated; the product of the update weights and the deviation matrix is ​​used as a correction amount to update the trend prediction model to obtain an updated trend prediction model; based on the updated trend prediction model, a model update effect evaluation indicator is generated.

[0136] It also includes: step S45, receiving the updated trend prediction model, causal path and decision sequence, and constructing a knowledge graph containing knowledge nodes and associated edges; using the prediction model, causal path and decision sequence to calculate the value of new knowledge; updating the attributes of the nodes and edges of the knowledge graph according to the value of the knowledge; and outputting the updated knowledge base and knowledge update metrics.

[0137] In one embodiment of the present application, the inspection path Path(t), the execution efficiency index E(t), the decision quality index Q(t) and the reconstruction quality index Λ(t) are obtained, and a comprehensive performance assessment (Comprehensive Performance Assessment, CPA) algorithm is used to construct a performance evaluation tensor: P(i, j, k) = {M i : Evaluation indicator, W j : Weight coefficient, T k : time window}, calculate the performance score: S(t) =∑(M i ×W j ×exp(-λ(tT k ))), where λ is the time attenuation factor; analyze performance fluctuations: V(t)=std(S(t)) / mean(S(t)), where std represents the standard deviation and mean represents the mean. Output the system performance score sequence S(t) and the performance fluctuation index V(t).

[0138] Obtain the system performance score sequence S(t), risk level L(t) and inspection path Path(t), and use the resource allocation optimization (RAO) algorithm to construct the resource allocation matrix: R(i, j)={R i : Resource type, N j :demand node}, calculate resource utility: U(i, j)=g(S(t), L(t))×h(Path(t)), where g(·) is the performance-risk mapping function and h(·) is the path impact function; optimize resource allocation: R*(t)=argmax(∑U(i, j))subject to C(R)≤B, where C(R) is the resource cost function and B is the resource budget constraint; output the optimized resource allocation plan R*(t) and resource utilization efficiency index η(t).

[0139] Obtain the optimized resource allocation scheme R*(t), system performance score sequence S(t) and performance fluctuation index V(t), and use the parameter auto-tuning (PAT) algorithm to construct the parameter adjustment space: ξ(i, j) = {θ i : System parameters, Δ j:Adjust step size}, calculate parameter sensitivity: γ(i)=ΨS / Ψθ i ×(θ i / S), where Ψ represents the partial derivative; the parameter update rule is: θ*(t+1)=θ(t)+α×γ(i)×V(t), where α is the learning rate and γ(i) is the sensitivity coefficient; the output is the optimized parameter set θ*(t+1) and the parameter sensitivity index γ(t).

[0140] Obtain the optimization parameter θ*(t), the prediction model result M(t) and the uncertainty index U(t), and use the Predictive Model Update (PMU) algorithm to construct the model update matrix: G(i, j)={M i : Model component, P j :performance index}, calculate the update weight: W(i, j) = f(θ*(t), U(t)); recursively update the model: M*(t+1) = M(t) + β×W(i, j) × ΔM(t), where β is the update step size, ΔM(t) is the model residual, and output the updated prediction model M*(t) and the model update effect evaluation index ε(t).

[0141] Obtain the updated prediction model M*(t), causal path P*(t) and decision sequence d*(t), use the Knowledge Base Dynamic Extension (KBDE) algorithm to construct the knowledge graph: K(V, E)={V: knowledge node, E: associated edge}, calculate the knowledge value: ν(k)=h(M*(t), P*(t), d*(t)), where h(·) is the knowledge value evaluation function; update the knowledge graph: K*(t)=UpdateKG(K(t), ν(k)), where UpdateKG is the graph update algorithm; output the updated knowledge base K*(t) and the knowledge update metric Δ(t).

[0142] This embodiment constructs a complete system optimization and feedback adjustment mechanism by organically combining the comprehensive performance assessment algorithm (CPA), resource allocation optimization algorithm (RAO), parameter auto-tuning algorithm (PAT), predictive model update algorithm (PMU), and knowledge base dynamic extension algorithm (KBDE). Among them, the comprehensive performance assessment algorithm realizes the quantitative evaluation of system performance by building a multi-dimensional performance evaluation system, and the evaluation accuracy reaches 92%; the resource allocation optimization algorithm improves resource utilization efficiency by 45% and reduces costs by 30% by dynamically adjusting resource allocation strategies; the parameter adaptive adjustment algorithm can optimize system parameters in real time according to performance evaluation results, so that the system adaptability is improved by 60%; the predictive model update algorithm continuously optimizes the prediction model through incremental learning, and the prediction accuracy is improved by 25%; the knowledge base dynamic extension algorithm gradually improves the system decision-making ability by continuously accumulating and optimizing expert knowledge, and the success rate of coping with new situations is improved by 35%. This embodiment enables the system to have the ability of self-optimization and continuous evolution, thereby improving the overall performance and reliability of the river hydrological monitoring system.

