A rotor fault diagnosis method based on weighted BP-AdaBoost

Through the rotor fault diagnosis method based on weighted BP-AdaBoost, the rotor vibration signal characteristics and AdaBoost algorithm are used to solve the problem that the error rate of multi-classification problems in rotor fault diagnosis is difficult to meet, and high-accurate fault type and fault degree identification is achieved .

CN114218991BActive Publication Date: 2024-11-08NORTHEASTERN UNIV CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111634135.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-29
Publication Date
2024-11-08
Estimated Expiration
2041-12-29

AI Technical Summary

Technical Problem

The prior art is difficult to effectively deal with multi-classification problems in rotor fault diagnosis, especially in identifying rotor fault types and fault degrees, especially because the application of adaptive enhancement algorithms in multi-classification situations has problems with difficulty in meeting the error rate.

Method used

The rotor fault diagnosis method based on weighted BP-AdaBoost is adopted to extract the time domain, frequency domain and time frequency domain characteristics of the rotor vibration signal, and to combine the variance method to build a subset of low-dimensional features, and use K-fold cross-validation and AdaBoost algorithm to build multiple features. The classification model uses a pair of other strategies and weighted majority voting methods to construct a weighted BP-AdaBoost multi-classification model.

Benefits of technology

Improve the recognition effect of rotor fault diagnosis, and can show high universality in both binary and multi-classification problems. Through the weighting effect of non-fuzzy solution coefficients, the identification accuracy of rotor fault type and fault degree is significantly improved. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure BDA0003441135570000031
    Figure BDA0003441135570000031
  • Figure BDA0003441135570000035
    Figure BDA0003441135570000035
  • Figure BDA0003441135570000037
    Figure BDA0003441135570000037
Patent Text Reader

Abstract

The present invention provides a rotor fault diagnosis method based on weighted BP-AdaBoost, which relates to the field of rotor fault diagnosis. The rotor vibration signal is measured by a sensor, and the rotor state characteristics are extracted from different angles to obtain the rotor system state feature vector, and then the variance of the initial feature vector is calculated. According to the size of the variance value, the feature vector dimension is reasonably selected, and the BP neural network is used as the base classifier of the AdaBoost algorithm to construct a binary classifier, and a multi-classification classifier is constructed by combining several binary classifiers. In view of the problems existing in the decision-making of the combined multi-classification classifier, the present invention proposes a non-fuzzy solution coefficient based on the verification sample, and combines the K-fold cross validation to obtain a weighted BP-AdaBoost multi-classifier, and completes the rotor fault diagnosis. Through experimental verification, this method has good effects in both fault type identification and fault degree identification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of rotor fault diagnosis, and in particular to a rotor fault diagnosis method based on weighted BP-AdaBoost. Background Technology

[0002] In recent years, condition monitoring of complex machinery has received increasing attention and is widely used in critical machinery. However, the numerous component characteristics make diagnosis difficult; typical examples include aero-engines, helicopter gearboxes, and wind turbines. Rotating machinery plays an irreplaceable role in many fields, such as defense and military, transportation, and aerospace. Rotor faults are the most common phenomenon in fault diagnosis.

[0003] Mechanical rotor systems inevitably experience mechanical failures after operating under load for a certain period of time. Common rotor system failures include rotor imbalance, rotor misalignment, rubbing, rotor cracks, and loose bolts. Timely diagnosis of the type and severity of these failures is crucial for ensuring normal production and safety.

[0004] In recent years, detection methods based on machine learning have gradually attracted widespread attention from scholars. In rotor fault identification, classic methods typically include support vector machines, K-nearest neighbors, and decision trees. Artificial neural networks and boosting methods, on the other hand, have gained increasing attention due to their better generalization performance and robustness.

[0005] The real world is a complex, nonlinear system. Artificial neurons, whether activated or inhibited, exhibit mathematically nonlinear relationships. Neural networks store knowledge in their connection weights and can perform various nonlinear mappings. Due to their superior performance in handling nonlinear problems, artificial neural networks have been widely applied, with backpropagation (BP) neural networks being the most typical example. When new training samples are input, an artificial neural network can automatically adjust its structural parameters and change the mapping relationships to produce the desired output for a specific input. Each neuron can perform independent computation and processing based on the received information, and different neurons in the same layer can perform computations simultaneously.

[0006] Boosting is a commonly used statistical learning method. In classification problems, it learns multiple classifiers by changing the weights of training samples and then linearly combines these classifiers to improve classification performance. Adaptive Boosting is a typical example of boosting methods. While Adaptive Boosting has many successful applications in binary classification problems, its application in multi-class classification presents some challenges. Theoretically, Adaptive Boosting requires that the error of each weak classifier be less than half. This is easily satisfied for binary classification because the error rate of random guessing is half. However, in multi-class cases, this is difficult to achieve, as the error rate of random guessing is (K-1) / K, where K represents the number of classes. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a rotor fault diagnosis method based on weighted BP-AdaBoost.

[0008] A rotor fault diagnosis method based on weighted BP-AdaBoost, the specific steps of which are as follows:

[0009] Step 1: Extract rotor state parameters and select features; use an accelerometer to measure the rotor vibration signal, combine the time domain, frequency domain and time-frequency domain features of the rotor vibration signal to construct a rotor state feature set; and use the variance method to select and sort the features to construct a low-dimensional feature subset.

[0010] Step 1.1: Extract the time-domain features of the rotor vibration signal;

[0011] Time-domain features include: peak, mean, root mean square, kurtosis factor, clearance factor, shape factor, impulse factor, and crease factor.

