A multi-model fusion water turbine vibration analysis method

Abstract

The application discloses a kind of multi-model fusion's water turbine vibration analysis method, comprising the following steps: S1 obtains historical unit vibration relevant data, and carries out pretreatment;S2 unit vibration relevant data is carried out feature engineering and obtains training set;S3 extract the data associated with vibration in training set, respectively input SVR model, LightGBM model and XGBoost model training;S4 the results of three models are least square fitting, obtain the weight distribution of three models, form fusion model;S5 vibration relevant data of on-line monitoring are input into fusion model, obtain the vibration value predicted and as the standard value of vibration under this operating condition;S6 compare the vibration data collected with standard value, mark abnormal data and abnormal grade.The application considers multidimensional influence factor, considers water turbine operating condition, trains using three models respectively, ensures the accuracy of prediction model by automatic weighting method, improves the accuracy and scientificity of prediction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of operation analysis of hydroelectric generators, and particularly relates to a multi-model fusion water turbine vibration analysis method. BACKGROUND

[0002] As large rotating machinery, the vibration of the hydroelectric generator set in operation is a common phenomenon that cannot be completely avoided and eliminated. Severe vibration of the hydroelectric generator set affects the power supply quality, safe operation and service life of the hydroelectric generator set. Under the combined influence of mechanical, hydraulic, electromagnetic factors and aging of mechanical parts, the faults of the hydroelectric generator set are mostly manifested in the form of vibration and swing, and therefore the vibration and swing signal can directly characterize the operation state of the hydroelectric generator set.

[0003] At present, the important indicators are monitored by the hydroelectric generator set monitoring system and the online monitoring system, and the alarm limit value is set for the indicators. However, in order to avoid false alarms, the limit is set to be high, and when the hydroelectric generator set reaches the alarm limit value, a serious fault may have occurred. Even in the stable operation area, the monitoring indicators of the hydroelectric generator set are also affected by the working conditions such as water head and excitation current, and fluctuate up and down, so the change rate of the monitoring indicators directly collected and calculated cannot reflect the real situation of the equipment state. With the application of new technologies such as artificial intelligence and big data analysis, it is necessary to rely on intelligent algorithms and technologies to make trend analysis possible, and to change the production mode of the hydropower plant from traditional manual monitoring and manual decision-making to information-based, automated and intelligent machine decision-making.

[0004] Defects and deficiencies of the prior art:

[0005] 1. At present, a considerable number of online monitoring systems have been installed in most hydroelectric generator sets, but there is no standardized operation, use and maintenance method, and the application of the online monitoring system is not paid enough attention, and the collected data lack in-depth analysis by professionals and technical support of professional technicians.

[0006] 2. The state monitoring and early warning of the hydroelectric generator currently adopts the method of setting a fixed threshold and calculating the change rate, and there are problems of false reporting and untimely early warning.

[0007] 3. Considering the coupling influence of various factors such as operating environment and local impact, the vibration and swing monitoring signal of the hydroelectric generator set often presents complex non-stationary and nonlinear characteristics, and it is difficult to obtain satisfactory prediction accuracy by using the existing method to predict the development trend of the vibration and swing signal of the hydroelectric generator set. SUMMARY

[0008] The present application aims at the deficiencies of the prior art, and provides a multi-model fusion water turbine vibration analysis method. The method can accurately alarm the abnormal vibration and swing value in real time, so as to achieve the purpose of fault diagnosis and auxiliary decision-making.

[0009] The technical purpose of the present application is achieved by the technical solution of a multi-model fusion water turbine vibration analysis method, specifically comprising the following steps:

[0010] S1 obtains historical unit vibration related data and pre-processes the unit vibration related data;

[0011] S11 collects historical online and offline monitoring data, removes data irrelevant to vibration and error data, and preliminarily integrates a data set used for training;

[0012] S12 cleans repeated data, zero value data and missing data in the data set, and performs resampling;

[0013] S2 further performs feature engineering on the pre-processed unit vibration related data to obtain a training set;

[0014] S21 performs correlation analysis on the data in the data set, uses the Pearson correlation coefficient to perform correlation analysis on the feature attributes and target attributes, and selects the required feature values according to the correlation ranking;

[0015] S22 finds the maximum value max and the minimum value min of each feature according to each column of features;

[0016] S23 if min >= 0, then the following normalization is performed on each column of data:

[0017]

[0018] wherein x is the feature value of the i-th column and the j-th row, x' is the normalized value, min is the minimum value of the i-th column, and max is the maximum value of the i-th column; i,j i,j i i

