Bearing fault diagnosis method, device and storage medium based on joint distribution adaptation

Abstract

The present invention discloses a bearing fault diagnosis method, device, and storage medium based on improved joint distribution adaptation, and relates to the field of bearing fault diagnosis methods. The method comprises the following steps: Step 1, establishing a source domain feature set and a target domain feature set based on a bearing vibration signal; Step 2, establishing a source domain feature sample set and a target domain feature sample set after constructing a feature transferability quantification index; Step 3, performing manifold subspace learning on the source domain feature sample set to obtain a source domain feature subset and a feature mapping matrix; Based on the feature mapping matrix, the target domain feature sample set is converted into a target domain feature subset and then solved by a joint distribution adaptation algorithm to obtain a distribution-adapted source domain feature set and a target domain feature set; Step 4, training a classifier using the source domain feature set after distribution adaptation, and using the trained classifier to obtain a fault diagnosis result based on the target domain feature set after distribution adaptation; The device and storage medium are used to implement the method.

Description

Technical Field

[0001] The present invention relates to the field of bearing fault diagnosis methods, and in particular to a bearing fault diagnosis method, device and storage medium based on joint distribution adaptation. Background Art

[0002] The reliability and stability of coal mining machinery and its key components directly impact the efficiency of coal mining and are key to achieving unmanned coal mining. As a key piece of equipment in coal mines, the operating status of hoists directly affects the safety of production personnel and the mine's production capacity. Bearings, a core component of hoists, often operate in harsh and complex conditions, which can easily lead to bearing damage. Therefore, research on intelligent fault diagnosis for hoist bearings is crucial for ensuring safe coal production and reducing economic losses. Similarly, intelligent bearing fault diagnosis is also crucial for other mechanical equipment that utilize bearings.

[0003] Currently, researchers often use vibration signals for mechanical equipment fault status analysis and diagnosis. On this basis, they combine signal processing methods, machine learning, and deep feature transfer learning to build intelligent fault diagnosis systems. Existing intelligent bearing fault diagnosis methods based on deep feature transfer learning generally follow the following steps:

[0004] Step 1: Domain division. The source domain contains labeled samples to provide diagnostic information, while the unlabeled samples in the target domain are expected to be classified using the diagnostic information in the source domain.

[0005] Step 2: Feature extraction: Nonlinear feature mapping is used to extract transferable features from the data in the source and target domains. The data in the source and target domains are processed by the same nonlinear feature mapping at the same time.

[0006] Step 3: Domain Adaptation. First, a non-parametric distance metric, MMD (Maximum Mean Discrepancy, MMD), is used to measure the distributional variance of the learned transferable features. Then, using the distributional variance of the learned transferable features as the optimization objective, backpropagation is used to train the parameters of the nonlinear feature mapping. The goal of optimization is to minimize the distributional variance of the learned transferable features, thereby obtaining features with minimal cross-domain variance. By minimizing the MMD of the learned transferable features, the parameters of the domain-sharing CNN and the Convolutional Neural Network (CNN) are trained.

[0007] Step 4: Fault identification: Use the domain-shared classifier to correctly classify the unlabeled samples in the target domain.

[0008] The existing bearing fault intelligent diagnosis methods based on deep feature transfer learning have the following main problems:

[0009] (1) The non-parametric distance metric MMD is used to measure the distribution difference of the learned transferable features. The optimization goal is to minimize the distribution difference of the learned features, thereby obtaining features with smaller cross-domain differences. This does not fully consider the impact of feature discrimination ability on the accuracy of cross-domain fault pattern recognition, which in turn affects the improvement of the model's cross-domain fault diagnosis accuracy.

[0010] (2) The domain adaptation process is only implemented by domain-sharing CNN and MMD constraints. The extracted features are not very interpretable, and the inter-domain conditional probability distribution and marginal probability distribution are not reasonably considered at the same time. The feature distortion problem that exists in the process of reducing distribution differences in high-dimensional feature space is not considered, which leads to insufficient generalization ability of the fault diagnosis model.

[0011] (3) The features extracted based on deep learning methods have problems such as complex models, hyperparameters and poor interpretability. Summary of the Invention

[0012] The present invention provides a bearing fault diagnosis method, device and storage medium based on joint distribution adaptation to solve the problems existing in the above-mentioned prior art bearing fault intelligent diagnosis method based on deep feature transfer learning.