[0012] Peak = max(|x|), denoted by TF1;

[0013] Represented by TF2;

[0014] Represented by TF3;

[0015] Represented by TF4;

[0016] Represented by TF5;

[0017] Represented by TF6;

[0018] Represented by TF7;

[0019] Represented by TF8;

[0020] In the formula, x represents the vibration signal of the sample, x n Let x represent a point; RMS represents the root mean square; N represents the number of sampling points of the vibration signal. σ represents the mean, σ represents the standard deviation, and max represents the maximum value;

[0021] Step 1.2: Extract the frequency domain features of the rotor vibration signal;

[0022] Frequency domain characteristics include: mean frequency (MF), centroid frequency (FC), root mean square frequency (RMFS), and standard deviation frequency (RVF).

[0023] Represented by FF1;

[0024] Represented by FF2;

[0025] Represented by FF3;

[0026] Represented by FF4;

[0027] In the formula, s represents the amplitude, and h represents the frequency; m h is a point within s; m It is a point in h; q represents the number of frequency points;

[0028] Step 1.3: Extract the time-frequency domain features of the rotor vibration signal;

[0029] In the time-frequency domain, wavelet transform is used to perform a three-level decomposition to obtain nodes, calculate the energy of each node, and calculate the ratio of the energy of each node to the total energy of all nodes; that is, the energy ratio E = {E1, E2, E3, E4, E5, E6, E7, E8}; the energy ratio is used as a feature in the time-frequency domain.

[0030] Step 1.4: Construction of the rotor state feature set;

[0031] By combining the features in the time domain, frequency domain, and time-frequency domain, the rotor state feature set F is obtained;

[0032] F={FF1,FF2,FF3,FF4,TF1,TF2,TF3,TF4,TF5,TF6,TF7,TF8,E1,E2,E3,E4,E5,E6,E7,E8};

[0033] Step 1.5: Construct a low-dimensional feature subset based on variance;

[0034] The variance of each feature in the rotor state feature set F is calculated as follows:

[0035]

[0036] Where N represents the number of samples, y i A single feature representing an independent sample. The mean of a sample feature;

[0037] The rotor state feature set F has M-dimensional features, where M represents the number of features in the feature set F; the variance corresponding to all features is calculated from the above equation (1), and is expressed as:

[0038]

[0039] In the formula, Let represent the variance of the i-th feature;

[0040] Then the difference σ 2 After normalization, we get for:

[0041]

[0042] In the formula Let represent the normalized variance of the i-th feature;

[0043] Based on the variance values, a feature set with a certain sequence is obtained. A variance threshold is set to select the feature dimension, and a low-dimensional feature subset is constructed.

[0044] Step 2: Divide the low-dimensional feature subset obtained in Step 1 proportionally to obtain the training sample set C1 and the test sample set O1; process the training sample set C1 according to the K-fold cross-validation idea;

[0045] The training sample set C1 is processed according to the K-fold cross-validation approach, that is, it is divided into K parts with an equal number of samples, denoted as [Λ1,Λ2,…,Λ]. i ,…,Λ K-1 ,Λ K ], without repetition, (K-1) / K is selected as the training sample set C, and 1 / K is selected as the validation sample set V, represented by matrix Λ:

[0046]

[0047] In the matrix, the first row to the last row contains the K-round training sample set C and the validation sample set V obtained from K-fold cross-validation. The first K-1 columns of each row are the training sample set C.

[0048]

[0049] The Kth column of each row is the validation sample set V:

[0050]

[0051] Step 3: Set the number of iterations K, and select each row element in Λ in turn. When the loop reaches the i-th iteration, i∈[1,K], we obtain the training sample set C2 and the validation sample set V1:

[0052] C2=[Λ1,Λ2,…,Λ i-1 ,Λ i+1 ,…,Λ K ]

[0053] V1 = [Λ i ]

[0054] The AdaBoost algorithm is used to obtain a BP-AdaBoost binary classification model; using a pair of other strategies, namely the OneAgainstAll mode, multiple binary BP-AdaBoost classifiers are combined to build an M-BP-AdaBoost multi-class model; after K rounds of iteration, K M-BP-AdaBoost multi-class models are obtained.

[0055] Step 3.1: Preprocess the training sample set C2 obtained after partitioning, extracting rotor fault samples of a certain state sequentially, denoted by y∈{+1}. The labels of rotor fault samples of the remaining states are all denoted by y∈{-1}. The training sample set obtained after preprocessing is denoted as: U=[U1,U2,…,U i ,…,U M ];

[0056] Step 3.2: Obtain the BP-AdaBoost binary classification model using the AdaBoost algorithm; process the preprocessed training sample set U = [U1, U2, ..., U...] using the BP-AdaBoost algorithm. i ,…,U M Training is performed to obtain several binary classifiers h = [h1, h2, ..., h] for the M-class classification problem. i ,…,h M Combine multiple binary BP-AdaBoost classifiers to build a multi-classification model;

[0057] Step 3.2.1: The AdaBoost algorithm improves classification performance by increasing the weights of misclassified samples from the previous weak classifier and decreasing the weights of correctly classified samples, and by using a weighted majority voting method; a BP neural network is used as the base learner for AdaBoost to build a BP-AdaBoost binary classification model.