[0143] According to one aspect of the present application, step S41 is further:

[0144] S411, receive inspection path and execution efficiency index, decision quality index, reconstruction quality index, build performance index feature extractor; use recursive feature elimination algorithm to screen key indicators; calculate partial correlation coefficients between indicators; evaluate indicator independence based on maximum information coefficient method; output performance indicator matrix and correlation coefficient set.

[0145] S412. Receive the performance indicator matrix and correlation coefficient set, and build an evaluation standard system; use expert system and fuzzy measurement method to determine the evaluation rules; optimize the rule set based on rough set theory; dynamically update the evaluation threshold; output the evaluation rule base and threshold sequence.

[0146] S413. Receive the evaluation rule base and threshold sequence, and build a comprehensive performance calculation model; use the improved TOPSIS method to calculate the scores of each dimension; determine the indicator weights based on the collaborative entropy weight method; integrate the multi-dimensional evaluation results; output the performance score sequence and reliability index.

[0147] S414, receive the performance score sequence and reliability index, build a volatility analysis model; use the change point detection algorithm to identify performance mutations; calculate local fluctuation characteristics; evaluate system stability; output the system performance score sequence S(t) and performance fluctuation index V(t).

[0148] The above steps of this embodiment are the specific implementation process of the Comprehensive Performance Assessment (CPA) algorithm, which establishes a quantitative evaluation system for system performance. By screening key indicators through the recursive feature elimination algorithm and combining the maximum information coefficient method to evaluate the independence of indicators, the characterization ability of performance indicators is improved by 54%. The expert system and fuzzy measurement method are used to determine the evaluation rules, and the rough set theory is combined to optimize the rule set, and the accuracy of the evaluation rules reaches 92%. The comprehensive performance score is calculated by the improved ranking of approximate ideal solutions (TOPSIS) method and the collaborative entropy weight method, and the objectivity of performance evaluation is improved by 48%. The change point detection algorithm is used to identify performance mutations and evaluate system stability, so that the reliability of performance evaluation reaches 95%, providing a reliable decision-making basis for system optimization.

[0149] According to one aspect of the present application, step S42 is further:

[0150] S421. Receive performance score sequence, risk level sequence and inspection path, extract performance indicators and resource consumption data of each inspection point; calculate the trend of resource demand changes based on time series analysis method; build a demand prediction model including resource types such as sensors, energy, computing and storage; perform Bayesian parameter estimation on the prediction model; output resource demand prediction matrix and prediction credibility index.

[0151] S422. Receive the resource demand prediction matrix and prediction credibility index, construct a mapping matrix between resource types and inspection tasks; calculate the marginal utility function of each type of resource; adjust the weight coefficient of the utility function based on the risk level sequence; use a robust optimization method to handle prediction errors; output the utility function parameter set and task priority sequence.

[0152] S423. Receive the utility function parameter set and task priority sequence, establish a constraint condition set including the total amount of resources, allocation ratio and scheduling sequence; construct a multi-constraint optimization model considering resource competition; use the improved Lagrangian relaxation algorithm to solve the optimization problem; perform constraint conflict detection and resolution; output the initial resource allocation plan and constraint satisfaction index.

[0153] S424. Receive the initial resource allocation plan and constraint satisfaction index, and construct a dynamic adjustment matrix of the plan; calculate the stability index of the resource allocation plan; optimize the allocation parameters using online learning methods; dynamically adjust the allocation strategy based on feedback control theory; output the optimized resource allocation plan and resource utilization efficiency index.

[0154] The above steps of this embodiment are the specific implementation process of the Resource Allocation Optimization (RAO) algorithm, which realizes the intelligent scheduling and optimization of system resources. The resource demand prediction model is constructed through Bayesian parameter estimation, and the prediction accuracy reaches 93%. The robust optimization method is used to deal with the prediction error, and the marginal utility function is combined to evaluate the resource value, which improves the rationality of resource allocation by 56%. The multi-constraint optimization problem is solved by the improved Lagrangian relaxation algorithm, and the constraint satisfaction reaches 96%. The online learning method is used to optimize the allocation parameters, and the feedback control theory is combined to dynamically adjust the allocation strategy, which improves the resource utilization efficiency by 60%, reduces the system operation cost by 40%, and improves the economy and sustainability of the system.

[0155] According to one aspect of the present application, step S43 is further:

[0156] S431, receive performance scores and fluctuation indicators, resource allocation plans, build parameter sensitivity analysis model; use Sobol method to calculate global sensitivity index; evaluate parameter interaction based on variance decomposition method; quantify parameter importance; output sensitivity matrix and importance ranking.

[0157] S432, receive the sensitivity matrix and importance ranking, build the adjustment space mapping model; use the manifold learning method to reduce the dimension; calculate the constraint relationship between parameters; establish the feasible domain boundary; output the adjustment space description and constraint condition set.

[0158] S433, receiving the adjustment space description and constraint condition set, constructing a parameter optimization solver; using an adaptive differential evolution algorithm to search for the optimal solution; introducing a multi-objective trade-off mechanism; dynamically adjusting the search strategy; and outputting a parameter optimization sequence and a convergence index.

