Fault diagnosis method of enhanced extreme learning machine based on small sample imbalanced learning
By selecting sensitive features and weighted output errors on small sample imbalance data sets, the enhanced limit learning machine model is constructed, which solves the problem of low accuracy in fault diagnosis under small sample imbalance data in the existing technology, and realizes the failure of rotating machinery Efficient identification.
Patent Information
- Application Number
- CN202311491198.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-09
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2043-11-09
AI Technical Summary
Existing machine learning methods are difficult to accurately identify fault samples under small sample imbalanced data sets, resulting in low accuracy of fault diagnosis and easy missed fault status samples, neglecting the diagnosis of fault categories.
By obtaining the fault feature set of small sample imbalanced data, selecting sensitive features using the maximum correlation and minimum redundancy method, calculating the mutual information value weighted feature matrix, constructing an enhanced limit learning machine model, combining sample weighted output error, the balanced model learning is normal and fault status samples to improve diagnostic performance.
It realizes accurate diagnosis of rotating mechanical faults under small sample imbalance data, improves diagnostic performance, overcomes the instability caused by random weights, and enhances the ability to identify fault categories.
Smart Images

Figure CN117540288B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a fault diagnosis method of an enhanced extreme learning machine based on small sample imbalance learning. Background Art
[0002] As an important part of large-scale equipment, rotating machinery has become increasingly complex in structure and large in scale. Due to long-term continuous operation in harsh environments such as high temperature, high pressure, high speed, and heavy load, faults are inevitable. Usually, rotating machinery operates in normal conditions most of the time, and the condition monitoring system can often collect a large number of normal state samples. However, when a rotating machinery fails, due to the inherent characteristics of enterprise production, the unit is not allowed to continue to operate in a faulty state. In addition, the labeling of monitoring data is time-consuming and laborious, making it difficult to collect an appropriate amount of fault samples, resulting in a scarcity of samples in the faulty state, and a high degree of imbalance in the number of samples in the faulty state. The sample data set presents a typical small sample imbalance characteristic. In recent years, the problem of fault diagnosis under small sample imbalance of rotating machinery has received widespread attention. Fault diagnosis under small sample imbalance refers to the use of a large number of normal state samples and a small number of fault samples to establish a diagnostic model to achieve accurate identification of fault samples. This leads to the following problem: how to improve the diagnostic ability of fault samples when normal state samples are sufficient, fault state samples are insufficient, or even scarce.
[0003] Most existing machine learning methods are designed based on balanced sample sets, and usually can show good recognition performance under balanced sample sets, but the recognition performance tends to decline under small sample imbalanced data sets. When the existing diagnostic model processes small sample imbalanced data sets, it is easy to over-learn the normal state samples and under-learn the fault state samples, resulting in the diagnostic accuracy of the fault state samples being lower than that of the normal state samples. The diagnostic results tend to be biased towards the normal state, resulting in misdiagnosis of normal state samples and missed diagnosis of fault state samples. Consider such a problem. Assuming that the small sample imbalanced data set contains 99 normal category samples and 1 fault category sample, even if the diagnostic model identifies all samples as normal samples, it can still obtain an accuracy of 99%. On the surface, this result is ideal, but in fact, it ignores the diagnosis of the most concerned fault state samples. Considering that the cost of missing the fault state sample is much higher than the cost of misjudging the normal state sample, we should focus on the diagnosis of the fault category samples, that is, we need to identify the fault category to which the small sample belongs as accurately as possible. Summary of the invention
[0004] The present invention achieves the technical effect of improving the fault diagnosis performance under small sample imbalance by providing an enhanced extreme learning machine fault diagnosis method based on small sample imbalance learning.