[0013] In order to achieve the above object, the technical solution adopted by the present invention is:

[0014] The bearing fault diagnosis method based on joint distribution adaptation includes the following steps:

[0015] Step 1: Obtain bearing vibration signals with known working conditions and bearing vibration signals with unknown working conditions. Extract multiple time-frequency domain statistical feature data based on the bearing vibration signals with known working conditions as a labeled source domain feature set. Extract multiple time-frequency domain statistical feature data based on the bearing vibration signals with unknown working conditions as an unlabeled target domain feature set.

[0016] Step 2: Based on the labeled source domain feature set obtained in step 1, the feature importance FI of each statistical feature data is obtained. The feature importance FI is also used as the feature importance of the corresponding class statistical feature data in the target domain feature set. The feature importance FI represents the discriminative ability of each statistical feature data in the source domain feature set and the target domain feature set.

[0017] Calculate the KL divergence KLD of each statistical feature data in the source domain feature set in the source domain feature set and the target domain feature set, and calculate the KL divergence KLD of each statistical feature data in the target domain feature set in the source domain feature set and the target domain feature set, to measure the difference of each statistical feature data in different domain feature sets. The KL divergence KLD represents the adaptability of each statistical feature data in different domain feature sets;

[0018] Based on the feature importance FI and KL divergence KLD, a feature transferability quantitative index RFK of each feature statistical data in the source domain feature set and the target domain feature set is constructed;

[0019] Finally, based on the feature transferability quantitative index RFK, some statistical feature data are selected from the source domain feature set to construct a labeled source domain feature sample set D S , and based on the feature transferability quantitative index RFK, select some statistical feature data from the target domain feature set to construct an unlabeled target domain feature sample set D T ;

[0020] Step 3: The source domain feature sample set D obtained in step 2 S The statistical feature data in the manifold subspace is learned to obtain the source domain feature subset Z S and the feature map matrix;

[0021] Based on the feature mapping matrix, the target domain feature sample set D obtained in step 2 is T The statistical feature data in the target domain is mapped and transformed to obtain the target domain feature subset Z T ;

[0022] Based on the source domain feature subset Z S , target domain feature subset Z T The marginal probability distribution and conditional probability distribution between the statistical feature data are calculated, and the joint distribution adaptation algorithm JDA model is established and solved to obtain the source domain feature set Z′ after the distribution adaptation with labels. S , and the unlabeled distribution-adapted target domain feature set Z′ T ;

[0023] Step 4: Use the labeled distribution-adapted source domain feature set Z′ obtained in step 3 S Train the machine learning classifier and use the trained classifier to adapt the target domain feature set Z′ obtained in step 3 to the unlabeled distribution. T Processing is performed to obtain the bearing fault diagnosis results.

[0024] In the further step 1, the wavelet packet transform method is used to perform multi-layer decomposition on the bearing vibration signal with known working conditions and the bearing vibration signal with unknown working conditions, respectively, to obtain corresponding reconstructed signals, and then a variety of statistical feature parameters are calculated based on the reconstructed signals, respectively, to obtain a labeled source domain feature set and an unlabeled target domain feature set.

[0025] Furthermore, there are 11 statistical characteristic parameters, namely range, mean, standard deviation, kurtosis, energy, energy entropy, kurtosis, impulse factor, shape factor, skewness, and extreme value.

[0026] In the further step 2, the random forest algorithm is used to process the labeled source domain feature set obtained in step 1 to obtain the feature importance FI of each statistical feature data.

[0027] In the further step 2, the feature transferability quantitative index RFK of each statistical feature data is the ratio of the feature importance FI of the statistical feature data to the KL divergence KLD.

[0028] In the further step 2, according to the feature transferability quantitative index RFK from large to small, the same number of statistical feature data are selected from the source domain feature set and the target domain feature set, and the labeled source domain feature sample set D is constructed with the statistical feature data selected from the source domain feature set. S , construct an unlabeled target domain feature sample set D using statistical feature data selected from the target domain feature set T .

[0029] In the further step 3, the neighborhood preserving embedding algorithm NPE is used to embed the source domain feature sample set D obtained in step 2. S The statistical feature data in is processed to realize manifold subspace learning.

[0030] In the further step 3, the source domain feature subset Z is established S , target domain feature subset Z T The maximum mean difference MMD matrix of the marginal probability distribution between the statistical feature data in the source domain and the source domain feature subset Z S , target domain feature subset Z T The maximum mean difference MMD matrix of the conditional probability distribution between the statistical feature data in the marginal probability distribution is obtained; the maximum mean difference MMD matrix of the conditional probability distribution and the maximum mean difference MMD matrix of the conditional probability distribution are both minimized as the optimization goal, and the joint distribution adaptation algorithm JDA model is established and solved.