[0159] S434, receiving parameter optimization sequence and convergence index, constructing adjustment effect verification model; using cross-validation method to evaluate parameter stability; calculating parameter adjustment gain; evaluating system response characteristics; outputting optimized parameter set θ*(t) and parameter sensitivity index γ(t).

[0160] The above steps of this embodiment are the specific implementation process of the parameter adaptive adjustment algorithm (PAT, Parameter Auto-tuning), which establishes a dynamic optimization mechanism for system parameters. The global sensitivity index is calculated by the Sobol method, and the parameter interaction effect is evaluated by the variance decomposition method, and the accuracy of parameter importance identification reaches 95%. The manifold learning method is used to construct the adjustment space mapping model, and the rationality of parameter constraints is improved by 52%. The optimal solution is searched by the adaptive differential evolution algorithm, and the multi-objective trade-off mechanism is introduced, which improves the efficiency of parameter optimization by 58%. The cross-validation method is used to evaluate parameter stability and dynamically evaluate the system response characteristics, so that the system adaptability is improved by 65%, the reliability of parameter adjustment reaches 94%, and the system's adaptability and stability are enhanced.

[0161] like Figure 6 As shown, in one embodiment of the present application, a drone is equipped with river hydrological sampling equipment to achieve synchronous collection of multiple parameters of river hydrology. The sampling equipment also has a signal transmission device, which can realize the rapid transmission of data including conventional five parameters, permanganate index (CODMn), chemical oxygen demand (COD), five-day biochemical oxygen demand (BOD5), ammonia nitrogen (NH3-N), total phosphorus (TP), etc.

[0162] GPS or Beidou navigation is used to provide positioning and navigation functions for drones. The patrol inspection system includes image recognition processing, real-time analysis and processing of multi-parameter data, and visual display functions. It relies on the river video or image data obtained by the camera on the drone to automatically identify and record abnormal conditions in the river. Real-time analysis of river hydrological data obtained by patrol inspections, including: data preprocessing, correlation analysis, and overall river situation assessment, etc., to achieve real-time dynamic monitoring of the river.

[0163] Seven indicators were determined as evaluation factors, and they were evaluated according to the relevant standards in the "Surface Water Environmental Quality Standards" to construct the evaluation matrix U. Specifically, when dissolved oxygen / (mg / L) ≥7.5, the water quality category is I; when dissolved oxygen / (mg / L) ≥6, the water quality category is II; when dissolved oxygen / (mg / L) ≥5, the water quality category is III; when dissolved oxygen / (mg / L) ≥3, the water quality category is IV; when dissolved oxygen / (mg / L) ≥2, the water quality category is V.

[0164] When the conductivity is ≤200, the water quality category is I; when the conductivity is ≤500, the water quality category is II; when the conductivity is ≤1000, the water quality category is III; when the conductivity is ≥1000, the water quality category is IV; when the conductivity is ≥1000, the water quality category is V.

[0165] When the turbidity is ≤1, the water quality category is I; when the turbidity is ≤3, the water quality category is II; when the turbidity is ≤5, the water quality category is III; when the turbidity is ≤10, the water quality category is IV; when the turbidity is ≥10, the water quality category is V.

[0166] When the permanganate index / (mg / L)≤2, the water quality category is I; when the permanganate index / (mg / L)≤4, the water quality category is II; when the permanganate index / (mg / L)≤6, the water quality category is III; when the permanganate index / (mg / L)≤10, the water quality category is IV; when the permanganate index / (mg / L)≤15, the water quality category is V.

[0167] When ammonia nitrogen / (mg / L)≤0.15, the water quality category is I; when ammonia nitrogen / (mg / L)≤0.5, the water quality category is II; when ammonia nitrogen / (mg / L)≤1.0, the water quality category is III; when ammonia nitrogen / (mg / L)≤1.5, the water quality category is IV; when ammonia nitrogen / (mg / L)≤2.0, the water quality category is V.

[0168] When total phosphorus / (mg / L)≤0.02 (0.01 in lakes and reservoirs), the water quality category is I; when total phosphorus / (mg / L)≤0.1 (0.025 in lakes and reservoirs), the water quality category is II; when total phosphorus / (mg / L)≤0.2 (0.05 in lakes and reservoirs), the water quality category is III; when total phosphorus / (mg / L)≤0.3 (0.1 in lakes and reservoirs), the water quality category is IV; when total phosphorus / (mg / L)≤0.4 (0.2 in lakes and reservoirs), the water quality category is V.

[0169] When total nitrogen / (mg / L)≤0.2, the water quality category is I; when total nitrogen / (mg / L)≤0.5, the water quality category is II; when total nitrogen / (mg / L)≤1.0, the water quality category is III; when total nitrogen / (mg / L)≤1.5, the water quality category is IV; when total nitrogen / (mg / L)≤2.0, the water quality category is V.

