Rotary machinery cross-domain fault diagnosis method based on enhanced heterogeneous domain self-adaption

Abstract

The invention provides a rotating machine cross-domain fault diagnosis method based on enhanced heterogeneous domain self-adaption, and the method comprises the following steps: firstly, obtaining an IMF component through employing a segmentation type self-correlation noise reduction method NRC and inherent time scale decomposition ITD; and selecting an energy value index to carry out sample refining on the effective layer to finally improve the anti-noise capability and the characterization capability of the sample. An enhanced cross-domain projection ECDP algorithm is provided by adopting a semi-supervised heterogeneous domain adaptive method, and the method projects a source domain sample and a target domain sample into the same subspace through subspace solution, so that the intra-class distance of the samples in the subspace is minimized, and the inter-class distance of the samples in the subspace is maximized. And finally, the classification precision of the model is improved by using a classification iteration model, and accurate diagnosis under the cross-domain task of the rotating machinery is realized. According to the method, a sample intrinsic expression mode with high robustness can be provided, a new optimization solution is provided for the heterogeneous migration method with classification performance as guidance, and the method can be extended to fault diagnosis of heterogeneous signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating machinery fault diagnosis, and in particular relates to a classification iterative model of an enhanced-heterogeneous domain adaptive algorithm. Background Art

[0002] In the modern equipment manufacturing industry, rotating machinery is developing towards heavy load, automation and intelligence. Its structure is more precise and its function is more complex, which has become an indispensable part of modern equipment. Therefore, it is of great significance to detect the state and diagnose the faults of the key transmission components of the rotating machinery system and give maintenance strategies in time. Timely diagnosis of the operating status of rotating machinery mainly involves the collection of vibration data and the analysis and calculation of data. However, in actual industrial production, the working conditions of rotating machinery are more complex and the operating conditions are more severe, which will cause the vibration signals of rotating machinery to have the characteristics of amplitude change, pulsation impact interval, sampling phase instability and signal noise pollution. These messy distributed signals will cause problems such as unbalanced sample distribution, small number of target samples and sample heterogeneity, resulting in poor classification accuracy. However, most fault diagnosis methods under variable working conditions are based on the idea of ​​isomorphic transfer and assume that the source domain and the target domain are in the same label space or the same feature dimension. If faced with the migration of different fault category spaces, migration across acquisition channels and migration across physical signals, ordinary domain adaptation methods cannot effectively correct the deviation of signal distribution. In addition, most heterogeneous domain adaptation methods align data distribution by solving the projected subspace, but fail to consider the impact of the scattered distribution of samples in the subspace on the classification performance, resulting in low reliability of the classification results. Therefore, in the subspace solution process, the key to improving the classification accuracy of the diagnostic model lies in how to optimize the samples in the subspace to minimize the intra-class distance and maximize the inter-class distance. In addition, how to make the intrinsic expression of the sample more robust in a strong noise background is also a problem to be solved. Summary of the invention

[0003] In view of the above technical problems, the present invention provides a rotating machinery variable cross-domain fault diagnosis model based on enhanced heterogeneous domain adaptation. First, the segmented autocorrelation noise reduction method NRC and the intrinsic time scale decomposition ITD are used to obtain the rotation component IMF, and the energy value index is selected to refine the samples of the effective layer to finally improve the noise resistance and characterization ability of the samples. In view of the problems such as data distribution drift and sample category imbalance under variable working conditions, an enhanced cross-domain projection ECDP algorithm is proposed by using a semi-supervised heterogeneous domain adaptation method. This method projects the source domain and target domain samples into the same subspace through subspace solution, and optimizes some edge confusion samples in the subspace so that the samples can achieve the minimum intra-class distance and the maximum inter-class distance in the subspace. Finally, the classification iteration model is used to improve the classification accuracy of the model and realize accurate diagnosis under the cross-domain task of rotating machinery. This method can not only provide a sample intrinsic expression method with strong robustness, but also provide a new optimization solution for heterogeneous migration methods guided by classification performance, and can be extended to the fault diagnosis of heterogeneous signals.

[0004] The present invention achieves the above technical objectives through the following technical means.

