Drip irrigation pipeline leakage point positioning method based on simulated annealing and extreme learning machine

Abstract

The invention discloses a drip irrigation pipeline leakage point positioning method based on simulated annealing and an extreme learning machine. The method comprises the following steps: S1, obtaining a standardized monitoring data set; s2, performing combination to generate a leakage correlation feature vector set; s3, constructing a dual-channel topological attention extreme learning machine; s4, starting a simulated annealing optimization module, and outputting a leakage point positioning model according to the optimal hyper-parameter set solution; s5, deploying the leakage point positioning model in an online inference engine, and outputting a leakage positioning result; and S6, when the leakage positioning result is output each time, a leakage alarm module is triggered, and the leakage positioning result, the model version identifier and the timestamp are sent to a drip irrigation system maintenance scheduling center. According to the invention, the response capability and the scheduling guidance value of the system to the sudden leakage event are obviously enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline leakage, and in particular to a method for locating leakage points in a drip irrigation pipeline based on simulated annealing and an extreme learning machine. Background Art

[0002] With the continuous development of modern agricultural water resource management technologies, drip irrigation systems have been widely deployed in large and medium-sized farmland irrigation areas due to their high water-saving efficiency and precise nutrient transfer. The stable operation of drip irrigation systems has a significant impact on crop yield and resource utilization efficiency. Among them, leak detection and positioning of drip irrigation pipelines are key tasks in the maintenance process, which are directly related to irrigation uniformity, water and fertilizer utilization rate, and operation and maintenance costs. Traditional leakage point detection methods mainly rely on manual inspections or judgment based on sensor threshold alarm mechanisms. The judgment is usually based on abnormal fluctuations in a single water pressure or flow rate value.

[0003] At present, manual inspection methods have obvious limitations: first, drip irrigation pipe networks are often distributed over large areas or in areas with complex terrain, and manual inspection efficiency is extremely low. It is also easily affected by subjective experience and may lead to missed inspections. Second, it is impossible to achieve real-time response to leakage. It is often discovered only after serious leakage or abnormal crop moisture performance, which has caused great losses.

[0004] On the other hand, in recent years, some studies have attempted to introduce machine learning models for data-driven anomaly detection, such as analyzing sensor data based on support vector machines, random forests, or shallow neural networks. However, existing models mostly rely on manually selected features or static structural configurations, making it difficult to fully explore the spatial correlation and temporal evolution of monitoring data. In particular, in the context of the nonlinear hydraulic topology and dynamic flow disturbances of drip irrigation networks, model performance is often unstable. In addition, most methods lack a systematic optimization mechanism for model hyperparameters and are prone to falling into local optimality, affecting the accuracy and robustness of leak location.

[0005] During the model deployment phase, existing methods generally fail to fully consider the geographical distribution of sensors, pipeline topology information, and operational energy consumption constraints, resulting in large spatial offsets in prediction results, or a tendency to locate leakage points in invalid or sparsely covered areas, affecting the reliability of actual maintenance scheduling. At the same time, although some systems can achieve automatic alarms, their classification granularity is coarse and they are unable to provide fine-grained confidence grading outputs based on leakage probability and spatial error. They also lack dynamic adjustment mechanisms to deal with data drift caused by environmental changes.

[0006] In summary, the current drip irrigation leakage point location technology has significant deficiencies in terms of multidimensional correlation mining of feature extraction, dynamic optimization capability of model structure, fusion accuracy of prediction results and spatial geographic information, and robustness of confidence classification. It is difficult to meet the needs of modern smart agriculture for high-precision, low-energy, and real-time leakage monitoring of drip irrigation systems. Therefore, it is urgent to explore a new method that takes into account the analysis of time series dynamic features, spatial topology perception, and model parameter optimization capabilities to improve the accuracy and practicality of leakage detection. Summary of the Invention

[0007] One purpose of the present invention is to propose a method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machines. The present invention significantly enhances the system's response capability to sudden leakage events and its scheduling guidance value.

[0008] According to an embodiment of the present invention, a method for locating a leakage point in a drip irrigation pipeline based on simulated annealing and an extreme learning machine includes the following steps:

[0009] S1. Continuously collect raw monitoring data to form a raw monitoring data set and preprocess it to obtain a standardized monitoring data set;

[0010] S2. Perform sliding time window differential analysis based on the standardized monitoring data set to generate a set of leakage-related feature vectors;

[0011] S3. Construct a dual-channel topological attention extreme learning machine;

[0012] S4. Start the simulated annealing optimization module, optimize the hyperparameters of the dual-channel topological attention extreme learning machine, use the leakage localization error on the cross-validation set as the objective function, search for the optimal hyperparameter set solution for the model structure, and output the leakage point localization model based on the optimal hyperparameter set solution;

[0013] S5. Deploy the leak location model in an online inference engine, perform real-time window segmentation on the standardized monitoring dataset, generate a real-time leakage-related feature vector set, input this into the online inference engine, and obtain a preliminary leakage probability distribution and a set of candidate leakage point coordinates. Spatially fuse the candidate leakage point coordinates with the sensor geographic coordinate information and the drip irrigation pipeline hydraulic topology map, and output the leakage location results.

[0014] S6. Each time a leakage location result is output, the leakage alarm module is triggered, and the leakage location result, model version identifier, and timestamp are sent to the drip irrigation system maintenance and scheduling center.

[0015] Optionally, the S1 includes the following steps:

[0016] S11. Deploy water pressure sensors, water flow rate sensors, and temperature sensors at multiple key nodes of the drip irrigation pipeline to build a drip irrigation pipeline sensor monitoring network and continuously collect raw monitoring data during the operation of the drip irrigation system. The raw monitoring data includes the timestamp t i , water pressure value p i , water flow rate q i and temperature value θ i Four monitoring parameters are recorded in a unified structure and form the original monitoring data set D raw ,Among them, the timestamp is used to identify the moment of monitoring data collection, the ,water pressure value is used to reflect the current pipeline pressure state, the ,water flow rate is used to reflect the change of the flow speed, and the ,temperature value is used to reflect the thermal change state inside the pipeline or ,surrounding environment;

[0017] S12. Noise filtering is performed on the original monitoring data set. Sliding analysis is performed on the local changes in water pressure, water flow rate, and temperature values based on preset smoothing time window parameters. Monitoring data points exceeding a preset abnormal fluctuation threshold are identified and removed. The monitoring data that meets the stability condition is retained to form the noise-filtered monitoring data set.

