Training Method for Comprehensive Evaluation Model of Operating Efficiency of Irrigation Pumping Stations

Abstract

The present invention relates to the technical field of model training methods, and particularly relates to a method for training a comprehensive evaluation model of the operation efficiency of an irrigation pumping station. This method conducts model training based on incremental learning. The data of the irrigation pumping station collected in real time is input into the efficiency evaluation model, enabling the efficiency evaluation model to learn the features in the real-time collected data of the irrigation pumping station. On the basis of maintaining the original capabilities of the efficiency evaluation model, new knowledge and capabilities can be learned from the real-time irrigation pumping station data. Incremental learning can obtain knowledge from the inference tasks of the old efficiency evaluation model, enabling the efficiency evaluation model to learn to solve new inference tasks of the efficiency evaluation model while retaining the knowledge learned in the previous inference tasks of the efficiency evaluation model, avoiding retraining the parameters of the efficiency evaluation model when new irrigation pumping station data arrives. Through the incremental learning method, the present invention can obtain knowledge from the old comprehensive evaluation model of the operation efficiency of the irrigation pumping station, thereby improving the operation efficiency of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of model training methods, and particularly to a method for training a comprehensive evaluation model for the operation efficiency of irrigation pumping stations. Background Art

[0002] An integrated prefabricated pumping station is a pumping station that integrates main components such as a wellbore, pumps, pipelines, a control system, and a ventilation system in a factory and is pre-assembled and tested before leaving the factory. It is applicable to water supply, drainage, and waterlogging prevention projects. In recent years, integrated prefabricated pumping stations have been widely used in China. In the irrigation system where the integrated prefabricated pumping station is located, as the core equipment, the operation efficiency of the pumping station directly affects the water resource utilization rate and energy consumption level of agricultural irrigation. However, the accurate evaluation of the operation efficiency of the pumping station faces various challenges: diverse data sources, complex and dynamically changing feature distributions, lack of labels for some data, and redundancy problems of high-dimensional data, etc. These problems lead to low data utilization efficiency, insufficient model generalization ability, and poor adaptability to real-time changes in the prior art during the efficiency evaluation process. Traditional methods usually rely on static models trained offline and cannot adapt to the dynamic changes of real-time operation data of the pumping station. At the same time, due to the multi-source, multi-dimensional, and highly non-linear characteristics of irrigation pumping station data, existing generation methods are difficult to expand the sample space of real data. When the model has insufficient training data, it often exhibits poor performance. In addition, high-dimensional data processing technologies are prone to losing important information during the dimensionality reduction process, resulting in inaccurate description of the operation characteristics of the pumping station.

[0003] Therefore, the present invention proposes a method for training a comprehensive evaluation model for the operation efficiency of irrigation pumping stations to solve the above problems. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention develops a method for training a comprehensive evaluation model for the operation efficiency of irrigation pumping stations. Through a comprehensive evaluation method for the efficiency of irrigation pumping stations that can dynamically learn new data features, expand the sample space, and efficiently process high-dimensional data, the present invention can improve the accuracy, real-time performance, and adaptability of the model and meet the actual needs of the irrigation system.

[0005] The technical solution for the present invention to solve the technical problem is a method for training a comprehensive evaluation model for the operation efficiency of irrigation pumping stations, which performs model training based on incremental learning, and the training method is as follows:

[0006] Import the data in the original irrigation pump station dataset into the existing comprehensive evaluation model for the operating efficiency of the irrigation pump station in the offline state for training to obtain a basic efficiency evaluation model. Input the data in the new irrigation pump station dataset into the basic efficiency evaluation model, and determine whether the data in the newly input irrigation pump station dataset is available. If it is not available, terminate the training of the model and output the basic efficiency evaluation model. If it is available, fuse the data in the new irrigation pump station dataset with the prediction results of the model. The fused data is input into the basic efficiency evaluation model for training to generate an efficiency evaluation model. Input the data in the new irrigation pump station dataset into the newly generated efficiency evaluation model again, and determine whether the data in the new irrigation pump station dataset is available. If it is not available, terminate the training of the model and output the newly generated efficiency evaluation model. If it is available, fuse the data in the newly input irrigation pump station dataset with the prediction results of the model. The fused data is input into the newly generated efficiency evaluation model for training to generate a new efficiency evaluation model again. Iterate and loop multiple times until the final efficiency evaluation model is output. The final efficiency evaluation model outputs the prediction results for the operating efficiency of the irrigation pump station;

[0007] When the data in the irrigation pump station dataset is insufficient for multiple trainings, a generative adversarial network based on feature distribution adversarial loss is used for data sample augmentation. After sample augmentation, a self-encoder algorithm based on dynamic projection is used for data dimensionality reduction.

[0008] In the specific implementation manner, the incremental learning process is as follows:

[0009] The incremental learning process is a process of continuously learning through training. When the irrigation pump station executes a task, the comprehensive evaluation model for the operating efficiency of the irrigation pump station evaluates the operating efficiency of the pump station according to the data collected during the task execution, and at the same time, manually annotates the operating efficiency of the irrigation pump station. The annotated labels are low operating efficiency, medium operating efficiency, and high operating efficiency;

[0010] When the irrigation pump station executes the first task, the irrigation pump station dataset input into the basic efficiency evaluation model is denoted as , and it is judged whether the data in is available. When it is available, a new efficiency evaluation model is obtained. The efficiency prediction result of is used as a pseudo-label to fuse with the dataset to obtain a new dataset and is used for the training of . The dataset represents the irrigation pump station dataset input into the new efficiency evaluation model when the irrigation pump station executes the second task, and it is judged Whether the data is available, if available, obtain a new efficiency evaluation model , and the efficiency prediction result is used as a pseudo-label to fuse with the dataset to obtain a new dataset and is used for the training of . The dataset represents the irrigation pump station dataset input to the new efficiency evaluation model when the irrigation pump station executes the third task. Determine whether the data in is available. After multiple iterations, when executing tasks, if the dataset is unavailable, stop the iteration, that is, terminate the training process of the incremental learning for the model, and output the final efficiency evaluation model , which represents the irrigation pump station dataset input to the new efficiency evaluation model when the irrigation pump station executes the th task. The efficiency prediction result output by the final efficiency evaluation model is expressed as , and the dataset used for training is expressed as . , incorporates the efficiency prediction result of as a pseudo-label and the dataset , .

