Rolling bearing unknown fault diagnosis method based on frequency characteristic analytic hierarchy process

Abstract

The invention relates to the technical field of rolling bearing unknown fault diagnosis, and discloses a rolling bearing unknown fault diagnosis method based on frequency characteristic analytic hierarchy process, which comprises the following steps: S1, frequency division characteristic extraction: extracting frequency division characteristics based on an input original signal, namely a vibration signal of a rolling bearing; feature representations of different frequencies are obtained through TQWT frequency division feature extraction based on Bayesian optimization, the feature representations of the different frequencies serve as input of different frequency branches, and a high-frequency branch, a low-frequency branch and an original branch are constructed; s2, establishing a frequency characteristic hierarchical relationship; and S3, weighted fusion based on information entropy: finally outputting a clear and specific diagnosis conclusion. Discriminative features can be extracted from high-frequency, low-frequency and original signals, and more comprehensive and robust feature representation is constructed; according to the method, the hierarchical relationship between the frequency features can be constructed, high-frequency and low-frequency features are used as assistance, learning of the trunk features of the original signals is enhanced, and fuzziness and uncertainty in the features are eliminated.

Description

Technical Field

[0001] This application belongs to the field of unknown fault diagnosis technology for rolling bearings, and specifically relates to a method for diagnosing unknown faults in rolling bearings based on frequency feature hierarchical analysis. Background Technology

[0002] Rolling bearing fault detection can identify bearing failures at an early stage and prevent them from developing further, making it widely used in machinery, automotive, and nuclear power industries. However, due to the high safety and reliability requirements of the equipment, the number of failure samples is scarce.

[0003] Currently, many deep learning-based fault diagnosis methods have been researched to address this problem. However, these methods do not consider how the model identifies unknown faults when new faults occur. Furthermore, these methods suffer from the following issues: First, the discriminative features of vibration signals in multiple frequency domains have not been fully explored. For example, high-frequency components can effectively reflect local details such as transient impacts caused by faults, while low-frequency components carry more information about the overall system operating state and macroscopic fault trends. Most existing fault diagnosis methods only use raw time-domain or frequency-domain signals for learning, failing to systematically separate and enhance these complementary features distributed across different frequency bands; Second, there is no explicit modeling of the hierarchical relationships between different frequency features. For example, during feature learning, there is an inherent contextual dependency between the original signal features, high-frequency derived features, and low-frequency derived features. Local anomalies in high frequencies and macroscopic trends in low frequencies jointly determine the true category of the fault. Existing methods lack explicit modeling of this hierarchical relationship between frequencies, resulting in relatively isolated feature learning for each frequency branch. This prevents information synergy and complementarity, thus limiting the model's ability to generalize to complex and unknown faults.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To address or at least alleviate one or more of the above problems, a method for diagnosing unknown faults in rolling bearings based on frequency feature hierarchical analysis is provided. This method can extract discriminative features from high-frequency, low-frequency, and original signals to construct a more comprehensive and robust feature representation. It can also construct hierarchical relationships between frequency features, using high-frequency and low-frequency features as aids to enhance the learning of the main features of the original signal and eliminate ambiguity and uncertainty in the features.

[0006] To achieve the above objectives, according to the first aspect of this application, a method for diagnosing unknown faults in rolling bearings based on frequency characteristic hierarchical analysis is provided, comprising the following steps: S1. Frequency-division feature extraction: Based on the input raw signal, i.e. the vibration signal of the rolling bearing, feature representations of different frequencies are obtained by TQWT frequency division feature extraction based on Bayesian optimization. The feature representations of different frequencies are used as inputs to different frequency branches to construct high-frequency branches, low-frequency branches and the original branch. S2. Frequency Feature Hierarchical Relationship Modeling: Each frequency branch obtains a feature representation of a different frequency through feature encoding. The similarity matrix between the high-frequency branch and the original branch and the similarity matrix between the low-frequency branch and the original branch are calculated respectively to obtain two hierarchical relationship matrices. The two hierarchical relationship matrices are directly applied to the feature representation of the original branch to realize the context enhancement of the original branch features by high-frequency and low-frequency features. S3. Weighted fusion based on information entropy: The feature information entropy of each frequency branch is calculated separately. A fusion weight is generated based on the calculated feature information entropy. The fusion weight is used to achieve the final feature fusion. Finally, a clear and specific diagnostic conclusion is output, that is, a specific and newly discovered fault category label is assigned to the input and unknown vibration signal.