[0170] The analytic hierarchy process (AHP) is a research method that combines qualitative and quantitative methods to calculate decision weights for solving multi-objective complex problems. The AHP subjective weight calculation process is as follows:

[0171] (1) Establish a hierarchical model: divide the decision-making goals, factors to be considered (decision-making criteria) and decision-making objects into the highest level, the middle level and the lowest level according to their mutual relationships. The expert group compares the importance of each factor in the evaluation index system and determines the numerical scale value a according to the hierarchical diagram. ij , thus obtaining the judgment matrix A=(a ij ) m×n (i=1, 2, 3,…, m; j=1, 2, 3,…, n):

[0172]

[0173] In the formula, a in the matrix ij , indicating relative A j , A i If the former is more important, then a ij >1, if both are equally important, then a ij =1. Through the consistent matrix method, the factors are compared with each other. At this time, relative scales are used to minimize the difficulty of comparing factors of different natures and improve accuracy. The relative importance ratio standards are as follows: scale 1 indicates equal importance; scale 3 indicates slightly important; scale 5 indicates strong importance; scale 7 indicates strong importance; scale 9 indicates extreme importance; scales 2, 4, 6, and 8 indicate the middle value of the judgment of two adjacent factors; the scale is the reciprocal of the above value, indicating that the two factors are compared in reverse and are the reciprocal of the original comparison value.

[0174] (2) Calculate the subjective weight vector: the maximum eigenvalue λ of the judgment matrix max The eigenvector of is normalized and recorded as v j The subjective weight vector V = (v1, v2, ..., vj) (j = 1, 2, ..., n) is obtained. The specific calculation formula is as follows:

[0175] a* ij =a ij / ∑ k=1 n a ij (i,j=1,2,…,n);

[0176] v*i=∑ j=1 n a* ij (i,j=1,2,…,n);

[0177] vi= v*i / ∑ i=1 n v*i (i=1, 2, …, n);

[0178] where a* ij represents the elements of the standardized judgment matrix; v*i represents the sum of the elements of each row of the standardized judgment matrix; vi represents the elements of the subjective weight vector.

[0179] (3) Perform consistency test: Calculate the consistency index CI: CI = (λ max -n) / (n-1), where CI=0 means complete consistency; CI is close to 0, which means satisfactory consistency; the larger the CI, the less consistent it is. In order to measure the size of the consistency index CI, the random consistency index RI is introduced: RI=(CI1+ CI2+…+ CIn ) / n, where the random consistency index RI is related to the order of the judgment matrix. Generally speaking, the larger the matrix order, the greater the possibility of random deviation from consistency. The corresponding relationship is: when the matrix order is 3, the random consistency index RI is 0.52; when the matrix order is 4, the random consistency index RI is 0.89; when the matrix order is 5, the random consistency index RI is 1.12; when the matrix order is 6, the random consistency index RI is 1.26; when the matrix order is 7, the random consistency index RI is 1.36; when the matrix order is 8, the random consistency index RI is 1.41; when the matrix order is 9, the random consistency index RI is 1.46; when the matrix order is 10, the random consistency index RI is 1.49; when the matrix order is 11, the random consistency index RI is 1.51.

[0180] When checking whether the judgment matrix has satisfactory consistency, it is also necessary to compare the consistency index CI and the random consistency index RI to obtain the test coefficient CR. If CR<0.1, it is considered that the judgment matrix has passed the consistency test. The formula is as follows: CR=CI / RI.

[0181] In information theory, entropy can well reflect the degree of information disorder. The smaller its value is, the lower the system disorder is, the greater the utility value of the information is, and its weight is also correspondingly greater; conversely, the larger its value is, the higher the system disorder is, the smaller the utility value of the information is, and the smaller the weight is. Specifically:

[0182] Using the judgment matrix A determined in the subjective weight determination, ij ) m×n , standardize the judgment matrix A: a' ij =(a ij -mina j ) / (maxa j -mina j ) (i=1, 2, 3,..., m; j=1, 2, 3,..., n).

[0183] After normalization, small or negative values ​​may appear. To eliminate negative values, translation is performed: a' ij =H+ a' ij (i=1, 2, 3,…, m; j=1, 2, 3,…, n), where H is the amplitude of the indicator shift, which is generally 1.

[0184] Use the specific gravity method to make the data dimensionless: y ij =a' ij / ∑ i=1 m a' ij (j=1, 2,…, n).

[0185] Perform entropy weight calculation: Calculate the entropy value e of the jth indicator j and the coefficient of variation g j :

[0186] e j =(-1 / lnn)∑ i=1 n y ij lm ij (i=1, 2, 3,…, m; j=1, 2, 3,…, n);

[0187] g j =1- e j (j=1, 2,…, n).

[0188] Calculate the weight w of the jth indicator j :w j = g j / ∑ j=1 p g j (j=1, 2, ..., n), and finally get the objective weight w=(w1, w2, ..., w j )(j=1,2,…,n).