[0011] In the specific implementation manner, the conditions for determining whether the data in the new irrigation pump station dataset is available are as follows:

[0012] Manually annotate the operating efficiency of the irrigation pump station. The content of the annotation is the operating efficiency of the irrigation pump station, which are low operating efficiency, medium operating efficiency, and high operating efficiency respectively. Input the data in the new irrigation pump station dataset after annotation into the basic efficiency evaluation model for training, and compare the accuracy of the trained basic efficiency evaluation model with the accuracy of the untrained basic efficiency evaluation model. If the accuracy of the trained basic efficiency evaluation model has improved, the data in the new irrigation pump station dataset is available; otherwise, it is unavailable.

[0013] In the specific implementation manner, construct the irrigation pump station dataset:

[0014] Monitor the relevant data during the operation of the irrigation pumping station through multiple sensors, and record the relevant operation events and environmental data at the same time. The sensors include flow meters, pressure sensors, current sensors, and temperature sensors. The data collected by the sensors during the monitoring process include the electrical, mechanical, and hydraulic data collected in real time. The operation events include the manual intervention of the pumping station operators, the start and stop of the pumping station, and the fault records of the pumping station. The environmental data includes climate, meteorology, and soil humidity;

[0015] Uniformly store the collected data in a distributed database system in a standardized storage format.

[0016] In the specific implementation manner, the process of data sample augmentation based on the feature distribution adversarial loss of the generative adversarial network is as follows:

[0017] Perform data sample augmentation through the generator and discriminator. Specifically, optimize the generator through the feature distribution adversarial loss and polynomial kernel function, train the discriminator using the weighted feature distribution adversarial loss and adaptive weighting coefficients, and achieve the collaborative optimization of the generator and discriminator through the feedback control mechanism, thereby generating more realistic irrigation pumping station data.

[0018] In the specific implementation manner, the specific process of the data sample augmentation process is as follows:

[0019] (1) Correct the feature distribution of the data in the collected irrigation pumping station dataset through an adaptively calculated mapping matrix. The calculation process is as follows:

[0020] ,

[0021] ,

[0022] Among them, represents the mapping matrix of the irrigation pumping station data features, represents the weighting coefficient of the mapping matrix, represents the standard deviation of the input irrigation pumping station data, represents the mean value of the input irrigation pumping station data, represents the number of features of the irrigation pumping station data input in the current batch, represents the index of, represents the th irrigation pumping station data feature, represents the original pumping station data feature, , represents the corrected irrigation pumping station data feature, represents the mapping function, represents the mapping function with respect to the Jacobian matrix of, the mapping function Use the principal component analysis method function;

[0023] (2) Optimize the generator network:

[0024] Optimize the generator through the feature distribution adversarial loss and the polynomial kernel function to capture the nonlinear data features of the irrigation pumping station. Use the polynomial kernel function to fit the flow characteristic curve of the pumping station, and adjust to correctly depict the nonlinear changes under the actual working conditions, and then generate irrigation pumping station data that conforms to the actual situation. The basic adversarial loss calculation formula of the generator is as follows:

[0025] ,

[0026] Among them, represents the basic adversarial loss of the generator , represents subordinate to the expected value of the feature distribution of the irrigation pumping station data , represents the discriminator, represents the random noise vector input to the generator subordinate to the expected value of the noise distribution ;

[0027] At the same time, use the feature distribution adversarial loss for constraint, and perform random generation constraint loss by perturbing the randomness factor of the generator input, and then calculate the total loss. The calculation formula is as follows:

[0028] ,

[0029] ,

[0030] ,

[0031] ,

[0032] Among them, represents the weighting coefficient for adjusting the intensity of the feature distribution adversarial loss, represents the weighting coefficient for adjusting the intensity of the random generation constraint loss, represents the feature distribution adversarial loss of the generator , represents the random generation constraint loss, represents the coefficient for controlling the weight of the gradient multiplication term, represents the gradient calculation, represents the polynomial kernel function used to measure the nonlinear relationship between the generated data samples and the target irrigation pumping station data, represents the order of the polynomial kernel function, represents Norm represents the intensity coefficient for controlling the strength of randomly generated constraint terms represents the adoption of a randomness factor and the variation in the output of the generator after generation, where the randomness factor , represents the hyperparameter for controlling the noise amplitude represents the normal distribution noise with a mean of zero and a variance of the identity matrix ;

[0033] (3) Train the discriminator network:

[0034] The discriminator uses a weighted feature distribution adversarial loss to dynamically adjust the adaptive weighting coefficients through a feedback mechanism, calculates the total loss of the discriminator with enhanced discrimination ability, and then calculates the total loss function of the discriminator. The calculation formula is as follows:

[0035] ,

[0036] ,

[0037] where represents the total loss function of the discriminator represents the weighting coefficient of the discriminator loss represents the feature distribution adversarial loss function of the discriminator represents the first adaptive weighting coefficient of the discriminator represents the second adaptive weighting coefficient of the discriminator;

[0038] Dynamically update the adaptive weighting coefficients of the discriminator based on the feedback control method. The calculation formula is as follows:

[0039] ,

[0040] ,

[0041] where represents the parameter update operation and respectively represent the first and second adaptive weighting coefficients of the updated discriminator represents the learning rate of the adaptive weighting coefficient represents taking the partial derivative;

[0042] (4) Co-optimize the generator and the discriminator:

[0043] Co-optimize the generator and the discriminator through a feedback control mechanism. The calculation formula is as follows:

[0044] ,

[0045] ,

[0046] Among them, and respectively represent the optimized generator parameters and discriminator parameters, and respectively represent the generator parameters and discriminator parameters before optimization, represents the learning rate of the generator, represents the learning rate of the discriminator, represents the gradient of the generator loss function with respect to , represents the gradient of the discriminator loss function with respect to ;

