High-precision Diagnosis Method for the Number of Broken Rotor Bars of a 15-Phase Asynchronous Motor Based on ESPRIT-PSA and LGBM

Abstract

The present invention is a high-precision diagnostic method for the number of broken rotor bars of a fifteen-phase asynchronous motor based on ESPRIT-PSA and LGBM. It aims at the instantaneous reactive power signal sampled for a short time (only 2 seconds), and uses ESPRIT (Estimation of Signal Parameters via Rotational Invariance Techniques) and PSA (Pattern Search Algorithm) to obtain the accurate amplitude A of the characteristic components therein q ; then 71 features such as the corresponding voltage amplitude, current amplitude, average active power P, and average reactive power Q are obtained; then the A q and the top 5 features with the highest weight ratios among the above 71 features - A q , P, Q, the voltage amplitude U of the first phase m , and the current amplitude I of the first phase m are used to form a data set; then the LGBM model is trained and saved with it, and the number of broken rotor bars is diagnosed with high precision (the training and testing accuracies are both 100%, and the 5-fold cross-validation accuracy is 99.38%).

Description

Technical Field

[0001] The present invention relates to a method for diagnosing the number of broken rotor bars of a fifteen-phase asynchronous motor, belonging to the technical field of fault diagnosis. Background Technique

[0002] Due to its many advantages such as high reliability and fault-tolerant operation, the fifteen-phase asynchronous motor has been applied to special fields such as ships and submarines. Rotor bar breakage is a typical fault mode of the fifteen-phase asynchronous motor. Therefore, the diagnosis of rotor bar breakage faults plays an important role in improving the reliability of the operation of the fifteen-phase asynchronous motor.

[0003] When a fifteen-phase asynchronous motor has a rotor bar breakage fault, side frequency components with a frequency of (1±2s)f 1 (s is the slip ratio, f 1 is the power supply frequency) will appear in its stator current. The method of achieving the purpose of rotor bar breakage diagnosis by examining this side frequency component is called the method of analyzing the stator current signal of the motor. The research on this type of method has been mature, but in the case of low slip ratio, the side frequency components of the stator current may be submerged by the f 1 frequency components, which poses a severe challenge to this type of method.

[0004] Therefore, a method of performing spectral analysis on the instantaneous reactive power signal and then realizing the diagnosis of rotor bar breakage faults has gradually developed and formed. Its essence is: when a rotor bar breakage fault occurs, a characteristic component with a frequency of 2sf 1 will appear in its instantaneous reactive power. This type of method is usually called the rotor bar breakage fault diagnosis method of analyzing the instantaneous reactive power signal of the motor. Its advantage is that even in the case of low slip ratio, this type of method can still accurately judge whether a rotor bar breakage fault has occurred.

[0005] After accurately judging that a rotor bar breakage fault has occurred, it is necessary to further diagnose the number of broken rotor bars. This is because: rotor bar breakage is a progressive fault. Usually, 1 bar breaks initially, and then other bars adjacent to the broken bar continue to break, and the output of the fifteen-phase asynchronous motor will drop significantly or even cause shutdown. If the number of broken rotor bars can be diagnosed, naturally the severity of the rotor bar breakage fault can be grasped, so as to arrange maintenance in time. Therefore, the diagnosis of the number of broken rotor bars is of great significance.

[0006] At present, the method of analyzing the instantaneous reactive power signal of the motor has provided a diagnostic formula for the number of broken rotor bars, but in actual application, there is a large deviation between its diagnostic result and the actual number of broken rotor bars. In view of this, the present invention uses a method combining ESPRIT (Estimation of Signal Parameters via Rotational Invariance Techniques), PSA (Pattern Search Algorithm) and LGBM (Light Gradient Boosting Machine) to diagnose the number of broken rotor bars of a fifteen-phase asynchronous motor. Summary of the Invention

[0007] The object of the present invention is to provide a method for diagnosing the number of broken rotor bars of a fifteen-phase asynchronous motor based on ESPRIT (Estimation of Signal Parameters via Rotational Invariance Techniques), PSA (Pattern Search Algorithm) and LGBM (Light Gradient Boosting Machine), which uses the instantaneous reactive power signal sampled in a short time (only 2 seconds) as the analysis medium, has high diagnostic accuracy, and is applicable to the case of low slip ratio; in addition, because it only requires short-time sampled signals, it is particularly applicable to the case of severe interference such as load fluctuations and noise.

