Rolling bearing fault diagnosis method, device and equipment based on digital-analog driving

Abstract

The embodiment of the invention provides a rolling bearing fault diagnosis method, device and equipment based on digital-analog driving. According to the method, vibration signals of the rolling bearing under different working conditions are collected, the vibration signals are preprocessed and then input into a trained multi-source domain fault diagnosis model, and a fault diagnosis result of the rolling bearing is obtained. The training process of the multi-source-domain fault diagnosis model comprises the following steps: generating simulation data according to a rolling bearing dynamical model, and forming a multi-source-domain data set together with measured data; performing multi-scale feature extraction on the multi-source domain data set, and performing feature dynamic matching; identifying and discarding a specific domain sensitive channel by using a domain sensitive feature suppression strategy; a classification aggregation loss function is introduced, and representation learning with invariant cross-domain class conditions is realized on a feature space level; loss back propagation is carried out according to dynamic matching feature loss, domain sensitive feature suppression loss and classification aggregation loss, model parameters are optimized, and the accuracy and reliability of a rolling bearing fault diagnosis result are improved.

Description

Technical Field

[0001] The present application relates to the technical field of mechanical fault diagnosis, and in particular to a rolling bearing fault diagnosis method, device and equipment based on digital-analog drive. Background Art

[0002] As a key component in rotating machinery, the operating condition of rolling bearings is directly related to the reliability and safety of the entire mechanical system. In practice, rolling bearings often need to operate continuously under complex conditions such as high speeds, heavy loads, and variable operating conditions, resulting in a significantly higher failure rate than other mechanical components. Therefore, developing effective fault diagnosis and health monitoring methods is crucial for ensuring and maintaining the safety of rotating machinery.

[0003] With the rapid advancement of artificial intelligence algorithms, fault diagnosis methods based on deep learning have been proposed. By constructing a multi-layer neural network model, rolling bearing fault diagnosis methods based on deep learning can automatically extract deep-level fault features from multi-source monitoring data such as vibration and noise, and classify and locate fault types. However, fault diagnosis methods based on deep learning are highly dependent on a large number of labeled samples. In scenarios where labeled fault samples are scarce, the performance of deep learning fault diagnosis models may significantly decline. When the distribution of training data and test data is inconsistent (such as across working conditions and cross-equipment scenarios), the insufficient generalization ability of deep learning fault diagnosis models makes it difficult to adapt to the actual needs of cross-domain diagnosis.

[0004] To address the poor generalization capabilities of deep learning-based fault diagnosis methods when data distribution is uneven, a rolling bearing fault diagnosis method based on transfer learning was proposed. This method effectively transfers fault diagnosis knowledge from the source domain dataset to the target domain by constructing a cross-domain knowledge transfer mechanism, thereby improving the adaptability and generalization performance of the fault diagnosis model under different operating conditions. However, traditional transfer learning methods typically require the fault label spaces of the source and target domains to be consistent, resulting in insufficient recognition of unknown fault types in the target domain. In scenarios with insufficient fault sample data, the model's ability to identify fault samples in the target domain is insufficient, significantly reducing fault diagnosis accuracy. Furthermore, the effectiveness of knowledge transfer is highly dependent on the distribution similarity between the source and target domains. When the sample space feature distribution differences between the source and target domains are too large, negative transfer may occur, resulting in reduced fault diagnosis performance. Summary of the Invention

[0005] The present application provides a rolling bearing fault diagnosis method, device and equipment based on digital analog drive to solve the problems of sample imbalance, data scarcity and insufficient model generalization ability in existing fault diagnosis methods.

[0006] In a first aspect, the present application provides a rolling bearing fault diagnosis method based on digital-analog drive, the method comprising:

[0007] The vibration signals of rolling bearings under different working conditions are collected. After fast Fourier transform and sliding window segmentation, the collected vibration signals are input into the trained multi-source domain fault diagnosis model for analysis and processing to obtain the fault diagnosis results of the rolling bearings.

[0008] The training process of the multi-source domain generalized fault diagnosis model includes:

[0009] Obtain real fault data;

[0010] Generate simulation fault data based on rolling bearing dynamics model;

[0011] Perform fast Fourier transform and sliding window segmentation processing on the real fault data and the simulated fault data to obtain the pre-processed real fault data and the simulated fault data;

[0012] A multi-source domain dataset is constructed based on the pre-processed real fault data and simulated fault data. The multi-source domain dataset includes: a source domain and a target domain. Each source domain includes real fault data and simulated fault data, and the target domain is the vibration signal to be diagnosed.

