Water delivery pump set health state monitoring method and system based on graph neural network

Abstract

The invention discloses a method and system for monitoring the health state of a water delivery pump group based on a graph neural network, and the method comprises the following steps: collecting the multi-modal data of the water delivery pump group, the multi-modal data comprising sound data and vibration data; preprocessing the multi-modal data, and constructing a training data set; training a preset graph neural network model based on the training data set to obtain a water delivery pump group health state monitoring model; and the real-time multi-modal data of the water feeding pump set is input into the water feeding pump set health state monitoring model, the health state of the water feeding pump set is predicted, and the accuracy of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a method and system for monitoring the health status of a water pump group based on a graph neural network. Background Art

[0002] With the rapid development of social economy and the rapid development of information technology, the application of intelligent detection equipment in the water supply industry is becoming more and more extensive. Its main purpose is to improve the maintenance level of equipment and optimize water supply efficiency, so as to achieve effective management and utilization of limited water resources. However, the existing health monitoring technology has some obvious limitations, especially in the fusion of multimodal signals. At present, most health monitoring methods fail to effectively integrate various types of signals (such as vibration signals, acoustic signals, etc.) generated during the operation of water pump groups, resulting in insufficient accuracy of the diagnostic model, which further affects the comprehensive evaluation of the overall health status of water pump groups.

[0003] Hefei Sanyi Jianghai Intelligent Technology Co., Ltd. proposed a water pump health monitoring method based on curve similarity evaluation in its patent application "An intelligent control system for water pump units with health monitoring and fault diagnosis" (application patent number: 202310578835.4, application publication number: CN116557328B). It takes vibration, sound, temperature and other signals as input, calculates the monitoring index of each signal and generates a monitoring curve based on it, and calculates the similarity with the standard monitoring curve to evaluate the health status of the water pump unit. However, the above-mentioned literature method only integrates the monitoring index of each signal, and does not use artificial intelligence technology to deeply integrate each signal, so the accuracy of the diagnostic model is not enough. Summary of the invention

[0004] The main purpose of the present invention is to provide a method and system for monitoring the health status of a water pump group based on a graph neural network, aiming to solve the technical problem of insufficient accuracy of the diagnostic model in the background technology.

[0005] To achieve the above object, the first aspect of the present invention provides a method for monitoring the health status of a water pump group based on a graph neural network, which is characterized by comprising the following steps:

[0006] Step S100: collecting multimodal data of a water pump group, wherein the multimodal data includes sound data and vibration data;

[0007] Step S200: preprocessing the multimodal data to construct a training data set;

[0008] Step S300: training a preset graph neural network model based on the training data set to obtain a water pump group health status monitoring model;

[0009] Step S400: inputting the real-time multimodal data of the water pump group into the water pump group health status monitoring model to predict the health status of the water pump group.

[0010] Furthermore, in step S300, the step of training the preset graph neural network model based on the training data set to obtain the water delivery pump group health status monitoring model includes:

[0011] Step S310: extracting the health features of the water pump group of each modal data from the pre-processed multi-modal data, wherein the health features of the water pump group include the global features of the sound signal and the local features of the vibration signal;

[0012] Step S320: fusing the health features of the water pump group through a feature extraction network of a graph neural network model to obtain a fused health feature of the water pump group;

[0013] Step S330: predicting the health status of the water pump group through the regression prediction network of the graph neural network model according to the fused health characteristics of the water pump group;

[0014] Step S340: Calculate the gradient value of the model parameter according to the predicted value and the true value of the graph neural network model, and update the model parameter according to the gradient value;

[0015] Step S350: input the training data set into the graph neural network model training to obtain a water pump group health status monitoring model.

[0016] Furthermore, in step S340, the step of calculating the gradient value of the model parameter according to the predicted value and the true value of the graph neural network model includes the following steps:

[0017] Step S341: A first loss function is constructed using a pseudo-Huber function to evaluate the difference between the model prediction value and the true value, as shown in the following formula:

[0018]

[0019] Where, L PseudoHuber is the loss value of the training data set, i is the sample of the training data set, N offline is the number of samples in the training data set, r i is the true value of the health index of the water pump group of sample i, p i is the predicted value of the health index of the water pump group of sample i, δ is the hyperparameter for switching the control function from quadratic to linear;

[0020] Step S342: Calculate the loss value using the first loss function according to the predicted value and the true value of the graph neural network model;

[0021] Step S343: Calculate the gradient value of the model parameter according to the loss value through the back propagation algorithm.

[0022] Furthermore, in step S350, the step of inputting the training data set into the graph neural network model training to obtain the water pump group health status monitoring model includes:

[0023] Step S351: randomly initializing the parameters of the water pump group health status monitoring model;

[0024] Step S352: setting the optimizer of the water pump group health status monitoring model to adopt an adaptive moment estimation Adam optimizer;

[0025] Step S353: randomly collecting a preset number of sample subsets from the training data set, inputting them into the water pump group health status monitoring model, and generating predicted values ​​of the sample subsets;

[0026] Step S354: according to the predicted value and the true value of the sample subset, the loss value of the sample subset is calculated using the first loss function, the gradient value of the model parameter is calculated according to the loss value of the sample subset through the back propagation algorithm, and the model parameter is updated according to the calculated gradient value using the Adam optimizer;

[0027] Step S355: determine whether the sample subset traverses the entire training data set. If so, increase the number of iterative optimization times by one. If not, return to step S353 to continue collecting new sample subsets.