[0005] The present invention provides a fault diagnosis method of an enhanced extreme learning machine based on small sample imbalance learning, comprising:
[0006] Obtain fault feature sets for small sample unbalanced data;
[0007] By formula Select feature a with the greatest correlation with fault category c from the fault feature set of the small sample unbalanced data i , the feature a i Into the feature space S m-1 ; Where D is the vector of mutual information values between the features and fault category vectors of the fault feature set containing the small sample imbalanced data, |S| is the dimension of the feature space s, and I(a i ; c) is the feature a i and the mutual information value between the vector of the fault category c;
[0008] By formula Obtain m sensitive feature subsets from the fault feature set of the small sample unbalanced data; wherein feature a j is the remaining feature space {sS m-1}, I(a j ; c) is the feature a j and the mutual information value between the vector of the fault category c, I(a i ; a j ) is the feature a i and the feature a j The mutual information value between ;
[0009] Based on the sensitive feature subset, the feature vector X of the sample to be diagnosed is calculated by the function The output of the enhanced extreme learning machine is obtained; wherein m(X) is the feature mapping vector of the sample to be diagnosed, M is the hidden layer node output matrix of the enhanced extreme learning machine, C is a positive coefficient, I is the unit matrix, W s is the weight of the samples in the sensitive feature subset, T is the expected output matrix, N is the number of samples in the sensitive feature subset, and L is the number of hidden layer nodes of the enhanced extreme learning machine;
[0010] By the formula Label(X) = arg max f i (X), i=1, ..., d, to obtain the fault diagnosis label of the sample to be diagnosed.
[0011] Specifically, the method of obtaining a fault feature set of small sample unbalanced data includes:
[0012] Collect vibration signals of machinery under different fault conditions;
[0013] Using a wavelet packet decomposition method to obtain decomposed signals of the vibration signal in different frequency bands;
[0014] Calculate the fault feature based on the vibration signal and the decomposed signal to form a fault feature set;
[0015] The fault feature set of the small sample imbalance data is divided from the fault feature set according to a preset small sample imbalance rate.
[0016] Specifically, the I(a i ; The calculation formula of c) is Among them, p(x, y) represents the feature a i and the joint probability density function of the vector of the fault category c, p(x) represents the feature a i , p(y) represents the marginal probability density function of the vector of fault category c.
[0017] Specifically, the I(a j ; The calculation formula of c) is Among them, p(z, y) represents the feature a j and the joint probability density function of the vector of the fault category c, p(z) represents the feature a j , p(y) represents the marginal probability density function of the vector of fault category c.
[0018] Specifically, the I(a i ; a j ) is calculated as Among them, p(x, z) represents the feature a i and the feature a j The joint probability density function of p(x) represents the feature a i The marginal probability density function of p(z) represents the feature a j The marginal probability density function of .
[0019] Specifically, the calculation formula of m(X) is m(X)=[g( <W F ⊙W1,X>+b1)g( <W F ⊙W2,X>+b2)...g( <W F ⊙W L , X>+b L )]; where W F =(w f1 , w f2 , ..., w fn ) T is the input feature weight, W j=(w j1 , w j2 , ..., W jn ) T is the input weight of the jth hidden layer node of the enhanced extreme learning machine, j=1, ..., L; n is the number of neurons in the input layer of the enhanced extreme learning machine, b j is the jth hidden layer bias of the enhanced extreme learning machine, g(x) is the activation function Sigmoid, W F ⊙W j W F and W j The Hadamard product between <W F ⊙W2, X> represents W F ⊙The inner product between W2 and the feature vector X of the sample to be diagnosed.
[0020] Specifically, the Among them, t i =(t i1 , t i2 , ..., t id ) T ∈R d is the expected output of the i-th input sample, and d is the number of neurons in the output layer of the enhanced extreme learning machine.
[0021] Specifically, the W s =diag(w s1 , w s2 , ..., w sN ), where w si =1 / #(t i ), #(t i ) means it belongs to category t i , i=1,..., the number of samples of d.