[0047] (5) Optimize the generated data samples:

[0048] Optimize the generated data samples by means of non - linear fitting, capture and improve the non - linear feature distribution, correct the generated irrigation pumping station data by means of local weighting to make it more conform to the real irrigation pumping station data feature distribution, and optimize and correct the distribution of the generated data samples and real samples in the high - dimensional space through cosine similarity and difference penalty term. The calculation formula is as follows:

[0049] ,

[0050] Among them, represents the optimized irrigation pumping station data sample, represents the cosine similarity calculation function, represents the adjustment coefficient for punishing the difference between the generated irrigation pumping station data and the original irrigation pumping station data, represents the corrected irrigation pumping station data feature, that is, the irrigation pumping station data feature generated by the generator, represents, that is, the real irrigation pumping station data feature;

[0051] (6) Repeat the above steps (1) - (5) and set the stop iteration condition until the preset iteration condition is met and then stop to complete the data sample expansion.

[0052] In the specific implementation manner, the data dimensionality reduction process based on the dynamic projection auto - encoder algorithm is as follows:

[0053] Optimize the autoencoder algorithm based on dynamic projection to improve the dimensionality reduction effect. Optimize the autoencoder by randomly initializing the parameters of the autoencoder. Implement the forward propagation of the autoencoder through the ReLU activation function, and then iteratively update the parameters through the gradient descent method to minimize the reconstruction error and dynamically adjust the projection direction. Classify the data after dimensionality reduction through the Softmax function, with types being low operating efficiency, medium operating efficiency, and high operating efficiency. Take the data feature category with the highest class probability as the category for the comprehensive evaluation of the operating efficiency of the irrigation pumping station to obtain the prediction result of the operating efficiency of the irrigation pumping station.

[0054] In the specific implementation manner, the specific process of data dimensionality reduction is as follows:

[0055] Randomly initialize the weight matrix and bias term of the autoencoder. The initialized parameters follow a normal distribution with a mean of 0 and a variance of the identity matrix. Implement the forward propagation of the irrigation pumping station data in the autoencoder through the ReLU activation function, and the calculation formula is as follows:

[0056] ,

[0057] where, represents the feature after dimensionality reduction, represents the autoencoder function, adopts the ReLU activation function, represents the irrigation pumping station data input into the autoencoder, represents the weight matrix of the autoencoder, represents the bias term of the autoencoder;

[0058] The weight matrix and the bias term are the final results obtained after iterative updates. The update process is as follows:

[0059] ,

[0060] ,

[0061] where, represents the index of the iterative update times of, , and respectively represent the weight matrices of the autoencoder after the th and th iterative updates, and respectively represent the bias terms of the autoencoder after the th and th iterative updates. represents the learning rate, represents the loss function of the autoencoder, represents the gradient of the autoencoder function with respect to the weight matrix, represents the gradient of the autoencoder with respect to the bias term;

[0062] By optimizing the projection of the high-dimensional features of the input irrigation pumping station data, the projection direction of the autoencoder is dynamically adjusted, and the irrigation pumping station data input to the autoencoder is calculated and the reconstruction error of the irrigation pumping station data after dimensionality reduction , and then the total loss of the autoencoder is calculated , and the calculation formula is as follows:

[0063] ,

[0064] ,

[0065] ,

[0066] ,

[0067] where, represents the inverse operation of the autoencoder function, represents the transpose of the weight matrix of the decoder, represents the transpose of the bias term of the decoder, represents the projection matrix of the regularization weight, represents the projection matrix of the Frobenius norm, represents the relaxation term constraint, represents the relaxation term constraint weight, represents the projection matrix of the transpose, represents the regularization term weight of the features after dimensionality reduction, represents the regularization term after dimensionality reduction, represents the number of features of the irrigation pumping station data input to the autoencoder, represents of the index, represents the th feature representation after dimensionality reduction;

[0068] The projection matrix is the final result after iterative updates. The projection matrix is updated according to the gradient calculation rule, and the calculation formula is as follows:

[0069] ,

[0070] , ,

[0071] Among them, represents the index of the iterative update times , , and respectively represent the projection matrices after the -th and -th iterative updates, , represents the loss function value of the autoencoder when inputting the -th feature, and respectively represent the learning rates of the projection matrix during the -th and -th iterative updates, represents the gradient of the total loss of the encoder with respect to the projection matrix, represents the hyperparameter that controls the adaptive step size, represents the gradient of the total loss of the autoencoder;

[0072] Judge whether the convergence state is reached according to the set convergence threshold of the autoencoder. If , the algorithm convergence is terminated, and the final stable dimensionality reduction result is output.

[0073] The effects provided in the invention content are only the effects of the embodiments, rather than all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0074] (1) In the comprehensive evaluation task of the operation efficiency of irrigation pumping stations, a generative adversarial network based on feature distribution adversarial loss is adopted. Aiming at the problem of inconsistent multi-source feature distributions of irrigation pumping station data, the generator is optimized through feature distribution adversarial loss, and the generated data is more in line with the actual pumping station characteristics;

[0075] (2) In the comprehensive evaluation task of the operation efficiency of irrigation pumping stations, the polynomial kernel function is used to capture the non-linear characteristics of the operation of pumping stations, generate high-fidelity data, and solve the problem of insufficient fitting of complex features by traditional methods;

[0076] (3) In the comprehensive evaluation task of the operation efficiency of irrigation pumping stations, the discriminator adopts weighted feature distribution adversarial loss and adaptive weighting coefficients to improve the discrimination ability of the generated samples;

[0077] (4) In the comprehensive evaluation task of the operation efficiency of irrigation pumping stations, an autoencoder based on dynamic projection is used for data dimensionality reduction. The dynamic projection is used to optimize the dimensionality reduction process of high-dimensional irrigation pumping station data, solve the problems of redundant information interference and feature loss, and ensure that the data after dimensionality reduction retains the core characteristics;

[0078] (5) In the comprehensive evaluation task of the operation efficiency of the irrigation pumping station, through regularization and relaxation constraint strategies, the distortion during the dimensionality reduction process is reduced, and multi-dimensional constraints are imposed on the dimensionality reduction results to ensure the physical interpretability of the data;

[0079] (6) In the comprehensive evaluation task of the operation efficiency of the irrigation pumping station, the adaptive learning rate of the projection matrix is dynamically adjusted to better adapt to the data feature distributions of different pumping stations and improve the model's processing ability for high-dimensional data. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The 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, but do not constitute a limitation to the present invention.

