A multi-pump parallel coordinated control method and system

Abstract

The present invention relates to the field of water pump system control technology, and discloses a method and system for coordinated control of multiple water pumps in parallel, wherein a method for coordinated control of multiple water pumps in parallel includes: collecting and preprocessing multi-sensor data to construct a multi-dimensional health status feature vector; performing fault feature learning, using a prototype network to construct a prototype representation of fault categories in the feature space, and maximizing the distance between different fault types through metric learning; establishing a transfer learning mechanism to transfer existing water pump historical fault data as source domain knowledge to water pumps that lack fault labeling data; analyzing the temporal state changes of the water pumps to predict the fault probability and remaining service life; eliminating the water hammer effect and pressure fluctuation during the switching process by precisely controlling the speed change rate and valve opening adjustment timing of each pump; the present invention solves the problems of fault prediction and smooth switching in the multi-pump parallel system through innovative small sample learning, knowledge transfer and smooth switching control technology.

Description

Technical Field

[0001] The present invention relates to the technical field of water pump system control, and more particularly, to a method and system for coordinated control of multiple water pumps in parallel. Background Art

[0002] Multi-pump parallel control systems are widely used in municipal water supply, industrial circulating water, agricultural irrigation, and other fields. They are key technologies for ensuring water supply security and stable system operation. With the advancement of industrial automation and the growth of water demand, higher requirements are being placed on the reliability, stability, and intelligence level of multi-pump parallel control systems.

[0003] At present, the multi-pump parallel control technology mainly has the following technical problems:

[0004] Traditional multi-pump control methods mainly adopt fixed-time period switching or control strategies based on simple threshold judgment. Such methods usually switch water pumps within a preset time interval, or trigger switching operations when certain operating parameters are detected to exceed the preset threshold. However, this passive response mechanism cannot predictively identify potential failures of water pumps, and often handles them when the failure has occurred or is about to occur, resulting in system water supply interruption or fluctuation in water supply quality. In addition, water hammer effect and pressure fluctuation are prone to occur during the traditional switching process, affecting the stability of the entire water supply system; existing fault prediction methods have poor adaptability when facing water pumps of different models and different working conditions. Due to differences in the working environment, load conditions, manufacturing process and other factors of the water pump, the failure modes and characteristics of different water pumps are also different. Traditional fault prediction methods are usually trained for specific models or specific working conditions, and it is difficult to effectively migrate to new application scenarios. At the same time, obtaining each working condition The complete fault data under the condition of fault labeling requires long-term operation monitoring and fault accumulation, which is costly and has a long cycle, resulting in insufficient generalization ability of the fault prediction model; for water pumps that lack fault labeling data, effective labeling data is usually extremely scarce, and traditional machine learning methods rely on a large amount of labeled data for model training. Under small sample conditions, overfitting often occurs, and the prediction accuracy decreases. Although transfer learning methods are also used in existing technologies, when there is a large difference between the source domain and the target domain, the knowledge transfer effect is not good, and it is impossible to fully utilize the existing historical fault data to solve problems in new scenarios; the existing control system lacks the ability to accurately evaluate the performance degradation trend and remaining service life of the water pump. Most systems can only provide monitoring of the current operating status and cannot predict the future performance change trend of the equipment. This leads to a lack of scientific basis for the formulation of maintenance plans, which are either too conservative and lead to waste of maintenance resources, or untimely maintenance leads to sudden equipment failures, affecting the normal operation of the system.

[0005] Therefore, it is urgent to develop a multi-pump parallel collaborative control method that can achieve accurate fault prediction under small sample conditions, has good migration capability, and can achieve smooth switching control, so as to improve the reliability, stability and intelligence level of the system. Summary of the Invention

[0006] The present invention provides a method and system for coordinated control of multiple water pumps in parallel, which solves the technical problems of insufficient fault prediction data and pressure fluctuations during switching in traditional multiple water pump parallel systems in related technologies.

[0007] The present invention provides a method for coordinated control of multiple water pumps in parallel, comprising:

[0008] Collecting and preprocessing multi-sensor data to construct a multi-dimensional health status feature vector, wherein the multi-sensor data includes vibration data, current data, temperature data, pressure data, and flow data, and the multi-dimensional health status feature vector includes time domain features, frequency domain features, and time-frequency features;

[0009] Fault feature learning is performed based on multi-dimensional health state feature vectors. Prototype networks are used to construct prototype representations of fault categories in feature space. Metric learning is used to maximize the distance between different fault types and minimize the distance within the same type.

[0010] A transfer learning mechanism is established based on the prototype representation of fault categories. The existing historical fault data of water pumps is used as source domain knowledge to transfer to water pumps that lack fault label data to achieve fault prediction.

[0011] Based on fault prediction results and historical monitoring data, a long-short-term memory network is used to analyze the temporal state changes of the water pump, predict the failure probability and remaining service life, and generate the optimal switching strategy;

[0012] Based on the optimal switching strategy, a seamless switching control algorithm with smooth transition is adopted. By accurately controlling the speed change rate of each pump and the valve opening adjustment timing, the water hammer effect and pressure fluctuation during the switching process are eliminated.

[0013] Furthermore, the step of learning fault characteristics includes:

[0014] Construct a feature embedding network to map the original multi-dimensional health status feature vector into a low-dimensional embedding space;

[0015] Based on the output of the feature embedding network, a prototype representation is constructed for each type of fault;

[0016] Optimizing the distance metric in the embedding space through metric learning methods;

[0017] Data augmentation, contrastive learning and meta-learning strategies are used to improve the model's adaptability under small sample conditions.

[0018] Furthermore, the feature embedding network includes an input layer, multiple hidden layers and an embedding layer. The metric learning method adopts a metric learning loss function to optimize the feature embedding network parameters so that similar fault samples are clustered in the embedding space and different fault samples are separated.

[0019] Furthermore, the step of establishing the transfer learning mechanism includes:

[0020] Construct a feature space mapping relationship between the source domain and the target domain;

[0021] Based on feature space mapping, a transfer learning model is constructed to achieve effective transfer of source domain knowledge to the target domain;

[0022] Adopt multi-source transfer learning methods to acquire knowledge from multiple source domains and transfer it to the target domain;

[0023] Design an incremental update mechanism to improve the model's adaptability to new data in the target domain.