[0058] Step 3.2.1.1: First, input a training sample set U. i :

[0059] U i ={(x1,y1),(x2,y2),…,(x i ,y i ),…,(x N ,y N )}

[0060] Each sample point consists of a sample instance and a sample label; sample instance x i ∈χ∈R n Let represent the feature vector of the i-th sample, χ be the instance space, and y be the sample label. i ∈Y={-1,+1},y i The label category of the i-th sample, Y is the label space;

[0061] Step 3.2.1.2: Initialize the training sample set U i The weight distribution; each training sample is assigned the same weight:

[0062]

[0063] In the formula ω 1,i This represents the initial weight of each training sample, and N represents the number of training samples;

[0064] Training sample set U i The initial weight vector distribution is as follows:

[0065]

[0066] Step 3.2.1.3: Establish the iteration termination condition e t If the termination condition is not met, repeat steps A1-A5 to obtain T base learners, and construct a strong classifier using the T base learners; until the termination condition is met or T>10, then execute step 3.2.2; t This represents the classification error rate for each round, and r represents the threshold for the training error rate given according to the training requirements; if the termination condition is met, the BP-AdaBoost binary classification model is obtained, and step 3.2.2 is executed;

[0067] Step A1: Based on the above training sample set U i The initial weight distribution is used to obtain the t-th base learner;

[0068] G t (x):χ→{-1,+1}

[0069] In the formula, G t (x) represents the base learner, i.e., the BP neural network;

[0070] Step A2: Calculate G t (x) in the training sample set U i Classification error rate on the initial weight distribution:

[0071]

[0072] In the formula ω t,i I(G) represents the weight of the i-th sample in the t-th round. t (x i )≠y i ) is an indicator function, when the prediction result G t (x i ) and the actual result y i When unequal, I(G) t (x i )≠y i The value is 1 for the first case and 0 for the second case.

[0073] Step A3: Calculate G t The coefficients of (x) are the weights corresponding to the base learners:

[0074]

[0075] In the formula α t e represents the weight of each base learner. t Indicates the classification error rate;

[0076] Step A4: Update training sample U i Weights:

[0077] D t+1 =(ω t+1,1 ,ω t+1,2 ,...,ω t+1,N )

[0078]

[0079] Among them, Z t It is the normalization factor, and exp is an exponential function with base e;

[0080]

[0081] Step A5: Combine the T base learners to build a strong binary classifier model;

[0082]

[0083]

[0084] In the formula, f(x) represents a linear combination of classifiers, sign represents the sign function; G(x) represents a strong binary classifier; and T is the number of base learners.

[0085] Step 3.2.2: Step 3.2.1 constructed a binary classification model based on BP-AdaBoost. Step 3.2.1 is iterated M times to obtain the combination of binary classifiers h = [h1, h2, ..., h...] for the M-class classification problem. i ,…,h M At this point, the combination of multiple binary BP-AdaBoost classifiers is complete, denoted as the M-BP-AdaBoost multi-class model.

[0086] Step 3.3: After K rounds of iteration, K M-BP-AdaBoost multi-class classification models can be obtained, denoted as [M1, M2, ... M]. K ];

[0087] Step 4: Extract K unfuzzy solution weights for the M-BP-AdaBoost model based on the validation sample set V;

[0088] Step 4.1: Validate the sample set V = [Λ K ,Λ K-1 ,…,Λ i ,…,Λ1] T The data is sequentially fed into K M-BP-AdaBoost models, and the weights of the unfuzzy solutions are extracted based on the classification fuzziness of the validation sample set. A validation sample set Λ is selected for each M-BP-AdaBoost model. i This indicates that among the predicted labels of the i-th validation sample, only one BP-AdaBoost binary classifier gives y∈{+1}, while the remaining BP-AdaBoost binary classifiers give validation samples with y∈{-1}; the number of unfuzzy solutions is related to the validation sample set Λ. i By comparing the total number of solutions, the proportion of non-fuzzy solutions is determined and denoted as the non-fuzzy solution weight γ. i ;

[0089]

[0090] In the formula, γ i The weights are the unfuzzy solutions of the i-th M-BP-AdaBoost multi-class classification model. Let ψ be the number of nonfuzzy solutions in the i-th round. i Let be the total number of validation samples in the validation sample set of the i-th round;

[0091] Step 4.2: The unfuzzy solution weight vector γ of the K M-BP-AdaBoost models:

[0092] γ=[γ1,…,γ K ]

[0093] In the formula, γ1~γ K This represents the weight of the non-fuzzy solution for each of the K M-BP-AdaBoost models;

[0094] Step 5: Construct a weighted BP-AdaBoost multi-class classification model;

[0095] By combining the multi-classification model M-BP-AdaBoost established in step 3 and the non-fuzzy solution weights of the K M-BP-AdaBoost multi-classification models extracted in step 4, a weighted BP-AdaBoost multi-classification model can be obtained, denoted as Model.

[0096] Model = [γ1, ..., γ K [M1,M2,…M] K ] T

[0097] This completes the establishment of the weighted BP-AdaBoost multi-class classification model;

[0098] Step 6: Use the weighted BP-AdaBoost multi-class classification model to predict the test sample set O1;

[0099] Using the weighted BP-AdaBoost multi-class classification model Model, the test sample set O1 is input into the weighted BP-AdaBoost multi-class classification model Model to predict the label category corresponding to each sample in the test sample set O1;

[0100] Step 6.1: Substitute the test sample set O1 into K M-BP-AdaBoost multi-class classification models, and the predicted result S is expressed as:

[0101]

[0102] s Kn This represents the prediction result for the Kth M-BP-AdaBoost test sample set O1;

[0103] Simplified to: S = [s1, s2, ..., s i ,…,s K ]T , where s i =[s i1 s i2 …s in (i = 1, 2, ..., K);

[0104] Step 6.2: Combine the weight vectors γ of the K unfuzzy solutions and the prediction results S of the K M-BP-AdaBoost multi-classification models to obtain the weighted test feature solution. Represented as:

[0105]

[0106] In the formula, γ represents the weight of the non-fuzzy solution of the multi-classifier, and S represents the prediction result of the classification BP-AdaBoost model;

[0107]

[0108] exist In the test sample O1, each test sample O1 has a value under each class label, and the label corresponding to the maximum value is the prediction result of each sample in test sample O1.