[0031] An electronic device includes a processor and a memory, wherein a storage medium of the memory stores program instructions, and when the program instructions are read and run, steps 1 to 4 of the above-mentioned bearing fault diagnosis method based on joint distribution adaptation are executed.

[0032] A storage medium stores program instructions that can be read and executed. When the program instructions are read and executed, steps 1 to 4 of the above-mentioned bearing fault diagnosis method based on joint distribution adaptation are executed.

[0033] The present invention proposes a bearing fault diagnosis method based on improved joint distribution adaptation, which performs an effective quantitative assessment of the transferability of each statistical feature and fully considers the impact of feature discrimination ability on the accuracy of cross-domain fault pattern recognition. Then, to address the feature distortion problem existing in the process of reducing distribution differences in high-dimensional feature space, inter-domain conditional probability distribution and marginal probability distribution are used, combined with joint distribution adaptation to adapt the distribution of source domain and target domain feature sets, so that the feature data used for fault judgment has good interpretability. Finally, the manifold subspace mined reduces the inter-domain distribution differences and improves the cross-domain fault recognition performance of the pattern recognition classifier. Therefore, the present invention can improve the accuracy and generalization ability of the machine learning classifier model used for bearing fault diagnosis, thereby improving the accuracy of intelligent diagnosis of bearing faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of a method according to a first embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of the NPE algorithm in Example 1 of the present invention. DETAILED DESCRIPTION

[0036] To help those skilled in the art better understand the present invention, the following detailed description of the embodiments of the present invention is provided in conjunction with the accompanying drawings and examples. This will help those skilled in the art to fully understand and implement the present invention by applying technical means to solve technical problems and achieve corresponding technical effects. The embodiments of the present invention and the various features therein may be combined with each other as long as they do not conflict with each other, and the resulting technical solutions are all within the scope of protection of the present invention.

[0037] Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0038] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or apparatuses.

[0039] Example 1

[0040] This embodiment discloses a bearing fault diagnosis method based on joint distribution adaptation, such as Figure 1 As shown, the following steps are included:

[0041] Step 1: Obtain bearing vibration signals for known and unknown operating conditions. Use wavelet packet transform to process these signals, respectively, to obtain corresponding reconstructed signals. Then, based on the reconstructed signals, 11 statistical feature parameters are calculated. This yields a labeled source domain feature set (labeled for known operating conditions) and an unlabeled target domain feature set.

[0042] Specifically, for bearing vibration signals with known operating conditions, the wavelet packet transform is first used to perform a four-layer decomposition of the signal, yielding a reconstructed signal with 16 terminal nodes at the fourth layer. The Hilbert envelope spectrum of the reconstructed signal with 16 terminal nodes at the fourth layer is then calculated. Based on this Hilbert envelope spectrum, 11 statistical feature parameters are calculated: range, mean, standard deviation, kurtosis, energy, energy entropy, kurtosis, impulse factor, shape factor, skewness, and extreme values. Finally, 16*11=176 time-domain statistical feature data are extracted from these calculated statistical feature parameters to construct a labeled source domain feature set.

[0043] The same processing is performed on the bearing vibration signal with unknown working conditions: first, the wavelet packet transform is used to decompose the bearing vibration signal with unknown working conditions into 4 layers in sequence to obtain the reconstructed signal of the 4th layer with a total of 16 terminal nodes. Then, the Hilbert envelope spectrum signal of the reconstructed signal of the 16 terminal nodes in the 4th layer is calculated, and a total of 11 statistical feature parameters are calculated based on the Hilbert envelope spectrum signal. Finally, a total of 16*11=176 types of time domain statistical feature data are extracted from the calculated statistical feature parameters to construct an unlabeled target domain feature set. Therefore, in this embodiment, a total of 352 statistical feature data of the source domain feature set and the target domain feature set are obtained.

[0044] Step 2: To reduce the interference and redundant features in the high-dimensional original feature set, select features that are beneficial to fault state discrimination and have small inter-domain distribution differences. Based on the source domain feature set and target domain feature set obtained in step 1, the discriminative ability of each statistical feature data, and the distribution differences of each statistical feature data in different domains, construct the feature transferability quantitative index RFK for each statistical feature data to be used in the subsequent steps to select statistical feature data. The process is as follows:

[0045] (2.1) Use the classic random forest algorithm to calculate the feature importance FI (Feature Importance, FI) of each statistical feature data in the labeled source domain feature set obtained in step 1. This feature importance FI is also used as the feature importance of the corresponding class statistical feature data in the target domain feature set. The feature importance FI represents the discriminative ability of each statistical feature data in the source domain feature set and the target domain feature set.