[0028] Step S356: Determine whether the number of iterative optimizations reaches a set threshold. If so, stop the iterative optimization. If not, return to step S353 to start a new round of iterative optimization.

[0029] Furthermore, the step S400 further includes the following steps:

[0030] Step S500: regularly checking the online update conditions of the water pump group health status monitoring model according to a preset statistical period, and performing online update or continuing monitoring according to the inspection result.

[0031] Furthermore, in step S500, the step of regularly checking the online update condition of the water pump group health status monitoring model according to a preset statistical period, and performing online update or continuing monitoring according to the inspection result includes:

[0032] Step S510: Check whether the water pump group health status monitoring model meets one of the following online update conditions: the model online time reaches the preset update cycle threshold, and the model performance index is lower than the preset threshold. If any of the conditions is met, step S520 is executed. If none of the conditions are met, the model status continues to be monitored and waits for the next periodic check.

[0033] Step S520: Acquire the latest multimodal data as training data, and perform online training on the water pump group health status monitoring model to obtain an updated model.

[0034] Furthermore, in step S520, the step of obtaining the latest multimodal data as training data and performing online training on the water pump group health status monitoring model to obtain an updated model includes:

[0035] Step S521: construct a second loss function, which includes an unsupervised learning part and a lifelong learning part, as shown in the following formula:

[0036]

[0037] In the formula, L is the loss value of the training data set, i is the sample of the training data set, and N online is the number of online samples, L USL is the loss value of the unsupervised learning part of the training dataset, L LL is the loss value of the lifelong learning part of the training data set; p i The predicted value of the health index output by the current model for sample i; p0 is the historical maximum value of the health index predicted by the model for sample i; is the predicted value of the health indicator output by the historical model of sample i; λ, β, and γ are all trade-off parameters that control the order of magnitude of each loss; α is a hyperparameter that controls the switch between the health indicator from slow change to sudden change;

[0038] Step S522: Calculate the loss value using the second loss function according to the predicted value and the true value of the graph neural network model;

[0039] Step S523: Calculate the gradient value of the model parameter according to the loss value through the back propagation algorithm;

[0040] Step S524: Update the model parameters according to the calculated gradient values.

[0041] Furthermore, in step S520, the step of obtaining the latest multimodal data as training data and performing online training on the water pump group health status monitoring model to obtain an updated model also includes:

[0042] Step S610: Acquire the latest multimodal data as training data and construct an updated training data set;

[0043] Step S620: randomly initializing the parameters of the water pump group health status monitoring model;

[0044] Step S630: setting the optimizer of the water delivery pump group health status monitoring model to adopt the Adam optimizer;

[0045] Step S640: randomly collecting a preset number of sample subsets from the updated training data set, inputting them into the water pump group health status monitoring model, and generating predicted values ​​of the sample subsets;

[0046] Step S650: according to the second loss function, calculate the loss value of the predicted value of the sample subset and the true label of the sample, calculate the gradient value of the second loss function with respect to the model parameters by back propagation algorithm, and use Adam optimizer to update the model parameters according to the calculated gradient value;

[0047] Step S660: Determine whether the sample subset traverses the entire updated training data set. If so, the number of iterations is increased by one. If not, return to step S640 to continue collecting new sample subsets.

[0048] Step S670: Determine whether the number of iterative optimizations reaches a set threshold. If so, stop the iteration; if not, return to step S640 to start a new round of iterative optimization.

[0049] A second aspect of the present invention provides a water pump group health status monitoring system based on a graph neural network, comprising:

[0050] A collection module, used for collecting multimodal data of the water pump group, wherein the multimodal data includes sound data and vibration data;

[0051] A processing module, used to pre-process the multimodal data and construct a training data set;

[0052] A generation module is used to train a preset graph neural network model based on the training data set to obtain a water pump group health status monitoring model;

[0053] The prediction module inputs the real-time multimodal data of the water pump group into the water pump group health status monitoring model to predict the health status of the water pump group.

[0054] Furthermore, the water pump group health status monitoring system based on graph neural network also includes:

[0055] The update module is used to regularly check the online update conditions of the water pump group health status monitoring model according to a preset statistical period, and perform online update or continue monitoring according to the inspection results.

[0056] The third aspect of the present invention provides a computer-readable storage medium, on which is stored a water pump group health status monitoring processing program based on a graph neural network. When the water pump group health status monitoring program based on a graph neural network is executed by a processor, the steps of the water pump group health status monitoring method based on a graph neural network as described in the first aspect are implemented.

[0057] Compared with the prior art, the beneficial effects of the present invention include at least:

[0058] The present invention provides a method and system for monitoring the health status of a water pump group based on a graph neural network. The method and system for monitoring the health status of a water pump group based on a graph neural network uses a graph neural network to fuse the health characteristics of multimodal data (such as sound data and vibration data) of the water pump group. Different from the traditional method of simply weighted averaging or directly splicing the monitoring indexes of each modal data, this deep fusion method can more effectively capture the correlation and complementary information between different modal data, and can capture complex relationships and patterns that cannot be displayed by single modal data, thereby improving the recognition ability and prediction accuracy of the health status monitoring model. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without creative work. The purpose, functional features and advantages of the present invention will be further explained in combination with the embodiments and with reference to the drawings.