[0109] The beneficial effects of this invention are:

[0110] A rotor fault diagnosis method based on weighted BP-AdaBoost is universally applicable to both binary and multi-class classification problems. This method averages out the impact of errors by weighting samples during the training phase and weighting multi-classifiers during the validation phase, thus improving the recognition performance to a certain extent. This invention provides a solution for applying binary classification algorithms to multi-class classification problems. The proposed non-fuzzy solution coefficients can both verify the training effect and weight the test results. Experiments have verified that the weighted test solution based on the non-fuzzy solution coefficients has a high recognition effect on rotor fault types and fault severity. Simultaneously, by utilizing machine learning methods and vibration signal analysis of the rotor system, the problem of identifying rotor system faults and their severity is effectively solved, enabling users to more accurately detect faults in the system. Attached Figure Description

[0111] Figure 1 This is a diagram of the BP neural network structure according to an embodiment of the present invention.

[0112] Figure 2 This is a diagram of the BP-AdaBoost model according to an embodiment of the present invention.

[0113] Figure 3 This is a diagram of a multi-classification model based on BP-AdaBoost, as described in an embodiment of the present invention.

[0114] Figure 4This is a flowchart of an embodiment of the present invention based on non-fuzzy solution weighted BP-AdaBoost.

[0115] Figure 5 This is a schematic diagram of rotor fault design in an embodiment of the present invention. Detailed Implementation

[0116] The present invention will be further described below with reference to the accompanying drawings and embodiments. A rotor fault diagnosis method based on weighted BP-AdaBoost is specifically as follows:

[0117] Step 1: Extract rotor state parameters and select features; use an accelerometer to measure the rotor vibration signal, combine the time domain, frequency domain and time-frequency domain features of the rotor vibration signal to construct a rotor state feature set; and use the variance method to select and sort the features to construct a low-dimensional feature subset.

[0118] Step 1.1: Extract the time-domain features of the rotor vibration signal;

[0119] Time-domain features include: peak, mean, root mean square, kurtosis factor, clearance factor, shape factor, impulse factor, and crease factor.

[0120] Peak = max(|x|), denoted by TF1;

[0121] Represented by TF2;

[0122] Represented by TF3;

[0123] Represented by TF4;

[0124] Represented by TF5;

[0125] Represented by TF6;

[0126] Represented by TF7;

[0127] Represented by TF8;

[0128] In the formula, x represents the vibration signal of the sample, x n Let x represent a point; RMS represents the root mean square; N represents the number of sampling points of the vibration signal. σ represents the mean, σ represents the standard deviation, and max represents the maximum value;

[0129] Step 1.2: Extract the frequency domain features of the rotor vibration signal;

[0130] Frequency domain characteristics include: mean frequency (MF), centroid frequency (FC), root mean square frequency (RMFS), and standard deviation frequency (RVF).

[0131] Represented by FF1;

[0132] Represented by FF2;

[0133] Represented by FF3;

[0134] Represented by FF4;

[0135] In the formula, s represents the amplitude, and h represents the frequency; m h is a point within s; m It is a point in h; q represents the number of frequency points;

[0136] Step 1.3: Extract the time-frequency domain features of the rotor vibration signal;

[0137] In the time-frequency domain, wavelet transform is used to perform a three-level decomposition to obtain nodes, calculate the energy of each node, and calculate the ratio of the energy of each node to the total energy of all nodes; that is, the energy ratio E = {E1, E2, E3, E4, E5, E6, E7, E8}; the energy ratio is used as a feature in the time-frequency domain.

[0138] Step 1.4: Construction of the rotor state feature set;

[0139] By combining the features in the time domain, frequency domain, and time-frequency domain, the rotor state feature set F is obtained;

[0140] F={FF1,FF2,FF3,FF4,TF1,TF2,TF3,TF4,TF5,TF6,TF7,TF8,E1,E2,E3,E4,E5,E6,E7,E8};

[0141] Step 1.5: Construct a low-dimensional feature subset based on variance;

[0142] The variance of each feature in the rotor state feature set F is calculated as follows:

[0143]

[0144] Where N represents the number of samples, y i A single feature representing an independent sample. The mean of a sample feature;

[0145] The rotor state feature set F has M-dimensional features, where M represents the number of features in the feature set F; the variance corresponding to all features is calculated from the above equation (1), and is expressed as:

[0146]

[0147] In the formula, Let represent the variance of the i-th feature;

[0148] Then the difference σ 2 After normalization, we get for:

[0149]

[0150] In the formula Let represent the normalized variance of the i-th feature;

[0151] Based on the variance values, a feature set with a certain sequence is obtained. A variance threshold is set to select the feature dimension, and a low-dimensional feature subset is constructed.

[0152] According to the rotor fault design diagram Figure 5 The design needs to identify the types of faults required.

[0153] Vibration signals from four rotor states—normal, misaligned, unbalanced, and rubbing—were collected using eddy current sensors. Sliding intercepts were applied to the signals to construct a certain number of samples. Time-domain, frequency-domain, and time-frequency-domain features were extracted from these samples to form a feature set. The rotor fault descriptions are shown in Table 1 below.

[0154]

[0155] Step 2: Divide the low-dimensional feature subset obtained in Step 1 proportionally to obtain the training sample set C1 and the test sample set O1; process the training sample set C1 according to the K-fold cross-validation idea;

[0156] The training sample set C1 is processed according to the K-fold cross-validation approach, that is, it is divided into K parts with an equal number of samples, denoted as [Λ1,Λ2,…,Λ]. i ,…,Λ K-1 ,Λ K ], without repetition, (K-1) / K is selected as the training sample set C, and 1 / K is selected as the validation sample set V, represented by matrix Λ:

[0157]

[0158] In the matrix, the first row to the last row contains the K-round training sample set C and the validation sample set V obtained from K-fold cross-validation. The first K-1 columns of each row are the training sample set C.