[0019] S24 if min < 0, then the following normalization is performed on each column of data:

[0020]

[0021] wherein x is the feature value of the i-th column and the j-th row, x' is the normalized value, min is the minimum value of the i-th column, and max is the maximum value of the i-th column; i,j i,j i i

[0022] S25 divides the normalized data into a training set and a test set, with a ratio of 8:2, and divides the feature values and the target values to obtain a training set;​​​​​​​​

[0023] S3 extracts the data associated with the whip in the training set, and inputs the SVR model, LightGBM model and XGBoost model respectively for training;

[0024] S4 least squares fitting is performed on the results of the three models to obtain the weight distribution of the three models and form a fusion model;

[0025] S5 inputs the whip-related data monitored online into the trained fusion model to obtain the predicted whip value and take it as the standard value of the whip under the working condition;

[0026] S6 compares the collected whip data with the standard value to mark abnormal data and abnormal levels.

[0027] Preferably, in step S6, when the vibration value and the swing value are less than 40pm, if the amplitude is greater than the predicted value by 10pm or more, it is judged as a secondary amplitude abnormality; if the amplitude is greater than the predicted value by 20pm or more, it is judged as a primary amplitude abnormality.

[0028] Preferably, in step S6, when the vibration value and the swing value are greater than 40pm, when the amplitude is greater than the predicted value by 10%-25%, it is judged as a secondary amplitude abnormality; when the amplitude is greater than the predicted value by 25%, it is judged as a primary amplitude abnormality.

[0029] Compared with the prior art, the present application has the following beneficial effects:

[0030] 1. The multi-model fusion water turbine whip analysis method provided by the present application considers the influence factors of multiple dimensions and takes into account the water turbine operating conditions, so that more accurate prediction results are obtained and more reliable alarm outputs are provided.

[0031] 2. The multi-model fusion water turbine whip analysis method provided by the present application uses three models for training respectively, and ensures the accuracy of the prediction model through automatic weighting, thereby improving the accuracy and scientificity of the prediction.

[0032] 3. The multi-model fusion water turbine whip analysis method provided by the present application can be integrated in an online monitoring system to perform real-time prediction on data, thereby making up for the limitations of the current online monitoring device which only uses the whip value overrun method for early warning. DETAILED DESCRIPTION

[0033] Figure 1 is a flowchart of an embodiment of the present application.

[0034] Figure 2 is a schematic diagram of the principle of the SVR model support vector regression in an embodiment of the present application.

[0035] Figure 3 is a water guide swing X direction model prediction result graph in an embodiment of the present application.

[0036] Figure 4 is a water guide swing Y direction model prediction result graph in an embodiment of the present application.

[0037] Figure 5 is a top cover vibration X direction model prediction result graph in an embodiment of the present application.

[0038] Figure 6 is a top cover vibration Y direction model prediction result graph in an embodiment of the present application.

[0039] Figure 7 is a top cover vibration Z direction model prediction result graph in an embodiment of the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0041] As a preferred embodiment of the present application, the embodiment provides a multi-model fusion water turbine vibration analysis method, referring to the accompanying drawings Figure 1 , and specifically includes the following steps:

[0042] S1 obtains historical unit vibration related data, including water head, power, excitation, guide vane opening and other possible related data, and pre-processes the unit vibration related data;

[0043] S11 collects historical online and offline monitoring data, removes data irrelevant to vibration and error data, and preliminarily integrates a data set used for training;

[0044] S12 cleanses repeated data, zero value data and missing data in the data set, and performs resampling;

[0045] S2 further performs feature engineering on the pre-processed unit vibration related data to obtain a training set;

[0046] S21 performs correlation analysis on the data in the data set, uses Pearson correlation coefficient to perform correlation analysis on feature attributes and target attributes, and selects required feature values according to correlation ranking;

[0047] S22 calculates the maximum value max and the minimum value min of each feature according to each column of features;

[0048] S23 If min>=0, normalize each column of data as follows:

[0049]

[0050] where x is the eigenvalue of the ith column and jth row, x' is the normalized value, min is the minimum value of the ith column, and max is the maximum value of the ith column. i,j i,j i i

[0051] S24 If min<0, normalize each column of data as follows:

[0052]

[0053] where x is the eigenvalue of the ith column and jth row, x' is the normalized value, min is the minimum value of the ith column, and max is the maximum value of the ith column. i,j i,j i i

[0054] S25 Split the normalized data into a training set and a test set with a ratio of 8:2, and divide the eigenvalues and target values to obtain a training set;

[0055] S3 Extract the data associated with the vibration from the training set and input it into the SVR model, LightGBM model, and XGBoost model for training;

[0056] S4 To increase the robustness of the model, perform least squares fitting on the results of the three models to obtain the weight distribution of the three models, form a fusion model, and obtain a model for predicting the vibration result. Through the weight distribution of the three models, a fusion model is formed, so that the predicted data is close to the true situation, and a model that can accurately predict the vibration result is obtained.