[0024] Furthermore, the step of analyzing the temporal state changes of the water pump using the long short-term memory network includes:

[0025] Combine fault prediction results with historical monitoring data and organize them into time series status data;

[0026] Construct a remaining life estimation network based on long short-term memory network to capture the long-term dependence of water pump performance degradation;

[0027] Introducing uncertainty quantification mechanisms to improve the reliability of remaining life estimates;

[0028] Based on the remaining life estimation and uncertainty quantification results, an intelligent switching strategy for water pumps is generated.

[0029] Furthermore, the remaining life estimation network based on the long short-term memory network includes an input layer, a multi-layer bidirectional long short-term memory layer, an attention mechanism layer, a fully connected layer and an output layer, and the output layer generates a remaining life estimation value and prediction uncertainty.

[0030] Furthermore, the steps of adopting a smooth transition seamless switching control algorithm include:

[0031] Build a hydraulic transient model of the pump system to predict pressure fluctuations during switching;

[0032] Optimize water pump switching trajectory based on hydraulic transient model and intelligent switching strategy;

[0033] Generate coordinated control sequences for each control actuator to achieve precise coordination;

[0034] Introduce a real-time feedback adjustment mechanism to deal with uncertainties in actual operations.

[0035] Furthermore, the optimized water pump switching trajectory is achieved by adopting a multi-objective optimization method, in which the objective function comprehensively considers pressure fluctuation index, energy consumption index and switching time index, and is solved through model predictive control, dynamic programming and heuristic algorithm.

[0036] The present invention provides a multi-pump parallel coordinated control system for executing the above-mentioned multi-pump parallel coordinated control method, comprising:

[0037] Multi-sensor data acquisition module, used to collect water pump operation data and perform pre-processing;

[0038] Fault feature learning module, used for fault feature learning based on prototype network;

[0039] Knowledge transfer module, used to establish a transfer learning mechanism to achieve fault prediction;

[0040] Remaining life estimation module, used to analyze the temporal state changes of the water pump based on long short-term memory network;

[0041] The switching control module is used to adopt a seamless switching control algorithm with smooth transition.

[0042] The beneficial effects of the present invention are: through accurate fault prediction and predictive maintenance, the system's unplanned downtime is reduced, the service life of the water pump is extended, and the maintenance cost is reduced. Predictive maintenance replaces the traditional passive response mode, thereby improving the reliability and availability of the system;

[0043] The seamless switching control algorithm with smooth transition reduces water supply pressure fluctuations, effectively avoiding damage to the system caused by water hammer. The smooth switching process ensures the stable operation of the water supply system and improves user satisfaction.

[0044] Under data-scarce conditions, the recognition accuracy of pumps lacking fault-labeled data has been improved, and warning times have been shortened. Small-sample learning and knowledge transfer technologies overcome the traditional method's reliance on large amounts of labeled data, providing maintenance personnel with ample response time.

[0045] Through transfer learning and small sample learning techniques, the present invention can quickly adapt to different types of water pumps and changing operating conditions, and improve the ability to recognize unseen fault patterns. This adaptability enables the system to cope with complex and changing actual operating environments and reduces dependence on large amounts of labeled data.

[0046] Based on accurate remaining life prediction, the present invention can automatically generate the optimal maintenance plan and realize predictive maintenance. Compared with the traditional regular maintenance mode, maintenance efficiency is improved, the demand for spare parts inventory is reduced, and the scientific maintenance plan optimizes resource allocation, avoiding resource waste caused by premature maintenance and equipment failure caused by untimely maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a method for coordinated control of multiple water pumps in parallel in the present invention;

[0048] Figure 2 It is a line graph comparing the fault identification accuracy of the proposed method and the traditional method under different numbers of training samples;

[0049] Figure 3 It is a line graph comparing the relative error of remaining life prediction using transfer learning and not using transfer learning under different target domain data ratio conditions;

[0050] Figure 4 It is a bar graph of the pressure fluctuation amplitude during the water pump switching process under different load conditions using the traditional switching method and the present method;

[0051] Figure 5 This is a radar chart showing the improvement of system reliability and maintenance efficiency of this method compared with traditional methods. DETAILED DESCRIPTION

[0052] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0053] At least one embodiment of the present invention discloses a method for controlling multiple water pumps in parallel. Figure 1 Shown, including:

[0054] Step 1: Collect and preprocess multi-sensor data to construct a multi-dimensional health status feature vector;

[0055] This step involves deploying multiple sensors in the water pump system to collect real-time water pump operation data, pre-processing it, and constructing a multi-dimensional health status feature vector. This step includes the following sub-steps:

[0056] Step 1.1, sensor network deployment;

[0057] Deploy various types of sensors at key locations in the pump system, including:

[0058] Vibration sensor: installed in key parts such as water pump bearings and motors to capture abnormal vibration signals;

[0059] Current sensor: monitors the current of each phase of the motor and is used to analyze electrical performance;

[0060] Temperature sensor: monitors bearing and motor winding temperatures for thermal status assessment;

[0061] Pressure sensor: measures pump inlet and outlet pressure for hydraulic performance analysis;

[0062] Flow sensor: measures the pump flow rate for working condition identification and efficiency calculation.

[0063] Sensor data is transmitted to the data acquisition unit via a fieldbus or wireless network. The sampling frequency is configured according to the characteristics of different parameters. Vibration signals usually use high-frequency sampling (such as 10kHz), while slowly changing parameters such as temperature can use low-frequency sampling (such as 1Hz).

[0064] Step 1.2, data preprocessing;

[0065] Preprocess the collected raw sensor data to eliminate noise and outliers and improve data quality. The preprocessing process includes:

[0066] Missing value processing: Time series interpolation method is used to fill missing data;

[0067] Outlier detection and correction: Use the moving median method or Z-score method to identify and handle outliers;

[0068] Noise filtering: Apply appropriate filtering algorithms based on different signal characteristics, such as using a bandpass filter for vibration signals and a low-pass filter for current signals;

[0069] Data normalization: Unify different types of sensor data to the same scale using Min-Max normalization or Z-score normalization methods.