[0046] The feature importance FI can measure the contribution of each statistical feature data to the classification accuracy. The larger the value of this indicator, the better the fault discrimination ability of the statistical feature data.

[0047] Assume that the source domain feature set There are M characteristic samples of K types of faults in total, among which the mth statistical characteristic data =The expression is shown in formula (1):

[0048]

[0049] In formula (1), Represents the nth statistical feature data of the kth fault, where N is the number of feature samples.

[0050] The source domain feature set is processed using the random forest algorithm to obtain the importance index of M features and construct the feature importance sequence S I As shown in formula (2):

[0051] S I ={i(1),i(2),…,i(M)} (2),

[0052] In formula (2), i(M) is the importance index FI of the Mth statistical feature data.

[0053] (2.2) Calculate the KL divergence KLD of each statistical feature data in the source domain feature set in the source domain feature set and the target domain feature set, and calculate the KL divergence KLD of each statistical feature data in the target domain feature set in the source domain feature set and the target domain feature set.

[0054] KL divergence (TFFK, Transferable feature selection based on feature importance and KLD) is used to measure the difference between each statistical feature data in different domain feature sets. KL divergence KLD represents the adaptability of each statistical feature data in different domain feature sets. KL divergence KLD, also known as KL distance, is a method commonly used in probability theory and information theory to describe the distribution difference between two probability distributions P and Q. It is also called relative entropy. The KL divergence between P and Q is defined as shown in formula (3):

[0055]

[0056] In formula (3), p(x i ) is the target distribution, q(x i ) is the matching distribution, x i is a discrete random variable, and L is the length of the probability distribution. KLD The smaller (P||Q) is, the smaller the distribution difference between P and Q is. KLD When (P||Q)=0, P and Q are completely matched.

[0057] Therefore, this embodiment uses the statistical feature data of the source domain feature set and the target domain feature set obtained in step 1 under normal conditions to calculate KLD, measure the distribution difference of each statistical feature data in different domains, and obtain the KLD sequence S of M features. KLD , as shown in formula (4):

[0058] S KLD ={d(1),d(2),…,d(M)} (4),

[0059] In formula (4), d(M) is the KL divergence KLD of the M-th statistical feature data.

[0060] (2.3) In order to comprehensively consider the performance of both feature importance and inter-domain distribution differences and achieve ideal feature transfer learning effects, this embodiment obtains the feature importance FI and KL divergence KLD of each statistical feature data based on the feature evaluation of the aforementioned content, and then constructs the feature transferability quantification index RFK of each feature statistical data in the source domain feature set and the target domain feature set based on the feature importance FI and KL divergence KLD. The feature transferability quantification index RFK is the ratio of the feature importance FL of each statistical feature data to the corresponding KL divergence KLD, which is used to characterize feature transferability. Therefore, this embodiment obtains the expression r(m) of the feature transferability quantification index RFK of each statistical feature data in the source domain feature set and the target domain feature set as shown in formula (5):

[0061]

[0062] Based on formula (5), we can obtain the transferability index sequence S of M kinds of statistical feature data: RFK As shown in formula (6):

[0063] S RFK ={r(1),r(2),…,r(M)} (6),

[0064] (2.4) Based on the feature transferability quantitative index RFK, some statistical feature data are selected from the source domain feature set to construct a labeled source domain feature sample set D S , and based on the feature transferability quantitative index RFK, select some statistical feature data from the target domain feature set to construct an unlabeled target domain feature sample set D T .

[0065] Specifically, for the source domain feature set, multiple data are selected to construct the source domain feature sample set D according to the feature transferability quantitative index RFK of the statistical feature data from large to small. S Similarly, for the target domain feature set, the same amount of data is selected to construct the target domain feature sample set D according to the order of the feature transferability quantitative index RFK of the statistical feature data from large to small. T .