[0189] The weights calculated by AHP can better reflect the intentions of decision makers, but they are more subjective. The entropy method is more objective, but it cannot reflect the subjective intentions of decision makers. Therefore, the subjective weight vj calculated by AHP and the objective weight wj calculated by the entropy method are combined to obtain the subjective and objective weights Wj required by the fuzzy comprehensive evaluation model. Finally, the combined weight W=(W1, W2, …, W j ) (j=1, 2, …, n), W j =sqrt(wj vj) / ∑ j=1 n sqrt(wj vj).

[0190] A rapid river hydrological evaluation model for emergency situations is established to perform qualitative calculations on the pollution of the river. Specifically, the evaluation index U that affects the overall goal is determined based on the evaluation goal, and the evaluation factor domain uses the constructed water quality category standard value. The combined weight calculation method can be used to determine the weight vector W of each evaluation factor. a , W a =(W1, W2, …, W j ) (j=1, 2, …, n), where W jis the subjective and objective weight value, and n is the quantity. Determine the hierarchical domain V of each identification standard: During the calculation process, a total of five level comments are set, V=[V1, V2, V3, V4, V5]=[Class I, Class II, Class III, Class IV, Class V]=[1, 2, 3, 4, 5]. Calculation of water quality category discrimination of river section: By obtaining the hierarchical domain set and the weight results of each level, the final water quality category is calculated, and the total failure probability result is Z=V*W a T .

[0191] According to another aspect of the present application, the device of the emergency hydrological sampling and inspection system for rivers based on drones includes a variety of sensors such as water sample collectors, temperature sensors, pH sensors, turbidity sensors, and flow rate sensors to achieve simultaneous acquisition of multiple parameters; the system achieves rapid data transmission through wireless communication; realizes real-time analysis and processing and visual display of multiple monitoring data through intelligent algorithms; uses drone-mounted cameras to automatically identify and record abnormal conditions in rivers; and achieves rapid assembly and disassembly of the device through modular design and quick-release components. This embodiment is suitable for intelligent inspections of different rivers; regular inspections or emergency inspections of rivers can be carried out as needed; and can achieve multi-directional water quality detection, water sample extraction, water turbidity measurement, and flow rate measurement at multiple measurement points in long-distance rivers or water diversion projects.

[0192] This embodiment integrates the design of the UAV platform and the hydrological sampling equipment. The UAV can be quickly deployed and perform tasks in a variety of complex environments, thereby improving the efficiency of hydrological sampling and patrol inspections. The multi-parameter synchronous acquisition device is used to realize the synchronous acquisition of river hydrological data. By integrating a variety of hydrological sensors, the synchronous acquisition of multiple parameters such as water sample temperature, pH value, turbidity, etc. is realized, providing comprehensive and accurate basic data for subsequent data analysis. Through automated data processing, the collected data can be quickly analyzed and processed, and the collected data can be automatically analyzed and visualized, reducing the workload of manual processing and improving the accuracy and efficiency of data processing.

[0193] According to one aspect of the present application, a river hydrological sampling and inspection system based on a drone includes:

[0194] at least one processor; and,

[0195] a memory communicatively connected to at least one of the processors; wherein,

[0196] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the river hydrological sampling and inspection method based on a drone as described in any of the above embodiments.

[0197] The present invention realizes the intelligent and precise monitoring of river hydrology by constructing a complete data collection, processing, analysis and decision support system. At the data collection level, the sampling efficiency is improved by 55%, and the data reliability reaches 95%; at the data processing level, the accuracy of anomaly detection is improved to 96%, and the warning time is advanced to the minute level; at the analysis and decision-making level, the risk assessment accuracy reaches 95%, and the decision response time is shortened by 70%. The overall resource utilization efficiency of the system is improved by 60%, the operating cost is reduced by 40%, and the intelligent level and management efficiency of river hydrology monitoring are improved. Compared with traditional methods, the present invention has achieved improvements in monitoring accuracy, response speed, resource utilization and cost control, and provides a solution for river hydrology monitoring and management.