[0070] Step 1.3, feature extraction;

[0071] Extract features that reflect the health status of the water pump from the preprocessed multi-sensor data. Different feature extraction methods are used for different types of sensor data:

[0072] Time domain characteristics: calculate statistics such as mean, standard deviation, peak, peak-to-peak value, skewness, and kurtosis;

[0073] Frequency domain features: Convert the time domain signal to the frequency domain through Fast Fourier Transform (FFT) to extract spectrum features such as frequency center and power spectrum density;

[0074] Time-frequency characteristics: Use methods such as wavelet transform to perform time-frequency analysis and extract energy distribution characteristics;

[0075] Trend characteristics: Perform trend analysis on long time series data to extract characteristics such as rate of change and periodicity.

[0076] Step 1.4, constructing multidimensional health status feature vector;

[0077] The features extracted by each sensor are fused to construct a multi-dimensional health status feature vector. The feature fusion adopts the following strategy:

[0078] Feature selection: Use recursive feature elimination or principal component analysis to select the most discriminative feature subset;

[0079] Feature combination: combining the features of different sensors into a unified feature vector according to a predefined structure;

[0080] Feature importance weighting: Assign appropriate weights to different features based on their contribution to fault diagnosis.

[0081] The final generated feature vector is expressed as:

[0082] ;

[0083] in represents the multidimensional health status feature vector, 、 、 The first, second, and Multidimensional health status characteristic values, is the dimension of the multidimensional health status feature vector.

[0084] Step 2: Fault feature learning is performed based on the multi-dimensional health state feature vector. The prototype network is used to construct the prototype representation of the fault category in the feature space and the distance between different fault types is maximized and the distance within the same type is minimized through metric learning.

[0085] This step uses the prototype network to establish a small sample learning mechanism, constructing a prototype representation for each type of fault in the feature space, and forming a highly discriminative fault feature space. It specifically includes the following sub-steps:

[0086] Step 2.1, feature embedding network construction;

[0087] Construct a feature embedding network to map the original multi-dimensional health status feature vector to a low-dimensional embedding space to enhance feature expression capabilities. The specific structure of the feature embedding network is as follows:

[0088] Input layer: receives the multidimensional health status feature vector generated in step 1 , the dimension is ;

[0089] The first hidden layer is a fully connected layer with 128 neurons, followed by a batch normalization layer and a ReLU activation function to map the input features into a high-dimensional space.

[0090] The second hidden layer is a fully connected layer with 256 neurons, followed by a batch normalization layer and a ReLU activation function to extract higher-level feature representations.

[0091] The third hidden layer is a fully connected layer with 128 neurons, followed by a batch normalization layer and a ReLU activation function to start the dimensionality reduction process.

[0092] Embedding layer: A fully connected layer containing 64 neurons that generates an embedding vector of fixed dimension, represented as , represents the parameterized feature embedding function, represents the parameters of the feature embedding network, Represents a multidimensional health status feature vector.

[0093] To improve the generalization ability of the feature embedding network, a dropout layer with a dropout rate of 0.3 is added between the first and second hidden layers to prevent overfitting. Furthermore, a residual connection structure is used to connect the input features directly to the output of the third hidden layer after linear transformation. This helps the network retain the original feature information and accelerates training convergence.

[0094] The feature embedding network is trained in an end-to-end manner and optimized together with subsequent prototype computation and metric learning. The initial learning rate is set to 0.001, and the Adam optimizer is used. The learning rate is decayed by a factor of 0.9 every 50 training cycles.

[0095] Step 2.2, fault category prototype construction;

[0096] Based on the output of the feature embedding network, a prototype representation is constructed for each type of fault. For the first type of failure Class Fault, its prototype The calculation is as follows:

[0097] ;

[0098] in, Indicates the Prototype of class fault; Indicates the The number of samples in the class fault sample set; Indicates the A sample set of class faults; Indicates the The multidimensional health status feature vector of samples; represents a parameterized feature embedding function; Represents the parameters of the feature embedding network; Represents the summation operator.

[0099] By calculating the center point of each type of fault sample in the embedding space, a prototype representation of this type of fault is formed.

[0100] In the small sample learning scenario, to improve the stability of the prototype representation, this implementation adopts the following improvement method:

[0101] Weighted prototype calculation: assign weights to each sample based on its quality , the revised prototype calculation formula is:

[0102] ;

[0103] in, Indicates the Weighted prototypes of fault-like classes; Indicates the The weight of the samples Indicates the The multidimensional health status feature vector of samples; represents a parameterized feature embedding function; Represents the parameters of the feature embedding network; Indicates the A sample set of class faults; Represents the summation operator.

[0104] The sample weight is determined based on its consistency with other samples in the class, reducing the impact of abnormal samples on the prototype.

[0105] Prototype enhancement: Use data augmentation technology to expand the sample set, and then calculate the enhanced prototype representation to improve the stability and representativeness of the prototype.

[0106] Prototype ensemble: Train multiple models at different training stages or with different initialization parameters, and integrate the prototypes generated by these models to obtain a more robust prototype representation.

[0107] After the prototype is built, each type of fault prototype is stored in the prototype library as a reference for fault identification. When a new sample is input into the system, the distance between its embedding vector and each prototype is calculated to determine its fault category.

[0108] Step 2.3, metric learning optimization;

[0109] According to an embodiment of the present application, a distance metric in an embedding space is optimized by a metric learning method, so that similar fault samples are aggregated and different fault samples are separated.

[0110] In some embodiments, metric learning optimization can be implemented using a triplet loss function. For each training example, a positive example (one of the same class) and a negative example (one of a different class) are selected to form a triplet. The goal of the triplet loss function is to ensure that the distance between the anchor example and the positive example is less than the distance between the anchor example and the negative example plus a threshold.

[0111] Alternatively, a prototype-based metric learning loss function can be used:

[0112] ;

[0113] in, represents the metric learning loss function; Represents the total number of training samples; Indicates the total number of fault categories; Represents the indicator function, when The samples belong to 1 when the class is valid, otherwise 0; represents the Euclidean distance metric function; represents a parameterized feature embedding function; Represents the parameters of the feature embedding network; Indicates the The multidimensional health status feature vector of samples; Indicates the Prototype of class fault; Indicates the preset boundary parameters; represents the maximum value function; Represents the summation operator.