[0022] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:
[0023] The enhanced extreme learning machine model based on the simultaneous weighting of fault samples and sensitive features can more accurately realize the fault diagnosis of rotating machinery under the fault feature set of small sample unbalanced data. Specifically, for the fault feature set of small sample unbalanced data obtained, the maximum relevance minimum redundancy method is first used to select important features sensitive to faults, and the mutual information values of each important feature are calculated. The feature weight matrix is constructed based on the mutual information value to achieve the weighting of the random weight of the extreme learning machine. Secondly, according to the distribution information of the unbalanced sample category, the sample weight matrix is constructed to achieve the weighting of the small sample output error. Finally, the optimal output weight is obtained to obtain the enhanced extreme learning machine model, and the fault category of the sample to be diagnosed is output by the enhanced extreme learning machine model to achieve the fault diagnosis of rotating machinery under small sample imbalance. Among them, the feature weighting is to adjust the random weights between the input layer and the hidden layer in the enhanced extreme learning machine model by enhancing the weights of the nodes with strong fault diagnosis capabilities in the input layer and the hidden layer, and reducing the weights of the nodes with weak fault diagnosis capabilities in the input layer and the hidden layer. This can not only improve the diagnostic performance of the model under small sample imbalance, but also overcome the unstable influence of random weights on the diagnostic results. Sample weighting balances the degree to which the model learns normal state samples and fault state samples by weighting the output errors of each small sample, and obtains the optimal output weight matrix by minimizing the weighted output errors of all training samples, further improving the diagnostic performance of the model under small sample imbalance. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 A schematic diagram of a fault diagnosis method of an enhanced extreme learning machine based on small sample imbalanced learning provided by an embodiment of the present invention;
[0025] Figure 2 A model diagram of an enhanced extreme learning machine in an enhanced extreme learning machine fault diagnosis method based on small sample imbalanced learning provided by an embodiment of the present invention;
[0026] Figure 3 A flowchart of a fault diagnosis method for an enhanced extreme learning machine based on small sample imbalanced learning provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0027] The embodiment of the present invention achieves the technical effect of improving the fault diagnosis performance under small sample imbalance by providing an enhanced extreme learning machine fault diagnosis method based on small sample imbalance learning.
[0028] The technical solution in the embodiment of the present invention is to achieve the above technical effects, and the overall idea is as follows:
[0029] The fault data sets collected from industrial sites often have typical small sample imbalance characteristics, which causes the performance of some existing diagnostic models to decrease. In view of the fault diagnosis problem under the small sample imbalance data set of rotating machinery, the embodiment of the present invention adopts the mutual information value matrix of sensitive features to realize the random weighting between the input layer and the hidden layer of the original extreme learning machine, enhances the weight of the nodes in the input layer that have strong fault diagnosis capabilities, and reduces the weight of the nodes in the input layer that have weak fault diagnosis capabilities. At the same time, the sample category matrix is used to realize the weighting of the output error of the training samples, balances the degree of the model learning fault state samples and normal state samples, and by minimizing the cumulative weighted output error of all training samples, finds the optimal output weight matrix, and constructs an enhanced extreme learning machine model (EELM model). The embodiment of the present invention can accurately diagnose the fault state of the unit under the small sample imbalance data set, and can still obtain better diagnostic effects when the small sample imbalance degree is higher.
[0030] The principle of the fault diagnosis method based on small sample imbalance learning provided by the embodiment of the present invention is as follows: Figure 1 As shown, the specific steps are as follows:
[0031] Step 1: Acquisition and division of data sets. Collect vibration signals of rotating machinery under different fault conditions, use wavelet packet decomposition method to obtain decomposed signals in different frequency bands, calculate fault features based on original signals and decomposed signals, and obtain fault feature sets. For all collected samples, determine the small sample unbalanced training set by setting the small sample unbalance rate, and the remaining samples are used as test sets.
[0032] Step 2: Feature weighting. The maximum relevance minimum redundancy method is used to obtain sensitive features under the small sample imbalanced training set, calculate the mutual information value of each sensitive feature, construct the mutual information value matrix, and realize the random weighting between the input layer and the hidden layer of the extreme learning machine.
[0033] Specifically, the mutual information value is used to measure the correlation between fault features and the correlation between fault features and fault category variables. According to the maximum correlation and minimum redundancy criteria, the feature subset that best describes the fault category information is selected. Let {x ij} N×p represents small sample imbalanced data, where x ij is the jth eigenvalue corresponding to the ith sample, N is the number of samples, and p is the number of features. The maximum correlation minimum redundancy method can select a feature subset with a dimension of m (m<p) to maximize the correlation between the features and the fault category variable, while minimizing the correlation between the features.