[0057] S5 Input the vibration-related data monitored online into the trained fusion model to obtain the predicted vibration value and use it as the standard value of the vibration under the working condition.

[0058] S6 Compare the collected vibration data with the standard value and mark the abnormal data and the abnormal level.

[0059] ​​​​​​​​In the above embodiment, the SVR model is full name support vector regression, which is a support vector machine (SVM) for regression problems. It uses a linear function to fit the sample in the vector space. The model takes the comprehensive distance of all sample actual positions to the linear function as the loss, and obtains the parameters of the linear function by minimizing the loss.

[0060] LightGBM (Light Gradient Boosting Machine) is a distributed gradient boosting framework based on decision tree algorithm. In order to meet the needs of the industry to shorten the model calculation time, the design idea of LightGBM is mainly two points: reduce the use of data to memory, ensure that a single machine can use more data without sacrificing speed; Reduce the cost of communication, improve the efficiency of multi-machine parallel, realize linear speedup in calculation.

[0061] The full name of XGBoost is eXtreme Gradient Boosting, which is an optimized distributed gradient boosting library, aiming at high efficiency, flexibility and portability. XGBoost is based on gradient boosting tree, redefines the loss function and weak evaluator, and improves the integration means of the boosting algorithm, realizes the balance of operation speed and model effect.

[0062] The training set after feature engineering is input into the SVR model LightGBM and XGBoost model for training, so that the model can find the relationship between the water head, power, excitation, guide vane and vibration data, and the output model can predict the vibration data to a certain extent.

[0063] In the above embodiment, set the sample data (x, y), and the model output value is f(x) and the true value is y. The traditional regression model takes the difference between f(x) and y as the loss value, and only when f(x) is equal to y, the model can determine the loss value as 0. While SVR can set a deviation value ∈, calculate the absolute value of the difference between f(x) and y, and the SVR model will only calculate the loss when the absolute value is greater than ∈. Referring to the attached Figure 3 In graphical representation, an interval band with a width of ∈ is set on both sides of f(x), and the values falling between the two interval bands are considered as correct values.

[0064] The difference between SVR and traditional SVR lies in that it considers the slack variable ξi, the penalty coefficient C, and the insensitive loss function ∈ in the derivation process. Through derivation, the function form of SVR can be finally solved:

[0065]

[0066]

[0067] where w represents the weight vector in high-dimensional space, b is the threshold, φ(x i ) is a nonlinear function, and ε is a set parameter of insensitive loss function. The slack variable ξ can play the role of gentle fitting function, further improving the generalization ability of the model; the penalty parameter C represents the control level of the sample points whose errors exceed the given value ε, mainly playing the role of balancing the generalization and complexity of the model, and the penalty factor is generally positive in most cases; ε represents the requirement for the error of the regression model; the introduced slack variable ξ * controls the upper and lower bounds of the output value.

[0068] When the sample space cannot be solved, the concept of kernel function K(x i ,x j ) is introduced, and the solution form is as follows:

[0069]

[0070] where α i and are the undetermined coefficients of the equation solved by the Lagrange method, K(x i ,x j ) = φ(x i )φ(x j ), which is a symmetric positive real function.

[0071] LightGBM belongs to the boosting class of ensemble learning methods, which is an efficient implementation of the gradient boosting decision tree (GBDT) algorithm framework. The execution process of the GBDT algorithm is as follows: input the training set {(x1,y1),(x2,y2),…(x N ,y N )}, initialize the classifier where h0(x) is the first base learner selected by the user, and the training target of T base learners is set, and the calculation method of each base learner is as follows:

[0072] 1) Calculate the negative gradient my i of the current loss function:

[0073]

[0074] 2) Fit the negative gradient to obtain the parameters of the current base learner h t :

[0075]

[0076] 3) Minimize the loss function to get the weight of the current base learner:

[0077]

[0078] Final classifier F t (x) is the weighted sum of each base learner:

[0079] F t (x) = F t-1 (x i )+ a t h t (x; w t ) (4)

[0080] From the calculation process of formula (1) - formula (4), it can be seen that in each iteration of the GBDT algorithm, the entire training data needs to be traversed multiple times. At this time, if the entire training data is loaded into the memory, it will limit the size of the training data, and if it is not loaded into the memory, it needs to repeatedly read and write the training data, which consumes a lot of computing time.