[0060] Figure 1 A flow chart of a method for monitoring the health status of a water pump group based on a graph neural network provided in an embodiment of the present invention;

[0061] Figure 2 A flow chart of a method for training a water pump group health status monitoring model provided by an embodiment of the present invention;

[0062] Figure 3 A flowchart of constructing a first loss function training model provided by an embodiment of the present invention;

[0063] Figure 4 A flow chart of a method for training a graph neural network model based on a preset first loss function provided in an embodiment of the present invention;

[0064] Figure 5 A flowchart of another method for monitoring the health status of a water pump group based on a graph neural network provided by an embodiment of the present invention;

[0065] Figure 6 A flow chart of a method for periodically checking whether the water pump group health status monitoring model meets the online update condition according to a preset statistical period provided in an embodiment of the present invention;

[0066] Figure 7 A flowchart of constructing a second loss function training model provided by an embodiment of the present invention;

[0067] Figure 8 A flow chart of a method for online training the water pump group health status monitoring model using a second loss function provided in an embodiment of the present invention;

[0068] Fig. 9 A schematic diagram of a water pump group health status monitoring system based on a graph neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] It should be noted that if the embodiments of the present invention involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0071] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the meaning of "and / or" appearing in the full text includes three parallel schemes. Taking "A and / or B" as an example, it includes scheme A, or scheme B, or a scheme that satisfies both A and B. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0072] In view of this, the main purpose of the present invention is to provide a method and system for monitoring the health status of a water pump group based on a graph neural network, aiming to solve the technical problem of insufficient accuracy of the diagnostic model in the background technology.

[0073] Embodiment 1

[0074] Reference Figure 1-Figure 8 The flowchart of the method of the present invention is as follows. The first aspect of the present invention provides a method for monitoring the health status of a water pump group based on a graph neural network, comprising the following steps:

[0075] Step S100: collecting multimodal data of a water pump group, wherein the multimodal data includes sound data and vibration data;

[0076] Step S200: preprocessing the multimodal data to construct a training data set;

[0077] Step S300: training a preset graph neural network model based on the training data set to obtain a water pump group health status monitoring model;

[0078] Step S400: inputting the real-time multimodal data of the water pump group into the water pump group health status monitoring model to predict the health status of the water pump group.

[0079] Specifically, the multimodal data of the water pump group can be collected by selecting sensors installed on the water pump group. These sensors include at least sound sensors and vibration sensors. It can be understood that they can also include temperature sensors, pressure sensors, humidity sensors, current sensors, flow sensors, current sensors, etc. For the convenience of explanation, the following mainly lists the scenarios of using sound sensors and vibration sensors, but the protection scope of this application is not limited to sound data and vibration data. The collected multimodal data is stored in a local server or cloud database according to the timestamp to facilitate subsequent data processing and analysis. Then the stored multimodal data is preprocessed, including: data cleaning, removing outliers and noise points, such as filtering out data points beyond the normal range by setting a threshold; data standardization / normalization, standardizing the sound data and vibration data respectively, ensuring that different modal data are at the same level, which is conducive to subsequent model training; data conversion, converting the original time series data into a form suitable for graph neural network input, and constructing nodes and edges to represent the connection relationship between different sensors; label generation, assigning health status labels to each sample according to the historical maintenance records of the pump group or expert knowledge, such as "normal", "minor fault", "serious fault", etc., and constructing a training data set. Then, a graph neural network model suitable for processing graph structure data is selected to extract the health characteristics of the water pump group from the preprocessed multimodal data, such as the global characteristics of the sound signal (such as frequency domain characteristics) and the local characteristics of the vibration signal (such as time domain and frequency domain characteristics). The selected graph neural network model is trained using the constructed training data set until the model converges to obtain the water pump group health status monitoring model. Finally, the multimodal data of the water pump group during operation is collected in real time through the deployed sensors and input into the water pump group health status monitoring model, and the current health status assessment results of the pump group are output. The prediction results are visualized to the operation and maintenance personnel, and threshold rules are set to issue an alarm in time when a potential fault is detected.

[0080] Furthermore, in step S300, the step of training the preset graph neural network model based on the training data set to obtain the water delivery pump group health status monitoring model includes:

[0081] Step S310: extracting the health features of the water pump group of each modal data from the pre-processed multi-modal data, wherein the health features of the water pump group include the global features of the sound signal and the local features of the vibration signal;

[0082] Step S320: fusing the health features of the water pump group through a feature extraction network of a graph neural network model to obtain a fused health feature of the water pump group;

[0083] Step S330: predicting the health status of the water pump group through the regression prediction network of the graph neural network model according to the fused health characteristics of the water pump group;

[0084] Step S340: Calculate the gradient value of the model parameter according to the predicted value and the true value of the graph neural network model, and update the model parameter according to the gradient value;

[0085] Step S350: input the training data set into the graph neural network model training to obtain a water pump group health status monitoring model.