[0018] S13. For monitoring data records with missing monitoring items in the noise-filtered monitoring data set, linear interpolation operations are performed based on the timestamp sequence, and numerical completion is performed based on the adjacent valid monitoring data to form an interpolated and completed monitoring data set;

[0019] S14. Perform unified time alignment on the interpolated and completed monitoring data set, set a unified time step parameter to align and synchronize all monitoring data along the time axis, and perform weighted averaging on the collected monitoring data at each time alignment point according to the sensor spatial layout and synchronization time to obtain a time-aligned monitoring data set;

[0020] S15. Normalize the water pressure values, water flow rate, and temperature values contained in the time-aligned monitoring dataset and convert them to the numerical range [0, 1] to form a standardized monitoring dataset D. norm ,In the normalization process, the range normalization method is adopted, that is, the normalized expression is calculated by the maximum and minimum values of ,each physical parameter in the monitoring data set to maintain the ,comparability between different physical quantities.

[0021] Optionally, the S2 includes the following steps:

[0022] S21. Set the sliding time window to the standardized monitoring data set D norm Sliding partitioning is performed along the time axis. Each sliding time window covers continuous standardized monitoring data within a time range, which is used to extract local change features.

[0023] S22. In each sliding time window, based on the temporal variation of the water flow rate, calculate the normalized water flow rate difference between the current time point and the previous time point to generate the water flow rate mutation feature Δq k , used to reflect the degree of local disturbance of flow velocity caused by leakage in drip irrigation pipelines;

[0024] S23. In each sliding time window, according to the time change of the water pressure value, the standardized water pressure difference between the current time point and the previous time point is calculated to generate the water pressure change trend feature Used to describe the water pressure gradient fluctuation caused by leakage in drip irrigation pipelines;

[0025] S24. In each sliding time window, based on the local change of the temperature value and the mean of all temperature values in the sliding time window, the temperature fluctuation degree in the time period is calculated to generate the temperature anomaly feature. Used to reflect the temperature instability caused by abnormal heat exchange in the leakage area of the drip irrigation pipeline;

[0026] S25. Change the water flow rate mutation characteristic Δq k , water pressure change trend characteristics and temperature anomaly characteristics Combine in a unified order to form the leakage correlation feature vector f at the current time point k , which is used to describe the operating status characteristics of the drip irrigation pipeline at the current moment, and extract the leakage correlation feature vector f at all time points in chronological order k , construct the leakage correlation feature vector set F.

[0027] Optionally, S3 includes the following steps:

[0028] S31. Based on the leakage correlation feature vector set F and the sensor geographic coordinate set Construct the sensor space topology connection matrix A:

[0029]

[0030] Among them, A ′ Represents the spatial topological connection matrix elements between pipeline sensors, representing the mth and mth, mm ′ Whether there is a physical connection between the sensors, d0 is the connection distance threshold of the pipeline segment, c m =[x m ,y m ] is the coordinate position of the mth sensor;

[0031] S32. Calculate the graph Laplacian matrix L = DA, where D is the diagonal matrix, D m,m =∑ m′ Am,m′ , and perform eigendecomposition on the graph Laplace matrix L to obtain the pipeline topology characteristic basis matrix R≤L is the number of selected characteristic bases;

[0032] S33. Leakage associated feature vector Weighted projection of pipeline topology characteristic matrix Cascade to construct multi-view input vectors And set the number of input layer nodes n in =3+R, where s k represents the spatial state vector of the drip irrigation pipe sensor corresponding to the kth time point;

[0033] S34. Randomly generate two sets of input weight matrices, including the input weight matrix W of the first channel (1) And the input weight matrix W of the second channel (2) And the bias vector b corresponding to the first channel hidden layer (1) and the bias vector b of the second channel hidden layer (1) , set the number of hidden layer nodes to n hid , construct the dual-channel topology attention extreme learning machine hidden layer output:

[0034]

[0035] Among them, g(·) is the candidate activation function, ⊙ is the element-wise multiplication, is the pressure gradient adaptive attention vector, 1 is the all-one vector, so that the feature is evenly expanded in the direction of each hidden layer node, γ is the adjustment coefficient, and σ(·) is the Sigmoid function;

[0036] S35. Perform topological sparse gating fusion on the dual-channel hidden layer output of the dual-channel topological attention extreme learning machine and define the gating coefficient vector:

[0037] g k =tanh(ηL s k );

[0038] Among them, η is the topological gating coefficient, and the hidden layer output of the fused dual-channel topological attention extreme learning machine is calculated:

[0039]

[0040] S36. Set the output weight matrix B, corresponding to the x-coordinate and y-coordinate of the leakage point space, and define the prediction output of the dual-channel topological attention extreme learning machine:

[0041]

[0042] in, The plane coordinates of the leakage point predicted at the kth time point;

[0043] S37. Set the number of input layer nodes n in , the number of hidden layer nodes n hid , dual-channel weight matrix W (1) ,W (2) , bias vector b (1) ,b (2) , attention parameter γ, topological gating coefficient η and output weight matrix B construct the hyperparameter set Θ.

[0044] Optionally, the S4 includes the following steps:

[0045] S41. Based on water pressure change trend characteristics Calculate the average value of the absolute value of the water pressure change at all time points, and define the average value as the pipeline water pressure comprehensive gradient index PGI. When setting the initial temperature of simulated annealing, the pipeline water pressure comprehensive gradient index PGI is used as the basis, multiplied by the preset temperature proportional factor κ T , used to dynamically set the initial temperature T0 of the simulated annealing process;

[0046] S42. All parameters to be optimized of the dual-channel topological attention extreme learning machine model constitute a hyperparameter set Θ, which is used as the state vector of simulated annealing to construct an optimization search space;

[0047] S43. Establish a multi-objective optimization objective function for the leakage point location model;

[0048] S44. Based on the water flow rate mutation characteristic Δq k , calculate the velocity change degree at all time points and normalize it to form the leakage response hotspot vector, in the current hyperparameter set state Θ curr Based on the above, the ratio of the current annealing temperature to the initial temperature is used as the perturbation amplitude, and the non-uniform perturbation term is constructed in combination with the leakage response hotspot vector. The new candidate parameter state Θ is generated by weighted perturbation. ′ , a local search strategy that simulates the preferential disturbance of key leakage areas;

[0049] S45. Design an adaptive cooling strategy coupled with the water pressure variation trend. In each annealing iteration, the average absolute value of the water pressure variation characteristics of the sample where the candidate solution is located is combined with the basic attenuation factor α. A and water pressure regulation coefficient β B , dynamically update the annealing temperature T k ,When the water pressure fluctuation amplitude is small, the annealing temperature drops faster;

[0050] S46. Define the acceptance probability mechanism that is dynamically coupled with the velocity mutation characteristic. If the current candidate state of the hyperparameter set Θ′ The corresponding total optimization target error is less than the current hyperparameter set state Θ curr If the error is less than 0, it is accepted unconditionally. Otherwise, the acceptance probability is calculated based on the current annealing temperature and the average velocity mutation degree of the candidate solution in the time period, and a random number that obeys a uniform distribution is generated for comparison to decide whether to accept it.