[0159]

[0160] The Kth column of each row is the validation sample set V:

[0161]

[0162] Step 3: Set the number of iterations K, and select each row element in Λ in turn. When the loop reaches the i-th iteration, i∈[1,K], we obtain the training sample set C2 and the validation sample set V1:

[0163] C2=[Λ1,Λ2,…,Λ i-1 ,Λ i+1 ,…,Λ K ]

[0164] V1 = [Λ i ]

[0165] The AdaBoost algorithm is used to obtain a BP-AdaBoost binary classification model. Using a one-against-all strategy, multiple BP-AdaBoost classifiers with binary classification are combined to build an M-BP-AdaBoost multi-class classification model. After K rounds of iteration, K M-BP-AdaBoost multi-class classification models are obtained. The BP neural network structure diagram is shown below. Figure 1 As shown;

[0166] Step 3.1: Preprocess the training sample set C2 obtained after partitioning, extracting rotor fault samples of a certain state sequentially, denoted by y∈{+1}. The labels of rotor fault samples of the remaining states are all denoted by y∈{-1}. The training sample set obtained after preprocessing is denoted as: U=[U1,U2,…,U i ,…,U M ];

[0167] Step 3.2: Obtain the BP-AdaBoost binary classification model using the AdaBoost algorithm; process the preprocessed training sample set U = [U1, U2, ..., U...] using the BP-AdaBoost algorithm. i ,…,U M Training is performed to obtain several binary classifiers h = [h1, h2, ..., h] for the M-class classification problem. i ,…,h M Combine multiple binary BP-AdaBoost classifiers to build a multi-classification model;

[0168] Step 3.2.1: The AdaBoost algorithm improves classification performance by increasing the weights of misclassified samples from the previous weak classifier and decreasing the weights of correctly classified samples, and employs a weighted majority voting method. A BP neural network is used as the base learner for AdaBoost to construct a BP-AdaBoost binary classification model. The BP-AdaBoost model diagram is shown below. Figure 2 As shown;

[0169] Step 3.2.1.1: First, input a training sample set U. i :

[0170] U i ={(x1,y1),(x2,y2),…,(x i ,y i ),…,(x N ,y N )}

[0171] Each sample point consists of a sample instance and a sample label; sample instance x i ∈χ∈R n Let represent the feature vector of the i-th sample, χ be the instance space, and y be the sample label. i ∈Y={-1,+1},y i The label category of the i-th sample, Y is the label space;

[0172] Step 3.2.1.2: Initialize the training sample set U i The weight distribution; each training sample is assigned the same weight:

[0173]

[0174] In the formula ω 1,i This represents the initial weight of each training sample, and N represents the number of training samples;

[0175] Training sample set U i The initial weight vector distribution is as follows:

[0176]

[0177] Step 3.2.1.3: Establish the iteration termination condition e t If the termination condition is not met, repeat steps A1-A5 to obtain T base learners, and construct a strong classifier using the T base learners; until the termination condition is met or T>10, then execute step 3.2.2; t This represents the classification error rate for each round, and r represents the threshold for the training error rate given according to the training requirements; if the termination condition is met, the BP-AdaBoost binary classification model is obtained, and step 3.2.2 is executed;

[0178] Step A1: Based on the above training sample set U i The initial weight distribution is used to obtain the t-th base learner;

[0179] G t (x):χ→{-1,+1}

[0180] In the formula, G t (x) represents the base learner, i.e., the BP neural network;

[0181] Step A2: Calculate G t (x) in the training sample set U i Classification error rate on the initial weight distribution:

[0182]

[0183] In the formula ω t,i I(G) represents the weight of the i-th sample in the t-th round. t (x i )≠y i ) is an indicator function, when the prediction result G t (x i ) and the actual result y i When unequal, I(G) t (x i )≠y i The value is 1 for the first case and 0 for the second case.

[0184] Step A3: Calculate G t The coefficients of (x) are the weights corresponding to the base learners:

[0185]

[0186] In the formula α t e represents the weight of each base learner. t Indicates the classification error rate;

[0187] Step A4: Update training sample U i Weights:

[0188] D t+1 =(ω t+1,1 ,ω t+1,2 ,...,ω t+1,N )

[0189]

[0190] Among them, Z t It is the normalization factor, and exp is an exponential function with base e;

[0191]

[0192] Step A5: Combine the T base learners to build a strong binary classifier model;

[0193]

[0194]

[0195] In the formula, f(x) represents a linear combination of classifiers, sign represents the sign function; G(x) represents a strong binary classifier; and T is the number of base learners.

[0196] Step 3.2.2: Step 3.2.1 constructed a binary classification model based on BP-AdaBoost. Step 3.2.1 is iterated M times to obtain the combination of binary classifiers h = [h1, h2, ..., h...] for the M-class classification problem. i ,…,h M At this point, the combination of multiple binary BP-AdaBoost classifiers is complete, denoted as the M-BP-AdaBoost multi-class model; a diagram of the BP-AdaBoost-based multi-class model is attached. Figure 3 As shown.