[0086] Specifically, the health characteristics of the water delivery pump group are extracted from the preprocessed multimodal data, such as the global characteristics of the sound signal (such as frequency domain characteristics, spectrum characteristics, Mel-frequency cepstral coefficients, zero-crossing rate, etc.) and the local characteristics of the vibration signal (such as time domain and frequency domain characteristics), and the characteristics are normalized. Then we start to build a graph structure, treating each sensor as a node in the graph, such as a sound sensor and a vibration sensor, and representing the extracted sound signal features and vibration signal features as node features; define edges according to the physical position and functional relationship between sensors, such as the connection relationship between adjacent sensors; use the extracted health features (global features of sound signals and local features of vibration signals) as feature vectors of nodes; select a graph neural network model suitable for processing graph structure data, such as graph convolutional networks (GCNs) or graph attention networks (GATs), and perform multi-layer propagation through the graph neural network model so that the feature information of each node can be transmitted and fused in the graph structure. In the last layer of the graph neural network, the feature vectors of all nodes are aggregated into a global feature vector as the fused health features of the water pump group, and the health status of the water pump group is predicted through the regression prediction network of the graph neural network model (such as a multi-layer perceptron MLP). Then, a first loss function is constructed to calculate the loss value based on the difference between the predicted value and the true value of the graph neural network model, and a back propagation algorithm is used to calculate the gradient value of the model parameters according to the loss value, and an optimizer (such as adaptive moment estimation Adam, stochastic gradient descent SGD, etc.) is used to update the model parameters according to the gradient value. Finally, the graph neural network model is trained based on the preset first loss function and the training data set, and the health status of the water pump group is predicted using the fused water pump group health features to obtain a water pump group health status monitoring model.

[0087] Since the water pump group includes at least sound data and vibration data, and the sound signal may have large dynamic changes in different time periods, especially in the early stage of equipment failure, the change of the sound signal may be relatively small, but as the failure develops, the change will gradually increase; the vibration signal may change suddenly in some cases, especially when the equipment fails, the change of the vibration signal may be very significant. In order to build an effective health status monitoring model, we need a loss function that can both smoothly process large error values ​​and sensitively process small error values. Specifically, when the error value is large, the loss function should be close to the mean absolute error (MAE) to reduce the penalty for large error values; when the error value is small, the loss function should be close to the mean squared error (MSE) to capture these small changes more sensitively.

[0088] Furthermore, in step S340, the step of calculating the gradient value of the model parameter according to the predicted value and the true value of the graph neural network model includes the following steps:

[0089] Step S341: A first loss function is constructed using a pseudo-Huber function to evaluate the difference between the model prediction value and the true value, as shown in the following formula:

[0090]

[0091] Where, L PseudoHuber is the loss value of the training data set, i is the sample of the training data set, N offline is the number of samples in the training data set, r i is the true value of the health index of the water pump group of sample i, p i is the predicted value of the health index of the water pump group of sample i, δ is the hyperparameter for switching the control function from quadratic to linear;

[0092] Step S342: Calculate the loss value using the first loss function according to the predicted value and the true value of the graph neural network model;

[0093] Step S343: Calculate the gradient value of the model parameter according to the loss value through the back propagation algorithm.

[0094] Specifically, the pseudo-Huberl function is used as the first loss function to avoid numerical instability problems. In particular, when the error is large, the square term may cause numerical overflow. The pseudo-Huberl loss function effectively alleviates this problem through the combination of square root and square, ensuring the stability and accuracy of the training process. Compared with the traditional mean square error (MSE), the pseudo-Huberl loss function has a lighter penalty for large error values, avoiding the problem of model overfitting caused by outliers; the pseudo-Huberl loss function adopts a form similar to the mean absolute error (MAE) when the error is large, and a form similar to the mean square error (MSE) when the error is small. This smooth transition makes the loss function continuous and differentiable in the entire error range, which is helpful for gradient calculation in the optimization process and improves the convergence speed and stability of the model. By adjusting the hyperparameter δ, the behavior of the loss function in different error ranges can be flexibly controlled. A smaller δ value makes the loss function closer to MSE and sensitive to small error values; a larger δ value makes the loss function closer to MAE and has a lighter penalty for large error values. This flexibility enables the model to adapt to different types of data distributions and application scenarios.

[0095] It is understandable that the protection scope of the present invention is not limited to the pseudo-Huber function. As long as the loss function designed based on the same idea and the characteristics of the multimodal data of the water pump group is within the protection scope of the present invention. In addition to the pseudo-Huber function, the first loss function can also consider other similar functions, as exemplified below:

[0096] Huber Loss:

[0097]

[0098] In the formula, r i is the true value of the health index of the water delivery pump group, p i is the predicted value of the health index of the water delivery pump group, N offline is the number of samples in the training data set, and δ is the hyperparameter that controls the switch from quadratic to linear function.

[0099] Smooth L1 Loss:

[0100]

[0101] In the formula, r i is the true value of the health index of the water delivery pump group, p i is the predicted value of the health index of the water delivery pump group, N offline is the number of samples in the training data set, and δ is the hyperparameter that controls the switch from quadratic to linear function.