[0005] A rotating machinery cross-domain fault diagnosis method based on enhanced heterogeneous domain adaptation comprises the following steps:

[0006] Step S1, sample denoising based on the noise-resistant correlation method (NRC): the collected vibration signal is separated into a periodic component and a residual component, and the average value is taken to construct a new signal to be compared with the previous signal to reduce the influence of noise, and then a correlation function between the original signal and the new signal is defined to enhance the periodic information of the sample under a strong noise background;

[0007] Step S2, intrinsic time scale decomposition (ITD) decomposition and index construction: the denoised samples are subjected to ITD decomposition to obtain a number of intrinsic rotation components (PRCs) with frequencies ranging from high frequency to low frequency, and the component signals are screened by combining weight index screening criteria; in order to reduce the computational load of the model, the energy values ​​of the selected components are extracted as indicators to reconstruct the samples;

[0008] Step S3: Construct the intra-class sample adjacency matrix: First, according to the source domain sample and a small part of the target domain sample labels, construct the intra-class adjacency matrix W based on the heat kernel function ij ; Secondly, construct the sample similarity weight matrix βc ij As a regularization term, it optimizes the intra-class distance of samples in the subspace;

[0009] Step S4, constructing the sample class adjacency matrix: First, set the parameter k to construct the nearest neighbor criterion of heterogeneous samples, calculate the Euclidean distance between heterogeneous samples, and use the nearest neighbor criterion of heterogeneous samples to determine whether the sample belongs to the boundary edge sample, thereby constructing the inter-class adjacent sample matrix D ij ;

[0010] Step S5, constructing an enhanced heterogeneous domain adaptive model: first, constructing an objective function with the purpose of minimizing the intra-class distance and maximizing the inter-class distance, respectively calculating partial derivatives of the source domain projection matrix and the target domain projection matrix, thereby converting them into a generalized eigenvalue solving problem, and finally calculating the projection matrices of the two domains, and projecting the samples into the same subspace for subsequent training;

[0011] Step S6, constructing a classification iteration model: multiplying the sample to be detected by the target domain projection matrix and performing the first classification, adding pseudo labels to the training samples according to the classification results, and repeating the above operations to achieve the best classification effect of the model;

[0012] Furthermore, the step S1 defines the composition of the collected vibration signal as a periodic component and a residual component, and defines the function expression of the new signal as follows:

[0013] y(t)=x(t)+∈(t)

[0014]

[0015] Among them, x(t) is a bounded periodic signal whose period is unknown and hidden in the original signal y(t); ∈(t) is independent and follows a normal distribution. The residual term of T (t) is a new periodic function, which is obtained by accumulating and averaging y(t) at time intervals of T; To perform floor operation, is the number of segments after the signal is separated; then based on the new periodic function f T (t), and construct a new autocorrelation function C(T) with it and y(t):

[0016]

[0017] According to the above steps, the periodic component S(T) in the strong noise background signal is extracted, where is the residual component, and σ is the residual parameter.

[0018] Furthermore, in order to better express the sample intrinsically in step S2, the periodic component S(T) is subjected to intrinsic time scale decomposition ITD, and the mathematical process is as follows:

[0019] S(T)=LS(T)+(1-L)S(T)=Lt +H t

[0020] Among them, L is the new operator defined, L t represents the extracted baseline signal, H t is the rotation component; then the baseline signal L t As the original signal, it is decomposed until the decomposed signal is less than the set value. The final decomposition result can be expressed as:

[0021]

[0022] Among them, H i (i=1, 2, ..., p) represents the i-th rotation component, L p represents the remaining baseline component after p decompositions, and k is the kth component. In order to reduce the redundant information of the sample, the energy value of the i rotation components before each decomposition is selected as the expression of the sample. The process of expressing source domain and target domain samples is as follows:

[0023]

[0024]

[0025] Where n is the number of samples, P i is the energy value of the i-th layer.

[0026] Further, in step S3, the source domain sample set is The target domain sample set is The adjacency matrix W of samples within a class ij for:

[0027] W ij =w ij +βc ij

[0028]

[0029]

[0030] where x i 、x j is the i-th or j-th sample from the source domain or the target domain, is the mean value corresponding to the sample, y i ,y j is the label corresponding to the sample; w ij The calculation method is the heat kernel function, which is based on the similarities and differences of label information. If they are of the same category, the weight is 1, and if they are of different categories, the weight is 0. ijis the similarity weight between samples, β is the regularization parameter, and the final sample class adjacency matrix W ij is the weighted sum of the two terms; in step S4, first set the point neighbor parameter k, the inter-class adjacency matrix D ij The composition is:

[0031]

[0032] Where N K (x j ) Sample x j A sample set of nearest neighbors based on Euclidean distance.