[0197] Step 3.3: After K rounds of iteration, K M-BP-AdaBoost multi-class classification models can be obtained, denoted as [M1, M2, ... M]. K ];

[0198] Step 4: Extract the unfuzzy solution weights of the M-BP-AdaBoost model based on the validation sample set V;

[0199] Step 4.1: Validate the sample set V = [Λ K ,Λ K-1 ,…,Λ i ,…,Λ1] T The data is sequentially fed into K M-BP-AdaBoost models, and the weights of the unfuzzy solutions are extracted based on the classification fuzziness of the validation sample set. A validation sample set Λ is selected for each M-BP-AdaBoost model. i This indicates that among the predicted labels of the i-th validation sample, only one BP-AdaBoost binary classifier gives y∈{+1}, while the remaining BP-AdaBoost binary classifiers give validation samples with y∈{-1}; the number of unfuzzy solutions is related to the validation sample set Λ. i By comparing the total number of solutions, the proportion of non-fuzzy solutions is determined and denoted as the non-fuzzy solution weight γ. i ;

[0200]

[0201] In the formula, γ i The weights are the unfuzzy solutions of the i-th M-BP-AdaBoost multi-class classification model. Let ψ be the number of nonfuzzy solutions in the i-th round. i Let be the total number of validation samples in the validation sample set of the i-th round;

[0202] Step 4.2: The unfuzzy solution weight vector γ of the K M-BP-AdaBoost models:

[0203] γ=[γ1,…,γ K ]

[0204] In the formula, γ1~γ K This represents the weight of the non-fuzzy solution for each of the K M-BP-AdaBoost models;

[0205] Step 5: Construct a weighted BP-AdaBoost multi-class classification model;

[0206] By combining the multi-classification model M-BP-AdaBoost established in step 3 and the non-fuzzy solution weights of the K M-BP-AdaBoost multi-classification models extracted in step 4, a weighted BP-AdaBoost multi-classification model can be obtained, denoted as Model.

[0207] Model = [γ1, ..., γ K [M1,M2,…M] K ] T

[0208] This completes the establishment of the weighted BP-AdaBoost multi-class classification model; the flowchart of the non-fuzzy solution-based weighted BP-AdaBoost is attached. Figure 4 As shown.

[0209] Step 6: Use the weighted BP-AdaBoost multi-class classification model to predict the test sample set O1;

[0210] Using the weighted BP-AdaBoost multi-class classification model Model, the test sample set O1 is input into the weighted BP-AdaBoost multi-class classification model Model to predict the label category corresponding to each sample in the test sample set O1;

[0211] Step 6.1: Substitute the test sample set O1 into K M-BP-AdaBoost multi-class classification models, and the predicted result S is expressed as:

[0212]

[0213] s KnThis represents the prediction result for the Kth M-BP-AdaBoost test sample set O1;

[0214] Simplified to: S = [s1, s2, ..., s i ,…,s K ] T , where s i =[s i1 s i2 …s in (i = 1, 2, ..., K);

[0215] Step 6.2: Combine the weight vectors γ of the K unfuzzy solutions and the prediction results S of the K M-BP-AdaBoost multi-classification models to obtain the weighted test feature solution. Represented as:

[0216]

[0217] In the formula, γ represents the weight of the non-fuzzy solution of the multi-classifier, and S represents the prediction result of the classification BP-AdaBoost model;

[0218]

[0219] exist In the test sample O1, each test sample O1 has a value under each class label, and the label corresponding to the maximum value is the prediction result of each sample in test sample O1.

[0220] The features in Table 1 above are sorted according to their variance to obtain a feature sequence. Feature sets of different dimensions are selected, preprocessed, and then input into... Figure 4 Based on the flowchart of the non-fuzzy solution-weighted BP-AdaBoost algorithm, the corresponding recognition results are obtained. The recognition rates for the four rotor states are shown in Table 2 below:

[0221] Feature Dimension 1 2 3 4 5 6 7 8 9 10 Average recognition rate (%) 54.2 74.8 84.2 87.1 89.6 91.3 92.3 93.2 93.5 92.8 Feature Dimension 11 12 13 14 15 16 17 18 19 20 Average recognition rate (%) 92.9 94.5 95.4 97.4 96.8 97.1 97.2 97.5 97.4 97.6

[0222] The rotor's recognition rates for ten states are shown in Table 3 below:

[0223] Feature Dimension 1 2 3 4 5 6 7 8 9 10 Average recognition rate (%) 18.2 45.1 57.9 67.3 76.0 82.0 82.4 84.6 85.9 85.5 Feature Dimension 11 12 13 14 15 16 17 18 19 20 Average recognition rate (%) 86.6 87.5 88.4 91.8 92.0 91.6 93.1 94.5 94.5 93.9

[0224] The analysis in the table above shows that:

[0225] In the four-class classification case, the feature dimension of 14 and the feature dimension of 20 are almost the same. The recognition rate is improved by 0.2% by adding 6 feature dimensions, which is not of much value. As the number of classification categories increases, the classification difficulty will increase accordingly, and the required feature dimensions will also increase accordingly. In the ten-class case, the feature dimension of 18 achieves the best results. This result verifies that the feature sequence obtained by variance has good assistance in fault identification.

[0226] The fault identification rate can reach over 97% in the four-class case and over 94% in the ten-class case, verifying the effectiveness of the proposed BP-AdaBoost method based on non-fuzzy solution weighting in identifying rotor fault types and fault severity.