[0051] S47. When the acceptance probability is greater than the random threshold, accept the current candidate state and update the current solution to the hyperparameter set candidate state Θ ′ If the multi-objective optimization objective function value of the current candidate state is better than the historical optimal solution, the historical optimal solution is updated synchronously. When the annealing temperature is lower than the set termination temperature threshold or there is no improvement in the optimization objective for several consecutive rounds, the simulated annealing process is terminated and the optimal hyperparameter set solution Θ is output. best ;

[0052] S48. Solve Θ with the optimal hyperparameter set best As the initialization hyperparameters, the leakage association feature vector set F and the historical leakage point coordinate label set Y are input into the dual-channel topological attention extreme learning machine model. The output weight matrix is calculated by the least squares closed-form solution to obtain the final trained drip irrigation pipeline leakage point location model M. leak .

[0053] Optionally, the multi-objective optimization objective function integrates the errors of the following three parts:

[0054] The first part is the mean square error of the leakage point location, that is, the difference in spatial coordinates between the leakage point predicted by the leakage point location model and the actual leakage point;

[0055] The second part is the topological structure error, which is the average spatial offset between the leakage point location model prediction result and its nearest known sensor location, which is used to constrain the prediction point to regress to the actual deployment area;

[0056] The third part is the sensor energy stability error, which is used to measure the degree of deviation of the leakage point location model from the low-energy sensor path during the reasoning process, and to prevent the leakage point location model from predicting the leakage point in the unstable area of the sensor operation.

[0057] The mean square error, topology error and sensor energy stability error are weighted and combined respectively, and the weight coefficients of topology error and sensor energy stability error are determined by the variance of water flow rate and temperature, respectively.

[0058] Optionally, the S5 includes the following steps:

[0059] S51. The trained leakage point location model M leak Deployed in the online inference engine, set the real-time sliding time window length wreal With step size, for the standardized monitoring data set D norm Perform real-time window segmentation to extract standardized monitoring data segments within the current window;

[0060] S52. Based on the standardized monitoring data segments in each real-time time window, calculate the velocity mutation characteristics Δq corresponding to each time point in the current window k , water pressure change trend characteristics Temperature anomaly characteristics Combined with the sensor space topology embedding vector z k , generate real-time leakage correlation feature vector set F real ;

[0061] S53. The real-time leakage associated feature vector set F real Input into the leakage point location model M deployed in the online inference engine in sequence leak , output the corresponding leakage prediction results, and obtain the candidate leakage point coordinate set at each time point

[0062] S54. Set the candidate leakage point coordinate set Y cand The preset sensor geographic coordinates set and drip irrigation pipeline hydraulic topology map G in the drip irrigation system pipe Perform spatial fusion, the fusion method includes calculating the nearest sensor node c based on Euclidean distance near (k) and correct the coordinates of the candidate leakage point by combining the hydraulically accessible paths of adjacent pipe sections in the topology map;

[0063] S55. The output result of the fused leakage point coordinates is defined as the final leakage location result Combined with the leakage probability value output by the leakage point location model and inference timestamp t k , construct the leakage location information structure, including coordinates, probability and time stamp.

[0064] Optionally, the classification rules for leakage location are as follows:

[0065] like and It is marked as a high confidence leakage point;

[0066] like and It is marked as a medium confidence leakage point;

[0067] like or It is marked as a low confidence point or an invalid positioning point;

[0068] in, is the confidence probability of the leakage point location model at the kth time point that the current predicted leakage point is a real leakage, Represents the final leakage point coordinates output by the leakage point location model and spatial fusion processing at the kth time point.

[0069] The beneficial effects of the present invention are:

[0070] (1) The present invention maps the spatial position of the sensor to a high dimension through the graph Laplacian eigenvalue matrix, and fuses it with the three-dimensional dynamic feature vector in the time window to construct an input feature representation with space-time coupling. A dual-channel parallel architecture is introduced on the structure of the extreme learning machine. One path captures the original dynamic features, and the other path constructs the attention weights through the pressure gradient to achieve adaptive enhancement of key features. A topological gating coefficient is introduced during the dual-channel fusion, and the hidden states are weighted combined through the graph structure to ensure that the model pays attention to the spatial nodes related to leakage.

[0071] (2) The present invention introduces the water pressure comprehensive gradient index into the temperature setting of simulated annealing, uses the system operation status to dynamically control the initial annealing temperature, and constructs an optimization objective function with multi-objective error as the core, covering three indicators: positioning error, topological structure offset error and energy consumption stability. During the annealing iteration process, the flow velocity perturbation is used to construct a non-uniform perturbation mechanism, focusing on perturbing the hyperparameter vector of the hot spot area where leakage occurs, and achieving a more accurate solution convergence process through the cooling strategy of water pressure control.

[0072] (3) The present invention spatially fuses the model prediction results with the geographic coordinates of the sensors in the system and the hydraulic topology map to avoid the positioning results from deviating from the effective maintenance area. At the same time, a three-layer confidence classification is set according to the predicted probability value and the Euclidean distance from the predicted point to the adjacent sensor node, and a ternary structure result containing coordinates, probability and timestamp is output, which significantly enhances the system's response capability to sudden leakage events and its scheduling guidance value. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0074] Figure 1 This is a flow chart of a method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machines proposed in the present invention. DETAILED DESCRIPTION

[0075] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0076] refer to Figure 1 A method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machines includes the following steps:

[0077] S1. Continuously collect raw monitoring data to form a raw monitoring data set and preprocess it to obtain a standardized monitoring data set;

[0078] S2. Perform sliding time window differential analysis based on the standardized monitoring data set to generate a set of leakage-related feature vectors;

[0079] S3. Construct a dual-channel topological attention extreme learning machine;

[0080] S4. Start the simulated annealing optimization module, optimize the hyperparameters of the dual-channel topological attention extreme learning machine, use the leakage localization error on the cross-validation set as the objective function, search for the optimal hyperparameter set solution for the model structure, and output the leakage point localization model based on the optimal hyperparameter set solution;

[0081] S5. Deploy the leak location model in an online inference engine, perform real-time window segmentation on the standardized monitoring dataset, generate a real-time leakage-related feature vector set, input this into the online inference engine, and obtain a preliminary leakage probability distribution and a set of candidate leakage point coordinates. Spatially fuse the candidate leakage point coordinates with the sensor geographic coordinate information and the drip irrigation pipeline hydraulic topology map, and output the leakage location results.

[0082] S6. Each time a leakage location result is output, the leakage alarm module is triggered, and the leakage location result, model version identifier, and timestamp are sent to the drip irrigation system maintenance and scheduling center.