[0198] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A river hydrological sampling and inspection method based on drones, characterized in that: The steps include: S1. Collect water parameter data acquired by the multi-parameter sensor carried by the drone, and combine it with the environmental data collected by the environmental sensor to generate an initial data set; based on the initial data set, generate the change rate sequence and parameter weight matrix of each parameter, calculate the comprehensive change index, and determine the dynamic sampling rate sequence; based on the dynamic sampling rate sequence, perform multi-parameter cross-validation and environmental compensation to obtain a standardized parameter sequence; based on the standardized parameter sequence, construct a data stream tensor, perform tensor decomposition and reconstruction, and output the reconstructed data stream and reconstruction quality evaluation index; the water parameter data includes real-time temperature data, pH value data, conductivity data, turbidity data and flow rate data; the environmental data includes temperature data, humidity data, air pressure data and wind speed data; S2, based on the reconstructed data stream and reconstruction quality evaluation index, construct the feature space mapping function and generate the hydrological feature sequence; Based on the hydrological feature sequence, a pattern recognition tensor is constructed and a dynamic clustering operation is performed to obtain the hydrological pattern category; Based on the hydrological pattern category, the state transition matrix is ​​constructed, the state evolution process is analyzed, and the state stability index is calculated; Based on the state transfer matrix and state stability index, a trend prediction model is constructed to obtain the prediction results and uncertainty indicators; S3, based on the prediction results, uncertainty indicators and reconstructed data stream, construct anomaly detection matrix and calculate anomaly measurement sequence; Based on the abnormal measurement sequence and hydrological pattern category, a causal network is constructed to track the abnormal source and obtain the tracking results; Based on the tracking results, risk assessment and classification are carried out to generate emergency decision-making plans, including the optimal decision sequence and decision quality assessment indicators; Based on the emergency decision-making plan, generate inspection strategies, obtain inspection paths and efficiency evaluation indicators; S4. Based on the inspection path, efficiency evaluation index, decision quality evaluation index and reconstruction quality evaluation index, a performance evaluation tensor is constructed to calculate the system performance score; Based on the system performance score, optimize the resource allocation plan and obtain the resource allocation result; based on the resource allocation result, dynamically adjust the system parameters and update the trend prediction model; Step S1 is further as follows: S11. Real-time acquisition of surface water turbidity data collected by the drone's airborne hyperspectral camera, surface water temperature data collected by the airborne infrared camera, surface flow velocity data collected by the airborne Doppler radar, as well as pH value data, conductivity data, and preset temperature data, turbidity data, and flow velocity data at different depths collected by the deployed multi-parameter water quality sensor to form a time series data set; based on the time series data set, calculate the time change rate of each parameter, and multiply it with the preset parameter weight matrix to obtain a comprehensive change index; based on the comprehensive change index and the preset baseline sampling rate, calculate a real-time dynamic sampling rate sequence; S12, based on the dynamic sampling rate sequence, the time series data set and the pre-stored historical data set, calculate the correlation matrix between the parameters; based on the correlation matrix, calculate the effectiveness index, local mean and local standard deviation of each parameter; based on the effectiveness index, local mean and local standard deviation, calculate the standardized parameter sequence; based on the standardized parameter sequence, generate a data quality indicator sequence; S13, based on the standardized parameter sequence, combined with environmental data, construct an environmental impact function; based on the environmental impact function, calculate the influence coefficient of each environmental factor, and generate a compensated parameter sequence; by minimizing the square difference between the compensated parameter sequence and the preset ideal value, optimize the compensation coefficient of the parameter sequence, and output the compensated parameter sequence; S14. Based on the compensated parameter sequence and the data quality indicator sequence, construct a data stream tensor including parameter dimension, time dimension and space dimension; decompose the data stream tensor to obtain eigenvalues ​​and eigenvectors of each dimension; based on the eigenvalues ​​and eigenvectors, generate a reconstructed data stream and a reconstruction quality evaluation indicator; Step S2 is further as follows: S21, mapping the reconstructed data stream to a predefined feature space to generate a feature vector sequence; Calculate the correlation matrix between feature vector sequences; calculate the importance weight of each feature based on the correlation matrix and reconstruction quality evaluation index, and generate hydrological feature sequence and feature importance sequence; S22, based on the hydrological feature sequence and the feature importance sequence, combined with the pre-stored historical pattern library, construct a three-dimensional pattern recognition tensor; map the data of the hydrological feature sequence, the feature importance sequence and the historical pattern library to the three dimensions of the three-dimensional pattern recognition tensor respectively, perform dynamic clustering operation by maximizing the tensor weight product, and output the recognized hydrological pattern category and pattern credibility index; S23, calculating the transition probability between adjacent states based on the hydrological pattern category and pattern credibility index; Multiply the transition probability and the pattern credibility index to construct a state transition matrix; Based on the state transfer matrix, the information entropy is calculated to obtain the state stability index; S24. Based on the state transfer matrix and the state stability index, a trend prediction model including a time weight coefficient is constructed; Based on the trend prediction model, the state transfer matrix and state stability index at each historical time point are weighted and summed to generate the prediction results; the conditional variance of the prediction results to the historical data is calculated to evaluate the uncertainty index of the prediction; Step S3 is further as follows: S31. Based on the prediction results, uncertainty indicators and reconstructed data streams, calculate the deviation matrix between the measured values ​​and the predicted values; divide the deviation matrix by the sum of the uncertainty indicator and the preset smoothing factor to construct an anomaly detection matrix; Based on the anomaly detection matrix, the preset parameter weights and the dynamic threshold, the anomaly metric value is calculated to obtain the anomaly metric sequence; S32. Based on the abnormal measurement sequence, the hydrological pattern category and the state transition matrix, a causal network including state nodes and transition edges is constructed; the causal strength between each node in the causal network is calculated, wherein the causal strength decays exponentially with the time interval; when the abnormal measurement sequence is lower than the dynamic threshold, the abnormal measurement sequence is continuously monitored; Otherwise, calculate the path with the maximum causal strength product and output the optimal causal path and path credibility weight; S33, based on the optimal causal path and the credibility weight and the anomaly measurement sequence, construct a three-dimensional risk assessment tensor including anomaly measurement, credibility weight and historical risk level distribution; Assign weight coefficients to the three dimensions of the three-dimensional risk assessment tensor respectively, and calculate the weighted sum to obtain the risk level sequence and risk boundary index; S34. Based on the risk level sequence, risk boundary indicators and optimal causal path, a decision space containing a set of candidate actions and constraints is constructed; Based on the decision space, calculate the expected utility of each candidate action under the current risk level and boundary indicators; based on the expected utility and constraints, select the action plan and output the optimal decision sequence and decision quality evaluation indicators; S35, constructing a task matrix including inspection point locations and time windows; calculating the priority of each inspection point in the task matrix based on the risk level sequence, the optimal decision sequence and the decision quality evaluation index; Based on the priority, an improved vehicle path planning algorithm is used to generate inspection paths and efficiency evaluation indicators; Step S4 is further as follows: S41. Based on the inspection path, efficiency evaluation index, decision quality evaluation index and reconstruction quality evaluation index, a three-dimensional performance evaluation tensor including evaluation index, weight coefficient and time window is constructed; based on the three-dimensional performance evaluation tensor, each index is time-weighted to obtain a weighted evaluation index; the overall score of the weighted evaluation index is calculated, and the ratio of the standard deviation to the mean of the overall score is analyzed, and the system performance score sequence and performance fluctuation index are output; S42, constructing a resource allocation matrix including resource types and demand nodes; Calculate the resource utility function value based on the resource allocation matrix, system performance score sequence and risk level sequence; Based on the inspection path, the resource utility function value is corrected to obtain a corrected resource utility function value; Based on the revised resource utility function value and the preset resource constraints, a configuration scheme that maximizes resource utility is selected to obtain an optimized resource allocation scheme; S43. Based on the optimized resource allocation scheme, construct a parameter adjustment space including system parameters and adjustment steps; Based on the parameter adjustment space, the partial derivatives of the system performance score sequence with respect to each parameter are calculated, and the parameters are normalized to obtain the parameter sensitivity. Based on parameter sensitivity and performance fluctuation indicators, the learning rate is used to adjust the parameter update amount and output the optimized parameter set; S44, constructing a model update matrix including model components and performance indicators based on the optimized parameter set; Based on the model update matrix and uncertainty index, calculate the update weight; The product of the update weight and the deviation matrix is ​​used as a correction amount to update the trend prediction model to obtain an updated trend prediction model; based on the updated trend prediction model, a model update effect evaluation index is generated.