[0013] Perform multi-scale feature extraction on multi-source domain datasets to obtain source domain features and target domain features;

[0014] The domain discriminator is used to identify domain-sensitive channels in domain features, and the probabilistic channel discarding mechanism is used to shield domain-sensitive channels.

[0015] By introducing the classification aggregation loss function, the model is trained across domain classes at the feature space level based on the extracted source domain features.

[0016] Use the fault classifier to classify and predict the target domain features based on the cross-domain class learning results;

[0017] The total loss is determined according to the feature extraction loss, domain adaptation loss, classification aggregation loss and classification prediction loss. The parameters of the multi-source domain generalized fault diagnosis model are adjusted through the back propagation algorithm until the total loss is less than the preset loss value, and the trained multi-source domain generalized fault diagnosis model is obtained.

[0018] Optionally, a simulated vibration signal is generated based on a rolling bearing dynamics model, including:

[0019] The dynamic models of local defects of the inner ring, outer ring and rolling element of the rolling bearing are established respectively;

[0020] Based on Hertz contact theory, calculate the time-varying displacement excitation generated when the rolling element contacts the inner and outer rings of the rolling bearing;

[0021] Based on the established dynamic model and the calculated time-varying displacement excitation, the nonlinear dynamic differential equation is constructed;

[0022] The constructed nonlinear dynamic differential equation is solved by the fourth-order Runge-Kutta numerical integration method to obtain the simulated fault data.

[0023] Optionally, multi-scale feature extraction is performed on multi-source domain datasets to obtain source domain features and target domain features, including:

[0024] Construct a multi-source domain dynamic matching feature extraction module, which includes: a convolutional feature extraction module and a dynamic feature extraction module. The convolutional feature extraction module consists of multiple convolutional layers, and the dynamic feature extraction module consists of a multi-scale block, a first cascade layer, an adaptive average pooling layer, a convolution layer, a softmax layer, multiple identical independent convolutional layers, and a second cascade layer.

[0025] Performing feature extraction on the source domain data and the target domain data through a convolutional feature extraction module to obtain a first source domain feature and a first target domain feature;

[0026] Extract multi-scale features of source domain data and target domain data through the multi-scale blocks of the dynamic feature extraction module;

[0027] The multi-scale features extracted by the multi-scale blocks are processed in sequence by the first cascade layer, the adaptive average pooling layer, and the convolution layer to obtain multiple dynamic matching coefficients, which are then normalized using the Softmax layer.

[0028] Assigning each dynamic matching coefficient to the corresponding convolutional layer to obtain the second source domain feature and the second target domain feature;

[0029] A source domain feature and a target domain feature are determined according to the first source domain feature, the first target domain feature, the second source domain feature, and the second target domain feature.

[0030] Optionally, domain-sensitive channels in domain features are identified through a domain discriminator, and domain-sensitive channels are shielded using a probabilistic channel discarding mechanism, including:

[0031] The domain classification prediction of each source domain feature is performed through the domain discriminator to obtain the domain sensitivity score;

[0032] Calculate the contribution of each feature channel to domain classification and generate channel importance weight;

[0033] According to the domain sensitivity score and channel importance weight, a probabilistic discarding mechanism is used to dynamically shield high domain sensitive feature channels.

[0034] Optionally, a classification aggregation loss function is introduced to perform cross-domain learning on the model in the feature space based on the extracted source domain features, including:

[0035] Maximize the mutual information between samples of the same category in different source domains through classification aggregation loss function;

[0036] The loss is calculated using the standard form of contrastive learning, where pairs of samples of the same category from different source domains are considered positive pairs, and pairs of samples of different categories are considered negative pairs.

[0037] By minimizing the contrastive learning loss, the model learns class-conditional invariant representations across multiple source domains.