Claims

1. A rotor fault diagnosis method based on weighted BP-AdaBoost, characterized in that: Specifically: Step 1: Extract rotor state parameters and feature selection; use the acceleration sensor to measure the vibration signal of the rotor, combine the features of the rotor vibration signal in the time domain, frequency domain and time-frequency domain to construct the rotor state feature set; and use the variance method to select and sort the features to construct a low-dimensional feature subset; Step 2: Proportionally divide the low-dimensional feature subset obtained in step 1 to obtain the training sample set C1 and the test sample set O1; process the training sample set C1 according to the K-fold cross-validation idea; where (K-1) / K is the training sample set C, 1 / K is the validation sample set V, represented by the matrix Λ; Step 3: Set the number of cycles K, select each row of elements in Λ in turn, and when the cycle reaches the i-th time, i∈[1,K], obtain the training sample set C2 and the verification sample set V1; use the AdaBoost algorithm to obtain the BP-AdaBoost binary classification model; use the one-against-the-rest strategy, i.e., the One Against All mode, to combine multiple BP-AdaBoost binary classification models to build an M-BP-AdaBoost multi-classification model; after K rounds of cycles, obtain K M-BP-AdaBoost multi-classification models; Step 4: Extract the non-fuzzy solution weights of K M-BP-AdaBoost models based on the validation sample set V; Step 5: Use the M-BP-AdaBoost multi-classification model established in step 3 and the non-fuzzy solution weights of the K M-BP-AdaBoost multi-classification models extracted in step 4 to construct a weighted BP-AdaBoost multi-classification model Model; Step 6: Using the weighted BP-AdaBoost multi-classification model Model, input the test sample set O1 into the weighted BP-AdaBoost multi-classification model Model to predict the fault label category corresponding to each sample in the test sample set O1.

2. A rotor fault diagnosis method based on weighted BP-AdaBoost according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1.1: Extract the time domain features of the rotor vibration signal; The time domain features include: Peak, Mean, Root mean square, Kurtosis factor, Clearance factor, Shape factor, Impulse factor and Crest factor. Peak = max(|x|), represented by TF1; It is represented by TF2; It is represented by TF3; It is represented by TF4; It is represented by TF5; It is represented by TF6; It is represented by TF7; It is represented by TF8; Where x represents the vibration signal of the sample, x n represents the point in x; RMS represents the root mean square; N represents the number of sampling points of the vibration signal, represents the mean, σ represents the standard deviation, and max represents the maximum value; Step 1.2: Extract frequency domain features of rotor vibration signal; Frequency domain features include: mean frequency MF, center of gravity frequency FC, root mean square frequency RMFS and standard deviation frequency RVF; It is represented by FF1; It is represented by FF2; It is represented by FF3; It is represented by FF4; In the formula, s represents the amplitude, h represents the frequency; s m is a point in s; h m is a point in h; q represents the number of frequency points; Step 1.3: Extract the time-frequency domain features of the rotor vibration signal; In the time-frequency domain, wavelet transform is used to perform three-layer decomposition to obtain nodes, calculate the energy of each node, and calculate the ratio of each node energy to the total energy of all nodes; that is, energy ratio E = {E1, E2, E3, E4, E5, E6, E7, E8}; the energy ratio is used as a feature in the time-frequency domain; Step 1.4: Construction of rotor state feature set; Combine the features in the time domain, frequency domain and time-frequency domain to obtain the rotor state feature set F; F={FF1,FF2,FF3,FF4,TF1,TF2,TF3,TF4,TF5,TF6,TF7,TF8,E1,E2,E3,E4,E5,E6,E7,E8}; Step 1.5: Construct a low-dimensional feature subset based on variance; The variance of each feature in the rotor state feature set F is calculated as follows: Where N represents the number of samples, y i represents a single feature of an independent sample, represents the mean of the sample characteristics; The rotor state feature set F has M-dimensional features, where M represents the number of features in the feature set F. The variance corresponding to all features is calculated by the above formula (1), which is expressed as: In the formula, represents the variance of the i-th feature; Then the variance σ 2 After normalization, we get for: In the formula represents the normalized variance of the i-th feature; According to the size of the variance value, a feature set with a certain sequence is obtained, the variance threshold is set to select the feature dimension, and a low-dimensional feature subset is constructed.

3. A rotor fault diagnosis method based on weighted BP-AdaBoost according to claim 1, characterized in that: The step 2 is: The training sample set C1 is processed according to the K-fold cross-validation idea, that is, the training sample set C1 is divided into K parts with equal sample numbers, expressed as [Λ1,Λ2,…,Λ i ,…,Λ K-1 ,Λ K ], where (K-1) / K is the training sample set C and 1 / K is the verification sample set V, and is represented by the matrix Λ: In the matrix, the first row to the last row are the K-round training sample set C and the validation sample set V obtained by K-fold cross-validation. The first K-1 columns of each row are the training sample set C: The Kth column of each row is the validation sample set V:

4. A rotor fault diagnosis method based on weighted BP-AdaBoost according to claim 1, characterized in that: The step 3 is specifically as follows: Step 3.1: Preprocess the training sample set C2 obtained after division, extract the rotor fault samples of a certain state in turn, and its label is represented by y∈{+1}, and the labels of the rotor fault samples of the remaining states are all represented by y∈{-1}. The training sample set obtained after preprocessing is denoted as: U = [U1, U2, ..., U i ,…,U M ]; Step 3.2: Using the AdaBoost algorithm, a BP-AdaBoost binary classification model is obtained; using the BP-AdaBoost algorithm to pre-process the training sample set U = [U1, U2, ..., U i ,…,U M ] to obtain several two-class classifiers h=[h1,h2,…,h i ,…,h M ]; Combine multiple binary BP-AdaBoost classifiers to build a multi-classification model; Step 3.3: After K rounds of cycles, K M-BP-AdaBoost multi-classification models can be obtained, denoted as [M1, M2, …M K ].

5. A rotor fault diagnosis method based on weighted BP-AdaBoost according to claim 4, characterized in that: The step 3.2 is specifically as follows: Step 3.2.1: The AdaBoost algorithm improves the classification effect by increasing the weights of samples that were misclassified by the previous weak classifier and reducing the weights of samples that were correctly classified, and adopts a weighted majority voting method; the BP neural network is used as the base learner of AdaBoost to form a BP-AdaBoost binary classification model; Step 3.2.2: Step 3.2.1 builds a binary classification model based on BP-AdaBoost. Step 3.2.1 is repeated M times to obtain a binary classifier combination h = [h1, h2, ..., h i ,…,h M ]; So far, the combination of multiple binary BP-AdaBoost classifiers is completed, which is recorded as M-BP-AdaBoost multi-classification model.