[0007] To achieve the above objectives, according to a second aspect of this application, a rolling bearing unknown fault diagnosis system based on frequency feature hierarchical analysis is provided, the rolling bearing unknown fault diagnosis system comprising: The frequency division feature extraction module is used to extract feature representations of different frequencies based on the input original signal, i.e., the vibration signal of the rolling bearing, through Bayesian-optimized TQWT frequency division feature extraction. The feature representations of different frequencies are used as inputs to different frequency branches to construct high-frequency branches, low-frequency branches and original branches. The frequency feature hierarchical relationship modeling module is used to obtain feature representations of different frequencies for each frequency branch through feature encoding, calculate the similarity matrix between the high-frequency branch and the original branch and the similarity matrix between the low-frequency branch and the original branch respectively, obtain two hierarchical relationship matrices, and directly apply the two hierarchical relationship matrices to the feature representation of the original branch to realize the context enhancement of the original branch features by high-frequency and low-frequency features. The weighted fusion module based on information entropy is used to calculate the feature information entropy of each frequency branch, generate fusion weights based on the calculated feature information entropy, use the fusion weights to achieve final feature fusion, and finally output a clear and specific diagnostic conclusion, that is, to assign a specific and newly discovered fault category label to the input, unknown vibration signal.

[0008] To achieve the above objectives, according to a third aspect of this application, a computer-readable storage medium is provided, storing a computer program that, when executed by a processor, implements the previously described method for diagnosing unknown faults in rolling bearings based on frequency feature hierarchical analysis.

[0009] By adopting the above technical solution, this application has the following beneficial effects compared with the prior art: This application simultaneously explores the complementary influence of high-frequency, low-frequency, and original signals on fault categories. A rolling bearing unknown fault diagnosis framework based on frequency feature hierarchical analysis adaptively extracts discriminative features of the signal in the high-frequency, low-frequency, and original domains using frequency-division feature extraction. Hierarchical similarity relationships are established between high-frequency features and original features, and between low-frequency features and original features, through frequency feature hierarchical relationship modeling. In this modeling, the hierarchical relationship matrix is ​​calculated, and information from high-frequency and low-frequency features is explicitly embedded into the original features, achieving interactive enhancement of features from different frequency branches. This fully utilizes the complementarity of multi-band information to comprehensively improve the ability to identify both known and unknown faults.

[0010] This application focuses on the main discriminative features related to faults in each frequency branch. Through entropy-based weighted fusion, the uncertainty of features in each frequency branch is quantified, and fusion weights are generated accordingly. Entropy-based weighted fusion draws on the inverse relationship between entropy and certainty in information theory, evaluating the discriminative ability of features by calculating their information entropy and assigning greater weight to features with high certainty and strong discriminative power. Weighted fusion according to the relative certainty of features in each branch effectively suppresses the interference of uncertain features in each frequency branch, highlighting the causal features most relevant to faults, enabling more accurate identification of fault categories, and exhibiting stronger robustness, particularly in the detection of unknown faults.

[0011] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. Attached Figure Description

[0012] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application. The illustrative embodiments and descriptions of the application are used to explain the application, but do not constitute an undue limitation of the application. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0013] In the attached diagram: Figure 1 This is a flowchart illustrating the method for diagnosing unknown faults in rolling bearings based on hierarchical analysis of frequency characteristics in this specific embodiment. Figure 2This is a logical diagram of the rolling bearing unknown fault diagnosis method based on frequency feature hierarchical analysis in this specific embodiment. Figure 3 This is a time-domain waveform diagram of the vibration acceleration as a function of time for normal data in this specific embodiment. Figure 4 This is a time-domain waveform diagram showing the vibration acceleration over time for a 0.007-inch inner ring fault in this specific embodiment. Figure 5 This is a time-domain waveform diagram showing the vibration acceleration over time for a 0.014-inch inner ring fault in this specific embodiment. Figure 6 This is a time-domain waveform diagram showing the vibration acceleration as a function of time for a 0.021-inch inner ring fault in this specific embodiment. Figure 7 This is a schematic diagram of the rolling bearing unknown fault diagnosis system based on frequency feature hierarchical analysis in this specific embodiment. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0015] The unknown fault diagnosis task aims to learn knowledge from labeled datasets with known categories and then use this knowledge to identify new categories in unlabeled datasets. Unlike traditional multi-class classification tasks, in this setting, the unlabeled dataset may contain unlabeled samples of known categories as well as samples of new fault categories. That is, the categories in the unlabeled dataset may overlap with the categories in the labeled dataset, or they may be entirely new categories. Specifically, this method divides the dataset into labeled and unlabeled datasets. and unlabeled datasets .in, and These are labeled datasets. and unlabeled datasets Data samples in It is a labeled dataset In the dataset, the class labels are given, and N and M are the number of samples in the labeled and unlabeled datasets, respectively. In this context, the unlabeled dataset... The samples in the dataset may belong to either a known class or an unknown class. An effective method is needed to utilize the labeled dataset. Knowledge discovery in unlabeled datasets The new category of faults. The key issue is how to correctly identify these unknown categories and avoid misclassifying them into known categories.