[0038] Optionally, a total loss is determined based on the feature extraction loss, domain adaptation loss, classification aggregation loss, and classification prediction loss, and the parameters of the multi-source domain generalized fault diagnosis model are adjusted through a backpropagation algorithm, including:

[0039] Determine a classification prediction loss, where the classification prediction loss indicates the difference between the prediction result and the true label;

[0040] determining a feature extraction loss, wherein the feature extraction loss is used to constrain the distribution characteristics of the generated source domain features and target domain features;

[0041] determining a domain adaptation loss, wherein the domain adaptation loss is used to reduce the feature distribution difference between the source domain and the target domain;

[0042] Determine a classification aggregation loss that maximizes the mutual information of features of similar categories in different domains and suppresses environmental variation information that is irrelevant to the task;

[0043] Construct a total loss function based on classification prediction loss, feature extraction loss, domain adaptation loss, and classification aggregation loss;

[0044] Through the back propagation algorithm, the parameters of the multi-source domain generalized fault diagnosis model are optimized based on the total loss function to realize the training of the multi-source domain generalized fault diagnosis model.

[0045] In a second aspect, the present application provides a rolling bearing fault diagnosis device based on digital-analog drive, comprising:

[0046] The first processing module is used to collect vibration signals of the rolling bearing under different working conditions;

[0047] The second processing module is used to input the collected vibration signal into the trained multi-source domain fault diagnosis model for analysis and processing after fast Fourier transform and sliding window segmentation to obtain the rolling bearing fault diagnosis result.

[0048] In a third aspect, the present application provides a rolling bearing fault diagnosis device based on digital-analog drive, comprising:

[0049] Memory;

[0050] processor;

[0051] wherein the memory stores computer-executable instructions;

[0052] The processor executes the computer-executable instructions stored in the memory to implement the rolling bearing fault diagnosis method based on digital-analog drive as described in the above-mentioned first aspect and various possible implementation methods of the first aspect.

[0053] In a fourth aspect, the present application provides a computer storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the rolling bearing fault diagnosis method based on digital-analog drive as described in the first aspect and various possible implementation methods of the first aspect.

[0054] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the rolling bearing fault diagnosis method based on digital-analog drive as described in the first aspect and various possible implementation methods of the first aspect.

[0055] The present application provides a rolling bearing fault diagnosis method, device and equipment based on digital analog drive. The method collects the vibration signals of the rolling bearing under different working conditions, and inputs them into the trained multi-source domain fault diagnosis model after fast Fourier transform and sliding window segmentation preprocessing to obtain the fault diagnosis results of the rolling bearing. The training process of the multi-source domain fault diagnosis model includes: constructing a local defect dynamics model of the rolling bearing to generate simulation data, which together with the measured data constitutes a multi-source domain data set; extracting multi-scale features through a multi-source domain dynamic matching feature extraction module, and adopting an adaptive parameter allocation strategy to achieve dynamic feature matching; using a domain-sensitive feature suppression strategy to identify and discard specific domain-sensitive channels to reduce domain-related feature interference; introducing a classification aggregation loss function to achieve cross-domain class condition-invariant representation learning at the feature space level; performing loss backpropagation based on dynamic matching feature loss, domain-sensitive feature suppression loss and classification aggregation loss to optimize model parameters, thereby improving the accuracy and reliability of the rolling bearing fault diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0057] Figure 1 A schematic diagram of the structure of a domain-sensitive feature suppression multi-domain generalization network based on digital-analog fusion drive provided in an embodiment of the present application;

[0058] Figure 2 A flowchart illustrating the training process of a multi-source domain generalized fault diagnosis model provided in an embodiment of the present application;

[0059] Figure 3 A schematic diagram of the radial displacement of the inner ring raceway of a defect-free rolling bearing under an external radial load provided by an embodiment of the present application;

[0060] Figure 4 Schematic diagram of the waveform and envelope spectrum of the simulated signal and the measured signal provided in the embodiment of the present application under the outer race fault;

[0061] Figure 5 Schematic diagram of the waveform and envelope spectrum of the simulated signal and the measured signal provided in the embodiment of the present application under the inner race fault;

[0062] Figure 6 Schematic diagram of the waveform and envelope spectrum of the simulated signal and the measured signal provided in the embodiment of the present application under rolling element fault;

[0063] Figure 7 A schematic diagram of the confusion matrix between the predicted labels and the true labels of the diagnostic model in the experimental verification stage provided in the embodiment of the present application;

[0064] Figure 8 Schematic diagram of the feature classification results of each diagnostic model under Task A1 in the experimental verification phase provided in an embodiment of the present application.