[0102] Log-Cosh Loss:

[0103]

[0104] In the formula, r i is the true value of the health index of the water delivery pump group, p i is the predicted value of the health index of the water delivery pump group, N offline is the number of samples in the training data set. Through the combination of the hyperbolic cosine function and the logarithmic function, the penalty for large error values ​​is lighter, and the penalty for small error values ​​is heavier.

[0105] Furthermore, in step S350, the step of inputting the training data set into the graph neural network model training to obtain the water pump group health status monitoring model includes:

[0106] Step S351: randomly initializing the parameters of the water pump group health status monitoring model;

[0107] Step S352: setting the optimizer of the water pump group health status monitoring model to adopt an adaptive moment estimation Adam optimizer;

[0108] Step S353: randomly collecting a preset number of sample subsets from the training data set, inputting them into the water pump group health status monitoring model, and generating predicted values ​​of the sample subsets;

[0109] Step S354: according to the predicted value and the true value of the sample subset, the loss value of the sample subset is calculated using the first loss function, the gradient value of the model parameter is calculated according to the loss value of the sample subset through the back propagation algorithm, and the model parameter is updated according to the calculated gradient value using the Adam optimizer;

[0110] Step S355: determine whether the sample subset traverses the entire training data set. If so, increase the number of iterative optimization times by one. If not, return to step S353 to continue collecting new sample subsets.

[0111] Step S356: Determine whether the number of iterative optimizations reaches a set threshold. If so, stop the iterative optimization. If not, return to step S353 to start a new round of iterative optimization.

[0112] Specifically, randomly initializing model parameters helps break the symmetry, allowing the model to find the right parameter space faster in the early stages of training and avoid local optimal solutions; using the Adam optimizer combines the advantages of the momentum method and RMSProp, and can adaptively adjust the learning rate, accelerate the convergence of the model, and improve the stability of training. Random small batch sampling helps break the sequential dependency of the data, allowing the model to better generalize to unseen data. Using different sample subsets for each training can improve the robustness and generalization ability of the model. By judging whether the sample subset traverses the entire training data set, it ensures that the model can see all the training samples in each epoch, so as to fully learn the characteristics of the data. The number of iterative optimizations is recorded after each epoch. By traversing the entire training data set and setting a threshold for the number of iterative optimizations, it ensures that the model fully learns the characteristics of the data and avoids overfitting problems.

[0113] Furthermore, the step S400 further includes the following steps:

[0114] Step S500: regularly checking the online update conditions of the water pump group health status monitoring model according to a preset statistical period, and performing online update or continuing monitoring according to the inspection result.

[0115] Specifically, a statistical cycle can be preset (such as once a week or once a month, adjusted as needed). Reasonable setting of the statistical cycle can ensure that the model is updated in time and avoid waste of resources caused by frequent updates. In each statistical cycle, the model performance is regularly evaluated to timely discover the degradation of the model. Setting clear online update conditions can ensure that the model is updated when necessary and avoid unnecessary update operations. If the online update conditions are met, the model is updated online using new data. Online updates can be performed through incremental learning or retraining. Online updates can adjust model parameters in time to adapt them to the new data distribution and improve the prediction accuracy and robustness of the model. Among them, incremental learning refers to incremental training based on the existing model using new data to update model parameters; retraining refers to retraining the model using a combination of new and old data to generate new model parameters. If the online update conditions are not met, continue to use the current model for monitoring and wait for the next statistical cycle.

[0116] Furthermore, in step S500, the step of regularly checking the online update condition of the water pump group health status monitoring model according to a preset statistical period, and performing online update or continuing monitoring according to the inspection result includes:

[0117] Step S510: Check whether the water pump group health status monitoring model meets one of the following online update conditions: the model online time reaches the preset update cycle threshold, and the model performance index is lower than the preset threshold. If any of the conditions is met, step S520 is executed. If none of the conditions are met, the model status continues to be monitored and waits for the next periodic check.

[0118] Step S520: Acquire the latest multimodal data as training data, and perform online training on the water pump group health status monitoring model to obtain an updated model.

[0119] Specifically, a statistical cycle can be preset (such as once a week or once a month, adjusted as needed), and a scheduled task scheduling tool can be used to regularly trigger online update condition checks. Update cycle threshold determination: record the current time when the model is online, and update the model's running time at each check. Set a fixed update cycle threshold (for example, update at least once every three months). At each check, calculate the current running time of the model and compare it with the preset update cycle threshold. If the model's running time reaches the preset update cycle threshold, the online update mechanism is triggered; Model performance indicator determination: In each statistical cycle, calculate the model's performance indicator and compare it with the preset threshold. If the performance indicator is lower than the preset threshold, the online update mechanism is triggered. It can be understood that common model performance indicators include response time, memory usage, CPU usage, etc. The health status monitoring of water supply pump groups usually requires real-time response. Models with short response time can detect abnormal conditions more quickly. When the response speed is lower than the preset threshold, it needs to be retrained and updated. In resource-constrained environments, such as embedded systems, models with low memory and CPU usage can run better. By monitoring the model's memory and CPU usage, the risk of system crashes can be reduced and the system's stability and reliability can be improved.