[0083] In this embodiment, S1 includes the following steps:

[0084] S11. Deploy water pressure sensors, water flow rate sensors, and temperature sensors at multiple key nodes of the drip irrigation pipeline to build a drip irrigation pipeline sensor monitoring network and continuously collect raw monitoring data during the operation of the drip irrigation system. The raw monitoring data includes the timestamp t i , water pressure value p i , water flow rate q i and temperature value θ i Four monitoring parameters are recorded in a unified structure and form the original monitoring data set D raw ,Among them, the timestamp is used to identify the moment of monitoring data collection, the ,water pressure value is used to reflect the current pipeline pressure state, the ,water flow rate is used to reflect the change of the flow speed, and the ,temperature value is used to reflect the thermal change state inside the pipeline or ,surrounding environment;

[0085] S12. Noise filtering is performed on the original monitoring data set. Sliding analysis is performed on the local changes in water pressure, water flow rate, and temperature values based on preset smoothing time window parameters. Monitoring data points exceeding a preset abnormal fluctuation threshold are identified and removed. The monitoring data that meets the stability condition is retained to form the noise-filtered monitoring data set.

[0086] S13. For monitoring data records with missing monitoring items in the noise-filtered monitoring data set, linear interpolation operations are performed based on the timestamp sequence, and numerical completion is performed based on the adjacent valid monitoring data to form an interpolated and completed monitoring data set;

[0087] S14. Perform unified time alignment on the interpolated and completed monitoring data set, set a unified time step parameter to align and synchronize all monitoring data along the time axis, and perform weighted averaging on the collected monitoring data at each time alignment point according to the sensor spatial layout and synchronization time to obtain a time-aligned monitoring data set;

[0088] S15. Normalize the water pressure values, water flow rate, and temperature values contained in the time-aligned monitoring dataset and convert them to the numerical range [0, 1] to form a standardized monitoring dataset D. norm ,In the normalization process, the range normalization method is adopted, that is, the normalized expression is calculated by the maximum and minimum values of ,each physical parameter in the monitoring data set to maintain the ,comparability between different physical quantities.

[0089] In this embodiment, S2 includes the following steps:

[0090] S21. Set the sliding time window to the standardized monitoring data set D norm Sliding partitioning is performed along the time axis. Each sliding time window covers continuous standardized monitoring data within a time range, which is used to extract local change features.

[0091] S22. In each sliding time window, based on the temporal variation of the water flow rate, calculate the normalized water flow rate difference between the current time point and the previous time point to generate the water flow rate mutation feature Δq k , used to reflect the degree of local disturbance of flow velocity caused by leakage in drip irrigation pipelines;

[0092] S23. In each sliding time window, according to the time change of the water pressure value, the standardized water pressure difference between the current time point and the previous time point is calculated to generate the water pressure change trend feature Used to describe the water pressure gradient fluctuation caused by leakage in drip irrigation pipelines;

[0093] S24. In each sliding time window, based on the local change of the temperature value and the mean of all temperature values in the sliding time window, the temperature fluctuation degree in the time period is calculated to generate the temperature anomaly feature. Used to reflect the temperature instability caused by abnormal heat exchange in the leakage area of the drip irrigation pipeline;

[0094] S25. Change the water flow rate mutation characteristic Δq k , water pressure change trend characteristics and temperature anomaly characteristics Combine in a unified order to form the leakage correlation feature vector f at the current time point k , which is used to describe the operating status characteristics of the drip irrigation pipeline at the current moment, and extract the leakage correlation feature vector f at all time points in chronological order k , construct the leakage correlation feature vector set F.

[0095] In this embodiment, S3 includes the following steps:

[0096] S31. Based on the leakage correlation feature vector set F and the sensor geographic coordinate set Construct the sensor space topology connection matrix A:

[0097]

[0098] Among them, A ′ Represents the spatial topological connection matrix elements between pipeline sensors, representing the mth and mth, mm ′ Whether there is a physical connection between the sensors, d0 is the connection distance threshold of the pipeline segment, c m =[x m ,y m ] is the coordinate position of the mth sensor;

[0099] S32. Calculate the graph Laplacian matrix L = DA, where D is the diagonal matrix, D m,m =∑ m′ A m,m′ , and perform eigendecomposition on the graph Laplace matrix L to obtain the pipeline topology characteristic basis matrix R≤L is the number of selected characteristic bases;

[0100] S33. Leakage associated feature vector Weighted projection of pipeline topology characteristic matrix Cascade to construct multi-view input vectors And set the number of input layer nodes n in =3+R, where s k represents the spatial state vector of the drip irrigation pipe sensor corresponding to the kth time point;

[0101] S34. Randomly generate two sets of input weight matrices, including the input weight matrix W of the first channel (1) And the input weight matrix W of the second channel (2) And the bias vector b corresponding to the first channel hidden layer (1) and the bias vector b of the second channel hidden layer (1) , set the number of hidden layer nodes to n hid , construct the dual-channel topology attention extreme learning machine hidden layer output:

[0102]

[0103] Among them, g(·) is a candidate activation function, which can be a Sigmoid function, a Tanh function (hyperbolic tangent) or a ReLU function, ⊙ is element-wise multiplication, is the pressure gradient adaptive attention vector, 1 is the all-one vector, so that the feature is evenly expanded in the direction of each hidden layer node, γ is the adjustment coefficient, and σ(·) is the Sigmoid function;

[0104] S35. Perform topological sparse gating fusion on the dual-channel hidden layer output of the dual-channel topological attention extreme learning machine and define the gating coefficient vector:

[0105] g k =tanh(ηL s k );

[0106] Among them, η is the topological gating coefficient, and the hidden layer output of the fused dual-channel topological attention extreme learning machine is calculated:

[0107]

[0108] S36. Set the output weight matrix B, corresponding to the x-coordinate and y-coordinate of the leakage point space, and define the prediction output of the dual-channel topological attention extreme learning machine:

[0109]

[0110] in, The plane coordinates of the leakage point predicted at the kth time point;

[0111] S37. Set the number of input layer nodes n in , the number of hidden layer nodes n hid , dual-channel weight matrix W (1) ,W (2) , bias vector b (1) ,b (2) , attention parameter γ, topological gating coefficient η and output weight matrix B construct the hyperparameter set Θ.