[0081] To solve this problem, LightGBM makes optimizations such as feature histogram, one-sided gradient sampling, mutually exclusive feature bundling, and Leaf-wise growth strategy in the traditional GBDT algorithm. These optimizations make the algorithm have faster training speed and lower memory consumption, so LightGBM algorithm is more suitable for processing massive data, and the vibration data has a large amount of data.

[0082] XGBoost is an ensemble learning method that improves the performance of the model by iteratively adding weak learners to the training data. XGBoost adds a new model to the original model at each iteration, and uses the new model to fit the residual between the predicted results of the previous model and the true results, in order to get better prediction results. In the problem of vibration prediction, XGBoost uses a basic regression tree model, and the ensemble model of the tree can be represented as follows:

[0083]

[0084] In the formula: x i is the i-th input feature vector; represents the predicted vibration value of the i-th sample; K represents the number of regression trees; R is the set space of regression trees; f k represents a function in the set R, which is the output result of the base learner.

[0085] By accumulating the results of the iteration process, the objective function of XGBoost can be transformed as follows:

[0086]

[0087] wherein: is the error between the predicted result and the true result, is the regularization term of the objective function, Ω(f k ) is expressed as:

[0088]

[0089] wherein: T is the number of leaf nodes; γ is the penalty function coefficient, used to control the number of leaf nodes; ω j is the weight of the leaf node; λ is the regularization penalty term coefficient. Finally, the optimal objective function value can be obtained by combining the iteration results of XGBoost and performing Taylor second-order expansion at f k

[0090] When performing least squares fitting on the results of the three models, assuming that the prediction result of SVR is y1, the prediction result of LightGBM is y2, and the prediction result of XGBoost is y3, then the prediction value y' of the fused model meets the following convention:

[0091] y' = βy1 + γy2 + λy3

[0092] Where β is the proportion of the weight of the SVR model in the combined weight, γ is the proportion of the weight of the LightGBM model in the combined weight, and λ is the proportion of the weight of the XGBoost model in the combined weight, satisfying:

[0093] β + γ + λ = 1

[0094] The optimal solution of the comprehensive weight is obtained according to the following formula:

[0095]

[0096] In the specific verification test, the guide swing of the water turbine and the vibration of the top cover are selected as the prediction target. After 49 relevant features are selected by expert experience, 25 highly correlated features are selected as the feature set by Pearson coefficient correlation analysis. The historical data is divided into training data and test data, and the SVR model, LightGBM model and XGBoost model are used for training and testing. Further, to increase the robustness of the model, the least squares method is used to weight the three models. According to the above weight fusion distribution method, the weight distribution of the fused model is shown in the following table:

[0097]

[0098] The prediction results of all models are selected from 100 points for plotting. The experimental results are shown in the following figures: Figure 3 -Appendix​Figure 7 .

[0099] From Figures 3-7 It can be seen that the prediction results of the four models on the test set are consistent with the trend of the true values, indicating that the four methods have good effects on the prediction of the vibration. In order to further evaluate the effect of the model, the determination coefficient R2 index is used to evaluate the three models, and the R2 score reflects the proportion of the total variation of the dependent variable that can be explained by the regression relationship of the independent variable, which is expressed as:

[0100]

[0101] Where y i represents the true observation value, y represents the average value of the true observation value, y represents the predicted value, MSE is the mean square error, and Var is the variance.

[0102] The R2 score ranges from 0 to 1, when R2 is 1, the predicted value and the true value in the sample are completely equal, without any error, indicating that the better the independent variable explains the dependent variable in the regression analysis, generally the larger the R2, the better the model fitting effect

[12] , and the results are shown in the following table:

[0103]

[0104] From the above table, it can be seen that the R2 coefficients of the three methods on the water guide swing and the top cover vibration are all greater than 0.95, indicating that the three models can effectively predict the vibration value, but through comparison, the R2 score of the fusion model after least square fusion can reach more than 0.98, and the prediction accuracy is significantly improved.

[0105] It should be noted that when the results of the three models are least square fitted, the weight distribution of the SVR model, the LightGBM model and the XGBoost model will be different according to the different feature sets, the different training data and the different test data. According to the above method, the accurate and unique weight distribution is obtained for the determined feature set and the determined training data and test data, which is the main reason for improving the prediction accuracy after model fusion, and is also the contribution of the present application to improving the vibration analysis and prediction of the water turbine.