[0016] Based on this, this application provides a rolling bearing unknown fault diagnosis method based on frequency feature hierarchical analysis. Through a learning strategy of multi-branch, multi-frequency feature hierarchical modeling, it can deeply explore the potential relationship of fault features in each frequency band and learn the contextual dependencies between frequency domain levels, so as to effectively identify and classify unknown types of faults.

[0017] Furthermore, this application provides a rolling bearing unknown fault diagnosis system based on frequency feature hierarchical analysis, comprising at least a frequency division feature extraction module, a frequency feature hierarchical relationship modeling module, and an information entropy-based weighted fusion module, to implement the rolling bearing unknown fault diagnosis method based on frequency feature hierarchical analysis provided in this application. For example, the vibration signal of the rolling bearing is input into the rolling bearing unknown fault diagnosis system for diagnostic analysis. The rolling bearing unknown fault diagnosis system executes the rolling bearing unknown fault diagnosis method based on frequency feature hierarchical analysis provided in this application, and finally outputs a clear and specific diagnostic conclusion, that is, ultimately assigning a specific, newly discovered fault category label to the input, unknown vibration signal.

[0018] Please see Figure 1 and Figure 2 This application provides a method for diagnosing unknown faults in rolling bearings based on frequency characteristic hierarchical analysis, including the following steps: S1. Frequency-division feature extraction: Based on the input raw signal, i.e. the vibration signal of the rolling bearing, feature representations of different frequencies are obtained by TQWT frequency division feature extraction based on Bayesian optimization. The feature representations of different frequencies are used as inputs to different frequency branches to construct high-frequency branches, low-frequency branches and the original branch. S2. Frequency Feature Hierarchical Relationship Modeling: Each frequency branch obtains a feature representation of a different frequency through feature encoding. The similarity matrix between the high-frequency branch and the original branch and the similarity matrix between the low-frequency branch and the original branch are calculated respectively to obtain two hierarchical relationship matrices. The two hierarchical relationship matrices are directly applied to the feature representation of the original branch to realize the context enhancement of the original branch features by high-frequency and low-frequency features. S3. Weighted fusion based on information entropy: The feature information entropy of each frequency branch is calculated separately. A fusion weight is generated based on the calculated feature information entropy. The fusion weight is used to achieve the final feature fusion. Finally, a clear and specific diagnostic conclusion is output, that is, a specific and newly discovered fault category label is assigned to the input and unknown vibration signal.

[0019] This application processes vibration signals generated during the operation of rolling bearings. The final output of this application is a clear and specific diagnostic conclusion. The final result is: assigning a specific, newly discovered fault category label to each input, unknown vibration signal. Specifically, this result manifests as: identifying new fault modes; discovering fault types that did not appear during the training phase, for example, successfully clustering unlabeled data into "New Fault Category #1" and "New Fault Category #2". Subsequent clustering results can be interpreted based on expert experience, for example, interpreting "New Fault Category #1" as "outer ring localized spalling fault" at the physical level, and "New Fault Category #2" as "inner ring and rolling element combined fault", etc.

[0020] Please see Figures 3-6 Taking the Case Western Reserve University Bearing Dataset (CWRU) as an example, the CWRU dataset provides fault and normal data related to various bearing fault types. This dataset includes 12kHz drive-end fault data, 12kHz fan-end fault data, 48kHz drive-end fault data, and normal baseline data. Bearing fault types cover ball bearing faults, inner race faults, outer race faults, and normal conditions. Each fault type is further subdivided into three different fault severity levels, corresponding to fault sizes of 0.007, 0.014, and 0.021 inches, respectively. Therefore, this dataset contains 10 independent labels. Furthermore, each fault category corresponds to different horsepower values: 1 hp, 2 hp, and 3 hp, associated with motor speeds of 1772 RPM, 1750 RPM, and 1730 RPM, respectively. To facilitate experimental analysis, this embodiment divides the dataset into two subsets: a labeled set, containing predefined, labeled data classes; and an unlabeled set, containing unlabeled data of unknown and known categories.