[0065] Figure 9 A schematic diagram of the structure of a rolling bearing fault diagnosis device based on digital-analog drive provided in an embodiment of the present application;

[0066] Figure 10 A schematic structural diagram of a rolling bearing fault diagnosis device based on digital-analog drive provided in an embodiment of the present application.

[0067] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0068] To make the objectives, technical solutions, and advantages of this application more clear, the technical solutions in this application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0069] The terms "first," "second," "third," "fourth," and so forth (if any) in the description and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequential sequence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in sequences other than those illustrated or described herein.

[0070] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0071] The accuracy of traditional data-driven fault diagnosis technology is usually subject to the completeness requirements of data samples. Therefore, the engineering practice of rolling bearing fault diagnosis often faces the challenge of scarce data samples. Existing rolling bearing fault diagnosis methods driven by digital-analog fusion widely adopt a single source domain data-model fusion drive. Since the fault types of data collected under different working conditions are different, that is, some fault types only occur under certain working conditions but not under other working conditions, the application scenarios of this method in rolling bearing fault diagnosis are limited. In the multi-source domain transfer learning framework, there may be shared fault categories between the source domains, or there may be unique fault categories specific to individual source domains, which leads to the occurrence of cross-domain category transfer. Category transfer between multiple source domains will hinder the transfer of fault knowledge in cross-domain multi-source domain generalized fault diagnosis, resulting in a decrease in fault diagnosis performance.

[0072] In response to the above problems, this application proposes a rolling bearing fault diagnosis method based on digital-analog drive. The method collects the actual vibration data of the rolling bearing and constructs a multi-source domain training set based on the simulation data generated by the rolling bearing dynamics model established based on multi-body dynamics theory. The actual vibration data and simulation data are pre-processed and input into the trained multi-source domain generalized fault diagnosis model to obtain the rolling bearing fault diagnosis results.

[0073] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0074] Figure 1 Schematic diagram of the structure of a domain-sensitive feature suppression multi-domain generalization network based on digital-analog fusion drive provided in an embodiment of the present application.

[0075] The rolling bearing fault diagnosis method based on digital-analog drive provided in this embodiment includes: using an acceleration sensor to collect the real vibration signal of the rolling bearing under different operating conditions, and inputting the collected real vibration signal into the trained multi-source domain fault diagnosis model for analysis and processing after fast Fourier transform and sliding window segmentation to obtain the rolling bearing fault diagnosis result.

[0076] Figure 2 The flowchart of the training process of the multi-source domain generalized fault diagnosis model provided in the embodiment of the present application is as follows: Figure 2 As shown in Figure 2, the training process of the multi-source domain generalized fault diagnosis model includes:

[0077] S1: Obtain real fault data.

[0078] The actual fault data is the vibration signal corresponding to the rolling bearing being in a fault state.

[0079] Specifically, the vibration signals of the rolling bearing under different speed and load conditions are collected by an acceleration sensor, and a real fault data set is constructed based on the collected historical vibration signals.

[0080] In an optional embodiment, the real bearing data set uses the public bearing fault data set of Case Western Reserve University (CWRU). By collecting the fault vibration signals of the motor drive end bearing at four different speeds of 1730 (0hp), 1750 (1hp), 1772 (2hp) and 1797 (3hp) rpm, the data under different operating conditions represent four different domains. The original signal is converted into a frequency domain amplitude signal by fast Fourier transform. In this data set, each fault category divides the original fault vibration signal into fault samples through a sliding window, and each fault type has 200 original fault samples with a length of 1024 points. The fault types simulated in the data set include health (H), inner ring fault (IRF), outer ring fault (ORF) and rolling element fault (BF), as shown in Table 1:

[0081] Table 1 Fault types and labels of bearing fault dataset

[0082]

[0083] S2: Generate simulation fault data based on the rolling bearing dynamics model.

[0084] Specifically, dynamic models of local defects of the inner ring, outer ring and rolling element of the rolling bearing are established respectively; based on the Hertz contact theory, the time-varying displacement excitation generated when the rolling element contacts the inner ring and outer ring of the rolling bearing is calculated; according to the established dynamic models of local defects of the inner ring, outer ring and rolling element of the rolling bearing and the calculated time-varying displacement excitation, a nonlinear dynamic differential equation is constructed; the constructed nonlinear dynamic differential equation is solved by the fourth-order Runge-Kutta numerical integration method to obtain simulated fault data, wherein the simulated fault data includes vibration data corresponding to three fault types: local defects of the inner ring, outer ring and rolling element of the rolling bearing.