[0034] Assume a irepresents the i-th feature of small sample imbalanced data, c represents the vector of fault categories, and the maximum correlation criterion is used to obtain features related to the fault category. The criterion is expressed as:
[0035]
[0036] Define the original feature space of small sample imbalanced data as s, |S| is the dimension of feature space s, D is the vector of mutual information values between features and fault category vectors of the fault feature set containing small sample imbalanced data, I(a i ; c) is the i-th feature a in the small sample imbalanced data i The mutual information value between the vector of and fault category c is defined as follows:
[0037]
[0038] Among them, p(x, y) represents feature a i The joint probability density function of the vector of and fault category c, p(x) represents feature a i The marginal probability density function of , p(y) represents the marginal probability density function of the vector of fault category c.
[0039] The essence of the above maximum correlation criterion is to calculate the mutual information value between each feature and the fault category, and arrange the fault features in descending order according to the size of the mutual information value. The maximum correlation criterion can find the fault feature with the greatest correlation with the fault category.
[0040] Considering the redundancy problem of the features selected based on the maximum correlation criterion, the minimum redundancy criterion is used to minimize the redundancy between features. The criterion is expressed as:
[0041]
[0042] Among them, I(a i ; a j ) is the i-th feature a i and the jth feature a j The mutual information value between is defined as follows:
[0043]
[0044] Among them, p(x, z) represents feature a i and feature a j The joint probability density function of p(x) represents feature a i The marginal probability density function of , p(z) represents the feature a j The marginal probability density function of . R is a vector containing the mutual information values between features.
[0045] Based on formula (1) and formula (3), the following formula is used to optimize D and R simultaneously.
[0046]
[0047] Formula (5) is used to transform the maximum correlation and minimum redundancy criterion problem into the problem of maximizing the mutual information difference, and an incremental search method is used to obtain a sensitive feature subset. Specifically, based on the maximum correlation criterion, the fault feature with the greatest correlation with the fault category is found, and other fault features are selected by combining formulas (3)-(5).
[0048] In practice, it is assumed that m-1 features have been selected and the sub-feature space S is constructed. m-1 , then from the remaining feature space {sS m-1 The following criteria should be followed when selecting the mth feature in}:
[0049]
[0050] Among them, a i is the selected feature space S m-1 The feature in, feature a j is the remaining feature space {sS m-1} features. I(a j ; c) is the jth feature a in the small sample imbalanced data j The mutual information value between the vector of and fault category c is defined as follows:
[0051]
[0052] Among them, p(z, y) represents feature a j The joint probability density function of the vector of and fault category c, p(z) represents feature a j The marginal probability density function of , p(y) represents the marginal probability density function of the vector of fault category c.
[0053] Formula (7) can maximize the correlation between features and categorical variables, and minimize the correlation between features. Formula (7) can finally obtain m sensitive feature subsets from small sample imbalanced data as input data for the subsequent enhanced extreme learning machine model.
[0054] Step 3: Sample weighting. Based on feature weighting, a cost-sensitive loss function is constructed to weight the output errors of each small sample, balance the degree to which the model learns normal category samples and fault category samples, and minimize the weighted cumulative output error of all small samples.
[0055] Specifically, based on the m sensitive features obtained from the small sample imbalanced data using the maximum correlation minimum redundancy method, assuming that there are N arbitrary samples (X i , t i ), for a single hidden layer feedforward neural network with L hidden layer nodes, the output can be expressed as Mβ=T, where M is the hidden layer node output matrix, β is the output weight matrix, and T is the expected output matrix. The specific formula of the EELM model is as follows:
[0056]
[0057]
[0058] Among them, X i is the feature vector of the i-th input sample, m(X i ) is the feature mapping vector of the i-th input sample, which maps the original spatial data into the hidden layer spatial data of dimension L, so it is also called the hidden layer output vector. F =(w f1 , w f2 , ..., w fn ) T is the input feature weight, W j =(w j1 , w j2 , ..., w jn ) T To enhance the input weight of the jth hidden layer node of the extreme learning machine, j = 1, ..., L; β j =(β j1 , β j2 , ..., β jd ) T To enhance the output weight between the jth hidden layer node and the output layer node of the extreme learning machine, b j To enhance the j-th hidden layer bias of the extreme learning machine, t i =(t i1 , t i2 , ..., t id ) T ∈R d is the expected output of the i-th input sample, n is the number of neurons in the input layer of the enhanced extreme learning machine, d is the number of neurons in the output layer of the enhanced extreme learning machine, and g(x) is the activation function Sigmoid. F ⊙W j W F and W j The Hadamard product between <W F ⊙W j , X i > indicates W F ⊙Wj and X i The inner product between .