[0081] Figure 1 It is a schematic flow chart of the training method of the efficiency evaluation model based on incremental learning in the present invention.

[0082] Figure 2 It is a comparison chart of the training time of the present invention under different learning rates of the adaptive weighting coefficient.

[0083] Figure 3 It is a comparison chart of the distribution of the data generated by the present invention and the real data.

[0084] Figure 4 It is a comparison chart of the convergence processes of the dynamic projection autoencoder and the traditional autoencoder used in the present invention.

[0085] Figure 5 It is a comparison of the reconstruction errors of two different dimensionality reduction algorithms, namely the autoencoder based on dynamic projection and the traditional PCA algorithm, in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0086] In order to clearly illustrate the technical features of the present solution, the present invention will be described in detail below through specific embodiments and in conjunction with its drawings.

[0087] Embodiment 1

[0088] A method for training a comprehensive evaluation model of the operation efficiency of an irrigation pumping station, based on incremental learning for model training, and the training method is as follows:

[0089] Import the data in the original irrigation pump station dataset into the comprehensive evaluation model for the operation efficiency of the irrigation pump station in the existing offline state for training to obtain a basic efficiency evaluation model. Input the data in the new irrigation pump station dataset into the basic efficiency evaluation model, and determine whether the data in the newly input irrigation pump station dataset is available. If it is not available, terminate the training of the model and output the basic efficiency evaluation model. If it is available, fuse the data in the new irrigation pump station dataset with the prediction results of the model. The fused data is input into the basic efficiency evaluation model for training to generate an efficiency evaluation model. Input the data in the new irrigation pump station dataset into the newly generated efficiency evaluation model again, and determine whether the data in the new irrigation pump station dataset is available. If it is not available, terminate the training of the model and output the newly generated efficiency evaluation model. If it is available, fuse the data in the newly input irrigation pump station dataset with the prediction results of the model. The fused data is input into the newly generated efficiency evaluation model for training to generate a new efficiency evaluation model again. Iterate and loop multiple times until the final efficiency evaluation model is output. The final efficiency evaluation model outputs the prediction results for the operation efficiency of the irrigation pump station;

[0090] When the data in the irrigation pump station dataset is not sufficient for multiple trainings, use a generative adversarial network based on feature distribution adversarial loss to perform data sample augmentation. After sample augmentation, use an autoencoder algorithm based on dynamic projection for data dimensionality reduction.

[0091] In the specific implementation manner, the incremental learning process is as follows:

[0092] The incremental learning process is a process of continuously learning through training. When the irrigation pump station is performing a task, the comprehensive evaluation model for the operation efficiency of the irrigation pump station evaluates the operation efficiency of the pump station based on the data collected during the task execution, and at the same time, manually annotates the operation efficiency of the irrigation pump station. The annotated labels are low operation efficiency, medium operation efficiency, and high operation efficiency;

[0093] When the irrigation pump station is performing the first task, the irrigation pump station dataset input into the basic efficiency evaluation model is denoted as , and determine whether the data in is available. When it is available, obtain a new efficiency evaluation model . Use the efficiency prediction result of as a pseudo-label to fuse with the dataset to obtain a new dataset and use it for the training of . The dataset represents the irrigation pump station dataset input into the new efficiency evaluation model when the irrigation pump station is performing the second task, and determine Whether the data is available, if available, obtain a new efficiency evaluation model , and the efficiency prediction result is used as a pseudo-label to be fused with the dataset to obtain a new dataset and is used for the training of . The dataset represents the irrigation pump station dataset input to the new efficiency evaluation model when the irrigation pump station executes the third task. Determine whether the data in is available. After multiple iterations and executing tasks, when the dataset is unavailable, stop the iteration, that is, terminate the training process of the incremental learning for the model, and output the final efficiency evaluation model , represents the irrigation pump station dataset input to the new efficiency evaluation model when the irrigation pump station executes the th task. The efficiency prediction result output by the final efficiency evaluation model is represented as . The dataset used for training is represented as . . The dataset , incorporates the efficiency prediction result of as a pseudo-label and the dataset , .

[0094] In the specific implementation manner, the conditions for determining whether the data in the new irrigation pump station dataset is available are as follows:

[0095] Manually annotate the operating efficiency of the irrigation pump station. The annotation content is the operating efficiency of the irrigation pump station, which are low operating efficiency, medium operating efficiency, and high operating efficiency respectively. Input the data in the new irrigation pump station dataset after annotation into the basic efficiency evaluation model for training. Compare the accuracy of the trained basic efficiency evaluation model with the accuracy of the untrained basic efficiency evaluation model. If the accuracy of the trained basic efficiency evaluation model has improved, the data in the new irrigation pump station dataset is available; otherwise, it is unavailable.

[0096] In the specific implementation manner, construct the irrigation pump station dataset:

[0097] Monitor relevant data during the operation of the irrigation pumping station through multiple sensors, and record relevant operation events and environmental data at the same time. The sensors include flow meters, pressure sensors, current sensors, and temperature sensors. The data collected by the sensors during the monitoring process includes real-time electrical, mechanical, and hydraulic data. The operation events include manual intervention by the pumping station operators, start-up and stop of the pumping station, and fault records of the pumping station. The environmental data includes climate, meteorology, and soil humidity;

[0098] Uniformly store the collected data in a distributed database system in a standardized storage format.