[0120] In one possible implementation, in each statistical period, test data with real labels are collected, and the model is used to predict the test data. The accuracy, precision, recall and F1 score of the model are calculated based on the prediction results, and the calculation results are compared with the preset thresholds. For example, the preset accuracy threshold is 90%, the precision threshold is 85%, the recall threshold is 85%, and the F1 score threshold is 85%. If the performance index is lower than the preset threshold, the online update mechanism is triggered. By regularly evaluating and updating the model, it is ensured that the model is always in the best state and can accurately identify the operating status of the device.

[0121] Furthermore, in step S520, the step of obtaining the latest multimodal data as training data and performing online training on the water pump group health status monitoring model to obtain an updated model includes:

[0122] Step S521: construct a second loss function, which includes an unsupervised learning part and a lifelong learning part, as shown in the following formula:

[0123]

[0124]

[0125] In the formula, L is the loss value of the training data set, i is the sample of the training data set, and N online is the number of online samples, L USL is the loss value of the unsupervised learning part of the training dataset, L LL is the loss value of the lifelong learning part of the training data set; p i The predicted value of the health index output by the current model for sample i; p0 is the historical maximum value of the health index predicted by the model for sample i; is the predicted value of the health indicator output by the historical model of sample i; λ, β, and γ are all trade-off parameters that control the order of magnitude of each loss; α is a hyperparameter that controls the switch between the health indicator from slow change to sudden change;

[0126] Step S522: Calculate the loss value using the second loss function according to the predicted value and the true value of the graph neural network model;

[0127] Step S523: Calculate the gradient value of the model parameter according to the loss value through the back propagation algorithm;

[0128] Step S524: Update the model parameters according to the calculated gradient values.

[0129] Specifically, the second loss function includes an unsupervised learning part and a lifelong learning part. The unsupervised learning part consists of three parts. Through cross entropy loss, absolute difference loss and logarithmic difference loss, the difference between the model's predicted value and the historical maximum value is minimized and the prediction stability is maintained. For cross entropy loss, the switch between the health indicator from slow change to sudden change can be controlled by adjusting the value of α. When α is close to 0, the cross entropy loss mainly focuses on p i When α is close to 1, the cross entropy loss focuses on 1-p i For the absolute difference loss, the weight of this loss can be controlled by β to ensure that the model's predicted value does not deviate too far from the historical maximum value, thereby maintaining the stability of the prediction. For the logarithmic difference loss, the weight of this loss can be controlled by γ to ensure that the model's predicted value p i It will not exceed the historical maximum value p0, thus avoiding over-prediction. The lifelong learning part is used to measure the prediction value p of the current model. i The predicted value of the historical model The difference between It can ensure that the predicted value p of the current model i The predicted value of the historical model Stay consistent, thereby achieving lifelong learning of the model, that is, the model can continue to learn and adapt to new data.

[0130] Furthermore, in step S520, the step of obtaining the latest multimodal data as training data and performing online training on the water pump group health status monitoring model to obtain an updated model also includes:

[0131] Step S610: Acquire the latest multimodal data as training data and construct an updated training data set;

[0132] Step S620: randomly initializing the parameters of the water pump group health status monitoring model;

[0133] Step S630: setting the optimizer of the water delivery pump group health status monitoring model to adopt the Adam optimizer;

[0134] Step S640: randomly collecting a preset number of sample subsets from the updated training data set, inputting them into the water pump group health status monitoring model, and generating predicted values ​​of the sample subsets;

[0135] Step S650: according to the second loss function, calculate the loss value of the predicted value of the sample subset and the true label of the sample, calculate the gradient value of the second loss function with respect to the model parameters by back propagation algorithm, and use Adam optimizer to update the model parameters according to the calculated gradient value;

[0136] Step S660: Determine whether the sample subset traverses the entire updated training data set. If so, the number of iterations is increased by one. If not, return to step S640 to continue collecting new sample subsets.

[0137] Step S670: Determine whether the number of iterative optimizations reaches a set threshold. If so, stop the iteration; if not, return to step S640 to start a new round of iterative optimization.

[0138] Specifically, the latest multimodal data (including sound and vibration data) is obtained from the real-time monitoring system, and an updated training data set is constructed to ensure that the model can learn the latest equipment operation status and improve the timeliness and accuracy of the model. Randomly initializing the model parameters helps to break the symmetry and enable the model to converge faster in the early stage of training. The optimizer of the model is set to Adam optimizer. Adam optimizer combines the advantages of momentum method and RMSProp, and can adaptively adjust the learning rate, accelerate the convergence speed of the model and improve the stability of training. Then, according to the second loss function, the loss value of the predicted value of the sample subset and the true label of the sample is calculated, and the gradient value of the loss function with respect to the model parameters is calculated by the back propagation algorithm, and the model parameters are updated according to the calculated gradient value using Adam optimizer. Since the second loss function combines the unsupervised learning part and the lifelong learning part, it can better capture the changing trend of the health indicators and improve the prediction accuracy and robustness of the model. Determine whether the current sample subset has traversed the entire updated training data set. If so, record the number of iterations plus one. If not, return to continue to collect new sample subsets to ensure that the model can see all training samples in each epoch, so as to fully learn the characteristics of the data. Finally, setting the threshold of the number of iterative optimizations can prevent the model from being overtrained and avoid overfitting problems. At the same time, through multiple iterative optimizations, the model can gradually approach the optimal solution and improve the performance of the final model.