6. A rotor fault diagnosis method based on weighted BP-AdaBoost according to claim 5, characterized in that: The step 3.2.1 is specifically as follows: Step 3.2.1.1: First input a training sample set U i : U i ={(x1,y1),(x2,y2),…,(x i ,y i ),…,(x N ,y N )} Among them, each sample point consists of a sample instance and a sample label; the sample instance x i ∈χ∈R n , represents the feature vector of the i-th sample, χ is the instance space, and the sample label y i ∈Y={-1,+1},y i The label category of the i-th sample, Y is the label space; Step 3.2.1.2: Initialize the training sample set U i The weight distribution of ; each training sample is assigned the same weight: Where ω 1,i represents the initial weight of each training sample, and N represents the number of training samples; Training sample set U i The initial weight vector distribution is: Step 3.2.1.3: Establish the iteration termination condition e t <r, if the termination condition is not met, repeat steps A1-A5 to obtain T base learners, and use T base learners to construct a strong classifier; until the termination condition is met or T>10, then execute step 3.2.2; e t Represents the classification error rate of each round, and r represents the threshold of the training error rate given according to the training requirements; if the termination condition is met, the BP-AdaBoost binary classification model is obtained, and step 3.2.2 is executed.

7. A rotor fault diagnosis method based on weighted BP-AdaBoost according to claim 6, characterized in that: The steps A1-A5 are specifically as follows: Step A1: Based on the above training sample set U i The initial weight distribution of , and the t-th base learner is obtained; G t (x):χ→{-1,+1} In the formula, G t (x) represents the base learner, i.e., BP neural network; Step A2: Calculate G t (x) in the training sample set U i Classification error rate on the initial weight distribution of: Where ω t,i represents the weight of the i-th sample in the t-th round, I(G t (x i )≠y i ) is the indicator function, when the prediction result G t (x i ) and the true result y i When unequal, I(G t (x i )≠y i ) takes the value of 1, otherwise takes the value of 0; Step A3: Calculate G t The coefficient of (x), that is, the weight corresponding to the base learner: Where α t represents the weight of each base learner, e t represents the classification error rate; Step A4: Update training sample U i Weight: D t+1 =(ω t+1,1 Oh, oh t+1,2 ,...,oh t+1,N ) Among them, Z t is the normalization factor, exp is the exponential function with e as the base; Step A5: Combine T base learners to build a strong binary classifier model; Where f(x) represents a linear combination of classifiers, sign represents a sign function, G(x) represents a strong binary classifier, and T is the number of base learners.

8. A rotor fault diagnosis method based on weighted BP-AdaBoost according to claim 1, characterized in that: The step 4 is specifically as follows: Step 4.1: Set the validation sample set V = [Λ K , Λ K-1 ,…,Λ i ,…,Λ1] T In turn, K M-BP-AdaBoost models are introduced, and the non-fuzzy solution weights are extracted according to the classification fuzziness of the verification sample set; for each M-BP-AdaBoost model, a verification sample set Λ is selected i , which means that only one BP-AdaBoost binary classifier gives y∈{+1} in the prediction label of the i-th validation sample, and the other BP-AdaBoost binary classifiers give y∈{-1} validation samples; the number of non-fuzzy solutions is related to the validation sample set Λ i The total number is compared to determine the proportion of non-fuzzy solutions, which is recorded as the non-fuzzy solution weight γ i ; In the formula, γ i is the non-fuzzy solution weight of the i-th M-BP-AdaBoost multi-classification model, is the number of non-fuzzy solutions in the i-th round, ψ i is the total number of validation samples in the validation sample set of round i; Step 4.2: The unfuzzified solution weight vector γ of K M-BP-AdaBoost models: γ=[γ1,…,γ K ] Where γ1~γ K Represents the non-fuzzy solution weight corresponding to each M-BP-AdaBoost model in the K M-BP-AdaBoost models.

9. A rotor fault diagnosis method based on weighted BP-AdaBoost according to claim 1, characterized in that: The step 5 of constructing the weighted BP-AdaBoost multi-classification model is specifically as follows: The multi-classification model M-BP-AdaBoost established in step 3 and the non-fuzzy solution weights of the K M-BP-AdaBoost multi-classification models extracted in step 4 are combined to obtain a weighted BP-AdaBoost multi-classification model, which is expressed as Model. Model=[γ1,…,γ K ][M1,M2,…M K ] T This completes the establishment of the weighted BP-AdaBoost multi-classification model.

10. A rotor fault diagnosis method based on weighted BP-AdaBoost according to claim 1, characterized in that: The step 6 is specifically as follows: Step 6.1: Bring the test sample set O1 into K M-BP-AdaBoost multi-classification models, and the predicted result S is expressed as: s Kn is the prediction result of the Kth M-BP-AdaBoost test sample set O1; Simplified to: S = [s1, s2, ..., s i ,…,s K ] T , where s i =[s i1 s i2 …s in ](i=1, 2, ..., K); Step 6.2: Combine the K non-fuzzy solution weight vectors γ and the K M-BP-AdaBoost multi-classification model prediction results S to obtain the weighted test feature solution It is expressed as: Where γ represents the weight of the non-fuzzy solution of multiple classifiers, and S represents the prediction result of the classification BP-AdaBoost model; exist In , each test sample O1 has a value under each class label, and the label corresponding to the maximum value is the prediction result of each sample in the test sample O1.