[0033] Furthermore, in step S5, in order to minimize the intra-class discreteness of samples in the subspace, based on the intra-class sample adjacency matrix W ij And the objective function constructed by minimizing the two-norm is as follows:

[0034]

[0035] Among them, P s , P t Represent the source domain projection matrix and the target domain projection matrix respectively, It is represented as the transpose of the source domain projection matrix and the target domain projection matrix; n s and n t is the number of samples in the source domain and the target domain; W s , W t Represents the sample intra-class adjacency matrix of the source domain and the target domain, W c is the intra-class adjacency matrix of cross-domain samples; in order to maximize the distance between classes of samples in the subspace, based on the inter-class sample adjacency matrix D ij And the objective function constructed by maximizing the two-norm is as follows:

[0036]

[0037] Where D s and D t Represent the inter-class adjacency matrix of the source domain and the target domain respectively; the above two objective functions are integrated into one calculation process, and the formula is as follows:

[0038]

[0039] Expanding and simplifying the bi-norm in the above objective function can obtain the following objective function:

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] Where X s and X t They are the sample set matrices of the source domain and the target domain respectively; is a diagonal matrix whose corresponding values ​​are the sums of fixed columns, is a diagonal matrix whose corresponding values ​​are the sums of fixed rows; and The diagonal matrices representing the sum of rows and columns across the domain respectively; and W s and W t The corresponding Laplace matrix; and D s and D t The corresponding Laplace matrix;

[0046] In order to simplify the objective function, we introduce the coefficient factors S c , and The equation is converted into the form of matrix trace:

[0047]

[0048]

[0049]

[0050]

[0051]

[0052]

[0053] For the objective function with respect to P s , P t Calculate the partial derivatives respectively and set the results to 0. Then the optimization problem of the final objective function can be transformed into solving the eigenvector corresponding to the generalized eigenvalue, where the optimal projection matrix P = (P s , P t ) can be composed of the eigenvectors corresponding to the d largest eigenvalues:

[0054] AP=(B+βI)PΛ

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] Where A and B are the coefficient matrices obtained after partial derivatives, and Λ is the regularization parameter term.

[0061] Furthermore, in step S6, the source domain and target domain samples are projected into the same subspace by the required projection matrix. Since the samples to be detected and the target domain samples obey the same distribution, the samples to be detected can be projected into the same subspace by the target domain projection matrix:

[0062]

[0063]

[0064]

[0065] in and They represent the original space source domain samples, target domain samples and samples to be detected respectively; and Respectively represent the source domain, target domain and the data set of the sample to be detected in the subspace. The training to be tested is labeled with pseudo labels:

[0066]

[0067] Among them, y * The pseudo labels given to the samples, Represents the mean of each label category sample, labels the sample to be tested and the sample with the closest category mean with corresponding pseudo labels, and adds the pseudo-labeled samples to the training set to participate in iterative training.

[0068] A method for realizing the cross-domain fault diagnosis model of rotating machinery based on enhanced heterogeneous domain adaptation, comprising a highly robust feature extraction method, construction of two adjacency matrices, a heterogeneous domain adaptation model, and a classification iteration module;

[0069] The highly robust feature extraction method performs noise reduction processing on samples under different working conditions through NRC, and then makes the sample features more obvious through time-intrinsic scale decomposition and indicator construction.

[0070] The construction of the two adjacency matrices refines the original manifold space of the samples, optimizes the intra-class manifold weights using the cosine similarity function, and refines the edge singular sample manifold using the close neighbor criterion.

[0071] The heterogeneous domain adaptive model projects cross-domain data into a subspace that obeys the same distribution, and optimizes the intra-class distance and inter-class distance of samples in the subspace to make the samples more separable.

[0072] The pseudo-label classification iterative model utilizes the iterative integration of training data and pseudo-labels to make the model training more complete, and utilizes the successfully trained model to complete the rotating machinery fault diagnosis under variable working conditions.

[0073] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0074] (1) In view of the influence of strong noise accompanying sample collection, the periodic components of the samples are enhanced by NRC, and the intrinsic expression of the samples is made more robust by combining ITD decomposition.