[0112] In this embodiment, S4 includes the following steps:

[0113] S41. Based on water pressure change trend characteristics Calculate the average value of the absolute value of the water pressure change at all time points, and define the average value as the pipeline water pressure comprehensive gradient index PGI. When setting the initial temperature of simulated annealing, the pipeline water pressure comprehensive gradient index PGI is used as the basis, multiplied by the preset temperature proportional factor κ T , used to dynamically set the initial temperature T0 of the simulated annealing process;

[0114] S42. All parameters to be optimized of the dual-channel topological attention extreme learning machine model constitute a hyperparameter set Θ, which is used as the state vector of simulated annealing to construct an optimization search space;

[0115] S43. Establish a multi-objective optimization objective function for the leakage point location model;

[0116] S44. Based on the water flow rate mutation characteristic Δq k , calculate the velocity change degree at all time points and normalize it to form the leakage response hotspot vector, in the current hyperparameter set state Θ curr Based on the above, the ratio of the current annealing temperature to the initial temperature is used as the perturbation amplitude, and the non-uniform perturbation term is constructed in combination with the leakage response hotspot vector. The new candidate parameter state Θ is generated by weighted perturbation. ′ , a local search strategy that simulates the preferential disturbance of key leakage areas;

[0117] S45. Design an adaptive cooling strategy coupled with the water pressure variation trend. In each annealing iteration, the average absolute value of the water pressure variation characteristics of the sample where the candidate solution is located is combined with the basic attenuation factor α. A and water pressure regulation coefficient β B , dynamically update the annealing temperature T k ,When the water pressure fluctuation amplitude is small, the annealing temperature drops faster;

[0118] S46. Define the acceptance probability mechanism that is dynamically coupled with the velocity mutation characteristic. If the current candidate state of the hyperparameter set Θ ′ The corresponding total optimization target error is less than the current hyperparameter set state Θ curr If the error is less than 0, it is accepted unconditionally. Otherwise, the acceptance probability is calculated based on the current annealing temperature and the average velocity mutation degree of the candidate solution in the time period, and a random number that obeys a uniform distribution is generated for comparison to decide whether to accept it.

[0119] S47. When the acceptance probability is greater than the random threshold, accept the current candidate state and update the current solution to the hyperparameter set candidate state Θ ′If the multi-objective optimization objective function value of the current candidate state is better than the historical optimal solution, the historical optimal solution is updated synchronously. When the annealing temperature is lower than the set termination temperature threshold or there is no improvement in the optimization objective for several consecutive rounds, the simulated annealing process is terminated and the optimal hyperparameter set solution Θ is output. best ;

[0120] S48. Solve Θ with the optimal hyperparameter set best As the initialization hyperparameters, the leakage association feature vector set F and the historical leakage point coordinate label set Y are input into the dual-channel topological attention extreme learning machine model. The output weight matrix is calculated by the least squares closed-form solution to obtain the final trained drip irrigation pipeline leakage point location model M. leak .

[0121] In this embodiment, the multi-objective optimization objective function integrates the errors of the following three parts:

[0122] The first part is the mean square error of the leakage point location, that is, the difference in spatial coordinates between the leakage point predicted by the leakage point location model and the actual leakage point;

[0123] The second part is the topological structure error, which is the average spatial offset between the leakage point location model prediction result and its nearest known sensor location, which is used to constrain the prediction point to regress to the actual deployment area;

[0124] The third part is the sensor energy stability error, which is used to measure the degree of deviation of the leakage point location model from the low-energy sensor path during the reasoning process, and to prevent the leakage point location model from predicting the leakage point in the unstable area of the sensor operation.

[0125] The mean square error, topology error and sensor energy stability error are weighted and combined respectively, and the weight coefficients of topology error and sensor energy stability error are determined by the variance of water flow rate and temperature, respectively.

[0126] In one embodiment, the water pressure variation trend characteristic sequence is used. Calculate the overall water pressure fluctuation level of the drip irrigation pipeline and define the comprehensive gradient index of pipeline water pressure as:

[0127]

[0128] Among them, PGI represents the gradient average used to measure the degree of change in the overall water pressure of the pipeline, M is the time series length of the standardized monitoring data, Represents the absolute value of the water pressure change at the kth time point; and the initial annealing temperature is set accordingly:

[0129] T0=κ T PGI;

[0130] Where T0 is the initial temperature of the simulated annealing process, κ T The temperature scaling factor is set empirically, and the initial temperature value is dynamically adjusted in combination with the water pressure fluctuation level, making the annealing search sensitive to the real pressure disturbance of the system.

[0131] All parameters to be optimized of the dual-channel topological attention extreme learning machine are encapsulated as a hyperparameter set Θ = {W (1) ,W (2) ,b (1) ,b (2) ,B,γ,η};

[0132] Construct a multi-objective optimization leakage localization cost function to comprehensively evaluate the advantages and disadvantages of different hyperparameter combinations:

[0133] E(Θ)=E mse (Θ)+λ L E topo (Θ)+μ E E energy (Θ);

[0134] in, represents the mean square error between the model predicted coordinates and the actual leakage coordinates, where c ne (k) represents the spatial coordinate of the sensor closest to the prediction result, which is used to constrain the spatial rationality of the leakage prediction result; in It represents the sensor energy stability error of the mth node under the current network conditions, which is used to reflect the operation stability of the sensor node. where σ q , σ θ are the normalized series variances of flow rate and temperature, respectively, which are used to adaptively weight the degree of constraint on energy dimension and spatial structure.

[0135] According to the flow rate mutation characteristic Δq k Construct the leakage response hotspot vector:

[0136]

[0137] Among them, P hot is the normalized average mutation response vector, which is used to emphasize the leakage sensitivity area in the current time period; on this basis, the current solution Θ curr Add response perturbation:

[0138]

[0139] Among them, ε k =T k / T0 is the disturbance amplitude factor, represents the standard normal distribution, and diag(·) represents the generation of a diagonal matrix, which is used to guide the perturbation to focus on the mutation-significant feature dimension.

[0140] An adaptive cooling strategy linked to the water pressure gradient fluctuation level is adopted, and the annealing temperature update formula is defined as:

[0141]

[0142] Among them, α A ∈(0,1) is the basic attenuation coefficient, β B >0 is the water pressure disturbance adjustment factor, The average absolute value of the water pressure change characteristics corresponding to the k-th round of accepted solutions is used to accelerate the cooling process when the system water pressure is stable and enhance the search efficiency.

[0143] Define a solution acceptance probability function that is dynamically coupled to the flow velocity characteristics:

[0144]

[0145] in, is the average velocity mutation degree in the time period corresponding to the candidate solution of the kth iteration, δ D is the mutation response adjustment coefficient, which increases the search tolerance in leakage risk areas;

[0146] According to the solution acceptance probability function P acc and generate random numbers Compare, if r <P acc , then accept the candidate solution, that is, update Θ curr =Θ ′ ; If E(Θ ′ ) <E(Θ best ), then record the current final optimal parameter Θ best =Θ ′ ; When T is satisfied k ≤T end or continuous z max When there is no improvement in the round, the simulated annealing process is terminated;

[0147] The final optimal parameter Θ best As initialization parameters, combined with the leakage correlation feature vector set F and the historical coordinate label set Input to the dual-channel topological attention extreme learning machine described in claim 4 to solve the output weight matrix Among them, H represents the fusion hidden layer output matrix, Y is the actual leakage point coordinate set, and finally forms the leakage point positioning model M of the drip irrigation pipeline after training. leak .