[0106] The above embodiment obtains the fusion model for analyzing and predicting the vibration of the hydraulic turbine by fusing multiple models, training respectively, and adopting the automatic weighting manner, ensures the accuracy of the prediction model, and considers the multi-dimensional influence factors, considers the operation conditions of the hydraulic turbine, including but not limited to the vibration related data such as the water head, the power, the excitation, the guide vane opening degree, can obtain more accurate prediction results, provides more reliable alarm output, realizes real-time prediction of the data, and makes up the limitation that the current online monitoring device only adopts the out-of-limit manner of the vibration value to provide early warning.

[0107] In some embodiments, on the basis of the above embodiment, in step S6, when the vibration value and the runout value are less than 40 mu m, if the amplitude is greater than the prediction value by 10 mu m or more, it is judged as a secondary amplitude abnormality; if the amplitude is greater than the prediction value by 20 mu m or more, it is judged as a primary amplitude abnormality.

[0108] In some other embodiments, on the basis of the above embodiment, in step S6, when the vibration value and the runout value are greater than 40 mu m, when the amplitude is greater than the prediction value by 10%-25%, it is judged as a secondary amplitude abnormality; when the amplitude is greater than the prediction value by 25%, it is judged as a primary amplitude abnormality.

[0109] The prior art part in the protection scope of the present application is not limited to the embodiments given in the present application file, all prior art not contradictory to the scheme of the present application, including but not limited to prior patent documents, prior published publications, prior public use, etc., can be included in the protection scope of the present application.

[0110] In addition, it should be further pointed out that the combination manner of the technical features in the present case is not limited to the combination manner recorded in the claims of the present case or the combination manner recorded in the specific embodiments, all technical features recorded in the present case can be freely combined or combined in any manner, unless contradictory to each other.

[0111] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any change, modification, substitution, combination, simplification made without departing from the spirit and principles of the present application should be equivalent replacement, and all are included in the protection scope of the present application.

Claims

1. A multi-model fusion water turbine vibration analysis method, characterized in that, Specifically comprising the following steps: S1 obtaining historical unit vibration related data, and preprocessing the unit vibration related data; S11 collecting historical online and offline monitoring data, removing data irrelevant to vibration and error data, and preliminarily integrating a data set for training; S12 cleaning repeated data, zero value data and missing data in the data set, and resampling; S2 further performing feature engineering on the preprocessed unit vibration related data to obtain a training set; S21 performing correlation analysis on the data in the data set, using Pearson correlation coefficient to perform correlation analysis on feature attributes and target attributes, and selecting required feature values according to correlation ranking; S22 calculating the maximum value max and the minimum value min of each feature according to each column of features; S23 if min≥0, then performing the following normalization on each column of data: where x i,j is the eigenvalue of the ith column and jth row, x' i,j is its normalized value, min i is the minimum value of the ith column, and max i is the maximum value of the ith column. S24 if min<0, then performing the following normalization on each column of data: where x i,j is the eigenvalue of the ith column and jth row, x' i,j is its normalized value, min i is the minimum value of the ith column, max i is the maximum value of the ith column; S25 dividing the normalized data into a training set and a test set, the ratio being 8:2, and dividing feature values and target values to obtain a training set; S3 extracting data related to vibration in the training set and inputting the data into SVR model, LightGBM model and XGBoost model for training; S4 performing least square fitting on the results of the three models to obtain weight distribution of the three models, forming a fusion model; S5 inputting vibration related data monitored online into the trained fusion model to obtain predicted vibration values and take the vibration values as standard values of vibration under the working condition; S6 comparing the collected vibration data with the standard values to mark abnormal data and abnormal levels.

2. The multi-model fused hydro-turbine hunting analysis method according to claim 1, wherein, In step S6, when the vibration value and the runout value are less than 40 μm, if the amplitude is greater than the predicted value by 10 μm or more, it is judged as a secondary amplitude abnormality; if the amplitude is greater than the predicted value by 20 μm or more, it is judged as a primary amplitude abnormality.

3. The multi-model fused hydro-turbine hunting analysis method according to claim 1, characterized in that: In step S6, when the vibration value and the runout value are greater than 40 um, when the amplitude is greater than 10%-25% of the predicted value, it is judged as a secondary amplitude abnormality; when the amplitude is greater than 25% of the predicted value, it is judged as a primary amplitude abnormality.

Images