[0075] (2) Based on the heterogeneous problems existing in actual rotating machinery fault diagnosis, the present invention constructs a heterogeneous domain adaptive model based on the ECDP algorithm to correct the cross-domain distribution deviation and optimize the intra-class distance and inter-class distance of samples in the subspace, so that the samples can achieve the best clustering form in the subspace.

[0076] (3) Considering the underfitting problem of samples in a limited number of classifications, a pseudo-label classification iterative model is constructed so that the samples can be fully trained in the model and ultimately complete the accurate diagnosis of cross-domain tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 is a flow chart of feature extraction NRC-ITD of the present invention;

[0078] Figure 2 is the overall flow chart of the present invention; (a) feature extraction; (b) manifold space construction and projection matrix calculation; (c) pseudo-label iteration model;

[0079] Figure 3 It is a flowchart of the construction of manifold space;

[0080] Figure 4 is the prediction result for bearing 1-1, where Figure 4 (a) is the signal before and after noise reduction of the inner circle fault in the source domain. Figure 4 (b) is the signal before and after denoising the inner circle fault in the target domain;

[0081] Figure 5 It is a scatter plot of the distribution of samples of migration tasks 1-6 in the subspace;

[0082] Figure 6 This is the classification result of the proposed algorithm after five iterations of migration tasks 1-6: DETAILED DESCRIPTION

[0083] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments described below with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be understood as limiting the present invention.

[0084] In order to verify the effectiveness and practicality of the present invention, the motor defect dataset provided by Southeast University is used to conduct verification using different fault categories, working conditions, acquisition channels, and physical signals.

[0085] Figure 2 The figure shows the process of the rotating machinery variable operating condition fault diagnosis model based on enhanced heterogeneous domain adaptation according to the present invention. The method comprises the following steps:

[0086] Step S1, noise reduction processing based on NRC samples: the collected vibration signal is separated into a periodic component and a residual component, and the average value is taken to construct a new signal to be compared with the previous signal to reduce the influence of noise, and then a correlation function between the original signal and the new signal is defined to enhance the periodic information of the sample under a strong noise background;

[0087] Step S2, ITD decomposition and index construction: The denoised samples are subjected to ITD decomposition to obtain several rotational PRC components with frequencies ranging from high to low. The component signals are screened by combining weight index screening criteria. In order to reduce the computational load of the model, the energy values ​​of the selected components are extracted as indicators to reconstruct the samples.

[0088] Step S3: Construct the intra-class sample adjacency matrix: First, based on the source sample and a small part of the target domain sample labels, construct the intra-class adjacency matrix W based on the heat kernel function ij . Secondly, construct the sample similarity weight matrix βc ij As a regularization term, it optimizes the intra-class distance of samples in the subspace;

[0089] Step S4, constructing the sample class adjacency matrix: First, set the parameter k to construct the nearest neighbor criterion of heterogeneous samples, calculate the Euclidean distance between heterogeneous samples, and use the nearest neighbor criterion of heterogeneous samples to determine whether the sample belongs to the boundary edge sample, thereby constructing the inter-class adjacent sample matrix D ij .

[0090] Step S5, construct an enhanced heterogeneous domain adaptive model: first, construct an objective function with the purpose of minimizing the intra-class distance and maximizing the inter-class distance, respectively obtain partial derivatives of the source projection matrix and the target domain projection matrix, and then convert it into a generalized eigenvalue solution problem, and finally obtain the projection matrices of the two domains, and project the samples into the same subspace for subsequent training.

[0091] Step S6, constructing a classification iteration model: multiplying the sample to be detected by the target domain projection matrix and performing the first classification, adding a pseudo label to the training sample according to the classification result, and repeating the above operation to make the model achieve the best classification effect.