[0066] Step 3: Get the source domain feature sample set D according to step 2 S , target domain feature sample set D T After that, the source domain feature sample set D S Perform manifold feature learning to obtain the source domain feature subset Z S And the feature mapping matrix A, and according to the feature mapping matrix A, the target domain feature sample set D T Convert to target domain feature subset Z T Then, the joint distribution adaptation algorithm JDA model is used to adapt the distribution to reduce the distribution difference, thereby achieving improved joint distribution adaptation (Improved Joint Distribution Adaptation, IJDA). The specific process is as follows:

[0067] (3.1) Using the neighborhood preserving embedding algorithm NPE, the source domain feature sample set D S The data is processed to realize manifold feature learning and the source domain feature subset Z is obtained S And the feature map matrix A.

[0068] Neighborhood Preserving Embedding (NPE) is a linear approximation of the classic manifold learning algorithm Locally Linear Embedding (LLE). The goal of the Neighborhood Preserving Projection (NPE) algorithm is to preserve the manifold structure of high-dimensional data when mapping it to a low-dimensional space. The operation process is as follows: Figure 2 As shown, the process is as follows:

[0069] (a) First, select the nearest neighbors and construct the neighbor graph G. Specifically, based on the K-nearest neighbor method, the source domain feature sample set D S The statistical feature data in the source domain is used as a node to S Each node in the constructs a neighboring graph. Among them, the i-th node corresponds to the data sample point x i (i.e., the statistical feature data corresponding to the node), calculate x based on the Euclidean distance i The data sample point x of the adjacent node j j If the distance between them is within the nearest neighbor range, the two nodes are connected to construct G.

[0070] (b) Calculate the neighbor graph reconstruction weight coefficient matrix W. The element w in the neighbor graph reconstruction weight coefficient matrix W ij is the weight of the edge from node i to node j, and the reconstruction loss function of the domain-preserving projection algorithm NPE As shown in formula (7):

[0071]

[0072] The neighborhood graph reconstruction weight coefficient matrix W can be obtained by minimizing the reconstruction loss function Solve, as shown in formula (8):

[0073]

[0074] y i =A T x i Bring in the minimized reconstruction loss function Solving, formula (8) can be further transformed into formula (9):

[0075] P(A)=min(tr(A T XZX T A)) (9),

[0076] Formula (9), yi is the result obtained after manifold learning; A is the feature mapping matrix, Z = (IW) T(IW), I = diag(1,…,1), I represents the unit diagonal matrix; X represents the input sample; Z represents the symmetric semi-positive definite matrix; P(A) represents the optimization problem of solving A.

[0077] (c) To further solve the eigenmapping matrix A, the optimization problem of formula (9) can be transformed into the problem of solving the minimum eigenvalue of the generalized eigenvector, as shown in formula (10):

[0078] XZX T a=λXX T a (10),

[0079] In formula (10), λ is the eigenvalue, a is the eigenvector corresponding to the eigenvalue, and the eigenvector corresponding to the eigenvalue is selected in ascending order according to the eigenvalue to construct a new feature subset. Thus, the source domain feature subset Z is obtained through the above-mentioned domain-preserving projection algorithm NPE. S And the feature map matrix A.

[0080] (3.2) Based on the obtained feature mapping matrix A, the target domain feature sample set D obtained in step 2 is T The statistical feature data in the target domain is mapped and transformed to obtain the target domain feature subset Z T .

[0081] (3.3) Establish the source domain feature subset Z S , target domain feature subset Z T The maximum mean difference MMD matrix of the marginal probability distribution between the statistical feature data in the source domain and the source domain feature subset Z S , target domain feature subset Z T The maximum mean difference (MMD) matrix of the conditional probability distribution between the statistical feature data is obtained. Then, the optimization goal is to minimize the maximum mean difference (MMD) matrix of the marginal probability distribution and the maximum mean difference (MMD) matrix of the conditional probability distribution. The joint distribution adaptation algorithm (JDA) model is established and solved.

[0082] The basic principle of the Joint Distribution Adaptation Algorithm (JDA) model is as follows: given a subset of labeled source domain features, and the unlabeled target domain feature subset where n S With n T are the number of samples in the source domain and target domain respectively, x ns represents the nth in the source domain feature subset s Sample data, y ns Represents x ns The corresponding category, x ns+nt represents the nth in the target domain feature subset sThere are differences in the marginal probability distribution and conditional probability distribution between the statistical feature data of the source domain and the target domain, that is, Q S (y S |x S )≠Q T (y T |x T ) and P S (x S )≠P T (x T ), where Q S (y S |x S ) represents the marginal probability distribution of statistical feature data in the source domain feature subset, Q T (y T |x T ) represents the marginal probability distribution of statistical feature data in the target domain feature subset, P S (x S ) represents the conditional probability distribution of statistical feature data in the source domain feature subset, P T (x T ) represents the conditional probability distribution of statistical feature data in the target domain feature subset, x S represents the sample data in the source domain feature subset, y S Represents x S Category, x T represents the sample data in the target domain feature subset, y T Represents x T Category.