[0139] Embodiment 2

[0140] Reference Fig. 9 The second aspect of the present invention provides a water pump group health status monitoring system based on graph neural network, comprising:

[0141] A collection module, used for collecting multimodal data of the water pump group, wherein the multimodal data includes sound data and vibration data;

[0142] A processing module, used to pre-process the multimodal data and construct a training data set;

[0143] A generation module is used to train a preset graph neural network model based on the training data set to obtain a water pump group health status monitoring model;

[0144] The prediction module inputs the real-time multimodal data of the water pump group into the water pump group health status monitoring model to predict the health status of the water pump group.

[0145] Furthermore, the water pump group health status monitoring system based on graph neural network also includes:

[0146] The update module is used to regularly check the online update conditions of the water pump group health status monitoring model according to a preset statistical period, and perform online update or continue monitoring according to the inspection results.

[0147] Embodiment 3

[0148] The present invention also provides a computer-readable storage medium, on which is stored a water pump group health status monitoring processing program based on a graph neural network. When the water pump group health status monitoring program based on a graph neural network is executed by a processor, the steps of the water pump group health status monitoring method based on a graph neural network as described in any of the above embodiments are implemented.

[0149] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the water pump group health status monitoring system based on the graph neural network, and uses various interfaces and lines to connect the various parts of the entire water pump group health status monitoring system based on the graph neural network to process the operable devices.

[0150] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the water pump group health status monitoring system based on graph neural network by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0151] Compared with the prior art, the beneficial effects of the present invention include at least:

[0152] The present invention provides a method and system for monitoring the health status of a water pump group based on a graph neural network, wherein the method for monitoring the health status of a water pump group based on a graph neural network fuses the health characteristics of multimodal data (such as sound data and vibration data) of a water pump group by using a graph neural network. Unlike the traditional method of simply weighted averaging or directly splicing the monitoring indexes of each modal data, this deep fusion method can more effectively capture the correlation and complementary information between different modal data, and can capture complex relationships and patterns that cannot be displayed by single modal data, thereby improving the recognition ability and prediction accuracy of the health status monitoring model. Furthermore, the present invention also provides a self-learning and online update mechanism, which regularly checks the performance indicators of the model and automatically triggers the online update of the model according to preset conditions, so that the model can continuously optimize its own parameters, adapt to changes in the equipment operating environment through a continuous learning process, ensure that the model is always in the best state, and maintain the prediction accuracy of the model.

[0153] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and the equivalent structural transformations made by the contents of the present invention specification and drawings, or direct / indirect application in other related technical fields are all included in the scope of patent protection of the present invention. Therefore, the scope of protection of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for monitoring the health status of a water pump group based on a graph neural network, characterized in that: The following steps are involved: Step S100: collecting multimodal data of the water pump group, wherein the multimodal data includes sound data and vibration data; Step S200: preprocessing the multimodal data to construct a training data set; Step S300: training a preset graph neural network model based on the training data set to obtain a water pump group health status monitoring model; Step S400: inputting the real-time multimodal data of the water pump group into the water pump group health status monitoring model to predict the health status of the water pump group.

2. The method for monitoring the health status of a water pump group based on a graph neural network according to claim 1, characterized in that: In step S300, the step of training the preset graph neural network model based on the training data set to obtain the water delivery pump group health status monitoring model includes: Step S310: extracting the health features of the water pump group of each modal data from the pre-processed multi-modal data, wherein the health features of the water pump group include the global features of the sound signal and the local features of the vibration signal; Step S320: fusing the health features of the water pump group through a feature extraction network of a graph neural network model to obtain a fused health feature of the water pump group; Step S330: predicting the health status of the water pump group through the regression prediction network of the graph neural network model according to the fused health characteristics of the water pump group; Step S340: Calculate the gradient value of the model parameter according to the predicted value and the true value of the graph neural network model, and update the model parameter according to the gradient value; Step S350: input the training data set into the graph neural network model training to obtain a water pump group health status monitoring model.

3. The method for monitoring the health status of a water pump group based on a graph neural network according to claim 2, characterized in that: In step S340, the step of calculating the gradient value of the model parameter according to the predicted value and the true value of the graph neural network model includes the following steps: Step S341: A first loss function is constructed using a pseudo-Huber function to evaluate the difference between the model prediction value and the true value, as shown in the following formula: Where, L PseudoHuber is the loss value of the training data set, i is the sample of the training data set, N offline is the number of samples in the training data set, r i is the true value of the health index of the water pump group of sample i, p i is the predicted value of the health index of the water pump group of sample i, δ is the hyperparameter for switching the control function from quadratic to linear; Step S342: Calculate the loss value using the first loss function according to the predicted value and the true value of the graph neural network model; Step S343: Calculate the gradient value of the model parameter according to the loss value through the back propagation algorithm.