[0099] The storage format adopts JSON or CSV. The specific JSON storage format is as follows:

[0100] {

[0101] "timestamp": "2025-01-01T12:00:00",

[0102] "Ra": 45.3, / / Current value (A)

[0103] "Rb": 3.2, / / Flow value (m^3 / h)

[0104] "Rc": 150.6, / / Pumping station pressure value (bar)

[0105] "Rd": 18.2, / / Temperature value (°C)

[0106] "Re": "Normal", / / Status: Normal / Fault

[0107] "Rf": 120, / / Working duration (minutes)

[0108] "Rg": 3, / / Pumping station mode: 1 - Automatic, 2 - Manual

[0109] "Rh": "Start", / / Operation event: Start / Stop

[0110] "Ri": 0.95, / / Electrical energy efficiency (%)

[0111] "Rj": 40.5 / / Soil humidity (%)

[0112] }

[0113] In the specific implementation manner, the process of data sample augmentation based on the feature distribution adversarial loss of the generative adversarial network is as follows:

[0114] Data sample augmentation is performed through a generator and a discriminator. Specifically, the generator is optimized by the feature distribution adversarial loss and the polynomial kernel function. The discriminator is trained using the weighted feature distribution adversarial loss and the adaptive weighting coefficient. The collaborative optimization of the generator and the discriminator is achieved through a feedback control mechanism, thereby generating more realistic irrigation pumping station data.

[0115] In the specific implementation manner, the specific process of data sample augmentation is as follows:

[0116] (1) The feature distribution of the data in the collected irrigation pumping station dataset is corrected through an adaptively calculated mapping matrix. The calculation process is as follows:

[0117] ,

[0118] ,

[0119] Among them, represents the mapping matrix of the irrigation pumping station data features, represents the weighting coefficient of the mapping matrix, represents the standard deviation of the input irrigation pumping station data, represents the mean of the input irrigation pumping station data, represents the number of irrigation pumping station data features input in the current batch, represents the index of, represents the th irrigation pumping station data feature, represents the original pumping station data feature, , represents the corrected irrigation pumping station data feature, represents the mapping function, represents the mapping function with respect to the Jacobian matrix of, and the mapping function adopts the principal component analysis method function;

[0120] (2) Optimize the generator network:

[0121] The generator is optimized by the feature distribution adversarial loss and the polynomial kernel function to capture the non - linear data features of the irrigation pumping station. The polynomial kernel function is used to fit the pumping station flow characteristic curve. By adjusting, the non - linear changes under the actual working conditions are correctly characterized, and then realistic irrigation pumping station data is generated. The basic adversarial loss calculation formula of the generator is as follows:

[0122] ,

[0123] Among them, represents the generator The basic adversarial loss denotes subordinate to the data feature distribution of the irrigation pumping station expectation denotes the discriminator denotes the random noise vector input to the generator subordinate to the noise distribution expectation;

[0124] At the same time, the feature distribution adversarial loss is used for constraint, and the random generation constraint loss is calculated by perturbing the randomness factor of the generator input, and then the total loss is calculated. The calculation formula is as follows:

[0125] ,

[0126] ,

[0127] ,

[0128] ,

[0129] Among them, denotes the weighting coefficient for adjusting the intensity of the feature distribution adversarial loss, set , denotes the weighting coefficient for adjusting the intensity of the random generation constraint loss, set , denotes the generator feature distribution adversarial loss denotes the random generation constraint loss denotes the coefficient for controlling the weight of the gradient multiplication term, set , denotes the gradient calculation denotes the polynomial kernel function used to measure the non-linear relationship between the generated data samples and the target irrigation pumping station data denotes the order of the polynomial kernel function denotes norm denotes the intensity coefficient for controlling the random generation constraint term denotes the use of the randomness factor the change in the generator output after , denotes the hyperparameter for controlling the noise amplitude denotes the normal distribution noise with mean zero and variance of the identity matrix ;

[0130] (3) Train the discriminator network:

[0131] The discriminator uses a weighted feature distribution adversarial loss to dynamically adjust the adaptive weighting coefficients through a feedback mechanism, calculates the total loss of the discriminator with enhanced discrimination ability, and then calculates the total loss function of the discriminator. The calculation formula is as follows:

[0132] ,

[0133] ,

[0134] where, represents the total loss function of the discriminator, represents the weighting coefficient of the discriminator loss, represents the feature distribution adversarial loss function of the discriminator, represents the first adaptive weighting coefficient of the discriminator, represents the second adaptive weighting coefficient of the discriminator;

[0135] Dynamically update the adaptive weighting coefficients of the discriminator based on the feedback control method. The calculation formula is as follows:

[0136] ,

[0137] ,

[0138] where, represents the parameter update operation, and respectively represent the first and second adaptive weighting coefficients of the updated discriminator, represents the learning rate of the adaptive weighting coefficient, set , represents taking the partial derivative;

[0139] (4) Co-optimize the generator and the discriminator:

[0140] Co-optimize the generator and the discriminator through the feedback control mechanism. The calculation formula is as follows:

[0141] ,

[0142] ,

[0143] where, and respectively represent the optimized generator parameters and discriminator parameters, and respectively represent the generator parameters and discriminator parameters before optimization, represents the learning rate of the generator, represents the learning rate of the discriminator, Denote the gradient of the generator loss function with respect to , Denote the gradient of the discriminator loss function with respect to ;

[0144] (5) Optimize the generated data samples:

[0145] Optimize the generated data samples by means of non - linear fitting to capture and improve the non - linear feature distribution, correct the generated irrigation pumping station data by means of local weighting to make it more conform to the real irrigation pumping station data feature distribution, and optimize and correct the distribution of the generated data samples and real samples in the high - dimensional space through cosine similarity and difference penalty terms. The calculation formula is as follows:

[0146] ,

[0147] where, represents the optimized irrigation pumping station data sample, represents the cosine similarity calculation function, represents the adjustment coefficient for punishing the difference between the generated irrigation pumping station data and the original irrigation pumping station data, represents the corrected irrigation pumping station data feature, that is, the irrigation pumping station data feature generated by the generator, represents, that is, the real irrigation pumping station data feature;

[0148] (6) Repeat the above steps (1)-(5) iteratively and set the stop iteration condition, and stop until the preset iteration condition is met to complete the data sample expansion.