[0083] The goal of the Joint Distribution Adaptation algorithm (JDA) is to use the source domain feature subset Z S and the target domain feature subset Z T Learn to obtain a feature mapping transformation matrix W, so that the distance between the marginal probability distribution and the conditional probability distribution of the statistical feature data in the transformed source domain and target domain feature subsets is reduced as much as possible, that is, P S (W T x S ) and P T (W T x T ), Q S (y S |W T x S ) and Q T (y T |W T x T ) as much as possible, where P S (W T x S ) represents the source domain feature subset ZS The marginal probability distribution after W mapping transformation, P T (W T x T ) represents the target domain feature subset Z T The marginal probability distribution after W mapping transformation, Q S (y S |W T x S ) represents the source domain feature subset Z S The conditional probability distribution after W mapping transformation, Q T (y T| W T x T ) represents the target domain feature subset Z T Conditional probability distribution after W mapping transformation.

[0084] Therefore, the JDA algorithm model in this embodiment includes two optimization objectives:

[0085] (A) Realize the source domain feature subset Z S and the target domain feature subset Z T The marginal probability distribution adaptation of the statistical feature data, that is, P S (W T x S ) and P T (W T x T ) is the smallest, and the optimization objective expression is shown in formula (11):

[0086]

[0087] in, Represents the statistical feature data in the source domain feature subset, represents the statistical feature data in the target domain feature subset, and X represents the combined source domain and target domain feature subsets.

[0088] In formula (11), M0 is P S (W T x S ) and P T (W T x T ), the maximum mean difference MMD matrix between them, the calculation expression of M0 is shown in formula (12):

[0089]

[0090] (B) Adapt the conditional probability distribution of statistical feature data in the source domain feature subset and the target domain feature subset, i.e. Q S (y S|W T x S ) and Q T (y T |W T x T ) is the smallest, and the optimization objective expression is shown in formula (13):

[0091]

[0092] In formula (13), and are the source domain feature subset Z S and the target domain feature subset Z T The number of samples of category c, and are the source domain feature subset Z S and the target domain feature subset Z T The c-th sample in , C is the total number of sample categories.

[0093] M c Q S (y S |W T x S ) and Q T (y T |W T x T ) is the maximum mean difference MMD matrix, M c The calculation expression of is shown in formula (14):

[0094]

[0095] Based on the above optimization objectives (A) and (B), the overall optimization objective of the Joint Distribution Adaptation (JDA) algorithm is as shown in formula (15):

[0096]

[0097] Formula (15) unifies the two distances of formulas (11) and (13) by c = 0, 1, 2, ..., C. is the regularization term, is a trade-off parameter, W T XHX T W=I is the constraint condition, I is the unit diagonal matrix, is the combined feature subset of the source and target domains.

[0098] Therefore, the improved joint distribution adaptation IJDA algorithm is realized through the above process. First, the source domain data and the target domain data are subjected to manifold feature learning, and then the maximum mean difference MMD matrix M0 of the marginal probability distribution and the maximum mean difference MMD matrix M of the conditional probability distribution are used to calculate the manifold feature. c The joint distribution adaptation JDA algorithm model is solved by taking the minimum as the optimization goal, and then the feature mapping transformation matrix W is obtained. Based on the feature mapping transformation matrix W, the source domain feature subset Z is S , target domain feature subset Z T Perform mapping transformation to obtain the source domain feature set Z′ after the labeled probability distribution is adapted S And the target domain feature set Z′ after adaptation of the unlabeled probability distribution T .

[0099] In step 3 of this embodiment, a manifold feature learning component is added to the joint distribution adaptation JDA algorithm model, thereby obtaining an improved joint distribution adaptation IJDA algorithm, which can make up for the limitation of the traditional feature transfer learning method that only considers the adaptation of the marginal probability distribution, and comprehensively considers the two probability distributions, thereby improving the transfer learning effect. However, there are still limitations in directly adapting the two distributions to the source domain and target domain samples, that is, when adapting the two distributions in the original data space, the problem of feature distortion cannot be overcome, resulting in poor distribution adaptation effect. Therefore, this embodiment proposes to first perform manifold feature learning on the original data space, and then adapt the two distributions after mapping the original data space to the manifold subspace, thereby overcoming the problem of poor distribution adaptation effect caused by feature distortion. After step 3 of this embodiment, feature transfer learning is achieved after mapping the source domain and target domain data to the manifold space, thereby achieving inter-domain distribution adaptation while overcoming the defect of feature distortion.