[0008] The problems of the present invention are realized by the following technical solutions:

[0009] A method for diagnosing the number of broken rotor bars of a fifteen-phase asynchronous motor, which aims at the instantaneous reactive power signal of a fifteen-phase asynchronous motor sampled in a short time (only 2 seconds). First, ESPRIT is used to calculate the accurate frequency value, rough amplitude and initial phase angle of the characteristic component with a frequency of 2sf 1 (s is the slip ratio, f 1 is the supply frequency); then the result calculated by ESPRIT is used as the initial value and substituted into PSA, so as to calculate the accurate amplitude A q and initial phase angle of the characteristic component, and A q is placed into the data set X as the first characteristic variable; then, the instantaneous signals of the fifteen-phase voltage and fifteen-phase current of the fifteen-phase asynchronous motor are subjected to refined Fourier spectrum analysis to obtain the fifteen-phase voltage amplitude, fifteen-phase current amplitude, fifteen-phase voltage phase and fifteen-phase current phase of the stator, a total of 60 characteristics corresponding to A q ; then, according to these 60 characteristics, the effective values of positive-sequence, negative-sequence and zero-sequence voltages, the effective values of positive-sequence, negative-sequence and zero-sequence currents, the modulus values of positive-sequence, negative-sequence and zero-sequence impedances, as well as the average active power and average reactive power are obtained, a total of 11 derivative characteristics corresponding to A q ; then, the above 71 characteristics corresponding to A q are placed into the data set X to form a 72-dimensional data set X 1 ; furthermore, X 1 is input into the LGBM classifier for feature weight calculation, and the top 5 features with the highest weight ratio are determined and selected - the amplitude A q of the characteristic component of the instantaneous reactive power, the average active power P, the average reactive power Q, the amplitude U m of the first-phase voltage, and the amplitude I m of the first-phase current, to form a new data set X 2 ; then, X 2The training set and the test set are divided in the ratio of 8:2 and input into the LGBM model to classify and diagnose the number of broken bars in the rotor of a fifteen-phase induction motor - normal (0 broken bars), 1 broken bar, and 2 broken bars; finally, the optimal hyperparameters of the LGBM model are selected by using GridSearchCV in the scikit-learn library to maximize the training accuracy of the model. The scikit-learn library is a free machine learning toolkit for the Python language, which has various classification, regression, and clustering algorithms, covering almost all mainstream machine learning algorithms including LGBM. GridSearchCV is a parameter automatic adjustment module in the scikit-learn library that systematically traverses various parameter combinations using the exhaustive method and determines the optimal parameters through cross-validation.

[0010] It should be noted that although LGBM only uses the amplitude of the instantaneous reactive power feature component and does not use its frequency and initial phase angle when diagnosing the number of broken bars in the rotor, during the application of PSA, both the amplitude and the initial phase angle must be calculated simultaneously.

[0011] The training accuracy of the present invention is 100%, the test accuracy is 100%, and the 5-fold cross-validation accuracy is 99.38%, successfully realizing the application of LGBM in the field of diagnosing broken bar faults in the rotor of a fifteen-phase induction motor. Accordingly, the number of broken bars in the rotor of a fifteen-phase induction motor can be diagnosed, and the trained LGBM model can be saved for subsequent diagnosis.

[0012] The above method for diagnosing the number of broken bars in the rotor of a fifteen-phase induction motor includes the following steps:

[0013] a. Conduct systematic and extensive experiments to measure the instantaneous stator fifteen-phase voltage signal u sn and the instantaneous stator fifteen-phase current signal i sn (n represents the phase number, n = 1, 2,..., 15);

[0014] This work is carried out one by one for the three states of the fifteen-phase induction motor: normal, with 1 broken bar fault in the rotor, and with 2 broken bar faults in the rotor, and the experiments in each state include load changes (three cases of full load, half load, and no load).

[0015] The above three states of the fifteen-phase induction motor (normal, with 1 broken bar fault in the rotor, and with 2 broken bar faults in the rotor) are respectively and sequentially labeled as state 0, 1, and 2.