[0148] In this embodiment, S5 includes the following steps:

[0149] S51. The trained leakage point location model M leak Deployed in the online inference engine, set the real-time sliding time window length w real With step size, for the standardized monitoring data set D norm Perform real-time window segmentation to extract standardized monitoring data segments within the current window;

[0150] S52. Based on the standardized monitoring data segments in each real-time time window, calculate the velocity mutation characteristics Δq corresponding to each time point in the current window k , water pressure change trend characteristics Temperature anomaly characteristics Combined with the sensor space topology embedding vector z k , generate real-time leakage correlation feature vector set F real ;

[0151] S53. The real-time leakage associated feature vector set F real Input into the leakage point location model M deployed in the online inference engine in sequence leak , output the corresponding leakage prediction results, and obtain the candidate leakage point coordinate set at each time point

[0152] S54. Set the candidate leakage point coordinate set Y cand The preset sensor geographic coordinates set and drip irrigation pipeline hydraulic topology map G in the drip irrigation system pipe Perform spatial fusion, the fusion method includes calculating the nearest sensor node c based on Euclidean distance near (k) and correct the coordinates of the candidate leakage point by combining the hydraulically accessible paths of adjacent pipe sections in the topology map;

[0153] S55. The output result of the fused leakage point coordinates is defined as the final leakage location result Combined with the leakage probability value output by the leakage point location model and inference timestamp t k , construct the leakage location information structure, including coordinates, probability and time stamp.

[0154] In this embodiment, the classification rules for leakage location are as follows:

[0155] like and It is marked as a high confidence leakage point;

[0156] like and It is marked as a medium confidence leakage point;

[0157] like or It is marked as a low confidence point or an invalid positioning point;

[0158] in, is the confidence probability of the leakage point location model at the kth time point that the current predicted leakage point is a real leakage, Represents the final leakage point coordinates output by the leakage point location model and spatial fusion processing at the kth time point.

[0159] Example 1:

[0160] In mid-June 2024, during routine summer irrigation operations at Farm A, some cotton planting areas experienced leaf drying and slowed growth. Onsite maintenance personnel suspected leakage in the drip irrigation pipeline in this area. However, due to the large area of 186 mu (approximately 166 acres), the complexity of the pipelines, and the varying depths of burial, the traditional method of relying on manual inspections combined with the judgment of sudden drops in water pressure could not quickly locate the problem.

[0161] To address this issue, the farm partnered with an agricultural information technology service company to conduct a pilot run of the system in the area. The farm already had 78 dual-mode water pressure and flow rate sensors deployed across the main pipes, branch pipes, and end-of-pipe drippers. This pilot also included 14 temperature sensors to measure local pipe wall temperature fluctuations. The entire drip irrigation system operates three times daily, each lasting 110 minutes, with data sampling every 20 seconds.

[0162] After collecting data continuously for 7 days (June 12 to June 18, 2024), a total of 2,139,600 valid monitoring data were obtained, covering four-dimensional parameters: timestamp, water pressure value, flow rate and temperature value. The system first preprocessed the raw data, including high-frequency noise filtering (applying a median filter with a sliding window size of 5 minutes), linear interpolation repair of missing data, and normalization of the unified time step (set to 60 seconds). Finally, a standardized monitoring data set was obtained, containing a total of 21,396 groups of time-aligned records.

[0163] Based on a standardized data set, the system calculates feature triplets for the flow velocity, water pressure, and temperature differences within each 60-second sliding time window, and constructs a spatial topological connection matrix A (with the threshold d0 set to 8 meters) based on the sensor's geographic coordinates. Based on this, the graph Laplacian matrix L is generated and the feature basis matrix Φ (R=4) is extracted, ultimately forming a multidimensional nested input vector for extreme learning machine training.

[0164] The model uses a dual-channel architecture during initialization. Channel 1 processes the three basic features, while channel 2 automatically weights the water pressure gradient amplitude using an attention mechanism and employs a topological gating strategy for fusion. A simulated annealing algorithm dynamically optimizes the following hyperparameters: the number of hidden layer nodes, the topological gating coefficient η, the attention adjustment coefficient γ, and the channel weight initialization matrix. The objective function comprehensively considers three factors: leakage point prediction error, spatial offset error, and energy stability error.

[0165] The model training data selected the data of the past 5 days as the training set (a total of 15,284 groups), and the remaining 2 days as the validation set (6,112 groups). Eleven actual leakage points were manually marked from the maintenance records for model evaluation.

[0166] During the verification period, a horizontal comparison was conducted with the farm's traditional baseline methods (threshold differential alarm method, static ELM model). Key performance data are shown in the following table:

[0167] Table 1 Comparison of data between the present invention and the static ELM model

[0168] Indicator Method of the present invention Static ELM model Threshold alarm method Average positioning error (m) 0.84 2.31 3.94 Accuracy (Top-1) 92.3％ 76.1％ 59.4％ Mean reversion residuals 0.12 0.47 0.83 Number of high-confidence identification points 9 5 2 Average leakage response time (min) 12.4 29.5 43.2

[0169] At 10:30 am on June 19, 2024, the system automatically identified a high-probability leakage in the area near sensor number #32. The predicted coordinates were [x=231.4m, y=118.7m], with a prediction confidence of 0.87. This coordinate point is only 0.7 meters away from the nearest sensor node, making it a high-confidence point. Operations and maintenance personnel went to the site for excavation verification and found a crack about 3.2 cm long in the pipe wall at this point, with visible signs of water seepage.

[0170] At 12:45, the system identified another medium-confidence point at the end of the pipe network with a confidence level of 0.65 and a coordinate deviation of 1.6 meters. Although the point deviated from the sensor, it was automatically corrected through the spatial fusion module and the pipe network map Gpipe. The prediction range accurately covered the pipe section where the dripper was loose and disconnected. Although this problem did not cause a large-scale water supply failure, if it continued to develop, it would easily lead to uneven water supply at the tail end.

[0171] A total of 13 leakage points were identified during this field trial, 12 of which were confirmed to exist, with a prediction accuracy of 92.3%, significantly better than the 59.4% of the traditional method. During the high temperature period in the afternoon when temperature disturbances were significant (such as 13:00-15:00 on June 17), this method captured weak heat flow disturbances and effectively located two low-speed leakage points that were not alarmed by the traditional method.

[0172] The present invention performs high-dimensional mapping of the spatial position of the sensor through the graph Laplacian eigenvalue matrix, and fuses it with the three-dimensional dynamic feature vector in the time window to construct an input feature representation with space-time coupling. A dual-channel parallel architecture is introduced into the structure of the extreme learning machine. One path captures the original dynamic features, and the other path constructs attention weights through the pressure gradient to achieve adaptive enhancement of key features. Topological gating coefficients are introduced during dual-channel fusion, and hidden states are weightedly combined through the graph structure to ensure that the model focuses on spatial nodes related to leakage.