[0100] Step 4: Adapt the source domain feature set Z′ using the labeled probability distribution obtained in step 3 S Train the classifier of intelligent fault diagnosis machine learning, and then adapt the unlabeled probability distribution obtained in step 3 to the target domain feature set Z′ T Input it into the trained classifier and finally obtain the bearing fault diagnosis result.

[0101] Example 2

[0102] This embodiment discloses an electronic device for implementing the bearing fault diagnosis method based on joint distribution adaptation described in Embodiment 1, the electronic device comprising a processor and a memory. The storage medium of the memory stores program instructions that can be read and executed by the processor.

[0103] The program instructions include a data pre-processing module, a data post-processing module, a joint distribution adaptation module, and a classifier module, wherein:

[0104] The data pre-processing module obtains the vibration signals of bearings with known working conditions and bearing vibration signals with unknown working conditions, and processes the signals to obtain statistical feature data in the time-frequency domain, and then constructs a labeled source domain feature set and an unlabeled target domain feature set.

[0105] The data post-processing module obtains the source domain feature set and the target domain feature set from the data pre-processing module, and calculates the feature importance FI and KL divergence KLD to construct the feature transferability quantitative index RFK. Then, based on the feature transferability quantitative index RFK, the statistical feature data is selected to construct the labeled source domain feature sample set D. S , unlabeled target domain feature sample set D T .

[0106] The joint distribution adaptation module obtains the source domain feature sample set D from the data post-processing module S , target domain feature sample set D T , and the source domain feature sample set D S Perform manifold subspace learning to obtain the source domain feature subset Z S And feature mapping matrix, and the target domain feature sample set D based on the obtained feature mapping matrix T Perform mapping transformation to obtain the target domain feature subset Z T ;Finally calculate the source domain feature subset Z S , target domain feature subset Z T The marginal probability distribution and the maximum mean difference MMD matrix of the conditional probability distribution are used to establish the joint distribution adaptation algorithm JDA model to solve it and obtain the source domain feature set Z′ after the distribution adaptation of the label S , and the unlabeled distribution-adapted target domain feature set Z′ T .

[0107] The classifier module pre-stores a machine learning classifier model, and the classifier module obtains the source domain feature set Z′ from the joint distribution adaptation module S , target domain feature set Z′ T , and use the source domain feature set Z′ S After training the classifier model, the trained classifier model is used to identify the target domain feature set Z′. T A bearing fault is detected.

[0108] Therefore, when the processor reads and runs the program instructions in the storage medium of the memory, steps 1 to 4 of the bearing fault diagnosis method based on joint distribution adaptation described in the first embodiment are executed.

[0109] The preferred embodiments of the present invention are described in detail above with reference to the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner unless there is any contradiction. Such combinations should also be regarded as the contents disclosed in this disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0110] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the scope of the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by those skilled in the art should fall within the scope of protection of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.