[0114] In some implementations, to further improve metric learning, a hybrid loss function can be constructed by combining center loss or contrast loss. The hybrid loss function comprehensively considers prototype-based metric loss, center loss, and contrast loss, and adjusts the contribution ratio of each component by using weighting parameters.

[0115] Step 2.4, small sample adaptability is enhanced;

[0116] To improve the model's adaptability under small sample conditions, the following techniques are used:

[0117] Data augmentation technology can use the following data augmentation methods to expand limited labeled samples:

[0118] Signal transformation: scaling, translation, flipping, and other transformations of time domain signals;

[0119] Noise injection: add Gaussian noise, impulse noise or colored noise;

[0120] Mixed enhancement: different samples of the same type are mixed in different proportions to generate new samples;

[0121] Time window sliding: By sliding the time window, multiple subsequences can be extracted from long sequence data.

[0122] For example, for a vibration signal , you can generate enhanced samples:

[0123] ;

[0124] in, represents the enhanced vibration signal; represents the first mixing coefficient; represents the second mixing coefficient; represents the original vibration signal; Indicates the time offset; represents the injected noise signal; Represents a time variable.

[0125] Contrastive learning is implemented through the following steps:

[0126] Generate multiple enhanced views for each original sample;

[0127] Different views of the same sample are considered as positive sample pairs, and views of different samples are considered as negative sample pairs;

[0128] Construct a contrast loss function to minimize the distance between positive sample pairs in the feature space and maximize the distance between negative sample pairs;

[0129] The contrast loss function can be expressed as:

[0130] ;

[0131] in, represents the contrast loss function; represents the natural logarithm function; represents the exponential function; represents the similarity function; Indicates the Feature representation of samples; Indicates The first sample of the same type Feature representation of samples; Indicates Samples of different classes Feature representation of samples; represents the temperature parameter; Represents the summation operator.

[0132] Meta-learning strategies can be implemented in the following ways:

[0133] Model-agnostic meta-learning: Construct a set of meta-tasks, each containing a support set and query set of a small number of samples. By training the model on multiple meta-tasks, it can quickly adapt to new tasks.

[0134] Gradient descent meta-learning: This uses model-sensitive meta-learning methods, such as Model-Agnostic Meta-Learning (MAML), to optimize the model's initial parameters so that it can quickly adapt to new tasks with a small number of gradient updates.

[0135] Memory-enhanced meta-learning: Introducing an external memory module to store and retrieve historical experience related to the current task to assist in small-sample learning.

[0136] By combining the above technologies, this embodiment can effectively identify different fault types under small sample conditions and provide reliable fault diagnosis capabilities for the multi-pump parallel system.

[0137] Through these sub-steps, a prototype network model was constructed that can effectively identify different fault types under small sample conditions. This model can map the health status feature vector of the water pump into the feature space and identify the fault type by comparing the distance with the prototype of each fault category.

[0138] Step 3: Establish a transfer learning mechanism based on the prototype representation of fault categories, and transfer the existing historical fault data of water pumps as source domain knowledge to water pumps that lack fault label data to achieve fault prediction;

[0139] This step establishes a transfer learning mechanism, using existing historical pump failure data as source domain knowledge to transfer to pumps in the target domain that lack fault labeling data, thereby achieving fault prediction. This includes the following sub-steps:

[0140] Step 3.1: Mapping the source domain and target domain feature spaces;

[0141] Construct a feature space mapping relationship between the source domain and the target domain so that the knowledge in the source domain can be effectively transferred to the target domain. The specific implementation method is as follows:

[0142] Define the source domain feature space and the target domain feature space , corresponding to the pumps with sufficient data and the pumps lacking fault marking data, respectively;

[0143] Construct the feature mapping function by sharing the underlying feature extraction layer and adding an adaptation layer:

[0144] ;

[0145] in, represents the feature mapping function; Represents the source domain feature space; Represents the target domain feature space; Indicates the mapping relationship symbol.

[0146] Use Maximum Mean Discrepancy (MMD) or domain adversarial training methods to reduce the difference in feature distribution between the source domain and the target domain.

[0147] Through feature space mapping, the consistency of feature representation between the source domain and the target domain is achieved, laying the foundation for knowledge transfer.

[0148] Step 3.2, transfer learning model construction;

[0149] The transfer learning model construction method proposed in this application first pre-trains a fault prediction model on source domain data and then transfers this knowledge to the target domain through a specific transfer mechanism. The transfer learning loss function consists of two parts: task loss and adaptability loss. The task loss is calculated based on the limited labeled data in the target domain, ensuring that the model maintains accuracy in the target domain. The adaptability loss measures the difference in model parameters between the source and target domains, controlling the degree of similarity between the two.

[0150] The transfer learning loss function can be expressed as:

[0151] ;

[0152] in, represents the total loss function of the target domain; represents the target domain task loss; represents a loss of fitness; represents the trade-off parameter; Represents the source domain model parameters; Represents the target domain model parameters.

[0153] In some implementations, transfer learning models can be implemented using domain adversarial training. This method introduces a domain discriminator and uses adversarial learning to reduce the difference in feature distributions between the source and target domains. Additionally, methods such as feature adjustment, pseudo-labeling, and adversarial generative network assistance can also be used to achieve effective knowledge transfer.

[0154] Step 3.3, multi-source transfer learning enhancement;

[0155] To further improve the transfer learning effect, a multi-source transfer learning method is used to acquire knowledge from multiple source domains and transfer it to the target domain. The specific implementation includes:

[0156] Select multiple related source domains, such as pump failure data of different models and operating conditions;

[0157] Assign a weight to each source domain, reflecting its relevance to the target domain;

[0158] Weighted fusion of knowledge from multiple source domains to form richer prior knowledge;

[0159] Construct a multi-source transfer learning model to achieve the collaborative transfer of multi-source knowledge to the target domain.

[0160] Multi-source transfer learning can comprehensively utilize the complementary information of different source domains to improve the fault prediction performance in the target domain.