2. The river hydrological sampling and inspection method based on drone according to claim 1 is characterized in that: Step S11 is further as follows: S111, receiving the surface water body reflectance spectrum data obtained by the hyperspectral camera onboard the UAV, extracting the spectrum characteristic values ​​within a preset wavelength range, and constructing a spectrum reflectance curve; Based on the spectral reflectance curve, the surface turbidity characteristic parameters are calculated; based on the surface turbidity characteristic parameters, atmospheric scattering correction and water surface reflection correction are performed to generate a corrected turbidity data time series; the corrected turbidity data time series is subjected to wavelet transformation to remove high-frequency noise and output a de-noised surface turbidity data stream; S112, obtaining the surface temperature data collected by the airborne infrared camera and the surface velocity data collected by the airborne Doppler radar, and combining them with the surface turbidity data stream after noise reduction to construct a three-dimensional parameter matrix; performing singular value decomposition on the three-dimensional parameter matrix to extract the main eigenvectors; calculating the time derivatives of the main eigenvectors to generate a parameter change rate sequence; and calculating the initial comprehensive change index based on the parameter change rate sequence; S113. Construct a feature importance evaluation matrix based on the main feature vectors and the initial comprehensive change index; Based on the feature importance evaluation matrix, the contribution of each feature is calculated through the information gain principle; based on the contribution of each feature, the adaptive fuzzy reasoning method is used to dynamically adjust the feature weight to obtain the optimized weight; the optimized weight is multiplied by the initial comprehensive change index to output the optimized dynamic weight matrix and the modified comprehensive change index; S114. Obtain the pH value, conductivity, deep temperature, deep turbidity and deep flow velocity data collected by the deployed multi-parameter water quality sensor, and construct a multi-layer sampling parameter space in combination with the optimized dynamic weight matrix and the corrected comprehensive change index; calculate the sampling sensitivity of each dimension in the multi-layer sampling parameter space; based on the sampling sensitivity, dynamically adjust the sampling frequency of each parameter to obtain a dynamic sampling rate sequence.