[0059]
[0060] In order to solve Mβ=T, it is transformed into minimizing the sum of the weighted output errors of all samples, that is, minimizing ||Mβ-T|| 2 and ||β|| 2 To solve, the specific optimization objective function is as follows:
[0061]
[0062] The positive coefficient c is introduced to improve the generalization ability of the model. S =diag(w s1 , w s2 , ..., w sN ) is the sample weight, where w si =1 / #(t i ), #(t i ) means it belongs to category t i , i = 1, ..., the number of samples of d. i =[ξ i,1 ,ξ i,2 , ..., ξ i,d ] T represents the i-th input sample X i The error between the actual output and the expected output.
[0063] According to the Karush-Kuhn-Tucker (KKT) theory, formula (11) can be transformed into a dual optimization problem, which is as follows:
[0064]
[0065] Among them, the Lagrange multiplier γ i is the input sample feature vector X in the linear combination i For formula (12), we can respectively obtain the constant factors of variables β and ξ. i , γ i The partial derivative of and set it equal to zero, that is, require
[0066]
[0067] By combining the three equations in formula (13), two different β solutions are derived. The specific form of the solution is shown in formula (14):
[0068]
[0069] Where, when N<L, I represents the identity matrix with dimension N×N. When N>L, I represents the identity matrix with dimension L×L. Based on the calculation of m(X i ) method and formula (14), for a given feature vector X of a sample to be diagnosed, its output function is f(X) = sign m(X)β:
[0070]
[0071] The output vector of the EELM model is f(X)=[f1(X),...,f d (X)], the fault diagnosis label of the sample to be diagnosed is obtained according to the following formula.
[0072] Label(X)=arg max f i (X), i=1,...,d (16)
[0073] Step 4: Construction of EELM model, such as Figure 2 As shown in the figure, the relevant parameters of the EELM diagnostic model are set, including the number of hidden layer nodes, regularization parameters, the number of model repetitions (the number of random training sample extractions), etc. The model is trained using a small sample unbalanced training set to obtain the model's optimal output weight matrix. Based on the trained EELM model, the test set is used to verify the effectiveness of the model.
[0074] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0075] See also Figure 3 The embodiment of the present invention provides an enhanced extreme learning machine fault diagnosis method based on small sample imbalance learning, including:
[0076] Step S110: obtaining a fault feature set of small sample unbalanced data;
[0077] This step is specifically described to obtain the fault feature set of small sample unbalanced data, including:
[0078] Collect vibration signals of machinery under different fault conditions;
[0079] The decomposition signals of vibration signals in different frequency bands are obtained by using wavelet packet decomposition method;
[0080] Fault features are calculated based on the vibration signal and the decomposed signal to form a fault feature set;
[0081] A fault feature set of small sample imbalanced data is divided from the fault feature set according to a preset small sample imbalance rate as a training data set for the enhanced extreme learning machine model.
[0082] Step S120: By formula Select feature a with the greatest correlation with fault category c from the fault feature set of small sample imbalanced data i , feature a i Into the feature space S m-1 ; Where D is the vector of mutual information values between the features and fault category vectors of the fault feature set containing small sample imbalanced data, |S| is the dimension of the feature space s, and I(a i ; c) is feature a i and the mutual information value between the vector of fault category c; specifically, I(a i ; The calculation formula of c) is Among them, p(x, y) represents feature a i The joint probability density function of the vector of and fault category c, p(x) represents feature a i The marginal probability density function of , p(y) represents the marginal probability density function of the vector of fault category c.
[0083] Step S130: By formula Obtain m sensitive feature subsets from the fault feature set of small sample imbalanced data; among them, feature a j is the remaining feature space {sS m-1}, I(a j ; c) is feature a j The mutual information value between the vector of fault category c, I(a i ; a j ) is feature a i and feature a j The mutual information value between j ; The calculation formula of c) is Among them, p(z, y) represents feature a j The joint probability density function of the vector of and fault category c, p(z) represents feature a j The marginal probability density function of p(y) represents the marginal probability density function of the vector of fault category c. i ; a j ) is calculated as Among them, p(x, z) represents feature a i and feature a j The joint probability density function of p(x) represents feature a i The marginal probability density function of , p(z) represents the feature a j The marginal probability density function of .