[0161] Step 3.4: incremental learning optimization of the target domain;

[0162] In order to improve the model's adaptability to new data in the target domain, an incremental learning method is used for optimization:

[0163] Design an incremental update mechanism to enable the model to effectively utilize newly collected data in the target domain;

[0164] Adopting an incremental learning strategy to balance learning from new data and retaining existing knowledge;

[0165] Set up a knowledge distillation mechanism to retain useful knowledge in the original model and avoid catastrophic forgetting;

[0166] Regularly evaluate the performance of the model on the target domain and dynamically adjust the transfer learning parameters.

[0167] Through incremental learning optimization, the model can continuously improve itself as the target domain data accumulates, thereby improving the accuracy of fault prediction.

[0168] Through the above sub-steps, a fault prediction mechanism based on transfer learning is implemented, enabling the model to effectively predict water pump failures using source domain knowledge when target domain data is limited, providing a basis for subsequent remaining life estimation and intelligent switching strategy generation.

[0169] Step 4: Based on the fault prediction results and historical monitoring data, the long short-term memory network is used to analyze the temporal state changes of the water pump, predict the failure probability and remaining service life, and generate the optimal switching strategy;

[0170] This step uses the Long Short-Term Memory (LSTM) network to analyze the temporal state changes of the water pump, predict the failure probability and remaining service life, and generate the optimal switching strategy. It includes the following sub-steps:

[0171] Step 4.1, time series state data organization;

[0172] Combine the fault prediction results in the previous steps with the historical monitoring data and organize them into time series status data for remaining life estimation:

[0173] Construct a sequence of time windows, where each window contains state data of a fixed duration (such as one day or one week);

[0174] Extract statistical features and trend features within the window, such as mean, variance, slope, etc.;

[0175] Combined with the fault prediction results, each time window is marked with a corresponding health index or fault probability;

[0176] The data is organized in chronological order to form a time series state sequence that reflects the performance degradation process of the equipment.

[0177] The time series status data provides the input basis for the subsequent remaining life estimation.

[0178] Step 4.2, LSTM network construction;

[0179] Construct an LSTM-based remaining life estimation network to capture the long-term dependencies of water pump performance degradation:

[0180] The specific structure of the LSTM network is as follows:

[0181] Input layer: Receives time series state data with the dimensions [batch size, time step, feature dimension], where the time step is set to 48 (representing two days of data, one sampling point per hour), and the feature dimension is determined by the number of features extracted in step 1;

[0182] The first LSTM layer: contains 128 LSTM units with a bidirectional structure, which is used to capture the forward and backward dependencies of time series data. The output dimension is [batch size, time step, 256] (bidirectional output connection);

[0183] The second LSTM layer: contains 64 LSTM units with a bidirectional structure to further extract time series features. The output dimension is [batch size, time step, 128];

[0184] Attention mechanism layer: Implements a multi-head self-attention mechanism with 4 heads and 32 dimensions per head. It is used to learn the dependencies between different time steps and focus on state changes at key time points.

[0185] Global pooling layer: uses attention-weighted average pooling to aggregate features in the time dimension into a vector with an output dimension of [batch size, 128];

[0186] The first fully connected layer contains 64 neurons, followed by a batch normalization layer and a ReLU activation function;

[0187] The second fully connected layer contains 32 neurons, followed by a batch normalization layer and a ReLU activation function;

[0188] Output layer: Contains 2 neurons, which output the estimated remaining life and the prediction uncertainty respectively. The uncertainty is achieved by learning the logarithm of the prediction variance.

[0189] The multi-head self-attention mechanism implemented in the embodiment of the present application is as follows: first, the hidden state sequence of the LSTM output is transformed into a query matrix, a key matrix and a value matrix through three different linear transformations; then the attention weight is calculated, and the weight calculation is based on the similarity between the query and the key, and is normalized by softmax; then the attention output is obtained, and the weight is multiplied by the value matrix; finally, for the multi-head mechanism, the outputs of multiple independent attention calculation units are spliced ​​and the final result is obtained through linear transformation.

[0190] It's important to note that the attention mechanism can adaptively focus on important parts of time series data, improving the model's efficiency in utilizing information at key time points. In estimating the remaining life of a water pump, state data from different time periods contributes differently to the prediction results. The attention mechanism can effectively identify and focus on time points that contain important degradation information.

[0191] According to an embodiment of the present application, the LSTM network is trained using a negative log-likelihood loss function, taking into account both prediction error and uncertainty. The loss function can be expressed as:

[0192] ;

[0193] in, Represents the loss function of the LSTM network; represents the exponential function; represents the uncertainty of the prediction (logarithm of the variance); Indicates the true remaining lifespan; Represents the predicted remaining life value.

[0194] In addition, this application also adopts a series of training strategies, including Adam optimizer, cosine annealing learning rate scheduling, Dropout regularization, early stopping strategy and gradient clipping, to improve the generalization ability and training stability of the model.

[0195] In order to improve the reliability of the remaining life estimation, an uncertainty quantification mechanism is introduced:

[0196] Monte Carlo dropout (MCDropout) is used to keep the dropout layer activated during the inference phase to obtain multiple prediction results;

[0197] Calculate the mean of multiple predictions as the final remaining life estimate ;

[0198] Calculate the standard deviation of the predictions as a measure of uncertainty ;

[0199] Based on uncertainty information, the credibility of the prediction results is evaluated to guide subsequent decision-making.

[0200] Uncertainty quantification provides reliability assurance for remaining life estimation and avoids the risks that may be caused by single-point prediction.

[0201] Step 4.4, intelligent switching strategy generation;

[0202] Generate intelligent pump switching strategies based on the remaining life estimation and uncertainty quantification results:

[0203] Setting safety thresholds ,When the estimated remaining life is less than the threshold, a switching warning is triggered;

[0204] Consider the impact of uncertainty, adjust the timing of switching warnings, and issue early warnings in cases of high uncertainty;

[0205] Optimize the start and stop sequence of water pumps according to the load status, efficiency curve and health status of each pump;

[0206] Considering system stability and energy efficiency requirements, a multi-step predictive switching plan is generated, including switching timing, sequence, and transition method.

[0207] The intelligent switching strategy pre-starts the backup pump before a fault occurs, ensuring continuous and stable system operation while avoiding unnecessary premature switching and achieving predictive maintenance.