[0085] Figure 3 The figure shows the radial displacement of the inner ring raceway of a defect-free rolling bearing under external radial load. In the figure, O and O1 represent the center of the outer ring and inner ring, respectively. Assume that the radial contact deformation between the rolling element and the inner ring and outer ring is δ i and δ o , whose algebraic sum is equivalent to the center distance OO1. Therefore, during the entire contact deformation process, the total radial contact deformation between the jth rolling element and the inner and outer rings can be expressed as:

[0086]

[0087] Where x and y are the vibration displacements of the inner ring of the rolling bearing in the X-axis and Y-axis directions under the external radial load, respectively. is the internal radial clearance of the bearing, φ j is the angular position of the jth rolling element.

[0088] Based on the Hertz contact mechanics theory framework, the components of the contact force between the rolling element and the inner and outer ring raceways along the X-axis and Y-axis directions of the coordinate system can be expressed as:

[0089]

[0090] Among them, f x and f y They are defined as the orthogonal components of the contact force between the rolling element and the inner and outer ring raceways of the bearing in the X-axis and Y-axis directions, respectively, N b is the total number of rolling elements, K t is the total contact stiffness of the inner and outer rings, γ j is the discriminant parameter. Since the rolling element can only bear pressure, when the contact deformation δ j When it is a non-positive value, no contact force will be generated. In order to determine whether the j-th rolling element has contact deformation, the discriminant parameter γ can be introduced j As a quantitative characterization indicator of mechanical state, it is specifically defined as follows:

[0091]

[0092] Among them, δ j Indicates the jth rolling element at any angle position φ j The total radial contact deformation between the inner and outer rings.

[0093] Based on the above theories and assumptions, and according to the concentrated spring-mass model of rolling bearings, the following 2-DOF ball bearing dynamic equations are constructed:

[0094]

[0095] Where m represents the equivalent inertial mass of the inner ring and the shaft, C s is the viscous damping coefficient of the bearing system; x and y are the time-varying displacement responses of the inner ring in the X-axis and Y-axis directions respectively; and are the first-order derivatives of the time-varying displacements in the x and y directions, respectively, indicating their vibration velocities; and are the second-order derivatives of the time-varying displacements in the x and y directions, respectively, representing their vibration accelerations; F x and F y are the radial contact forces borne by the inner ring of the rolling bearing in the X-axis and Y-axis directions respectively; Δδ represents the additional displacement under different defect conditions, that is, the additional displacement caused by defects in the inner ring, outer ring, and rolling element.

[0096] To verify the accuracy and rationality of the dynamic models for the three defect conditions mentioned above, three dynamic models of rolling bearings were simulated separately. The rolling bearing simulation analysis was performed using the CWRU dataset, and the detailed parameters are shown in Table 2:

[0097] Table 2 CWRU bearing parameters

[0098]

[0099] According to the geometric parameters and speed of the CWRU bearing fault in Table 2, the failure frequency of the rolling bearing is calculated, as shown in Table 3:

[0100] Table 3 CWRU bearing fault characteristic frequency

[0101]

[0102] According to the CWRU bearing simulation parameters set above, the fourth-order Runge-Kutta numerical integration method based on Matlab software is used to solve the time-frequency diagrams of the simulated bearing outer ring, inner ring and rolling element defect faults, as shown below: Figure 4 、 Figure 5 and Figure 6As shown. Each figure contains the simulation signal and the measured fault signal of the corresponding part. The first line is the original simulation signal (left) and its envelope spectrum (right); the second line is the measured local fault signal (left) and its envelope spectrum (right). It can be observed from the figure that the rolling bearing fault produces periodic impulses in the simulated and measured time domain signals, and the pulse trend of the simulated signal of each fault is similar to the measured signal. For the envelope spectrum, it can be found that the main azimuth frequency components of the simulation signal are consistent with the theoretical value, and the measured signal and the simulation signal have the same frequency components, which proves the effectiveness of the rolling bearing dynamics model established in this embodiment.