[0092] According to this embodiment, preferably, in step S1, the motor stator current data and vibration data are used as two different domains to construct a fault diagnosis model including cross-channel fault diagnosis and cross-physical signal fault diagnosis. The working conditions and acquisition channels are shown in Table 2. According to different working conditions, different acquisition channels, and different physical signals, six migration tasks are specified, as shown in Table 1:

[0093] Table 1 Migration task parameters

[0094]

[0095]

[0096] Table 2 Working conditions

[0097]

[0098] In the above scheme, step S1 defines the composition of the collected vibration signal as a periodic component and a residual component, and defines a new signal whose function expression is as follows:

[0099] y(t)=x(t)+∈(t)

[0100]

[0101] Where x(t) is a bounded periodic signal whose period is unknown and hidden in the original signal y(t). ∈(t) is independent and normally distributed The residual term of T (t) is a new periodic function defined by accumulating and averaging y(t) at time intervals of T. To perform floor operation, The number of segments after the signal is separated. Then based on the new periodic function f T (t), and construct a new autocorrelation function C(T) with it and y(t):

[0102]

[0103] According to the above steps, the periodic component S(T) in the strong noise background signal is extracted, where is the residual component

[0104] In this example, for migration task 1, the original vibration signal and the noise-reduced vibration signal of the inner race fault are compared. Figure 4 As shown, it can be seen that the periodicity of the impact position is significantly improved.

[0105] In the above scheme, in order to better express the sample intrinsically, the periodic component S(T) is subjected to intrinsic time scale decomposition ITD in step S2, and the mathematical process is as follows:

[0106]

[0107] In this example, according to the screening criteria, the first four layers of itd components are extracted from the denoised data, and the energy value is calculated for each layer. Therefore, in the subsequent classification task, the dimension of the sample is 4.

[0108] In the above scheme, in step S3, the source domain sample set is The target domain sample set is The adjacency matrix W of samples within a class ij for:

[0109] W ij =w ij +βc ij

[0110]

[0111]

[0112] Among them, w ij The calculation method is the heat kernel function, and its principle number is based on the label information. If the same category has a weight of one heterogeneous weight, c ij is the similarity weight between samples, β is the regularization parameter, and the final sample class adjacency matrix W ij is the weighted sum of the two items.

[0113] In the above scheme, in step S4, firstly, the point neighbor parameter k and the inter-class adjacency matrix D are set. ij The composition is:

[0114]

[0115] Where N K (x j ) Sample x j The nearest neighbor sample set based on Euclidean distance. The final adjacency matrix W ij and Dij The calculation process is as follows Figure 3 shown.

[0116] In the above scheme, in step S5, in order to minimize the intra-class discreteness of samples in the subspace, based on the intra-class sample adjacency matrix W ij And the objective function constructed by minimizing the two-norm is as follows:

[0117]

[0118] Among them, P s , P t Represent the source domain projection matrix and the target domain projection matrix respectively, It is represented as the transpose of the source domain projection matrix and the target domain projection matrix; n s and n t is the number of samples in the source domain and the target domain; W s , W t Represents the sample intra-class adjacency matrix of the source domain and the target domain, W c is the intra-class adjacency matrix of cross-domain samples; in order to maximize the distance between classes of samples in the subspace, based on the inter-class sample adjacency matrix D ij And the objective function constructed by maximizing the two-norm is as follows:

[0119]

[0120] Where D s and D t Represent the inter-class adjacency matrix of the source domain and the target domain respectively; the above two objective functions are integrated into one calculation process, and the formula is as follows:

[0121]

[0122] Expanding and simplifying the bi-norm in the above objective function can obtain the following objective function:

[0123]

[0124]

[0125]

[0126]

[0127]

[0128] Where X s and X t They are the sample set matrices of the source domain and the target domain respectively; is a diagonal matrix whose corresponding values ​​are the sums of fixed columns, is a diagonal matrix whose corresponding values ​​are the sums of fixed rows; and The diagonal matrices representing the sum of rows and columns across the domain respectively; and W s and W t The corresponding Laplace matrix; and D s and D t The corresponding Laplace matrix;

[0129] In order to simplify the objective function, we introduce the coefficient factors S c , and The equation is converted into the form of matrix trace:

[0130]

[0131]

[0132]

[0133]

[0134]

[0135]

[0136] For the objective function with respect to P s , P t Calculate the partial derivatives respectively and set the results to 0. Then the optimization problem of the final objective function can be transformed into solving the eigenvector corresponding to the generalized eigenvalue, where the optimal projection matrix P = (P s , P t ) can be composed of the eigenvectors corresponding to the d largest eigenvalues:

[0137] AP=(B+βI)PΛ

[0138]

[0139]

[0140]

[0141]

[0142]

[0143] In this example, based on the migration tasks 1-6 given in Table 1, the samples are projected into the same subspace through this algorithm. Figure 5 The sample distribution scatter plot in the subspace reflects the sample distribution of each task in the subspace. It can be seen that the proposed algorithm has a strong clustering effect for each migration task.