[0208] Through the above sub-steps, the remaining life estimation and intelligent switching strategy generation based on time series data analysis are realized, providing a decision-making basis for the final smooth transition and seamless switching control.

[0209] Step 5: Based on the optimal switching strategy, a smooth transition seamless switching control algorithm is adopted to eliminate water hammer effects and pressure fluctuations during the switching process by precisely controlling the speed change rate of each pump and adjusting the valve opening timing.

[0210] This step uses a smooth transition seamless switching control algorithm to eliminate water hammer effects and pressure fluctuations during the switching process by precisely controlling the speed change rate of each pump and the valve opening adjustment timing. It specifically includes the following sub-steps:

[0211] Step 5.1, hydraulic transient model construction;

[0212] Build a hydraulic transient model of the pump system to predict pressure fluctuations during switching:

[0213] A mathematical model based on the characteristic line method, which is used to establish a mathematical model of hydraulic transients in a pipe network. This method is based on the mass conservation equation and momentum conservation equation in fluid mechanics. For one-dimensional flow in a pipe, water hammer can be described using the continuity equation and momentum equation.

[0214] The continuity equation represents the principle of conservation of fluid mass, taking into account the relationship between factors such as head, flow velocity, wave velocity, and pipe cross-sectional area. The momentum equation represents the principle of conservation of fluid momentum, taking into account the influence of factors such as flow velocity, head, friction coefficient, and pipe diameter.

[0215] It should be understood that through the characteristic line method, the above partial differential equation can be transformed into an ordinary differential equation along the positive and negative characteristic lines, so that the complex hydraulic transient problem can be transformed into a form that is easy to solve numerically. For the positive characteristic line, its propagation speed is the wave speed , and the propagation speed of the negative characteristic line is .

[0216] The numerical solution method provided in this application uses the finite difference method to solve the characteristic line equation. The calculation steps are as follows:

[0217] Discrete the pipe network into multiple nodes and the time dimension into multiple time steps;

[0218] At each internal node, using the known values ​​of the previous time step, trace back to two points of the previous time step along the positive and negative characteristic lines, and obtain the head and flow velocity values ​​of these two points by interpolation;

[0219] According to the characteristic line equation, calculate the water head and flow velocity of the node in the current time step;

[0220] For boundary nodes (such as water pumps, valves, etc.), solve them in combination with specific boundary condition equations;

[0221] Repeat the above steps and gradually advance the calculation to obtain the hydraulic transient process in the entire time domain.

[0222] Regarding the pump boundary, it should be noted that this application uses a complete four-quadrant characteristic curve to describe the behavior of the pump under various operating conditions, including normal operation, deceleration, reversal, and braking. The pump characteristic curve is usually expressed as a head equation and a torque equation, as shown below:

[0223] ;

[0224] in, Indicates lift; Indicates rated head; Indicates relative speed; Indicates the pump characteristic coefficient; Indicates flow rate; Represents the square operation.

[0225] ;

[0226] in, Indicates torque; Indicates rated torque; Indicates relative speed; Indicates the torque characteristic coefficient; Indicates flow rate; Represents the square operation.

[0227] The change in pump speed during the switching process can be described by the relationship between the motor torque, pump torque, and moment of inertia. The rate of change in speed is proportional to the torque difference and the moment of inertia, a relationship that is crucial for modeling the dynamic characteristics of the pump's start-stop process.

[0228] To improve the prediction accuracy of the model, the following steps are used for calibration and validation:

[0229] Utilize the actual operating data of the water pump system under different working conditions, including time series data such as pressure, flow, speed, etc. under normal start-stop and fault switching conditions;

[0230] Use parameter identification methods to optimize key parameters in the model, such as wave speed, friction coefficient, water pump characteristic parameters, etc.

[0231] The optimal parameter combination is obtained by minimizing the error function between the model prediction value and the measured value;

[0232] An independent validation dataset was used to evaluate the model performance and ensure the model’s prediction accuracy under different operating conditions.

[0233] The calibrated hydraulic transient model can accurately predict the pressure fluctuations during the pump switching process, providing a reliable simulation environment for subsequent switching trajectory optimization.

[0234] Step 5.2, switch trajectory optimization;

[0235] Optimize pump switching trajectory based on hydraulic transient model and intelligent switching strategy:

[0236] The handover process is modeled as a multi-stage optimization problem, including pre-start, load transfer, and smooth transition.

[0237] Define the objective function of the switching process, taking into account the pressure fluctuation amplitude, energy consumption and switching time;

[0238] Set constraints, including pump operating parameter ranges, system pressure limits, flow balance, etc.;

[0239] Sequential quadratic programming or model predictive control methods are used to solve the optimization problem and obtain the optimal switching trajectory.

[0240] In some embodiments, the switching trajectory optimization can be implemented using a multi-objective optimization method, and the objective function is designed as follows:

[0241] ;

[0242] in, represents the overall objective function; Indicates the weight coefficient of the pressure fluctuation index; Indicates the weight coefficient of energy consumption index; Represents the weight coefficient of the switching time indicator; Indicates pressure fluctuation index; Indicates energy consumption index; Indicates the switching time indicator.

[0243] Optionally, for different application scenarios, the switching trajectory optimization can adopt the following different methods:

[0244] Optimization based on Model Predictive Control (MPC):

[0245] At each control step, the future system behavior is predicted based on the current system state and the hydraulic transient model;

[0246] Solve the optimization problem in the prediction time domain and obtain a series of control actions;

[0247] Execute the first control action, then slide the time window and repeat the optimization process;

[0248] This method can respond to system state changes in real time and has good robustness.

[0249] Optimization based on dynamic programming:

[0250] Discretize the switching process into a finite state and action space;

[0251] Define the state transition equation and immediate reward function;

[0252] Solve the optimal strategy through reverse recursion;

[0253] This method is suitable for situations where the system dynamic characteristics are clear and the state space is small.

[0254] Optimization based on heuristic algorithms:

[0255] Use heuristic methods such as genetic algorithms and particle swarm optimization to search for the optimal switching trajectory;

[0256] Finding approximate optimal solutions in large-scale search spaces by simulating evolution or swarm intelligence mechanisms;

[0257] This method has high computational complexity, but it can handle nonlinear and non-convex optimization problems.