[0173] The present invention introduces the comprehensive water pressure gradient index into the temperature setting of simulated annealing, uses the system operation status to dynamically control the initial annealing temperature, and constructs an optimization objective function with multi-objective error as the core, covering three indicators: positioning error, topological structure offset error and energy consumption stability. During the annealing iteration process, the flow velocity perturbation is used to construct a non-uniform perturbation mechanism, focusing on perturbing the hyperparameter vectors of the hot spot area where leakage occurs, and achieving a more accurate solution convergence process through the cooling strategy of water pressure control.

[0174] The present invention spatially fuses and corrects the model prediction results with the geographic coordinates of the sensors in the system and the hydraulic topology map to prevent the positioning results from deviating from the effective maintenance area. At the same time, a three-layer confidence classification is set according to the predicted probability value and the Euclidean distance from the predicted point to the adjacent sensor node, and a ternary structure result containing coordinates, probability and timestamp is output, which significantly enhances the system's response capability to sudden leakage events and its scheduling guidance value.

[0175] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machine, characterized in that: The steps include: S1. Continuously collect raw monitoring data to form a raw monitoring data set and preprocess it to obtain a standardized monitoring data set; S2. Perform sliding time window differential analysis based on the standardized monitoring data set to generate a set of leakage-related feature vectors; S3. Construct a dual-channel topological attention extreme learning machine; S4. Start the simulated annealing optimization module, optimize the hyperparameters of the dual-channel topological attention extreme learning machine, use the leakage localization error on the cross-validation set as the objective function, search for the optimal hyperparameter set solution for the model structure, and output the leakage point localization model based on the optimal hyperparameter set solution; S5. Deploy the leak location model in an online inference engine, perform real-time window segmentation on the standardized monitoring dataset, generate a real-time leakage-related feature vector set, input this into the online inference engine, and obtain a preliminary leakage probability distribution and a set of candidate leakage point coordinates. Spatially fuse the candidate leakage point coordinates with the sensor geographic coordinate information and the drip irrigation pipeline hydraulic topology map, and output the leakage location results. S6. Each time a leakage location result is output, the leakage alarm module is triggered, and the leakage location result, model version identifier, and timestamp are sent to the drip irrigation system maintenance and scheduling center.

2. The method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machine according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Deploy water pressure sensors, water flow rate sensors, and temperature sensors at multiple key nodes of the drip irrigation pipeline to build a drip irrigation pipeline sensor monitoring network and continuously collect raw monitoring data during the operation of the drip irrigation system. The raw monitoring data includes the timestamp t i , water pressure value p i , water flow rate q i and temperature value θ i Four monitoring parameters are recorded in a unified structure and form the original monitoring data set D raw ; S12. Noise filtering is performed on the original monitoring data set. Sliding analysis is performed on the local changes in water pressure, water flow rate, and temperature values based on preset smoothing time window parameters. Monitoring data points exceeding a preset abnormal fluctuation threshold are identified and removed. The monitoring data that meets the stability condition is retained to form the noise-filtered monitoring data set. S13. For monitoring data records with missing monitoring items in the noise-filtered monitoring data set, linear interpolation operations are performed based on the timestamp sequence, and numerical completion is performed based on the adjacent valid monitoring data to form an interpolated and completed monitoring data set; S14. Perform unified time alignment on the interpolated and completed monitoring data set, set a unified time step parameter to align and synchronize all monitoring data along the time axis, and perform weighted averaging on the collected monitoring data at each time alignment point according to the sensor spatial layout and synchronization time to obtain a time-aligned monitoring data set; S15. Normalize the water pressure values, water flow rate, and temperature values contained in the time-aligned monitoring dataset and convert them to the numerical range [0, 1] to form a standardized monitoring dataset D. norm .

3. The method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machine according to claim 2, characterized in that: The S2 comprises the following steps: S21. Set the sliding time window to the standardized monitoring data set D norm Sliding division is performed according to the time axis, and each sliding time window covers continuous standardized monitoring data within the time length range; S22. In each sliding time window, based on the temporal variation of the water flow rate, calculate the normalized water flow rate difference between the current time point and the previous time point to generate the water flow rate mutation feature Δq k , used to reflect the degree of local disturbance of flow velocity caused by leakage in drip irrigation pipelines; S23. In each sliding time window, according to the time change of the water pressure value, the standardized water pressure difference between the current time point and the previous time point is calculated to generate the water pressure change trend feature Used to describe the water pressure gradient fluctuation caused by leakage in drip irrigation pipelines; S24. In each sliding time window, based on the local change of the temperature value and the mean of all temperature values in the sliding time window, the temperature fluctuation degree in the time period is calculated to generate the temperature anomaly feature. Used to reflect the temperature instability caused by abnormal heat exchange in the leakage area of the drip irrigation pipeline; S25. Change the water flow rate mutation characteristic Δq k , water pressure change trend characteristics and temperature anomaly characteristics Combine in a unified order to form the leakage correlation feature vector f at the current time point k , which is used to describe the operating status characteristics of the drip irrigation pipeline at the current moment, and extract the leakage correlation feature vector f at all time points in chronological order k , construct the leakage correlation feature vector set F.

4. The method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machine according to claim 3, characterized in that: The S3 includes the following steps: S31. Based on the leakage correlation feature vector set F and the sensor geographic coordinate set Construct the sensor space topology connection matrix A: Among them, A ′ Represents the spatial topological connection matrix elements between pipeline sensors, representing the mth and mth,m m ′ Whether there is a physical connection between the sensors, d0 is the connection distance threshold of the pipeline segment, c m =[x m ,y m ] is the coordinate position of the mth sensor; S32. Calculate the graph Laplacian matrix L = DA, where D is the diagonal matrix, D m,m =∑ m′ A m,m′ , and perform eigendecomposition on the graph Laplace matrix L to obtain the pipeline topology characteristic basis matrix R≤L is the number of selected characteristic bases; S33. Leakage associated feature vector Weighted projection of pipeline topology characteristic matrix Cascade to construct multi-view input vectors And set the number of input layer nodes n in =3+R, where s k represents the spatial state vector of the drip irrigation pipe sensor corresponding to the kth time point; S34. Randomly generate two sets of input weight matrices, including the input weight matrix W of the first channel (1) And the input weight matrix W of the second channel (2) And the bias vector b corresponding to the first channel hidden layer (1) and the bias vector b of the second channel hidden layer (1) , set the number of hidden layer nodes to n hid , construct the dual-channel topology attention extreme learning machine hidden layer output: Among them, g(·) is the candidate activation function, ⊙ is the element-wise multiplication, is the pressure gradient adaptive attention vector, 1 is the all-one vector, so that the feature is evenly expanded in the direction of each hidden layer node, γ is the adjustment coefficient, and σ(·) is the Sigmoid function; S35. Perform topological sparse gating fusion on the dual-channel hidden layer output of the dual-channel topological attention extreme learning machine and define the gating coefficient vector: g k =tanh(ηLs k ); Among them, η is the topological gating coefficient, and the hidden layer output of the fused dual-channel topological attention extreme learning machine is calculated: S36. Set the output weight matrix B, corresponding to the x-coordinate and y-coordinate of the leakage point space, and define the prediction output of the dual-channel topological attention extreme learning machine: in, The plane coordinates of the leakage point predicted at the kth time point; S37. Set the number of input layer nodes n in , the number of hidden layer nodes n hid , dual-channel weight matrix W (1) ,W (2) , bias vector b (1) ,b (2) , attention parameter γ, topological gating coefficient η and output weight matrix B construct the hyperparameter set Θ.