[0103] S3: Perform fast Fourier transform and sliding window segmentation processing on the real fault data and the simulated fault data to obtain the processed real fault data and the simulated fault data.

[0104] Specifically, the time domain signal is converted into the frequency domain signal by fast Fourier transform, and the sliding window sampling technology is used to perform data segmentation to obtain the pre-processed real fault data and simulated fault data.

[0105] S4: Construct a multi-source domain dataset based on the preprocessed real fault data and simulated fault data. The multi-source domain dataset includes: a source domain and a target domain. Each source domain includes real fault data and simulated fault data, and the target domain is the vibration signal to be diagnosed.

[0106] Specifically, the three working condition data in the CWRU dataset are used as the source domain. Each working condition contains four types of samples: normal (0), inner ring fault (1), outer ring fault (2) and rolling element fault (3). Each type of fault contains only 5% of the real labeled data, and the rest is supplemented by the simulation data generated by the dynamic model; the target domain is set to the four types of fault data under another working condition other than the source domain, and all of them use unlabeled real working condition data.

[0107] For example, after empirical data collection and simulation, the source domain data and target domain data of the bearing fault dataset are shown in Table 4:

[0108] Table 4 Source domain data and target domain data of bearing fault dataset

[0109]

[0110] S5: Perform multi-scale feature extraction on multi-source domain datasets to obtain source domain features and target domain features.

[0111] Specifically, it is divided into two parts: convolutional feature extraction module and dynamic feature extraction module, and the expression is:

[0112] ρ(x)=ρ s (x)+ρ d (x)

[0113] Among them, ρ s (x) represents the output of the convolutional feature extraction module, which can be obtained through the convolution layer. d (x) represents the output of the dynamic propagation unit, which is closely related to the input sample x of different data distributions. ρ(x) represents the output of the dynamic feature extraction module, which is ρ s (x) and ρ d The sum of the outputs of (x).

[0114] In the multi-source domain dynamic matching feature extraction module, the dynamic feature extraction module consists of a multi-scale block, a first cascade layer, an adaptive average pooling layer, a convolution layer, a softmax layer, a series of identical independent convolution layers, and a second-level connection. First, the multi-scale block uses some convolution kernels with different sizes to capture potential knowledge, so that the input sample x can be processed in the multi-scale block to obtain robust multi-scale features; secondly, these multi-scale features are processed in sequence by the first cascade layer, the adaptive average pooling layer, and the convolution layer to generate a set of dynamic matching coefficients. It is normalized to 0 to 1 by using the Softmax layer. Then, a series of identical independent convolutional layers are introduced, and each dynamic matching coefficient is assigned to the corresponding convolutional layer to obtain the output of the dynamic feature extraction module. The output acquisition process of the dynamic feature extraction module can be obtained by the following formula:

[0115]

[0116] Among them, μ i (x) represents the i-th independent convolutional layer, represents μ i (x), where W represents the total number of independent convolutional layers. By combining these independent convolutional layers with different coefficients in an input-dependent manner, the dynamic feature extraction module can dynamically adjust the degree of attention paid to different features to exploit various feature subspaces to mine the knowledge contained in different inputs.

[0117] S6: Identify domain-sensitive channels in domain features through the domain discriminator and use the probabilistic channel discarding mechanism to shield the domain-sensitive channels.

[0118] It can be understood that first, the correlation between each channel and specific domain information is quantified by calculating the weighted activation of the correct domain prediction; secondly, the domain discriminator is used to identify and discard unstable channels in the features of each layer of the network to continuously enhance the robustness of the feature channel to domain transformation.

[0119] In an optional embodiment, step S6 specifically includes:

[0120] S601: Assume that K source domains are given Each source domain follows a different distribution. For different source domains, where n k yes The number of samples in Represents the sample label pair of the i-th sample in the k-th domain. Using the source domain To train the expected unseen target domain In order to clearly inform the model to remove domain-specific features during training, a domain discriminator is introduced in multiple intermediate layers to locate domain-sensitive channels, which consists of global average pooling layers and fully connected layers. Input And the corresponding labels First, extract the features generated by the lth intermediate layer Where C is the number of channels and H is the sample length. Sample characteristics is fed to the domain discriminator F l d To predict the domain label and calculate the domain discrimination loss. In order to avoid the negative impact of the domain discrimination factor on the main network, a gradient reversal layer is used before the domain discriminator to truncate the gradient that minimizes the discrimination loss. The formula is as follows:

[0121]

[0122] in, is the domain discrimination loss, K is the total number of source domains, F d is the domain discriminator.