[0016] b. Calculate the instantaneous reactive power signal based on the stator fifteen-phase voltage and current instantaneous signals according to equations (1) and (2) and filter out its DC component according to equation (3) to obtain q A(Instantaneous reactive power signal after filtering out the DC component);

[0017]

[0018]

[0019] q A =q 0 -mean(q 0 )(3)

[0020] In equations (1), (2) and (3), represents the Hilbert transform of the instantaneous signal u of the nth-phase stator voltage; t represents time; τ represents delay; q sn represents the instantaneous reactive power; mean(q 0 ) represents the average value of q 0 (i.e., the DC component). 0

[0021] c. Perform ESPRIT analysis on q A to calculate the accurate frequency value, rough amplitude and initial phase angle of the characteristic components in the instantaneous reactive power signal;

[0022] d. Using the amplitude and initial phase angle of the instantaneous reactive power characteristic components calculated by ESPRIT as the initial reference values, use PSA to calculate the accurate amplitude A q and initial phase angle, and then put A q as the first characteristic variable into the data set X (X is the data set storing the value of A q );

[0023] e. Perform refined Fourier analysis on the instantaneous signals of the fifteen-phase voltage and fifteen-phase current to obtain the fifteen-phase stator voltage amplitude, fifteen-phase current amplitude, fifteen-phase voltage phase and fifteen-phase current phase, a total of 60 characteristics corresponding to A q , and then process these 60 characteristics to obtain the positive-sequence, negative-sequence and zero-sequence voltage effective values, positive-sequence, negative-sequence and zero-sequence current effective values, positive-sequence, negative-sequence and zero-sequence impedance modulus values, and average active power and average reactive power, a total of 11 derivative characteristics corresponding to A q , and then put the above 71 characteristics corresponding to A q into the data set X to form a 72-dimensional data set X 1 (X 1 is a 72-dimensional data set storing A q and 71 characteristic values corresponding to it), and then use the LGBM classifier to calculate the weights of the data set X 1 , and determine and select A qThe top 5 features with the highest weight ratios (the amplitude A of the instantaneous reactive power feature component q , the average active power P, the average reactive power Q, the amplitude U of the first-phase voltage m , and the amplitude I of the first-phase current m ) are used as the features for the training and learning of the LGBM model, thereby forming a new 5-dimensional dataset X 2 (X 2 is a 5-dimensional dataset composed of the top 5 features with the weight ratios calculated by LGBM);

[0024] The amplitude A of the instantaneous reactive power feature component q and the average active power P, the average reactive power Q, the amplitude U of the first-phase voltage m , and the amplitude I of the first-phase current m are all obtained by processing the sampling signals through the sliding window method (the window includes data with a duration of 2 seconds). Specifically, at state 0, 750 sets of sample data can be obtained for each of the full-load, half-load, and no-load conditions through the above operations. State 0 includes 2,250 sets of sample data. At states 1 and 2, 1,250 sets of sample data can be obtained for each of the full-load, half-load, and no-load conditions through the above operations. States 1 and 2 each include 3,750 sets of sample data. Then, the three states together contain 9,750 sets of sample data.

[0025] f. For the above dataset X 2 , mark it according to its corresponding state (0, 1, 2), corresponding to state 0, state 1, and state 2 respectively;

[0026] g. Divide the dataset X 2 into a training set and a test set in a ratio of 8:2, substitute it into the LGBM model for training, and perform hyperparameter tuning of the model through GridSearchCV to obtain the best-performing model;

[0027] h. Call the joblib package in the scikit-learn library to save the trained LGBM model as an executable code file (with the extension.m). When the LGBM model is needed, it can be read through the joblib package (the joblib package can save the trained model and can be directly called when needed, with the advantages of high efficiency and fast reading speed).

[0028] Furthermore, the ESPRIT is described as follows.

[0029] Applying ESPRIT can calculate the accurate frequency values, rough amplitudes, and initial phase angles of the feature components in the instantaneous reactive power signal with the shortest possible duration, as introduced below.