Claims

1. Bearing fault diagnosis method based on joint distribution adaptation, characterized in that: The following steps are involved: Step 1: Obtain bearing vibration signals with known working conditions and bearing vibration signals with unknown working conditions. Extract multiple time-frequency domain statistical feature data based on the bearing vibration signals with known working conditions as a labeled source domain feature set. Extract multiple time-frequency domain statistical feature data based on the bearing vibration signals with unknown working conditions as an unlabeled target domain feature set. Step 2: Based on the labeled source domain feature set obtained in step 1, the feature importance FI of each statistical feature data is obtained. The feature importance FI is also used as the feature importance of the corresponding class statistical feature data in the target domain feature set. The feature importance FI represents the discriminative ability of each statistical feature data in the source domain feature set and the target domain feature set. Calculate the KL divergence KLD of each statistical feature data in the source domain feature set in the source domain feature set and the target domain feature set, and calculate the KL divergence KLD of each statistical feature data in the target domain feature set in the source domain feature set and the target domain feature set, to measure the difference of each statistical feature data in different domain feature sets. The KL divergence KLD represents the adaptability of each statistical feature data in different domain feature sets; Based on the feature importance FI and KL divergence KLD, a feature transferability quantitative index RFK of each feature statistical data in the source domain feature set and the target domain feature set is constructed; Finally, based on the feature transferability quantitative index RFK, some statistical feature data are selected from the source domain feature set to construct a labeled source domain feature sample set D S , and based on the feature transferability quantitative index RFK, select some statistical feature data from the target domain feature set to construct an unlabeled target domain feature sample set D T ; Step 3: The source domain feature sample set D obtained in step 2 S The statistical feature data in the manifold subspace is learned to obtain the source domain feature subset Z S and the feature map matrix; Based on the feature mapping matrix, the target domain feature sample set D obtained in step 2 is T The statistical feature data in the target domain is mapped and transformed to obtain the target domain feature subset Z T ; Based on the source domain feature subset Z S , target domain feature subset Z T The marginal probability distribution and conditional probability distribution between the statistical feature data are calculated, and the joint distribution adaptation algorithm JDA model is established and solved to obtain the source domain feature set Z′ after the distribution adaptation with labels. S , and the unlabeled distribution-adapted target domain feature set Z′ T ; Step 4: Use the labeled distribution-adapted source domain feature set Z′ obtained in step 3 S Train the machine learning classifier and use the trained classifier to adapt the target domain feature set Z′ obtained in step 3 to the unlabeled distribution. T Processing is performed to obtain the bearing fault diagnosis results.

2. The bearing fault diagnosis method based on joint distribution adaptation according to claim 1 is characterized in that: In step 1, the wavelet packet transform method is used to perform multi-layer decomposition on the bearing vibration signal with known working conditions and the bearing vibration signal with unknown working conditions, respectively, to obtain the corresponding reconstructed signals. Then, based on the reconstructed signals, various statistical feature parameters are calculated respectively to obtain the labeled source domain feature set and the unlabeled target domain feature set.

3. The bearing fault diagnosis method based on joint distribution adaptation according to claim 2, characterized in that: There are 11 statistical characteristic parameters, namely range, mean, standard deviation, kurtosis, energy, energy entropy, kurtosis, impulse factor, shape factor, skewness, and extreme value.

4. The bearing fault diagnosis method based on joint distribution adaptation according to claim 2, characterized in that: In step 2, the random forest algorithm is used to process the labeled source domain feature set obtained in step 1 to obtain the feature importance FI of each statistical feature data.

5. The bearing fault diagnosis method based on joint distribution adaptation according to claim 1, characterized in that: In step 2, the feature transferability quantitative index RFK of each statistical feature data is the ratio of the feature importance FI of the statistical feature data to the KL divergence KLD.

6. The bearing fault diagnosis method based on joint distribution adaptation according to claim 1, characterized in that: In step 2, according to the feature transferability quantitative index RFK from large to small, the same number of statistical feature data are selected from the source domain feature set and the target domain feature set, and the labeled source domain feature sample set D is constructed using the statistical feature data selected from the source domain feature set. S , construct an unlabeled target domain feature sample set D using statistical feature data selected from the target domain feature set T .

7. The bearing fault diagnosis method based on joint distribution adaptation according to claim 1, characterized in that: In step 3, the neighborhood preserving embedding algorithm NPE is used to embed the source domain feature sample set D obtained in step 2. S The statistical feature data in is processed to realize manifold subspace learning.

8. The bearing fault diagnosis method based on joint distribution adaptation according to claim 1, characterized in that: In step 3, the source domain feature subset Z is established S , target domain feature subset Z T The maximum mean difference MMD matrix of the marginal probability distribution between the statistical feature data in the source domain and the source domain feature subset Z S , target domain feature subset Z T The maximum mean difference MMD matrix of the conditional probability distribution between the statistical feature data in the marginal probability distribution is obtained; the maximum mean difference MMD matrix of the conditional probability distribution and the maximum mean difference MMD matrix of the conditional probability distribution are both minimized as the optimization goal, and the joint distribution adaptation algorithm JDA model is established and solved.

9. An electronic device comprising a processor and a memory, wherein a storage medium of the memory stores program instructions, wherein: When the program instructions are read and executed, steps 1 to 4 of the bearing fault diagnosis method based on joint distribution adaptation according to any one of claims 1 to 8 are executed.

10. A storage medium storing program instructions that can be read and executed, characterized in that: When the program instructions are read and executed, steps 1 to 4 of the bearing fault diagnosis method based on joint distribution adaptation according to any one of claims 1 to 8 are executed.

Images