[0149] In the specific implementation manner, the data dimensionality reduction process based on the dynamic projection auto - encoder algorithm is as follows:

[0150] Optimize the dynamic projection auto - encoder algorithm to improve the dimensionality reduction effect. Optimize the auto - encoder by randomly initializing the parameters of the auto - encoder, implement the forward propagation of the auto - encoder through the ReLU activation function, then iteratively update the parameters by the gradient descent method to minimize the reconstruction error and dynamically adjust the projection direction, and classify the data after dimensionality reduction through the Softmax function. The types are low operating efficiency, medium operating efficiency, and high operating efficiency respectively. Take the data feature category with the maximum class probability as the category for the comprehensive evaluation of the irrigation pumping station operating efficiency to obtain the prediction result of the irrigation pumping station operating efficiency.

[0151] In the specific implementation manner, the specific process of the data dimensionality reduction process is as follows:

[0152] Randomly initialize the weight matrix and bias terms of the autoencoder. The initialized parameters follow a normal distribution with a mean of 0 and a variance of the identity matrix. The forward propagation of the irrigation pumping station data in the autoencoder is implemented through the ReLU activation function, and the calculation formula is as follows:

[0153] ,

[0154] where, represents the feature after dimensionality reduction, represents the autoencoder function, uses the ReLU activation function, represents the irrigation pumping station data input to the autoencoder, represents the weight matrix of the autoencoder, represents the bias term of the autoencoder;

[0155] The weight matrix of the autoencoder and the bias term are the final results obtained after iterative updates. The update process is as follows:

[0156] ,

[0157] ,

[0158] where, represents the index of the iterative update times of, , and respectively represent the weight matrix of the autoencoder after the -th and the -th iterative updates, and respectively represent the bias term of the autoencoder after the -th and the -th iterative updates, represents the learning rate, , represents the loss function of the autoencoder, represents the gradient of the autoencoder function with respect to the weight matrix, represents the gradient of the autoencoder with respect to the bias term;

[0159] By optimizing the projection of the high-dimensional features of the input irrigation pumping station data, the projection direction of the autoencoder is dynamically adjusted. Calculate the reconstruction error between the input irrigation pumping station data to the autoencoder and the dimensionality-reduced irrigation pumping station data, and then calculate the total loss of the autoencoder. The calculation formula is as follows:

[0160] ,

[0161] ,

[0162] ,

[0163] ,

[0164] Among them, represents the inverse operation of the autoencoder function, represents the transpose of the weight matrix of the decoder, represents the transpose of the bias term of the decoder, represents the projection matrix 's regularization weight, represents the projection matrix 's Frobenius norm, represents the relaxation term constraint, represents the relaxation term constraint weight, represents the projection matrix 's transpose, represents the regularization term weight of the features after dimensionality reduction, set , represents the regularization term after dimensionality reduction, represents the number of features of the irrigation pump station data input to the autoencoder, represents 's index, represents the th feature's representation after dimensionality reduction;

[0165] The projection matrix is the final result after iterative updates. The projection matrix is updated according to the gradient calculation rule, and the calculation formula is as follows:

[0166] ,

[0167] , ,

[0168] Among them, represents the index of the iterative update times 's index, , and respectively represent the projection matrices after the th and th iterative updates, , represents the loss function value of the autoencoder when inputting the th feature, and respectively represent the learning rates of the projection matrix during the -th and -th iterative updates. represents the gradient of the total loss of the encoder with respect to the projection matrix. represents the hyperparameter that controls the adaptive step size, and is set to , represents the gradient of the total loss of the autoencoder;

[0169] According to the set convergence threshold of the autoencoder judge whether the convergence state is reached. Set , if , then end the algorithm convergence and output the final stable dimensionality reduction result.

[0170] Example 2

[0171] To verify the influence of the learning rate of the adaptive weighting coefficient on training, as Figure 2 shown, by comparing different learning rates and batch sizes, it can be seen that the generator still maintains a low training time under the conditions of a large batch and a small learning rate, thereby improving the training efficiency of the model and solving the problem of slow convergence of traditional generative models on high-complexity data distributions.

[0172] Example 3

[0173] As Figure 3 shown, through the feature distribution adversarial loss and the adaptive mapping matrix correction, the generated data is closer to the real data in terms of feature distribution, avoiding the deviation phenomenon that is prone to occur in the generator, and the distribution curve of the generated data is highly consistent with the real data, indicating that the generator captures the characteristic distribution of the real irrigation pump station data.

[0174] Example 4

[0175] As Figure 4 shown, by comparing the reconstruction errors of different dimensionality reduction algorithms, the experimental data shows that the autoencoder based on dynamic projection can effectively reduce the reconstruction error during dimensionality reduction, and has lower reconstruction errors in each feature dimension compared with the traditional principal component analysis method.

[0176] Example 5

[0177] As Figure 5 shown, by comparing the reconstruction errors of different dimensionality reduction algorithms, the experimental data shows that the autoencoder based on dynamic projection can effectively reduce the reconstruction error during dimensionality reduction, and has lower reconstruction errors in each feature dimension compared with the traditional principal component analysis method.

[0178] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, they are not intended to limit the scope of protection of the present invention. Based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the scope of protection of the present invention.