[0030] The Estimation of Signal Parameters via Rotational Invariance Technique (ESPRIT) was proposed and developed by R. Roy, A. Paulraj, and T. Kailath, and has now become an effective tool for estimating the parameters (number and frequency) of sine (cosine) signals.

[0031] The sampled signal x(n) can be expressed as a combination of a series of cosine harmonic components, as shown in Equation (4).

[0032]

[0033] where T S represents the sampling period; N represents the number of sampling points; p represents the number of harmonics; A i , f i , represent the amplitude, frequency, and initial phase angle of the i-th harmonic, respectively.

[0034] Define y(n) = x(n + 1), and introduce the following m×N order matrix (ensuring m >> p):

[0035] X(n) = [x(n) x(n + 1) … x(n + m - 1)] T (5)

[0036] Y(n) = [y(n) y(n + 1) … y(n + m - 1)] T (6)

[0037] In Equations (5) and (6), T represents the transpose.

[0038] Then the autocorrelation matrix of X(n) is

[0039] R XX = E{X(n)X H (n)} (7)

[0040] And the cross-correlation matrix of X(n) and Y(n) is

[0041] R XY = E{X(n)Y H (n)} (8)

[0042] In Equations (7) and (8), E represents the mathematical expectation, and H represents the conjugate transpose.

[0043] The ESPRIT steps are as follows:

[0044] (a) Construct the correlation matrices R XX , R XY according to Equations (7) and (8);

[0045] (b) Perform eigenvalue decomposition on \(R\) XX to determine its minimum eigenvalue \(\sigma\) 2 ;

[0046] (c) Calculate \(R\) 1 = \(R\) XX - \(\sigma\) 2 \(I\), where \(I\) represents the \(m\times m\) identity matrix;

[0047] (d) Calculate \(R\) 2 = \(R\) XY - \(\sigma\) 2 \(Z\), where \(Z\) is an \(m\times m\) matrix, (here, \(I\) represents the \((m - 1)\times(m - 1)\) identity matrix);

[0048] (e) Perform singular value decomposition on \(R\) 1 \(R\) 1 = \(U\sum V\) H , where \(U = [U\) 1 \(U\) 2 ; (\(U\) 1 , \(U\) 2 , \(\sum\) 1 , \(\sum\) 2 , \(V\) 1 , \(V\) 2 are all the results obtained by performing singular value decomposition on \(R\) 1 , e.g., \(\sum\) 1 is a diagonal matrix composed of \(p\) main singular values);

[0049] (f) Calculate the matrix

[0050] (g) Perform generalized eigenvalue decomposition on to determine \(p\) generalized eigenvalues \(\lambda\) i (\(i = 1, 2, \cdots, p\)) (the remaining \(m - p\) generalized eigenvalues are always equal to 0);

[0051] (h) Determine the frequencies of the various components of the sampling signal based on the generalized eigenvalues \(\text{Im}(\lambda\) i ) and \(\text{Re}(\lambda\) i ) respectively represent the imaginary part and real part of the eigenvalue \(\lambda\) i ;

[0052] (i) Calculate the matrix

[0053] (j) Calculate the matrix \(c = (\lambda\) H \(\lambda)\) -1 \(\lambda\) H \(X\), where \(c\) is a column vector \(c = [c\) 1 \(c\) 2…c p T where X is a column vector [x(1) x(2) … x(N)] T ;

[0054] (k) Determine the amplitude and initial phase angle A of each component of the sampling signal i = 2|c i |,

[0055]

[0056] After a broken rotor bar fault occurs in a fifteen-phase induction motor, the instantaneous reactive power signal q after filtering the DC component A can be simulated by Equation (9) to analyze the performance of ESPRIT. The results are shown in Table 1. Here, the slip ratio s = 0.2% is selected to reflect the low slip ratio situation in engineering practice, and f 1 = 50 Hz, T s = 0.001 s, N = 2000, m = 200.

[0057]

[0058] According to Equation (9), s, f 1 , A 1 , , A 2 , are randomly and combinatorially transformed, and a large number of calculations are carried out, and the results are consistent.