[0021] It should be noted that the executing entity of the rolling bearing unknown fault diagnosis method based on frequency feature hierarchical analysis in this embodiment is a rolling bearing unknown fault diagnosis system based on frequency feature hierarchical analysis. This system can be an electronic device, a component in an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, etc., and non-mobile electronic devices can be servers and personal computers, etc., which are not specifically limited in this application. The following description uses a server as the executing entity to illustrate the rolling bearing unknown fault diagnosis method based on frequency feature hierarchical analysis in this embodiment.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0023] In some embodiments, S1, frequency division feature extraction: based on the input original signal, i.e. the vibration signal of the rolling bearing, the feature representation of different frequencies is obtained by TQWT frequency division feature extraction based on Bayesian optimization, and the feature representation of different frequencies is used as the input of different frequency branches to construct high frequency branch, low frequency branch and original branch.

[0024] The vibration signals of rolling bearings exhibit complex spectral characteristics, with signals in different frequency bands reflecting potential information about different types of faults. Due to the diverse distribution of fault features in the frequency domain, effective feature extraction targeting different frequency components of the signal is crucial. To comprehensively analyze the fault characteristics of rolling bearings across various frequency bands, this embodiment employs wavelet transform-based frequency division feature extraction with Bayesian optimization to extract more discriminative features from both high-frequency and low-frequency components.

[0025] Frequency division feature extraction based on Bayesian optimization wavelet transform can be understood through the following explanation. The TQWT wavelet transform with an adjustable Q-factor is a parameter-adjustable overcomplete wavelet transform. It can adaptively construct the optimal wavelet basis function to match the periodic impact characteristics of vibration signals by selecting appropriate parameters: quality factor Q, redundancy γ, and decomposition level J. Bayesian optimization is a global optimization strategy based on Bayesian statistical inference, well-suited for automated hyperparameter tuning in deep learning.

[0026] For example, TQWT uses the low-pass filter frequency response function and the frequency response function of the high-pass filter The formulas for obtaining the low-frequency and high-frequency components are as follows: ; ; Where α and β are the low-pass scale transformation coefficient and the high-pass scale transformation coefficient, respectively; α and β can be represented by the Q factor and redundancy γ of TQWT. , β= .

[0027] The decomposition and reconstruction of QWT are determined by three key parameters: the Q factor, redundancy γ, and the number of decomposition levels J. The maximum value of the number of decomposition levels J is [value missing]. , where N is the signal length.

[0028] The objective of this embodiment is to select a suitable set of parameter combinations x=(Q,γ,J) such that the optimization objective function f(x) reaches its maximum value, that is: ; in, This represents the parameter combination (Q, γ, J). Let f(x) be the combination of parameters that maximizes f(x), where X is the parameter space of x, and f(x) is the objective function.

[0029] For example, Bayesian optimization is typically modeled using a Gaussian process, expressed as: ; in, This represents the parameter combination (Q, γ, J). These represent two different combinations of parameters. It follows a Gaussian distribution. It is a mean function. It is the covariance function; The prediction formula for the probabilistic surrogate model, derived from Bayes' theorem, is expressed as follows: ; in, This represents the prior probability of the parameter combination x. This represents the prior probability of the objective function value f(x). Let f(x) represent the likelihood function, which expresses the probability of observing the parameter combination x given the objective function value f(x). Let f(x) represent the posterior probability, which represents the probability distribution of the objective function value f(x) given the combination of parameters x. To evaluate the location of the next sampling point, the expected improvement (EI) is used as the acquisition function. The expected improvement acquisition function aims to find the point that maximizes the improvement of the optimal objective function value, expressed as: ; in, This represents the best objective function value observed so far.

[0030] In Bayesian optimization, the optimization objective is the discriminative ability based on frequency-division features. Therefore, it is necessary to select appropriate metrics to measure the characteristics of frequency components. In this embodiment, two feature metrics are used: kurtosis (Kurt) and centroid frequency (FC). Kurtosis is highly sensitive to changes in signal vibration and can reflect the impulsive characteristics inherent in the signal; the larger the value, the more significant the impulsive characteristics, making it suitable for the analysis of high-frequency components. Centroid frequency describes the frequency of the signal components with larger components in the signal spectrum, reflecting the distribution of signal energy. It represents the main frequency components of the signal and is suitable for the analysis of low-frequency components.