[0084] Step S140: Calculate the feature vector X of the sample to be diagnosed based on the sensitive feature subset, through the function The output of the enhanced extreme learning machine is obtained; where m(X) is the feature mapping vector of the sample to be diagnosed, M is the hidden layer node output matrix of the enhanced extreme learning machine, C is a positive coefficient, I is the unit matrix, and W s is the weight of the sample in the sensitive feature subset, T is the expected output matrix, N is the number of samples in the sensitive feature subset, and L is the number of hidden layer nodes of the enhanced extreme learning machine; specifically, the calculation formula of m(X) is m(X)=[g( <W F ⊙W1,X>+b1)g( <W F ⊙W2,X>+b2)...g( <W F ⊙W L , X>+b L )]; where W F =(w f1 , w f2 , ..., w fn ) T is the input feature weight, W j =(w j1 , w j2 , ..., w jn ) T is the input weight of the jth hidden layer node of the enhanced extreme learning machine, j = 1, ..., L; n is the number of neurons in the input layer of the enhanced extreme learning machine, b j To enhance the j-th hidden layer bias of the extreme learning machine, g(x) is the activation function Sigmoid, W F ⊙W j W F and W j The Hadamard product between <W F ⊙W2, X> represents W F ⊙The inner product between W2 and the feature vector X of the sample to be diagnosed. Among them, t i =(t i1 , t i2 , ..., t id ) T ∈R d is the expected output of the i-th input sample, and d is the number of neurons in the output layer of the enhanced extreme learning machine. S =diag(w s1 , w s2 , ..., w sN ), where w si =1 / #(t i ), #(t i ) means it belongs to category t i, i=1,..., the number of samples of d.
[0085] Step S150: Use the formula Label(X)=arg max f i (X), i=1, ..., d, and obtain the fault diagnosis label of the sample to be diagnosed.
[0086] In order to verify the superiority of the EELM model provided by the embodiment of the present invention, the embodiment of the present invention compares and analyzes the K-Nearest Neighbor Algorithm (KNN), Naive Bayesian (NB) and Support Vector Machine (SVM). For fair comparison, all comparison models are repeated 50 times under different numbers of features. The training set used for each model run is randomly selected from the original data set. Finally, the average value of the 50 diagnosis results is taken as the final diagnosis result. The diagnosis results are shown in Table 1.
[0087] As can be seen from Table 1, the EELM model provided by the embodiment of the present invention has the best diagnostic results in most small sample imbalanced training sets. However, the diagnostic effects of the KNN model and the SVM model in all small sample imbalanced training sets are poor. For the NB model, the diagnostic effect is good in a few small sample imbalanced training sets, but the diagnostic effect is poor in most small sample imbalanced training sets.
[0088] Table 1 Diagnostic results of different models under 10 groups of small sample imbalanced training sets (%)
[0089]
[0090]
[0091] Based on the analysis of the above experimental results, the EELM model provided by the embodiment of the present invention can effectively realize fault diagnosis under small sample imbalance.
[0092] The embodiment of the present invention integrates the ideas of sensitive feature weighting and sample weighting into the original extreme learning machine, improves the random weights between the input layer nodes and the hidden layer nodes in the original extreme learning machine, learns small sample information from unbalanced sample data, and combines the cost-sensitive loss function to pay more attention to the misclassification loss of small samples, thereby improving the diagnostic accuracy of the model under small sample imbalance. Experimental results show that the EELM model provided by the embodiment of the present invention is very suitable for fault diagnosis problems under small sample imbalance, especially when the degree of small sample imbalance is higher, the EELM model can still obtain better diagnostic effects.