[0059] It can be inferred therefrom that applying ESPRIT to the analysis of instantaneous reactive power signals for rotor broken bar fault diagnosis is feasible and applicable to low slip ratio situations; in addition, it is particularly suitable for situations with severe interference such as load fluctuations and noise because only short-time sampling signals are required; however, ESPRIT cannot provide accurate results for the amplitude and initial phase angle of the fault characteristic components of the broken rotor bar.

[0060] According to the calculation results of ESPRIT, the amplitude and initial phase angle of the fault characteristic components of the broken rotor bar can be further accurately calculated by applying PSA, which is briefly introduced as follows.

[0061] The Pattern Search Algorithm (PSA) is a direct search optimization method. This method consists of "exploration moves" and "pattern moves", and can perform optimization iterations for multiple variables simultaneously, suitable for multi-variable search. Exploration moves explore along the axis with a certain step size, aiming to reveal the variation law of the objective function and detect the descending direction of the function; while pattern moves directly search along the favorable direction, aiming to use the discovered function variation law to find a better iteration point.

[0062] ​Consider the optimization problem

[0063] min[E(α)], α = [α 1 α 1 …α n T

[0064] where E(α) is the objective function, α is the undetermined state that minimizes E(α), and min represents finding the minimum value. For this problem, the basic steps of PSA are as follows:

[0065] (a) Given the initial state α 0 , the axial direction e 1 , e 1 , …e n , the step size δ, the reduction rate β ∈ (0, 1), the termination parameter ε, and let y 0 = α 0 .

[0066] (b) (Exploration move) For (i = 1, 2, …, n) perform the following axial searches in sequence:

[0067] Let If then let y 0 = y 0 + δ * e i ; otherwise, let If then let y 0 = y 0 - δ * e i .

[0068] (c) (Pattern move) If E(y 0 ) < E(α 0 ), then let α 1 = y 0 + (y 0 - α 0 ), take α 1 as the new initial state, go to (b), and obtain the new iteration point y 1 —— If E(y 1 ) < E(α 1 ), then let α 1 = y 1 ; otherwise, let δ = β * δ.

[0069] (d) If δ ≤ ε, stop; otherwise, go to (b).

[0070] For the sampled signal x(n) shown in Equation (1), first apply ESPRIT to determine the frequencies f i , amplitudes A i , initial phase angles​ i = 1, 2, … p. As can be seen from the above, f i is accurate, while A i and awaits PSA processing.

[0071] Applying PSA, the key is to construct a practical objective function, as follows.

[0072] The sampled signal x(n) shown in Equation (1) can be expressed as

[0073]

[0074] Generate the p×N matrices y 1 (n), y 2 (n), as follows:

[0075] y 1 (n) = [cos(2πf 1 nT S )cos(2πf 2 nT S )…cos(2πf p nT S )] T , n = 1, 2, …, N (11)

[0076] y 2 (n) = [sin(2πf 1 nT S )sin(2πf 2 nT S )…sin(2πf p nT S )] T , n = 1, 2, …, N (12)

[0077] Let the state α = [α 1 α 2 , where α 1 and α 2 are respectively

[0078]

[0079]

[0080] And the initial state α 0 can be set according to the calculation result of ESPRIT.

[0081] Construct the objective function

[0082] E(α) = (α 1 y 1 (n) - α2 y 2 (n)-X) 2 (15)

[0083] Here, X is the column vector [x(1) x(2) … x(N)] T 。

[0084] So far, PSA can be applied to determine the amplitudes A of the respective frequency components of the sampled signal x(n) i 、initial phase angle i = 1, 2, …, p.

[0085] For the instantaneous reactive power signal in the case of rotor broken bar fault of the fifteen-phase induction motor shown in Equation (9), applying PSA, the data shows that for a short-time sampled signal (only 2 seconds), based on the calculation results of ESPRIT, PSA can accurately calculate the amplitudes and initial phase angles of each frequency component.

[0086] According to Equation (9), randomly and combinatorially transform s, f 1 、A 1 、 A 2 、 The values of are taken, and a large number of calculations are carried out, and the results are consistent.

[0087] Based on this, it can be inferred that combining ESPRIT and PSA for instantaneous reactive power signal analysis to implement rotor broken bar fault diagnosis is feasible and applicable to the low slip ratio situation, so only a short-time sampled signal is required and it is especially applicable to situations with severe interference such as load fluctuations and noise.