[0031] For example, in the Bayesian optimization process, kurtosis Kurt and centroid frequency FC are chosen to measure the characteristics of the frequency components. Kurtosis Kurt is defined as: ; in, Indicates the kurtosis at scale kj. For subband signals at scale kj The ki-th element; Let kj be the length of the subband signal, and ki represent the traversal index of the subband signal; The average value of the sub-band signal; The variance of the subband signal; The center of gravity frequency is defined as b: ; in, This represents the frequency value corresponding to the fi-th frequency component. This represents the power spectral density (energy level) of that frequency component. Kurt kurtosis provides high-frequency impact characteristics, while the centroid frequency FC provides low-frequency main frequency components. By combining kurtosis and centroid frequency, both high-frequency impact characteristics and low-frequency energy characteristics of the signal can be considered simultaneously, thereby optimizing the performance of the frequency division feature extraction module.

[0032] For example, the objective function of the frequency division feature extraction module is: ; The objective function aims to maximize the signal's ability to distinguish high-frequency and low-frequency components by optimizing the parameter combination of TQWT, thereby providing a more accurate and reliable feature representation for subsequent fault diagnosis.

[0033] In some embodiments, S2, frequency feature hierarchical relationship modeling: each frequency branch obtains feature representations of different frequencies through feature encoding, calculates the similarity matrix between the high-frequency branch and the original branch and the similarity matrix between the low-frequency branch and the original branch respectively, and obtains two hierarchical relationship matrices. The two hierarchical relationship matrices are directly applied to the feature representation of the original branch to realize the context enhancement of the original branch features by high-frequency and low-frequency features.

[0034] To fully explore the potential complementary relationships between different frequency features, this embodiment proposes a frequency feature hierarchical relationship modeling module. This module explicitly models the hierarchical association between high-frequency, low-frequency, and original branches, strengthening the interaction of different frequency signal features and further improving the accuracy of unknown fault identification. Specifically, this module first uses the input of the frequency-division feature extraction module to obtain the feature representations of the high-frequency branch, the original branch, and the low-frequency branch, and constructs a mapping relationship between frequency features and fault categories in the three branches. However, if the features of each frequency branch are learned independently, the lack of explicit modeling of the contextual information between different frequencies will prevent the original branch from fully utilizing the auxiliary information of the high-frequency and low-frequency branches. Therefore, this embodiment designs a frequency feature hierarchical relationship modeling module to explicitly establish and utilize the hierarchical relationship between different frequency features to achieve information collaboration and contextual interaction between frequency branches.

[0035] In this embodiment, the definition of hierarchical relationships follows two principles: First, the principle of information directionality. The original signal is the fundamental signal, representing the overall operational information; high-frequency and low-frequency signals, as derived signals, highlight local and macroscopic information. Therefore, the original signal should be located at the center of the frequency hierarchy, bidirectionally assisted by high-frequency and low-frequency signals (i.e., high-frequency points to the original, and low-frequency points to the original). Second, the principle of complementarity. High-frequency and low-frequency signals provide different types of information and should serve as auxiliary information sources for the expression of the original signal's characteristics. High-frequency signals help detect abrupt changes and local faults, while low-frequency signals help detect overall macroscopic faults; both enhance different feature dimensions in the original signal. Therefore, the hierarchical relationship in this embodiment is based on the central position of the original signal and the influence of high-frequency and low-frequency signals on the original signal.

[0036] Specifically, each frequency branch uses an independent feature encoder based on a ResNet network, i.e., a high-frequency branch encoder. Original branch encoder and low-frequency branch encoder For example, for the i-th group of input samples, the high-frequency feature representations are obtained respectively. Original feature representation and low-frequency feature representation To enable contextual interaction between frequency branches, the predictions of the high-frequency / low-frequency branches are used as latent variables to aid in the prediction of the target hierarchical signal. This allows for the explicit establishment of similarity relationships across hierarchical structures and ensures consistency in classification context. This hierarchical relationship enables the high-frequency / low-frequency branches to help predict the target signal category.