Claims

1. A training method for a comprehensive evaluation model of irrigation pump station operation efficiency, characterized in that: Model training is based on incremental learning. The training method is as follows: Import the original irrigation pump station data set into the existing offline irrigation pump station operation efficiency comprehensive evaluation model for training to obtain a basic efficiency evaluation model, input the new irrigation pump station data set into the basic efficiency evaluation model, and judge whether the newly input irrigation pump station data set is available. If not, terminate the training of the model and output the basic efficiency evaluation model. If available, fuse the new irrigation pump station data set with the prediction result of the model, input the fused data into the basic efficiency evaluation model for training to generate the efficiency evaluation model, input the new irrigation pump station data set into the newly generated efficiency evaluation model again, and judge whether the new irrigation pump station data set is available. If not, terminate the training of the model and output the newly generated efficiency evaluation model. If available, fuse the new irrigation pump station data set with the prediction result of the model, input the fused data into the newly generated efficiency evaluation model for training, and generate a new efficiency evaluation model again. Iterate multiple times until the final efficiency evaluation model is output. The final efficiency evaluation model outputs the prediction result of the operation efficiency of the irrigation pump station. When the data in the irrigation pump station dataset is insufficient for multiple trainings, a generative adversarial network based on feature distribution adversarial loss is used to expand the data samples. After sample expansion, an autoencoder algorithm based on dynamic projection is used to reduce the data dimension. Constructing irrigation pump station data sets: Use multiple sensors to monitor relevant data during the operation of irrigation pump stations, and record relevant operation events and environmental data. Sensors include flow meters, pressure sensors, current sensors, and temperature sensors. The data collected by monitoring process sensors include real-time electrical, mechanical, and hydraulic data. Operation events include manual intervention by pump station operators, pump station start and stop, and pump station fault records. Environmental data include climate, meteorology, and soil moisture. The collected data is uniformly stored in a distributed database system in a standardized storage format. The process of data sample expansion based on the generative adversarial network with feature distribution adversarial loss is as follows: The data samples are expanded through the generator and discriminator. Specifically, the generator is optimized through feature distribution adversarial loss and polynomial kernel function, and the discriminator is trained using weighted feature distribution adversarial loss and adaptive weighting coefficient. The feedback control mechanism is used to achieve collaborative optimization of the generator and the discriminator, thereby generating more realistic irrigation pump station data. (1) The characteristic distribution of the data in the collected irrigation pump station data set is corrected through the adaptively calculated mapping matrix. The calculation process is as follows: ， ， in, The mapping matrix representing the characteristics of irrigation pump station data, represents the weight coefficient of the mapping matrix, represents the standard deviation of the input irrigation pump station data, represents the mean of the input irrigation pump station data, Indicates the number of irrigation pump station data features input in the current batch. express The index of Indicates Irrigation pump station data features, Represents the original pump station data characteristics, , represents the corrected irrigation pump station data characteristics, represents the mapping function, Represents the mapping function about The Jacobian matrix of The principal component analysis function was used; (2) Optimize the generator network: The nonlinear data characteristics of the irrigation pump station are captured by the feature distribution adversarial loss and polynomial kernel function optimization generator. The polynomial kernel function is used to fit the flow characteristic curve of the pump station. The nonlinear changes under actual working conditions are correctly characterized by adjustment, and then the irrigation pump station data that fits the actual situation is generated. The basic adversarial loss calculation formula of the generator is as follows: ， in, Representation Generator The basic adversarial loss, express Subject to the characteristic distribution of irrigation pump station data expectations, represents the discriminator, A random noise vector representing the input generator Subject to the noise distribution expectations; At the same time, the feature distribution is used to constrain the adversarial loss, and the randomness factor input by the perturbation generator is used to randomly generate the constraint loss, and then the total loss is calculated. The calculation formula is as follows: ， ， ， ， in, Represents the weighted coefficient for adjusting the feature distribution against the loss strength, represents the weighted coefficient for adjusting the strength of the randomly generated constraint loss, Representation Generator The feature distribution of adversarial loss, represents the randomly generated constraint loss, represents the coefficient that controls the weight of the gradient multiplication term, represents the gradient calculation, represents the polynomial kernel function used to measure the nonlinear relationship between the generated data samples and the target irrigation pump station data, represents the order of the polynomial kernel function, express norm, represents the strength coefficient of the randomly generated constraint term, Indicates the use of random factors The amount of change in the output of the post-generator, the randomness factor , represents the hyperparameter controlling the noise amplitude, Indicates that the mean is zero and the variance is the unit matrix Normally distributed noise; (3) Train the discriminator network: The discriminator uses weighted feature distribution adversarial loss to dynamically adjust the adaptive weighting coefficient through the feedback mechanism, calculate the total loss of the discriminator with enhanced distinguishing ability, and then calculate the total loss function of the discriminator. The calculation formula is as follows: ， in, represents the total loss function of the discriminator, represents the weighted coefficient of the discriminator loss, represents the feature distribution adversarial loss function of the discriminator, represents the first adaptive weighting coefficient of the discriminator, represents the second adaptive weighting coefficient of the discriminator; The adaptive weighting coefficient of the discriminator is dynamically updated based on the feedback control method. The calculation formula is as follows: ， ， in, Indicates a parameter update operation. and Respectively represent the first adaptive weighting coefficient and the second adaptive weighting coefficient of the updated discriminator, represents the learning rate of the adaptive weight coefficient, It means to find partial derivative; (4) Co-optimize the generator and discriminator: The generator and discriminator are collaboratively optimized through the feedback control mechanism. The calculation formula is as follows: ， ， in, and Represent the optimized generator parameters and discriminator parameters respectively, and Respectively represent the generator parameters and discriminator parameters before optimization, represents the learning rate of the generator, represents the learning rate of the discriminator, Represents the generator loss function pair The gradient of Denotes the discriminator loss function The gradient of (5) Optimize the generated data samples: The generated data samples are optimized by nonlinear fitting, the nonlinear feature distribution is captured and improved, the generated irrigation pump station data is corrected by local weighting to make it more consistent with the real irrigation pump station data feature distribution, and the distribution of the generated data samples and the real samples in high-dimensional space is optimized and corrected by cosine similarity and difference penalty terms. The calculation formula is as follows: ， in, represents the optimized irrigation pump station data sample, Represents the cosine similarity calculation function, represents the adjustment coefficient for penalizing the difference between the generated irrigation pump station data and the original irrigation pump station data, Represents the irrigation pump station data features generated by the generator, Represents the original pump station data characteristics; (6) Repeat the above steps (1)-(5) and set the stop iteration condition until the preset iteration condition is met, thus completing the data sample expansion.