[0123] S602: For input and feature extractor F l (·), the feature map The correlation coefficient score of the j channels in is defined as:

[0124]

[0125] in, is the weight parameter of the fully connected layer in the domain discriminator, and GAP(·) is the global average pooling.

[0126] S603: Using the correlation coefficient score s of each channel j , calculate the probability p of discarding the jth channel j , satisfying the following formula:

[0127]

[0128] Among them, s c is the total channel score, and C is the total number of channels.

[0129] Furthermore, a binary mask is generated based on the probability of each channel That is m j With probability p set to 1 j , whose channels have relatively high scores s j More likely to be discarded. To ensure the regularization effect of loss, try to use the probability p j A certain number of domain-specific channels are lost; to reduce computational cost, a weighted random selection algorithm is used to generate the binary mask m j , its time complexity is O(C). Specifically, for a relatively high score s j For the jth channel, first generate a random number r j ∈(0,1) and calculate the key value Then, select the M items with the largest key value and assign the m items corresponding to the mask m to j Set to 0, the specific formula is as follows:

[0130]

[0131] Among them, j is the feature channel index, TOP({k1,k2,...,k C},M) represents the M items with the largest key value k.

[0132] S7: By introducing the classification aggregation loss function, the model is trained across domain classes at the feature space level based on the extracted source domain features.

[0133] Specifically, given an anchor sample X u ∈B={b1,b2,...,b M}, where B is a set of batches from one of the M source domains. In order to maximize the anchor sample X u and the mutual information between the corresponding positive samples in B, while minimizing the similarity between it and the corresponding negative sample pairs in B. More specifically, the anchor sample X u The positive samples in B belong to the same class pos(u)={x v ∈B:y u =y v} all samples represent, x v represents the positive sample, y u Represents the anchor sample label, y v Represents the positive sample label, and the negative sample is the remaining sample in B belonging to different categories neg(u)={x k ∈B:y u ≠y k}, x k represents negative samples, y kRepresents the negative sample label. This method is applied to the class prediction level, enabling the model to find class-conditional invariant representations across multiple source domains. u The classification aggregation loss is recorded as Specifically expressed as:

[0134]

[0135] Among them, |pos(u)| is the cardinality of the positive sample set, positives represents the positive sample set, negatives represents the negative sample set, σ(h u , h v ) is h u 、h v The similarity scoring function is τ, which is the temperature parameter that controls the contrast. It's X u The output vector, g θ ⊙f θ represents the composition of the feature extractor and the predictor, and C is the dimension of the output vector (i.e., the number of classes), h v is sampled from the positive sample set. By optimizing Maximize anchor sample X u The mutual information between the classifier predictions of all corresponding positive samples is used to better distinguish samples from the same class, where h v Sample from the set of positive samples. At the same time, by minimizing the mutual information between the anchor sample and all negative samples. By minimizing the linear matrix inequality, the lower bound of the mutual information between positive samples can be maximized. The overall classification loss for all samples is calculated as follows:

[0136]

[0137] Where M is the total number of source domains, N i is the total number of samples in the i-th source domain, is the classification loss of the anchor sample.

[0138] S8: Use the fault classifier to perform classification prediction on the target domain features based on the cross-domain class learning results.

[0139] It is understandable that the fault classifier is used to classify and predict the optimized features and calculate the classification loss function To evaluate the generalization performance of the model on the target domain, the classification loss function is expressed as follows:

[0140]

[0141] in, is a sample, ω δ are the parameters of the feature extractor, is the feature extractor of layer l, is the label corresponding to the sample.

[0142] S9: Determine the total loss based on the feature extraction loss, domain adaptation loss, classification aggregation loss, and classification prediction loss. Adjust the parameters of the multi-source domain generalized fault diagnosis model through the back propagation algorithm until the total loss is less than the preset loss value, and obtain the trained multi-source domain generalized fault diagnosis model.