[0088] The present invention has two remarkable features:

[0089] First, even for a short-time signal (only 2 seconds), combining ESPRIT and PSA can still accurately estimate the characteristic component of the rotor broken bar fault - the amplitude of the 2sf component in the instantaneous reactive power, and use it as the first reliable classification feature for machine learning (LGBM is selected in the present invention); based on the analysis of ESPRIT and PSA, apply the refined Fourier transform to analyze the instantaneous signals of the fifteen-phase voltage and fifteen-phase current for a short time (only 2 seconds), so as to obtain the average active power P, the average reactive power Q, the amplitude U of the stator first-phase voltage 1 、the amplitude I of the stator first-phase current m 、 m , and use them as the 2nd to 5th reliable classification features for machine learning (LGBM is selected in the present invention). Note: Variables such as P, Q, U m 、I m are all corresponding to the f of the instantaneous signals of the fifteen-phase voltage and fifteen-phase current 1(Main frequency) component, so accurate results can also be obtained by using 2 seconds of short-time data for the refined Fourier transform.

[0090] Second, train and save the LGBM to obtain a diagnostic model based on the LGBM.

[0091] Due to the above two significant characteristics, the present invention has two unique advantages:

[0092] First, since the instantaneous reactive power signal sampled in a short time (only 2 seconds) is used as the analysis medium and ESPRIT and PSA are introduced, this method is applicable to situations with low slip rates and severe interferences such as load fluctuations and noise.

[0093] Second, since the LGBM is introduced, this method has high precision.

[0094] Verified by experiments, even in the case of a low slip rate, this method can still accurately diagnose the number of broken rotor bars. The model training accuracy is 100%, the test accuracy is 100%, and the 5-fold cross-validation accuracy is 99.38%.

[0095] The present invention will be further described below with reference to the accompanying drawings. Description of the Drawings

[0096] Figure 1 is the experimental wiring diagram;

[0097] Figure 2 is the decision tree growth strategy diagram of the LGBM. Specific Embodiments

[0098] The present invention proposes a method for diagnosing the number of broken rotor bars of a fifteen-phase asynchronous motor based on ESPRIT-PSA and LGBM, and the diagnostic accuracy rate of this method is as high as 99.38%. The characteristics of the present invention lie in the extraction of the amplitude A of the instantaneous reactive power fault component by applying ESPRIT-PSA q and the model training based on the LGBM, which will be specifically described below.

[0099] Figure 1This is the experimental wiring diagram. Among them, the experimental motor is a fifteen-phase asynchronous motor with a rated voltage of 380V, a rated power of 5.5kW, and a rated frequency of 50Hz. To conduct the rotor bar breaking experiment, in addition to the normal rotor, two faulty rotors (drilled holes at a distance of 10mm from the end ring on the bars, with a depth of 15mm and a diameter of 10mm) are equipped to simulate the bar breaking fault. These two faulty rotors have one broken bar and two consecutive broken bars respectively. The data acquisition system collects the instantaneous signals of the stator fifteen-phase voltage and the stator fifteen-phase current through current transducers and voltage transducers. The load uses a DC dynamometer, and by adjusting the DC dynamometer, the fifteen-phase asynchronous motor is respectively in the full-load, half-load, and no-load states.

[0100] A large number of experiments are carried out to measure the instantaneous signals of the stator fifteen-phase voltage and current. This work is carried out one by one for the three states of the fifteen-phase asynchronous motor being normal, having one rotor bar breaking fault, and having two rotor bar breaking faults, and the experiments in each state include the load change of the motor (full-load, half-load, and no-load). The above three states of the motor are respectively and sequentially marked as state 0, 1, and 2. Through this work, a large number of motor sample data are obtained. When the motor is in state 0, 750 groups of sample data can be obtained through the above work for each of the full-load, half-load, and no-load conditions, so state 0 includes 2250 groups of sample data. When the motor is in state 1 and state 2, 1250 groups of sample data can be obtained through the above work for each of the full-load, half-load, and no-load conditions, so state 1 and state 2 each include 3750 groups of sample data. Then the three states of the motor altogether contain 9750 groups of sample data.