2. The irrigation pump station operation efficiency comprehensive evaluation model training method according to claim 1 is characterized in that: The incremental learning process is as follows: The incremental learning process is a process of continuous learning through training. When the irrigation pump station is performing a task, the comprehensive evaluation model of the irrigation pump station operation efficiency evaluates the operation efficiency of the pump station based on the data collected during the task execution. At the same time, the operation efficiency of the irrigation pump station is manually labeled, and the labeled labels are low operation efficiency, medium operation efficiency and high operation efficiency. The irrigation pump station is input into the basic efficiency evaluation model when performing its first task. The irrigation pump station dataset is represented as ,judge Is the data available? If available, a new efficiency evaluation model is obtained. ,Will Efficiency prediction results As pseudo labels and datasets Fusion to obtain a new data set And used for Training,dataset Indicates that the irrigation pump station is input into the new efficiency evaluation model when performing the second task Irrigation pump station dataset, judge Is the data available? If available, a new efficiency evaluation model is obtained. ,Will Efficiency prediction results As pseudo labels and datasets Fusion to obtain a new data set And used for Training,dataset Indicates that the irrigation pump station is input into the new efficiency evaluation model when performing the third task Irrigation pump station dataset, judge Is the data available in multiple iterations? After the task, the dataset When it is unavailable, the iteration is stopped, that is, the training process of the model by incremental learning is terminated, and the final efficiency evaluation model is output. , Indicates that the irrigation pump station is in operation When the task is completed, it is input into the new efficiency evaluation model The irrigation pump station dataset and the final efficiency evaluation model The output efficiency prediction result is expressed as , used for training The data set is represented as , The pseudo-labels are integrated into Efficiency prediction results and dataset , .

3. The irrigation pump station operation efficiency comprehensive evaluation model training method according to claim 2 is characterized in that: The conditions for judging whether the data in the new irrigation pump station dataset is available are as follows: The operating efficiency of the irrigation pump station is manually labeled, and the labeled content is the operating efficiency of the irrigation pump station, which is low operating efficiency, medium operating efficiency and high operating efficiency. The data in the labeled new irrigation pump station dataset is input into the basic efficiency evaluation model for training, and the accuracy of the trained basic efficiency evaluation model is compared with the accuracy of the untrained basic efficiency evaluation model. If the accuracy of the trained basic efficiency evaluation model is improved, the data in the new irrigation pump station dataset is available, otherwise it is not available.

4. The irrigation pump station operation efficiency comprehensive evaluation model training method according to claim 3 is characterized in that: The process of data dimensionality reduction based on the dynamic projection autoencoder algorithm is as follows: The autoencoder algorithm based on dynamic projection is optimized to improve the dimensionality reduction effect. The autoencoder is optimized by randomly initializing its parameters. The forward propagation of the autoencoder is realized through the ReLU activation function. The parameters are iteratively updated by the gradient descent method to minimize the reconstruction error and dynamically adjust the projection direction. The Softmax function is used to classify the reduced data into low operating efficiency, medium operating efficiency and high operating efficiency. The data feature category with the largest class probability is used as the category for the comprehensive evaluation of the operating efficiency of the irrigation pump station, and the prediction result of the operating efficiency of the irrigation pump station is obtained.

5. The irrigation pump station operation efficiency comprehensive evaluation model training method according to claim 4 is characterized in that: The specific process of data dimensionality reduction is as follows: The weight matrix and bias term of the autoencoder are randomly initialized. The initialized parameters obey the normal distribution with a mean of 0 and a variance of the unit matrix. The forward propagation of the irrigation pump station data in the autoencoder is realized through the ReLU activation function. The calculation formula is as follows: ， in, represents the features after dimensionality reduction, represents the autoencoder function, Using ReLU activation function, represents the irrigation pump station data input to the autoencoder, represents the weight matrix of the autoencoder, Represents the bias term of the autoencoder; The weight matrix of the autoencoder and the bias term For passing The final result obtained by the iterative update, the update process is as follows: ， ， in, Indicates the number of iterations The index of , and Respectively represent Second and The weight matrix of the autoencoder after the iteration update, and Respectively represent Second and The bias term of the autoencoder after the iteration update, represents the learning rate, represents the loss function of the autoencoder, represents the gradient of the autoencoder function with respect to the weight matrix, represents the gradient of the autoencoder with respect to the bias term; By optimizing the projection of the high-dimensional features of the input irrigation pump station data, the projection direction of the autoencoder is dynamically adjusted to calculate the irrigation pump station data input to the autoencoder. Reconstruction error of irrigation pump station data after dimensionality reduction , and then calculate the total loss of the autoencoder , the calculation formula is as follows: ， ， ， ， in, represents the inverse operation of the autoencoder function, represents the transpose of the decoder weight matrix, represents the transposed bias term of the decoder, Represents the projection matrix The regularization weight of Represents the projection matrix The Frobenius norm of represents the relaxed term constraint, represents the slack constraint weight, Represents the projection matrix The transpose of represents the regularization term weight of the feature after dimensionality reduction, represents the regularization term after dimensionality reduction, represents the number of features of the irrigation pump station data input to the autoencoder, express The index of Indicates The representation of a feature after dimensionality reduction; Projection Matrix Update for iteration The final result after the projection matrix is ​​updated according to the gradient calculation rule. The calculation formula is as follows: ， ， ， in, Indicates the number of iterations The index of , and Respectively represent Second and The updated projection matrix after iterations is: , Indicates the input The loss function value of the autoencoder is and Respectively represent Second and The learning rate of the projection matrix at the iteration update, represents the gradient of the encoder’s total loss with respect to the projection matrix, represents the hyperparameter controlling the adaptive step size, represents the gradient of the total loss of the autoencoder; According to setting the convergence threshold of the autoencoder Determine whether the convergence state has been reached. If , the algorithm converges and outputs the final stable dimensionality reduction result.

Images