[0144] In the above scheme, in step S6, the source domain and target domain samples are projected into the same subspace by the required projection matrix. Since the samples to be detected and the target domain samples obey the same distribution, the samples to be detected can be projected into the same subspace by the target domain projection matrix:

[0145]

[0146]

[0147]

[0148] in and Respectively represent the source domain, target domain and the data set of the sample to be detected in the subspace. The training to be tested is labeled with pseudo labels:

[0149]

[0150] in Represents the mean of each label category sample, labels the sample to be tested and the sample with the closest category mean with corresponding pseudo labels, and adds the pseudo-labeled samples to the training set to participate in iterative training.

[0151] In this example, according to the proposed iterative method, Figure 6 This reflects that the classification results are flat after 5 iterations, proving that the model achieves the best classification performance.

[0152] Table 3 compares the classification performance of each transfer model for the above tasks. Classification accuracy refers to the proportion of correctly classified samples in the test samples, and its calculation method is shown in the formula.

[0153]

[0154] Among them, TCA is transfer component analysis, JDA is joint distribution adaptation, MRNA is multi-level representation adaptation network, DANN is adversarial training neural network and CDSPP cross-domain local preservation projection model. In order to ensure the fairness of the experiment, the necessary parameters are listed as follows: (1) The sample size of the source dataset is the same as that of the target dataset, with 200 samples for each fault type and a feature dimension of 1000 for each sample. (2) The classifier uniformly adopts 1-NN.

[0155] Table 3 Comparison of classification results of different models

[0156]

[0157]

[0158] The comparison results show that the proposed algorithm has the best classification effect with an average accuracy of 96.4%. It is worth mentioning that for migration task 5, the accuracy of the proposed algorithm for cross-channel fault diagnosis reaches 98.3%, which is a significant improvement compared to the common heterogeneous migration method CDSPP. For migration task 6, cross-physical signal fault diagnosis, the distribution deviation is large due to the different physical properties of the signals in the two domains, and the overall diagnosis accuracy is low in the comparative experiment. It can still be seen from the results that the proposed algorithm can correct the strong distribution deviation well and the classification accuracy can reach more than 90%.

[0159] A method for implementing the variable operating condition fault diagnosis model of rotating machinery based on enhanced heterogeneous domain adaptation, comprising a highly robust feature extraction method, construction of two adjacency matrices, a heterogeneous domain adaptation model, and a classification iteration module;

[0160] The highly robust feature extraction method performs noise reduction processing on samples under different working conditions through NRC, and then makes the sample features more obvious through time-intrinsic scale decomposition and indicator construction.

[0161] The construction of the two adjacency matrices refines the original manifold space of the samples, optimizes the intra-class manifold weights using the cosine similarity function, and refines the edge singular sample manifold using the close neighbor criterion.

[0162] The heterogeneous domain adaptive model projects cross-domain data into a subspace that obeys the same distribution, and optimizes the intra-class distance and inter-class distance of samples in the subspace to make the samples more separable.

[0163] The pseudo-label classification iteration model integrates pseudo-label iteration into classification training, making the model training more comprehensive, thereby completing accurate fault diagnosis of rotating machinery under variable working conditions.

[0164] In summary, the present invention discloses a cross-domain fault diagnosis model for rotating machinery based on enhanced heterogeneous domain adaptation. According to the heterogeneous problems existing in the actual rotating machinery fault diagnosis, the EDCP algorithm and the classification iteration model are combined to improve the accuracy of cross-domain diagnosis. Among them, the feature extraction method of NRC-ITD is used to improve the robustness of the sample in view of the influence of strong noise on the collected samples. The subspace projection method is used to correct the large deviation between heterogeneous signals, and the category distribution of samples in the subspace is optimized and integrated into the projection learning process. Finally, the pseudo-label classification iteration model is used to integrate the pseudo-label iteration into the classification training, so that the model training is more sufficient, thereby realizing the accurate judgment of the fault category. It can not only provide a sample intrinsic expression method with strong robustness, but also provide a new optimization solution for the heterogeneous migration method guided by classification performance, and can be extended to the fault diagnosis of heterogeneous signals. The present invention selects a variety of migration models to compare with this method. It can be seen from Table 3 that compared with the other five models, this method can effectively correct the sample distribution deviation brought about by variable working conditions, thereby accurately distinguishing the fault type of the sample to be detected. In addition, the accuracy of the proposed algorithm for cross-channel fault diagnosis of migration diagnosis task 5 reached 98.3%, which is a significant improvement compared to the common heterogeneous migration method CDSPP. In terms of the comprehensive average accuracy, the misjudgment rate of this method is the smallest. The experimental results show that compared with other diagnostic models, this method can accurately correct cross-domain distribution deviations and thus improve the performance of the cross-domain diagnostic model. Finally, the present invention uses different fault categories, working conditions, acquisition channels, and physical signals in the motor defect dataset provided by Professor Yan Ruqiang's team for verification. The results verify the effectiveness of the proposed method.