4. The method for monitoring the health status of a water pump group based on a graph neural network according to claim 3, characterized in that: In step S350, the step of inputting the training data set into the graph neural network model training to obtain the water delivery pump group health status monitoring model includes: Step S351: randomly initializing the parameters of the water pump group health status monitoring model; Step S352: setting the optimizer of the water pump group health status monitoring model to adopt an adaptive moment estimation Adam optimizer; Step S353: randomly collecting a preset number of sample subsets from the training data set, inputting them into the water pump group health status monitoring model, and generating predicted values ​​of the sample subsets; Step S354: according to the predicted value and the true value of the sample subset, the loss value of the sample subset is calculated using the first loss function, the gradient value of the model parameter is calculated according to the loss value of the sample subset through the back propagation algorithm, and the model parameter is updated according to the calculated gradient value using the Adam optimizer; Step S355: determine whether the sample subset traverses the entire training data set. If so, increase the number of iterative optimization times by one. If not, return to step S353 to continue collecting new sample subsets. Step S356: Determine whether the number of iterative optimizations reaches a set threshold. If so, stop the iterative optimization. If not, return to step S353 to start a new round of iterative optimization.

5. The method for monitoring the health status of a water pump group based on a graph neural network according to any one of claims 1 to 4, characterized in that: The step S400 further includes the following steps: Step S500: regularly checking the online update conditions of the water pump group health status monitoring model according to a preset statistical period, and performing online update or continuing monitoring according to the inspection result.

6. The method for monitoring the health status of a water pump group based on a graph neural network according to claim 5, characterized in that: In step S500, the step of regularly checking the online update condition of the water pump group health status monitoring model according to a preset statistical period, and performing online update or continuing monitoring according to the inspection result includes: Step S510: Check whether the water pump group health status monitoring model meets one of the following online update conditions: the model online time reaches the preset update cycle threshold, and the model performance index is lower than the preset threshold. If any of the conditions is met, step S520 is executed. If none of the conditions are met, the model status continues to be monitored and the next periodic check is waited for. Step S520: Acquire the latest multimodal data as training data, and perform online training on the water pump group health status monitoring model to obtain an updated model.

7. The method for monitoring the health status of a water pump group based on a graph neural network according to claim 6, characterized in that: In step S520, the step of obtaining the latest multimodal data as training data and performing online training on the water pump group health status monitoring model to obtain an updated model includes: Step S521: Construct a second loss function, which includes an unsupervised learning part and a lifelong learning part, as shown in the following formula: In the formula, L is the loss value of the training data set, i is the sample of the training data set, and N online is the number of online samples, L USL is the loss value of the unsupervised learning part of the training dataset, L LL is the loss value of the lifelong learning part of the training data set; p i The predicted value of the health index output by the current model for sample i; p0 is the historical maximum value of the health index predicted by the model for sample i; is the predicted value of the health indicator output by the historical model of sample i; λ, β, and γ are all trade-off parameters that control the order of magnitude of each loss; α is a hyperparameter that controls the switch between the health indicator from slow change to sudden change; Step S522: Calculate the loss value using the second loss function according to the predicted value and the true value of the graph neural network model; Step S523: Calculate the gradient value of the model parameter according to the loss value through the back propagation algorithm; Step S524: Update the model parameters according to the calculated gradient values.

8. The method for monitoring the health status of a water pump group based on a graph neural network according to claim 7, characterized in that: In step S520, the latest multimodal data is obtained as training data, and the water pump group health status monitoring model is trained online to obtain an updated model, and the step also includes: Step S610: Acquire the latest multimodal data as training data and construct an updated training data set; Step S620: randomly initializing the parameters of the water pump group health status monitoring model; Step S630: setting the optimizer of the water delivery pump group health status monitoring model to adopt the Adam optimizer; Step S640: randomly collecting a preset number of sample subsets from the updated training data set, inputting them into the water pump group health status monitoring model, and generating predicted values ​​of the sample subsets; Step S650: according to the second loss function, calculate the loss value of the predicted value of the sample subset and the true label of the sample, calculate the gradient value of the second loss function with respect to the model parameters by back propagation algorithm, and use Adam optimizer to update the model parameters according to the calculated gradient value; Step S660: Determine whether the sample subset traverses the entire updated training data set. If so, the number of iterations is increased by one. If not, return to step S640 to continue collecting new sample subsets. Step S670: Determine whether the number of iterative optimizations reaches a set threshold. If so, stop the iteration; if not, return to step S640 to start a new round of iterative optimization.

9. A water pump group health status monitoring system based on graph neural network, characterized in that: include: A collection module, used for collecting multimodal data of the water pump group, wherein the multimodal data includes sound data and vibration data; A processing module, used to pre-process the multimodal data and construct a training data set; A generation module is used to train a preset graph neural network model based on the training data set to obtain a water pump group health status monitoring model; The prediction module inputs the real-time multimodal data of the water pump group into the water pump group health status monitoring model to predict the health status of the water pump group.

10. The water pump group health status monitoring system based on graph neural network according to claim 9, characterized in that: Also includes: The update module is used to regularly check the online update conditions of the water pump group health status monitoring model according to a preset statistical period, and perform online update or continue monitoring according to the inspection results.

Images