[0101] The LGBM used in the present invention is a new member of the Boosting algorithm and is a machine learning method developed by Microsoft. It is an efficient implementation of the GBDT (Gradient Boosting Decision Tree) algorithm. Similar to the GBDT algorithm in principle, it both uses the negative gradient of the loss function as the residual approximation of the current decision tree to fit a new decision tree. However, compared with traditional machine learning algorithms, LGBM has obvious advantages: higher training efficiency, lower memory occupancy, higher accuracy, support for parallel learning, and can handle large-scale data.

[0102] Figure 2It is a decision tree growth strategy diagram of LGBM. LGBM generates decision trees through the leaf-wise (best-first) strategy. LGBM selects the leaf node with the largest loss to grow a new leaf. When growing the same number of leaves, the leaf-wise algorithm can reduce the loss compared to the level-wise algorithm. However, when the data volume is small, leaf-wise may cause overfitting. Therefore, LightGBM can use the additional parameter max_depth to limit the depth of the tree and avoid overfitting (the max_depth value set in the present invention is 3). In addition, the present invention customizes parameters such as min_data_in_leaf, max_bin, and bagging_fraction through GridSearchCV to handle the overfitting problem of the model and improve the training speed of the model.

[0103] In order to test the actual effect of the method of the present invention, for the normal state (state 0), 1 broken bar fault (state 1), and 2 broken bar faults (state 2), 45 groups of data were separately measured (15 groups for each state, and the load conditions were randomly set to no-load or half-load or full-load). Using the method of the present invention to perform "blind testing" on the above 45 groups of data, it can be seen from the results that the method of the present invention has high precision.