[0093] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0094] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0095] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0097] The parts not described in detail in the embodiments of the present invention are all known technologies to those skilled in the art. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A fault diagnosis method based on enhanced extreme learning machine with small sample imbalanced learning, characterized in that: include: Obtain fault feature sets for small sample unbalanced data; By formula Select feature a with the greatest correlation with fault category c from the fault feature set of the small sample unbalanced data i , the feature a i Into the feature space S m-1 ; Where D is the vector of mutual information values between the features and fault category vectors of the fault feature set containing the small sample imbalanced data, |S| is the dimension of the feature space S, and I(a i ; c) is the feature a i and the mutual information value between the vector of the fault category c; By formula Obtain m sensitive feature subsets from the fault feature set of the small sample unbalanced data; wherein feature a j is the remaining feature space {SS m-1 }, I(a j ; c) is the feature a j and the mutual information value between the vector of the fault category c, I(a i ; a j ) is the feature a i and the feature a j The mutual information value between ; Based on the sensitive feature subset, the feature vector X of the sample to be diagnosed is calculated by the function The output of the enhanced extreme learning machine is obtained; wherein m(X) is the feature mapping vector of the sample to be diagnosed, M is the hidden layer node output matrix of the enhanced extreme learning machine, C is a positive coefficient, I is the unit matrix, W s is the weight of the samples in the sensitive feature subset, T is the expected output matrix, N is the number of samples in the sensitive feature subset, and L is the number of hidden layer nodes of the enhanced extreme learning machine; By the formula Label(X)=argmaxf i (X), i=1,...,d, obtain the fault diagnosis label of the sample to be diagnosed; The calculation formula of m(X) is m(x)=[g( <W F ⊙W1,X>+b1)g( <W F ⊙W2,X>+b2)...g( <W F ⊙W L , X>+b L )]; where W F =(w f1 , w f2 , ..., w fn ) T is the input feature weight, W j =(w j1 , w j2 , ..., w jn ) T is the input weight of the jth hidden layer node of the enhanced extreme learning machine, j=1, ..., L; n is the number of neurons in the input layer of the enhanced extreme learning machine, b j is the jth hidden layer bias of the enhanced extreme learning machine, g(x) is the activation function Sigmoid, W F ⊙W j W F and W j The Hadamard product between <W F ⊙W2, X> represents W F ⊙The inner product between W2 and the feature vector X of the sample to be diagnosed; Specifically, the mutual information value of each sensitive feature is calculated, a mutual information value matrix is constructed, and random weighting between the input layer and the hidden layer of the extreme learning machine is realized.
2. The fault diagnosis method of enhanced extreme learning machine based on small sample imbalanced learning according to claim 1, characterized in that: The method of obtaining a fault feature set of small sample unbalanced data includes: Collect vibration signals of machinery under different fault conditions; Using a wavelet packet decomposition method to obtain decomposed signals of the vibration signal in different frequency bands; Calculate the fault feature based on the vibration signal and the decomposed signal to form a fault feature set; The fault feature set of the small sample imbalance data is divided from the fault feature set according to a preset small sample imbalance rate.
3. The fault diagnosis method of enhanced extreme learning machine based on small sample imbalance learning according to claim 1, characterized in that: The I(a i ; The calculation formula of c) is Among them, p(x,y) represents the feature a i and the joint probability density function of the vector of the fault category c, p(x) represents the feature a i , p(y) represents the marginal probability density function of the vector of fault category c.
4. The fault diagnosis method of enhanced extreme learning machine based on small sample imbalanced learning according to claim 1, characterized in that: The I(a j ; The calculation formula of c) is Among them, p(z, y) represents the feature a j and the joint probability density function of the vector of the fault category c, p(z) represents the feature a j , p(y) represents the marginal probability density function of the vector of fault category c.
5. The fault diagnosis method of enhanced extreme learning machine based on small sample imbalance learning according to claim 1, characterized in that: The I(a i ; a j ) is calculated as Among them, p(x,z) represents the feature a i and the feature a j The joint probability density function of p(x) represents the feature a i The marginal probability density function of p(z) represents the feature a j The marginal probability density function of .
6. The fault diagnosis method of enhanced extreme learning machine based on small sample imbalance learning according to claim 1, characterized in that: Said Among them, t i =(t i1 ,t i2 ,...,t id ) T ∈R d is the expected output of the i-th input sample, and d is the number of neurons in the output layer of the enhanced extreme learning machine.
7. The fault diagnosis method of enhanced extreme learning machine based on small sample imbalance learning according to claim 1, characterized in that: The W S =diag(w s1 ,w s2 ,...,w sN ), where w si =1 / #(t i ), #(t i ) means it belongs to category t i ,i=1,...,d is the number of samples.