[0037] Specifically, taking the high-frequency branch as an example, the similarity matrix between the high-frequency branch and the original branch is calculated. , is represented as: ; Taking the low-frequency-original branch as an example, the similarity matrix between the low-frequency branch and the original branch is calculated as follows: , is represented as: ; Subsequently, the hierarchical relationship matrix is ​​obtained through Softmax normalization, clarifying the relative contributions of high-frequency and low-frequency signals to the features of the original signal, expressed as: ; ; in, , These represent the hierarchical relationship between the high-frequency branch and the original branch, and the hierarchical relationship between the low-frequency branch and the original branch, respectively.

[0038] Normalized hierarchical relationship matrix and These represent the relative importance of high-frequency and low-frequency signals to the correlation between samples in the original signal features, respectively. Through matrix multiplication, the two hierarchical relationship matrices are directly applied to the feature representation of the original branch, thereby achieving contextual enhancement of the original branch features by high-frequency and low-frequency features, expressed as:

[0039] in, This represents the enhanced original branch features.

[0040] After the above processing, frequency context information is explicitly embedded into the original branch features, significantly enhancing the sensitivity of the original features to local minor faults and overall macroscopic abnormal states. Unlike the tree-like or category-hierarchical relationship modeling methods commonly used in previous studies, this embodiment explicitly constructs hierarchical relationships between frequency signals from the perspective of feature interaction, rather than category-based inclusion relationships. This feature-dimensional frequency hierarchical relationship modeling method provides a more flexible and comprehensive information interaction method, significantly improving the model's accuracy and generalization performance in diagnosing unknown category faults.

[0041] In some embodiments, S3, weighted fusion based on information entropy: calculate the feature information entropy of each frequency branch, generate fusion weights based on the calculated feature information entropy, and use the fusion weights to achieve final feature fusion.

[0042] First, calculate the feature information entropy for each frequency branch, as follows: ; in, This represents the probability after feature normalization, where ei and n represent the dimension index and the total number of dimensions of the feature vector, respectively. Secondly, fusion weights are generated based on the calculated feature information entropy. The lower the feature information entropy value, the larger the obtained fusion weight, as expressed as: ; in, The fusion weights are represented by j, and the scale is represented by j. , These represent the information entropy values ​​of each frequency branch feature; Finally, the final feature fusion is achieved using fusion weights, expressed as: ; in, This indicates the final fusion characteristics. , These represent the fusion weights of the high-frequency branch and the low-frequency branch, respectively.

[0043] To further enhance the richness and robustness of the original branch features, this embodiment utilizes information entropy to quantify the uncertainty of the features extracted from each frequency branch, and further dynamically fuses the features of high-frequency and low-frequency branches from the perspective of discriminative ability, so as to enhance the expression of the original features.

[0044] Although both the frequency feature hierarchical relationship modeling module and the information entropy-based weighted fusion module involve frequency feature fusion, they are not contradictory or redundant in essence and function, but rather complementary. The frequency signal hierarchical relationship modeling module enhances the richness and sensitivity of the features themselves by explicitly modeling the contextual relationship between high-frequency and low-frequency features and the original feature learning. Meanwhile, the information entropy-based weighted fusion module dynamically determines the final fusion weights from the perspective of statistical discriminative ability after feature learning, selectively fusing features from each frequency branch to further improve the accuracy of fault diagnosis. The former addresses how to learn features with high quality, while the latter addresses how to efficiently use the learned features. Together, they form a complete closed loop of feature learning and feature fusion, enabling the model to achieve better performance in unknown fault identification tasks.

[0045] Based on the same inventive concept, please see Figure 7This application also provides a rolling bearing unknown fault diagnosis system based on frequency feature hierarchical analysis, the rolling bearing unknown fault diagnosis system comprising: The frequency division feature extraction module is used to extract feature representations of different frequencies based on the input original signal, i.e., the vibration signal of the rolling bearing, through Bayesian-optimized TQWT frequency division feature extraction. The feature representations of different frequencies are used as inputs to different frequency branches to construct high-frequency branches, low-frequency branches and original branches. The frequency feature hierarchical relationship modeling module is used to obtain feature representations of different frequencies for each frequency branch through feature encoding, calculate the similarity matrix between the high-frequency branch and the original branch and the similarity matrix between the low-frequency branch and the original branch respectively, obtain two hierarchical relationship matrices, and directly apply the two hierarchical relationship matrices to the feature representation of the original branch to realize the context enhancement of the original branch features by high-frequency and low-frequency features. The weighted fusion module based on information entropy is used to calculate the feature information entropy of each frequency branch, generate fusion weights based on the calculated feature information entropy, use the fusion weights to achieve final feature fusion, and finally output a clear and specific diagnostic conclusion, that is, to assign a specific and newly discovered fault category label to the input, unknown vibration signal.