5. The method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machine according to claim 1, characterized in that: The S4 comprises the following steps: S41. Based on water pressure change trend characteristics Calculate the average value of the absolute value of the water pressure change at all time points, and define the average value as the pipeline water pressure comprehensive gradient index PGI. When setting the initial temperature of simulated annealing, the pipeline water pressure comprehensive gradient index PGI is used as the basis, multiplied by the preset temperature proportional factor κ T , used to dynamically set the initial temperature T0 of the simulated annealing process; S42. All parameters to be optimized of the dual-channel topological attention extreme learning machine model constitute a hyperparameter set Θ, which is used as the state vector of simulated annealing to construct an optimization search space; S43. Establish a multi-objective optimization objective function for the leakage point location model; S44. Based on the water flow rate mutation characteristic Δq k , calculate the velocity change degree at all time points and normalize it to form the leakage response hotspot vector, in the current hyperparameter set state Θ curr Based on the above, the ratio of the current annealing temperature to the initial temperature is used as the perturbation amplitude, and the non-uniform perturbation term is constructed in combination with the leakage response hotspot vector. The new candidate parameter state Θ is generated by weighted perturbation. ′ , a local search strategy that simulates the preferential disturbance of key leakage areas; S45. Design an adaptive cooling strategy coupled with the water pressure variation trend. In each annealing iteration, the average absolute value of the water pressure variation characteristics of the sample where the candidate solution is located is combined with the basic attenuation factor α. A and water pressure regulation coefficient β B , dynamically update the annealing temperature T k ,When the water pressure fluctuation amplitude is small, the annealing temperature drops faster; S46. Define the acceptance probability mechanism that is dynamically coupled with the velocity mutation characteristic. If the current candidate state of the hyperparameter set Θ ′ The corresponding total optimization target error is less than the current hyperparameter set state Θ curr If the error is less than 0, it is accepted unconditionally. Otherwise, the acceptance probability is calculated based on the current annealing temperature and the average velocity mutation degree of the candidate solution in the time period, and a random number that obeys a uniform distribution is generated for comparison to decide whether to accept it. S47. When the acceptance probability is greater than the random threshold, accept the current candidate state and update the current solution to the hyperparameter set candidate state Θ ′ If the multi-objective optimization objective function value of the current candidate state is better than the historical optimal solution, the historical optimal solution is updated synchronously. When the annealing temperature is lower than the set termination temperature threshold or there is no improvement in the optimization objective for several consecutive rounds, the simulated annealing process is terminated and the optimal hyperparameter set solution Θ is output. best ; S48. Solve Θ with the optimal hyperparameter set best As the initialization hyperparameters, the leakage association feature vector set F and the historical leakage point coordinate label set Y are input into the dual-channel topological attention extreme learning machine model. The output weight matrix is calculated by the least squares closed-form solution to obtain the final trained drip irrigation pipeline leakage point location model M. leak .

6. The method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machine according to claim 5, characterized in that: The multi-objective optimization objective function integrates the errors of the following three parts: The first part is the mean square error of the leakage point location, that is, the difference in spatial coordinates between the leakage point predicted by the leakage point location model and the actual leakage point; The second part is the topological structure error, which is the average spatial offset between the leakage point location model prediction result and its nearest known sensor location, which is used to constrain the prediction point to regress to the actual deployment area; The third part is the sensor energy stability error, which is used to measure the degree of deviation of the leakage point location model from the low-energy sensor path during the reasoning process, and to prevent the leakage point location model from predicting the leakage point in the unstable area of the sensor operation. The mean square error, topology error and sensor energy stability error are weighted and combined respectively, and the weight coefficients of topology error and sensor energy stability error are determined by the variance of water flow rate and temperature, respectively.

7. The method for locating drip irrigation pipeline leakage points based on simulated annealing and extreme learning machine according to claim 5, characterized in that: The S5 comprises the following steps: S51. The trained leakage point location model M leak Deployed in the online inference engine, set the real-time sliding time window length w real With step size, for the standardized monitoring data set D norm Perform real-time window segmentation to extract standardized monitoring data segments within the current window; S52. Based on the standardized monitoring data segments in each real-time time window, calculate the velocity mutation characteristics Δq corresponding to each time point in the current window k , water pressure change trend characteristics Temperature anomaly characteristics Combined with the sensor space topology embedding vector z k , generate real-time leakage correlation feature vector set F real ; S53. The real-time leakage associated feature vector set F real Input into the leakage point location model M deployed in the online inference engine in sequence leak , output the corresponding leakage prediction results, and obtain the candidate leakage point coordinate set at each time point S54. Set the candidate leakage point coordinate set Y cand The preset sensor geographic coordinates set and drip irrigation pipeline hydraulic topology map G in the drip irrigation system pipe Perform spatial fusion, the fusion method includes calculating the nearest sensor node c based on Euclidean distance near (k) and correct the coordinates of the candidate leakage point by combining the hydraulically accessible paths of adjacent pipe sections in the topology map; S55. The output result of the fused leakage point coordinates is defined as the final leakage location result Combined with the leakage probability value output by the leakage point location model and inference timestamp t k , construct the leakage location information structure, including coordinates, probability and time stamp.

8. The method for locating leakage points in drip irrigation pipelines based on simulated annealing and extreme learning machine according to claim 7, characterized in that: The classification rules for leakage location are as follows: like and It is marked as a high confidence leakage point; like and It is marked as a medium confidence leakage point; like or It is marked as a low confidence point or an invalid positioning point; in, is the confidence probability of the leakage point location model at the kth time point that the current predicted leakage point is a real leakage, Represents the final leakage point coordinates output by the leakage point location model and spatial fusion processing at the kth time point.

Images