[0258] Switching trajectory optimization ensures the smoothness and safety of the water pump switching process and minimizes the impact on system stability.

[0259] Step 5.3, collaborative control sequence generation;

[0260] Based on the optimized switching trajectory, the coordinated control sequence of each control actuator is generated:

[0261] Generate detailed speed change curve for each water pump and control the time constant of acceleration and deceleration;

[0262] Generate opening adjustment sequence for each valve to ensure flow balance and pressure stability;

[0263] Coordinate the action sequence of each actuator to avoid control conflicts;

[0264] Set up key point monitoring trigger mechanism to achieve dynamic adjustment based on real-time feedback.

[0265] The coordinated control sequence achieves precise coordination of multiple control actuators and ensures smooth transition of system parameters during the switching process.

[0266] Step 5.4, real-time feedback adjustment;

[0267] In order to cope with the uncertainty in actual operation, a real-time feedback adjustment mechanism is introduced:

[0268] Build a state estimator based on Kalman filtering to update the system status in real time;

[0269] Set safety thresholds for key parameters such as pressure, flow, and vibration;

[0270] When the actual measurement value deviates from the expected trajectory by more than a preset threshold, the adaptive adjustment of the control parameters is triggered;

[0271] According to the deviation size and change trend, the control sequence is dynamically modified to achieve closed-loop control.

[0272] The real-time feedback adjustment mechanism improves the robustness of switching control, can cope with the influence of model errors and external interference, and ensure the safety and reliability of the switching process.

[0273] Through the above sub-steps, smooth transition and seamless switching control of the water pump system is achieved, effectively eliminating the water hammer effect and pressure fluctuations during the switching process, and ensuring the stability of the system water supply pressure and flow.

[0274] A multi-pump parallel coordinated control system, used to execute the above-mentioned multi-pump parallel coordinated control method, comprising:

[0275] Multi-sensor data acquisition module, used to collect water pump operation data and perform pre-processing;

[0276] Fault feature learning module, used for fault feature learning based on prototype network;

[0277] Knowledge transfer module, used to establish a transfer learning mechanism to achieve fault prediction;

[0278] Remaining life estimation module, used to analyze the temporal state changes of the water pump based on long short-term memory network;

[0279] The switching control module is used to adopt a seamless switching control algorithm with smooth transition.

[0280] Here, the present invention provides an implementation example:

[0281] The municipal water supply system consists of four large centrifugal pumps operating in parallel, three in regular operation and one in standby. Each pump has a power of 315 kW, a rated flow rate of 1,200 m³ / h, and a rated head of 78 m. The system must maintain stable discharge pressure and flow to meet the city's water supply needs. The main challenges facing the water supply system include:

[0282] The aging degree of water pumps varies, making it difficult to obtain complete fault data;

[0283] Some pump models are newer and lack sufficient fault marking data;

[0284] The load fluctuates greatly, especially during peak water usage periods, and pumps need to be switched smoothly to maintain system stability.

[0285] Traditional periodic maintenance methods result in high maintenance costs and insufficient reliability.

[0286] A variety of sensors are installed on each water pump, including: bearing vibration sensor (sampling frequency 10kHz), motor current sensor (sampling frequency 1kHz), bearing and motor temperature sensors (sampling frequency 1Hz), inlet and outlet pressure sensors (sampling frequency 10Hz) and flow sensor (sampling frequency 10Hz).

[0287] Sensor data is collected and preprocessed through the data acquisition system. For example, using vibration data, a bandpass filter (10Hz-5kHz) is used to remove low-frequency interference and high-frequency noise. Outlier detection is then performed using a moving window method. Data points exceeding ±3 standard deviations from the mean are marked as abnormal and corrected.

[0288] After preprocessing, multidimensional feature vectors were extracted, including time domain features (RMS value, peak value, kurtosis, etc.), frequency domain features (characteristic frequency amplitude, spectral entropy, etc.) and time-frequency features (wavelet energy distribution, etc.). A total of 76-dimensional feature vectors were constructed for subsequent analysis.

[0289] Due to the limited fault labeling data for water pumps, this embodiment uses a prototype network for small sample learning. For a certain model of water pump, there are only 7 samples of bearing failure, 5 samples of impeller wear, 9 samples of seal leakage, and 30 samples of normal operation.

[0290] A feature embedding network was constructed to map the 76-dimensional original features into a 64-dimensional embedding space. Data augmentation techniques were used to expand the limited number of fault samples. These techniques included adding Gaussian noise (with a standard deviation of 10% of the original signal) and sliding the time window (with a step size of 1 / 4 the original window), increasing the number of fault samples by five times.

[0291] Metric learning is used to optimize the embedding space, clustering similar fault samples and separating different fault categories. After training, the model achieved a 91.2% fault identification accuracy on an independent test set, even with only five samples per fault category. This is a 23.5 percentage point improvement over traditional support vector machine methods.

[0292] For older pumps (source domain), where sufficient data is available, a complete fault prediction and remaining life estimation model was trained using historical operating data. Transfer learning was used to transfer the source domain knowledge to pumps (target domain), where labeled fault data is scarce.

[0293] In the implementation, the fault prediction model is first pre-trained in the source domain. Then, knowledge transfer is achieved through feature mapping and model parameter adjustment. During the transfer process, domain adversarial training is used to reduce the difference in feature distribution between the source and target domains while preserving task-related information.

[0294] Based on the transferred model, an LSTM network was combined to analyze the performance degradation trends of water pumps and predict their remaining useful life. Even with only 30% labeled data in the target domain, the average relative error of remaining useful life prediction remained within 12.5%, significantly outperforming a baseline method without transfer learning (relative error of 28.7%).

[0295] When the system predicts that the remaining life of a pump (for example, Pump 1) falls below a safety threshold (set at 300 hours) with low uncertainty, the intelligent switching strategy is triggered. The system selects the backup pump (Pump 4) as a replacement and generates an optimal switching trajectory.

[0296] The switching process consists of three phases:

[0297] Pre-start phase: Start the standby pump to a low speed (about 30% of the rated speed) to form initial flow in the pipe network;

[0298] Load transfer stage: According to the speed curve generated by the optimization algorithm, the speed of the standby pump is gradually increased, while the speed of the pump to be switched is reduced to achieve smooth load transfer;

[0299] Smooth transition stage: After the switching pump drops to the minimum speed and shuts down safely, the standby pump takes over all the load and the system resumes stable operation.