[0046] Based on the same inventive concept, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the previously described method for diagnosing unknown faults in rolling bearings based on frequency characteristic hierarchical analysis.

[0047] The program product of this application for implementing the above method may employ a portable compact disk read-only memory and include program code, and may run on a terminal device, such as a personal computer. However, the program product of this application is not limited thereto. In this application, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0048] It should be noted that a computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0049] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although this application has disclosed preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-mentioned technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of this application.

Claims

1. A method for unknown fault diagnosis of rolling bearings based on frequency feature hierarchical analysis, characterized in that, The method comprises the following steps: S1, frequency division feature extraction: Based on the input original signal, i.e., the vibration signal of the rolling bearing, different frequency feature representations are obtained through Bayesian optimization-based TQWT frequency division feature extraction, the different frequency feature representations are taken as inputs of different frequency branches, and a high-frequency branch, a low-frequency branch, and an original branch are constructed; S2, frequency feature hierarchical relationship modeling: Each of the frequency branches obtains different frequency feature representations through feature encoding, similarity matrices of the high-frequency branch and the original branch and similarity matrices of the low-frequency branch and the original branch are calculated respectively, two hierarchical relationship matrices are obtained, and the two hierarchical relationship matrices are directly applied to the feature representation of the original branch to realize context enhancement of high-frequency and low-frequency features on the original branch features; S3, information entropy-based weighted fusion: The feature information entropy of each of the frequency branches is calculated, a fusion weight is generated according to the calculated feature information entropy, the fusion weight is used to realize final feature fusion, and a clear and specific diagnosis conclusion is finally output, i.e., a specific and newly discovered fault category label is finally assigned to the input unknown vibration signal.

2. The method of claim 1, wherein, Based on the input original signal, i.e., the vibration signal of the rolling bearing, different frequency feature representations are obtained through Bayesian optimization-based TQWT frequency division feature extraction, including: TQWT uses a low-pass filter frequency response function and a high-pass filter frequency response function to obtain low-frequency components and high-frequency components, respectively, the decomposition and reconstruction of TQWT are determined by three parameters: Q factor, redundancy γ and decomposition layer number J, a set of suitable parameter combinations x=(Q, γ, J) is selected to make the objective function f(x) reach the maximum value, which is represented as: ； wherein denotes a parameter combination (Q, γ, J), denotes a parameter combination at which f(x) takes a maximum value, and X is a parameter space.

3. The method of claim 2, wherein, Based on the input original signal, i.e., the vibration signal of the rolling bearing, different frequency feature representations are obtained through Bayesian optimization-based TQWT frequency division feature extraction, further including: Bayesian optimization usually adopts Gaussian process modeling, which is represented as: ； wherein denotes a parameter combination (Q, γ, J), denote two different parameter combinations, respectively, is a Gaussian distribution, is a mean function, is a covariance function; According to the Bayesian formula, the prediction formula of the probability agent model is obtained, which is represented as: ； wherein denotes the prior probability of the parameter combination x, denotes the prior probability of the objective function value f(x), denotes the likelihood function, which denotes the probability of observing the parameter combination x given the objective function value f(x), denotes the posterior probability, which denotes the probability distribution of the objective function value f(x) given the parameter combination x; The expected improvement EI is used as the acquisition function, which is represented as: ； where, is the current observed best objective function value.

4. The method according to any one of claims 1 to 3, characterized in that, Based on the input original signal, i.e., the vibration signal of the rolling bearing, different frequency feature representations are obtained through Bayesian optimization-based TQWT frequency division feature extraction, further including: In the Bayesian optimization process, the kurtosis Kurt and the center of gravity frequency FC are selected to measure the features of the frequency components, the kurtosis Kurt is defined as: ； wherein denotes the kurtosis at scale kj, is the ki-th element of the subband signal at scale kj; is the length of the subband signal at scale kj, ki denotes the traversal index of the subband signal; is the mean of the subband signal; is the variance of the subband signal;​ The center of gravity frequency is defined as: ； wherein denotes the frequency value corresponding to the fi th frequency component, denotes the power spectral density of this frequency component; The kurtosis Kurt provides high-frequency impact features, and the center of gravity frequency FC provides low-frequency main frequency components, and the performance of the Bayesian optimization-based TQWT frequency division feature extraction is optimized through the combination of the kurtosis and the center of gravity frequency.