[0165] The above is a detailed introduction to the cross-domain fault diagnosis model of rotating machinery based on enhanced heterogeneous domain adaptation provided by the present invention. This article uses specific embodiments to briefly explain the basic principles and implementation methods of the present invention, but the protection scope of the present invention is not limited to this. For ordinary technicians in this field, it can be understood that without departing from the principles and spirit of the present invention, various changes, modifications, substitutions and variations can be made to the embodiments in this article. The scope of the present invention is defined by the attached claims and their equivalents.

[0166] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0167] The series of detailed descriptions listed above are only specific descriptions of feasible embodiments of the present invention. They are not intended to limit the scope of protection of the present invention. All equivalent embodiments or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for cross-domain fault diagnosis of rotating machinery based on enhanced heterogeneous domain adaptation, characterized in that: The following steps are involved: Step S1, sample denoising based on the noise-resistant correlation method (NRC): the collected vibration signal is separated into a periodic component and a residual component, and the average value is taken to construct a new signal to be compared with the previous signal to reduce the influence of noise, and then a correlation function between the original signal and the new signal is defined to enhance the periodic information of the sample under a strong noise background; Step S2, intrinsic time scale decomposition (ITD) decomposition and index construction: the denoised samples are subjected to ITD decomposition to obtain a number of intrinsic rotation components (PRCs) with frequencies ranging from high frequency to low frequency, and the component signals are screened by combining weight index screening criteria; in order to reduce the computational load of the model, the energy values ​​of the selected components are extracted as indicators to reconstruct the samples; Step S3: Construct the intra-class sample adjacency matrix: First, according to the source domain sample and a small part of the target domain sample labels, construct the intra-class adjacency matrix W based on the heat kernel function ij ; Secondly, construct the sample similarity weight matrix βc ij As a regularization term, it optimizes the intra-class distance of samples in the subspace; Step S4, constructing the sample class adjacency matrix: First, set the parameter k to construct the nearest neighbor criterion of heterogeneous samples, calculate the Euclidean distance between heterogeneous samples, and use the nearest neighbor criterion of heterogeneous samples to determine whether the sample belongs to the boundary edge sample, thereby constructing the inter-class adjacent sample matrix D ij ; Step S5, constructing an enhanced heterogeneous domain adaptive model: first, constructing an objective function with the purpose of minimizing the intra-class distance and maximizing the inter-class distance, respectively calculating partial derivatives of the source domain projection matrix and the target domain projection matrix, thereby converting them into a generalized eigenvalue solving problem, and finally calculating the projection matrices of the two domains, and projecting the samples into the same subspace for subsequent training; Step S6, constructing a classification iteration model: multiplying the sample to be detected by the target domain projection matrix and performing the first classification, adding a pseudo label to the training sample according to the classification result, and repeating the above operation to make the model achieve the best classification effect.

2. A method according to claim 1, characterized in that The step S1 defines the composition of the collected vibration signal as a periodic component and a residual component, and defines the function expression of the new signal as follows: y(T)=x(t)+∈(t) Among them, x(t) is a bounded periodic signal whose period is unknown and hidden in the original signal y(t); ∈(t) is independent and follows a normal distribution. The residual term of T (t) is a new periodic function, which is obtained by accumulating and averaging y(t) at time intervals of T; To perform floor operation, is the number of segments after the signal is separated; then based on the new periodic function f T (t), and construct a new autocorrelation function C(T) with it and y(t): According to the above steps, the periodic component S(T) in the strong noise background signal is extracted, where is the residual component, and σ is the residual parameter.