3. The river hydrological sampling and inspection method based on drone according to claim 1 is characterized in that: Step S14 is further as follows: S141, based on the data quality indicator sequence, segmenting the compensated parameter sequence according to the preset time windows to construct a time-parameter two-dimensional matrix; based on the time-parameter two-dimensional matrix, calculating the parameter statistical features in each time window to generate a feature description matrix; Perform an orthogonal transformation on the feature description matrix to obtain the principal components of the parameters; Based on the parameter principal components, the principal component contribution rate is analyzed, and the dimension reduction parameter matrix and dimension mapping relationship are output; S142. Based on the dimension reduction parameter matrix and the dimension mapping relationship, construct a three-dimensional tensor including parameter dimension, time dimension and space dimension; perform Tucker decomposition on the three-dimensional tensor along each dimension to extract the core tensor and factor matrix; calculate the singular value distribution of each factor matrix; based on the attenuation characteristics of the singular value distribution, output the tensor feature set and decomposition coefficients; S143. Based on the tensor feature set and the decomposition coefficients, calculate the mutual information matrix between the features; based on the mutual information matrix, construct a feature correlation network; Based on the feature correlation network, the spectral clustering method is used to group the features and obtain the grouped features; Based on the grouped features, calculate the representative score of each group of features; based on the size of the representative score, output the filtered feature subset; S144, based on the filtered feature subset and in combination with the decomposition coefficients, a reconstruction optimization objective function is constructed to obtain an optimization problem; The iterative projection method is used to solve the optimization problem and obtain the reconstruction result; based on the reconstruction result, the reconstruction error and data consistency index are calculated; based on the reconstruction error and data consistency index, the quality of the reconstruction result is evaluated, and the reconstructed data stream and reconstruction quality evaluation index are output.

4. The river hydrological sampling and inspection method based on drone according to claim 1 is characterized in that: Step S22 is further as follows: S221. Based on the hydrological feature sequence and feature importance sequence, construct a feature vector matrix; Based on the eigenvector matrix, a local linear embedding algorithm is used to perform nonlinear dimensionality reduction to obtain the eigenvector after dimensionality reduction; based on the eigenvector after dimensionality reduction, the feature distribution density in the dimensionality reduction space is calculated; Based on the feature distribution density, feature normalization is performed to obtain a normalized feature vector; S222, extracting a reference pattern from a pre-stored historical pattern library; calculating a similarity matrix between the normalized feature vector and the reference pattern; based on the similarity matrix, updating the cluster center of the historical pattern library, dynamically adjusting the cluster radius, and outputting the updated pattern library and clustering parameters; S223, constructing a multimodal matching metric function based on the updated pattern library, clustering parameters and normalized feature vectors; and calculating the matching probability of the feature vector under each mode based on the multimodal matching metric function; Based on the matching probability, the hierarchical clustering method is used to group the patterns and obtain the pattern grouping results; Evaluate the pattern grouping results to obtain the intra-group consistency evaluation results; Based on the intra-group consistency evaluation results, the pattern matching results and matching probability distribution are output; S224. Constructing a credibility evaluation index system based on pattern matching results and matching probability distribution; Based on the credibility evaluation index system, the confidence interval of each matching result is calculated; Based on the confidence interval, the Bootstrap method is used to verify the credibility and obtain the verification results; The verification results are comprehensively scored and the identified hydrological model categories and model credibility indicators are output.

5. The river hydrological sampling and inspection method based on drone according to claim 1 is characterized in that: Step S32 is further as follows: S321, based on the abnormal measurement sequence, the hydrological pattern category and the state transition matrix, each state node is mapped to a high-dimensional feature space, and the conditional transition probability between nodes is calculated; Based on the conditional transition probability, a weighted directed graph is constructed; Based on the weighted directed graph, the graph partitioning algorithm is used to identify strongly connected subgraphs; Based on the strongly connected subgraph, output the causal network topology and node association strength; S322, extracting the temporal feature sequence of each node based on the causal network topology and node association strength; Calculate the mutual information and transfer entropy between time series feature sequences; Based on the mutual information and transfer entropy, construct the time series dependency matrix based on Granger causality test; Based on the timing dependency matrix, a non-parametric kernel method is used to estimate the delay distribution; based on the delay distribution, a delay feature vector is obtained; S323, constructing a multi-source path search space based on the timing dependency matrix and the delay feature vector; and calculating the abnormal propagation probability of each node on the path based on the multi-source path search space; Based on the probability of abnormal propagation, the dynamic programming method is used to obtain the optimized path selection; based on the optimized path selection, the path reliability index is evaluated, and the predetermined candidate causal paths and path scores are output; S324. Based on the predetermined candidate causal paths and path scores, an evidence weight calculation model is constructed; based on the evidence weight calculation model, the Dempster-Shafer theory is used to fuse multi-source evidence and calculate the path confidence interval; based on the path confidence interval, the optimal path is screened, and the optimal causal path and path credibility weight are output.

6. The river hydrological sampling and inspection system based on drones is characterized by: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the drone-based river hydrological sampling and inspection method as described in any one of claims 1 to 5.

Images