[0300] During the switching process, the system monitored pipeline pressure fluctuations in real time and adjusted switching parameters based on feedback to ensure a smooth process. In practice, the switching process took approximately 120 seconds, and the pressure fluctuations throughout the process were controlled within ±0.05 MPa, with no significant water hammer effect.

[0301] This implementation method has achieved good results in the municipal water supply system, mainly in the following two aspects:

[0302] The system reliability improvement effect is shown in Table 1:

[0303] Table 1: System reliability improvement data

[0304]

[0305] The effect of improving water supply quality stability is shown in Table 2:

[0306] Table 2: Data on water supply quality stability improvement

[0307]

[0308] The above results demonstrate that this implementation not only improves the reliability and stability of a multi-pump parallel system, but also reduces maintenance costs and improves energy efficiency. Particularly under data-scarce conditions, this method demonstrates outstanding fault prediction and predictive maintenance capabilities, providing an efficient and reliable solution for multi-pump parallel applications such as municipal water supply and industrial circulating water.

[0309] like Figures 2 to 5 As shown, there are line graphs comparing the fault identification accuracy of the proposed method and the traditional method under different numbers of training samples; a line graph comparing the relative errors of remaining life prediction using transfer learning and not using transfer learning under different target domain data ratios; a bar graph of the pressure fluctuation amplitude during the water pump switching process under different load conditions using the traditional switching method and the proposed method; and a radar chart showing the improvement effect of the proposed method in terms of system reliability and maintenance efficiency compared with the traditional method.

[0310] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for controlling multiple water pumps in parallel, characterized in that: include: Collecting and preprocessing multi-sensor data to construct a multi-dimensional health status feature vector, wherein the multi-sensor data includes vibration data, current data, temperature data, pressure data, and flow data, and the multi-dimensional health status feature vector includes time domain features, frequency domain features, and time-frequency features; Fault feature learning is performed based on multi-dimensional health state feature vectors. Prototype networks are used to construct prototype representations of fault categories in feature space. Metric learning is used to maximize the distance between different fault types and minimize the distance within the same type. A transfer learning mechanism is established based on the prototype representation of fault categories. The existing historical fault data of water pumps is used as source domain knowledge to transfer to water pumps that lack fault label data to achieve fault prediction. Based on fault prediction results and historical monitoring data, a long-short-term memory network is used to analyze the temporal state changes of the water pump, predict the failure probability and remaining service life, and generate the optimal switching strategy; Based on the optimal switching strategy, a seamless switching control algorithm with smooth transition is adopted. By accurately controlling the speed change rate of each pump and the valve opening adjustment timing, the water hammer effect and pressure fluctuation during the switching process are eliminated.

2. A multi-pump parallel coordinated control method according to claim 1, characterized in that: The steps of fault feature learning include: Construct a feature embedding network to map the original multi-dimensional health status feature vector into a low-dimensional embedding space; Based on the output of the feature embedding network, a prototype representation is constructed for each type of fault; Optimizing the distance metric in the embedding space through metric learning methods; Data augmentation, contrastive learning and meta-learning strategies are used to improve the model's adaptability under small sample conditions.

3. A multi-pump parallel coordinated control method according to claim 2, characterized in that: The feature embedding network includes an input layer, multiple hidden layers and an embedding layer. The metric learning method adopts a metric learning loss function to optimize the feature embedding network parameters so that similar fault samples are clustered in the embedding space and different fault samples are separated.

4. The method for controlling multiple water pumps in parallel according to claim 1, characterized in that: The steps of establishing the transfer learning mechanism include: Construct a feature space mapping relationship between the source domain and the target domain; Based on feature space mapping, a transfer learning model is constructed to achieve effective transfer of source domain knowledge to the target domain; Adopt multi-source transfer learning methods to acquire knowledge from multiple source domains and transfer it to the target domain; Design an incremental update mechanism to improve the model's adaptability to new data in the target domain.

5. The method for controlling multiple water pumps in parallel according to claim 1, characterized in that: The step of analyzing the temporal state change of the water pump using the long short-term memory network includes: Combine fault prediction results with historical monitoring data and organize them into time series status data; Construct a remaining life estimation network based on long short-term memory network to capture the long-term dependence of water pump performance degradation; Introducing uncertainty quantification mechanisms to improve the reliability of remaining life estimates; Based on the remaining life estimation and uncertainty quantification results, an intelligent switching strategy for water pumps is generated.

6. The method for controlling multiple water pumps in parallel according to claim 5, characterized in that: The remaining life estimation network based on the long short-term memory network includes an input layer, a multi-layer bidirectional long short-term memory layer, an attention mechanism layer, a fully connected layer and an output layer, and the output layer generates a remaining life estimation value and prediction uncertainty.

7. The method for controlling multiple water pumps in parallel according to claim 1, characterized in that: The steps of adopting a smooth transition seamless switching control algorithm include: Build a hydraulic transient model of the pump system to predict pressure fluctuations during switching; Optimize water pump switching trajectory based on hydraulic transient model and intelligent switching strategy; Generate coordinated control sequences for each control actuator to achieve precise coordination; Introduce a real-time feedback adjustment mechanism to deal with uncertainties in actual operations.

8. The method for controlling multiple water pumps in parallel according to claim 7, characterized in that: The optimized water pump switching trajectory is achieved by adopting a multi-objective optimization method. The objective function comprehensively considers pressure fluctuation index, energy consumption index and switching time index, and is solved through model predictive control, dynamic programming and heuristic algorithm.

9. A multi-pump parallel coordinated control system, characterized in that: A method for implementing a multi-pump parallel coordinated control method according to any one of claims 1 to 8, comprising: Multi-sensor data acquisition module, used to collect water pump operation data and perform pre-processing; Fault feature learning module, used for fault feature learning based on prototype network; Knowledge transfer module, used to establish a transfer learning mechanism to achieve fault prediction; Remaining life estimation module, used to analyze the temporal state changes of the water pump based on long short-term memory network; The switching control module is used to adopt a seamless switching control algorithm with smooth transition.

Images