5. The method of claim 4, wherein, Based on the input original signal, i.e., the vibration signal of the rolling bearing, different frequency feature representations are obtained through Bayesian optimization-based TQWT frequency division feature extraction, including: ； wherein, denotes kurtosis, denotes the center of gravity frequency, the objective function maximizes the discriminative power of the signal on high and low frequency components by optimizing the parameter combination of the TQWT.

6. The method of claim 3, wherein, Each of the frequency branches obtains different frequency feature representations through feature encoding, similarity matrices of the high-frequency branch and the original branch and similarity matrices of the low-frequency branch and the original branch are calculated respectively, including: Each frequency branch uses an independent ResNet network based feature encoder, i.e., a high frequency branch encoder , an original branch encoder , and a low frequency branch encoder , to obtain a high frequency feature representation , an original feature representation , and a low frequency feature representation , respectively.

7. The method of claim 6, wherein, Each of the frequency branches obtains different frequency feature representations through feature encoding, similarity matrices of the high-frequency branch and the original branch and similarity matrices of the low-frequency branch and the original branch are calculated respectively, further including: Computing a similarity matrix of high frequency branches and original branches is represented as: ； The similarity matrix of the low-frequency branch and the original branch is calculated as is expressed as: ； Subsequently, the hierarchical relationship matrix is obtained through Softmax normalization processing, which is represented as: ； ； wherein, , respectively represent the hierarchical relationship between the high-frequency branch and the original branch, and the hierarchical relationship between the low-frequency branch and the original branch. The two hierarchical relationship matrices are directly applied to the original branch feature representation through matrix multiplication, thereby realizing context enhancement of high-frequency and low-frequency features on the original branch feature, and being expressed as: ； wherein, denotes the enhanced original branch feature.

8. The method of claim 1, wherein, The feature information entropy of each frequency branch is calculated respectively, and a fusion weight is generated according to the calculated feature information entropy, and the final feature fusion is realized by using the fusion weight: First, the feature information entropy of each frequency branch is calculated respectively, and being expressed as: ； wherein, denotes the normalized probability of a feature, e, n denote the dimension index and the total dimension number of the feature vector, respectively; Second, a fusion weight is generated according to the calculated feature information entropy, and the lower the feature information entropy value is, the greater the obtained fusion weight is, and being expressed as: ； wherein, denotes a fusion weight, j denotes a scale, , respectively denote information entropy values of respective frequency branch features; Finally, the final feature fusion is realized by using the fusion weight, and being expressed as: ； wherein, denotes the final fused feature, , denote the fusion weights for the high and low frequency branches, respectively.

9. A rolling bearing unknown fault diagnosis system based on frequency feature hierarchical analysis, characterized in that, The unknown fault diagnosis system of the rolling bearing comprises: The frequency division feature extraction module is used for obtaining different frequency feature representations by TQWT frequency division feature extraction based on Bayesian optimization based on the input original signal, i.e., the vibration signal of the rolling bearing, taking the different frequency feature representations as the input of different frequency branches, and constructing a high-frequency branch, a low-frequency branch and an original branch; The frequency feature hierarchical relationship modeling module is used for obtaining different frequency feature representations by feature coding for each frequency branch, calculating a similarity matrix of the high-frequency branch and the original branch and a similarity matrix of the low-frequency branch and the original branch respectively, obtaining two hierarchical relationship matrices, and directly applying the two hierarchical relationship matrices to the feature representation of the original branch, thereby realizing context enhancement of high-frequency and low-frequency features on the original branch feature. The information entropy-based weighted fusion module is used for calculating the feature information entropy of each frequency branch respectively, generating a fusion weight according to the calculated feature information entropy, realizing the final feature fusion by using the fusion weight, and finally outputting an explicit and specific diagnosis conclusion, i.e., finally assigning a specific and newly discovered fault class label to the input unknown vibration signal.

10. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to realize the rolling bearing unknown fault diagnosis method based on frequency feature hierarchical analysis according to any one of claims 1-8. The computer program is executed by the processor to realize the rolling bearing unknown fault diagnosis method based on frequency feature hierarchical analysis according to any one of claims 1-8.

Images