Claims

1. A high-precision diagnostic method for the number of broken rotor bars of a fifteen-phase induction motor based on ESPRIT-PSA and LGBM, characterized in that, it includes the following steps: For the instantaneous reactive power signal of a 15-phase induction motor with a 2-second short-time sampling, first use the rotational invariance signal parameter estimation technique ESPRIT to calculate the accurate frequency value, as well as the rough amplitude and initial phase angle of the characteristic component with a frequency of 2sf 1 , where s is the slip ratio and f 1 is the supply frequency; Then, the result calculated by ESPRIT is used as the initial value and substituted into the pattern search algorithm PSA, so as to calculate the accurate amplitude A of the characteristic component q and the initial phase angle, and A q is placed into the data set X as the first characteristic variable; Then, perform a refined Fourier spectrum analysis on the instantaneous signals of the fifteen-phase voltage and fifteen-phase current of the fifteen-phase asynchronous motor to obtain the fifteen-phase voltage amplitude, fifteen-phase current amplitude, fifteen-phase voltage phase, and fifteen-phase current phase, a total of 60 features corresponding to A q ; Then, based on these 60 features, the effective values of positive-sequence, negative-sequence, and zero-sequence voltages, the effective values of positive-sequence, negative-sequence, and zero-sequence currents, the magnitudes of positive-sequence, negative-sequence, and zero-sequence impedances, as well as the average active power and average reactive power are obtained, totaling 11 derivative features corresponding to A q ; Subsequently, the above 71 features corresponding to A q are placed into the dataset X to form a 72-dimensional dataset X 1 ; furthermore, X 1 is input into the Light Gradient Boosting Machine (LGBM) classifier to calculate the feature weights, and the top 5 features with the highest weight ratios are determined and selected, namely the amplitude A q of the instantaneous reactive power feature component, the average active power P, the average reactive power Q, the amplitude U m of the first-phase voltage, and the amplitude I m of the first-phase current, forming a new dataset X 2 ; Then, divide X 2 into a training set and a test set according to the ratio of 8:2 and input them into the LGBM model to classify and diagnose the number of broken bars in the rotor of a fifteen-phase induction motor, where 0 broken bars means normal, 1 broken bar, and 2 broken bars; finally, select the optimal hyperparameters of the LGBM model through GridSearchCV grid search built into the scikit-learn library. Accordingly, perform high-precision diagnosis on the number of broken rotor bars of the fifteen-phase induction motor, and save the trained LGBM model for subsequent diagnosis; Specifically including: a. Conduct an experiment to measure the instantaneous stator fifteen-phase voltage signal u sn and the instantaneous stator fifteen-phase current signal i sn , where n represents the phase number, n = 1, 2,..., 15. This work is carried out for a fifteen-phase induction motor under normal conditions, marked as state 0, with one broken rotor bar fault, marked as state 1, and with two broken rotor bar faults, marked as state 2. These three states are carried out one by one, and the experiments under each state include three load changes: full load, half load, and no load; b. Calculate the instantaneous reactive power signal from the stator fifteen-phase voltage and current instantaneous signals according to Equations (1) and (2), and filter out its DC component according to Equation (3), so as to obtain the instantaneous reactive power signal q after filtering out the DC component A ; q A = q 0 - mean(q 0 ) (3) In Formula (1), Formula (2) and Formula (3), represents the Hilbert transform of the instantaneous signal u of the nth-phase stator voltage sn ; t represents time; τ represents delay; q 0 represents the instantaneous reactive power; mean(q 0 ) represents the average value of q 0 , that is, the DC component: c. For q A Perform ESPRIT analysis to calculate the accurate frequency values, as well as the rough amplitude and initial phase angle, of the characteristic components in the instantaneous reactive power signal; d. Using the amplitude and initial phase angle of the instantaneous reactive power characteristic components calculated by ESPRIT as the initial reference values, use PSA to calculate the accurate amplitude A q and initial phase angle of the characteristic components, and then use A q as the first characteristic variable and place it into the data set X, where X is the data set storing Aq values; e. Perform refined Fourier analysis on the instantaneous signals of the fifteen-phase voltage and fifteen-phase current to obtain the fifteen-phase voltage amplitude, fifteen-phase current amplitude, fifteen-phase voltage phase, and fifteen-phase current phase, a total of 60 features corresponding to A q Then process these 60 features to obtain the positive-sequence, negative-sequence, and zero-sequence voltage effective values, positive-sequence, negative-sequence, and zero-sequence current effective values, positive-sequence, negative-sequence, and zero-sequence impedance modulus values, as well as the average active power and average reactive power, a total of 11 derivative features corresponding to A q Subsequently, place the above 71 features corresponding to A q into the dataset X to form a 72-dimensional dataset X 1 , where X 1 is a 72-dimensional dataset storing A q and the 71 corresponding feature values. Then use the LGBM classifier to calculate the weights of the dataset X 1 and determine and select the top 5 features with the highest weight ratios including the amplitude of the instantaneous reactive power feature component corresponding to A q , the average active power P, the average reactive power Q, the amplitude U q of the first-phase voltage, and the amplitude I m of the first-phase current as the features for the LGBM model to learn, thereby forming a new 5-dimensional dataset X m 2 , where X 2 is a 5-dimensional dataset composed of the top 5 features with the highest weight ratios calculated by LGBM. A q , P, Q, U m , and I m are all obtained by processing the sampling signals using the sliding window method. The window includes data with a duration of 2 seconds. Specifically, in state 0, 750 sets of sample data can be obtained for each of the full-load, half-load, and no-load conditions through the above work. State 0 includes 2250 sets of sample data; in states 1 and 2, 1250 sets of sample data can be obtained for each of the full-load, half-load, and no-load conditions through the above work. States 1 and 2 each include 3750 sets of sample data. Then the three states together contain 9750 sets of sample data;​ f. For the above dataset X 2 , mark it as 0, 1, 2 according to its corresponding status, corresponding to status 0, status 1, and status 2 respectively; g. Divide the dataset X 2 into a training set and a test set in a ratio of 8:2, substitute them into the LGBM model for training, and perform hyperparameter tuning of the model through GridSearchCV; h. Call the joblib package in the scikit-learn library, save the trained LGBM model as an executable code file with the extension.m, and read it through the joblib package when the LGBM model is needed. The joblib package can save the trained model and can be directly called when needed.

2. The high-precision diagnostic method for the number of broken rotor bars of a fifteen-phase induction motor based on ESPRIT-PSA and LGBM according to claim 1, characterized in that, use the Light Gradient Boosting Machine to classify the state of the fifteen-phase induction motor, specifically: normal is state 0, 1 broken rotor bar fault is state 1, and 2 broken rotor bar faults are state 2.

Images