3. A method according to claim 1, characterized in that In step S2, in order to better express the sample intrinsically, the periodic component S(T) is subjected to intrinsic time scale decomposition ITD, and the mathematical process is as follows: S(T)=LS(T)+(1-L)S(T)=L t +H t Among them, L is the new operator defined, L t represents the extracted baseline signal, H t is the rotation component; then the baseline signal L t As the original signal, it is decomposed until the decomposed signal is less than the set value. The final decomposition result can be expressed as: Among them, H i (i=1,2,…,p) represents the i-th rotation component, L p represents the remaining baseline component after p decompositions, and k is the kth component. In order to reduce the redundant information of the sample, the energy value of the i rotation components before each decomposition is selected as the expression of the sample. The process of expressing source domain and target domain samples is as follows: Where n is the number of samples, P i is the energy value of the i-th layer.

4. A method according to claim 1, characterized in that In step S3, the source domain sample set is The target domain sample set is Intra-class sample adjacency matrix W ij for: W ij =w ij +βc ij where x i 、x j is the i-th or j-th sample from the source domain or the target domain, is the mean value corresponding to the sample, y i ,y j is the label corresponding to the sample; w ij The calculation method is the heat kernel function, which is based on the similarities and differences of label information. If they are of the same category, the weight is 1, and if they are of different categories, the weight is 0. ij is the similarity weight between samples, β is the regularization parameter, and the final sample class adjacency matrix W ij is the weighted sum of the two terms; in step S4, first set the point neighbor parameter k, the inter-class adjacency matrix D ij The composition is: Where N K (x j ) Sample x j A sample set of nearest neighbors based on Euclidean distance.

5. A method according to claim 1, characterized in that: In step S5, in order to minimize the intra-class discreteness of samples in the subspace, based on the intra-class sample adjacency matrix W ij And the objective function constructed by minimizing the two-norm is as follows: Among them, P s , P t Represent the source domain projection matrix and the target domain projection matrix respectively, It is represented as the transpose of the source domain projection matrix and the target domain projection matrix; n s and n t is the number of samples in the source domain and the target domain; W s , W t Represents the sample intra-class adjacency matrix of the source domain and the target domain, W c is the intra-class adjacency matrix of cross-domain samples; in order to maximize the distance between classes of samples in the subspace, based on the inter-class sample adjacency matrix D ij And the objective function constructed by maximizing the two-norm is as follows: Where D s and D t Represent the inter-class adjacency matrix of the source domain and the target domain respectively; the above two objective functions are integrated into one calculation process, and the formula is as follows: Expanding and simplifying the bi-norm in the above objective function can obtain the following objective function: Where X s and X t are the sample set matrices of the source domain and the target domain respectively; D wt , is a diagonal matrix whose corresponding values ​​are the sums of fixed columns, D wt , is a diagonal matrix whose corresponding values ​​are the sums of fixed rows; and The diagonal matrices representing the sum of rows and columns across the domain respectively; and W s and W t The corresponding Laplace matrix; and D s and D t The corresponding Laplace matrix; In order to simplify the objective function, we introduce the coefficient factors S c , and The equation is converted into the form of matrix trace: For the objective function with respect to P s , P t Calculate the partial derivatives respectively and set the results to 0. Then the optimization problem of the final objective function can be transformed into solving the eigenvector corresponding to the generalized eigenvalue, where the optimal projection matrix P = (P s ,P t ) can be composed of the eigenvectors corresponding to the d largest eigenvalues: AP=(B+βI)PΛ Where A and B are the coefficient matrices obtained after partial derivatives, and ∧ is the regularization parameter term.

6. A method according to claim 1, characterized in that: In step S6, the source domain and target domain samples are projected into the same subspace by the required projection matrix. Since the samples to be detected and the samples in the target domain obey the same distribution, the samples to be detected can be projected into the same subspace by the target domain projection matrix: in and They represent the original space source domain samples, target domain samples and samples to be detected respectively; and Respectively represent the source domain, target domain and the data set of the sample to be detected in the subspace. The training to be tested is labeled with pseudo labels: Among them, y * The pseudo labels given to the samples, Represents the mean of each label category sample, labels the sample to be tested and the sample with the closest category mean with corresponding pseudo labels, and adds the pseudo-labeled samples to the training set to participate in iterative training.

Images