[0143] Specifically, first the loss Backpropagation is performed and the feature extractor parameters are optimized to mitigate the effects of category shift and domain gap between multiple source domains, as shown below:

[0144]

[0145] Furthermore, the feature extractor G fe and a domain discriminator for each specific domain Losses and Back propagation optimization parameters are performed as follows:

[0146]

[0147] in, is the feature extractor loss, G fe At the same time, the loss is used between each source domain Update the network parameters while taking the loss of classification aggregation The back propagation further updates G fe The network parameters of the whole model are as follows:

[0148]

[0149] in, is the domain discriminator of the source domain S1, The source domain S M The domain identifier.

[0150] It is understandable that in order to verify the effectiveness of the present invention, experiments were conducted using the CWRU bearing dataset and the bearing fault simulation dataset constructed in the embodiment. The specific task settings are as follows:

[0151] There are four known fault states, each of which is further divided into four different operating conditions. The four fault states under three of these operating conditions are used as the source domain, and the four fault states under another operating condition are used as the target domain. The source domain is assumed to contain only a small amount of data for a specific fault state under each operating condition, and all other data is supplemented by simulation data. This is shown in Table 5.

[0152] Table 5 Training and testing task settings

[0153]

[0154] Table 5 shows the A1 task settings. The four fault states under the three working conditions of 0hp, 1hp, and 2hp are used as the source domain, and the true labels only account for 5% of one type of fault. The remaining source domain label data is provided by the simulated signals simulated by the dynamic simulation model constructed in this paper. The four fault states under the 3hp working condition are used as the target domain. Similarly, in the A2 task, the four fault states under the three working conditions of 0hp, 1hp, and 3hp are set as the source domain, and the four fault states under the 2hp working condition are set as the target domain; in the A3 task, the four fault states under the three working conditions of 0hp, 2hp, and 3hp are set as the source domain, and the four fault states under the 1hp working condition are set as the target domain; in the A4 task, the four fault states under the three working conditions of 1hp, 2hp, and 3hp are set as the source domain, and the four fault states under the 0hp working condition are set as the target domain.

[0155] Based on the above bearing fault data, the proposed domain-sensitive feature suppression multi-domain generalization fault diagnosis model based on digital-analog fusion is compared with the Deep Domain Generalization Network for Fault Diagnosis (DDGFD) model proposed in the paper Deep domain generalization combining apriori diagnosis knowledge toward cross-domain fault diagnosis of rolling bearing. IEEE Transactions on instrumentation and measurement, 2020, the Intrinsic and Extrinsic Domain Generalization Network (IEDGNet) model proposed in the paper A hybrid generalization network for intelligent fault diagnosis of rotating machinery under unseen working conditions. IEEE Transactions on Instrumentation and Measurement, 2021, and the Causal Disentanglement Domain Generalization (CDDG) model proposed in the paper Causal Disentanglement Domain Generalization for time-series signal fault diagnosis. Neural Networks, 2024. To ensure the fairness of the experiment, the above methods all use the same feature extractor and classifier structure.

[0156] In order to minimize the influence of random errors, each task was repeated 5 times, and the model was trained 500 times each time. The sample batch size was set to 64, and Adam was selected as the optimization algorithm of the optimizer. The initial learning rate value was set to 0.001, and the cosine annealing algorithm was used to gradually adjust the learning rate as the number of training times increased. The specific fault diagnosis accuracy results of the compared methods and the proposed method are shown in Table 6. In the A1 task, the confusion matrix of the task was randomly drawn, as shown in Figure 7 As shown, the vertical axis represents the true label and the horizontal axis represents the label predicted by the model.

[0157] Table 6 Cross-domain fault diagnosis results of bearing fault dataset

[0158]

[0159] It can be seen from the experimental results that the proposed DSFIMDG method shows significant advantages. The average accuracy of this method in the CWRU dataset task reached 96.25%, which is significantly better than the comparison methods CDDG (84.12%), DDGFN (86.39%) and IEDGNet (93.94%). It is particularly noteworthy that the DSFIMDG method is 2.31 percentage points higher than the IEDGNet method with the highest average accuracy. This performance improvement fully verifies the effectiveness of the proposed method in solving the multi-source domain migration fault diagnosis problem. By analyzing the confusion matrix of task A1, it can be found that the DSFIMDG method maintains a relatively good balance in the identification of various types of faults, and significantly reduces the misjudgment rate. In order to further verify the effectiveness of the method proposed in this invention, Figure 8 Visualizations of feature classification results for each method in the random A1 task are presented. The first row, from left to right, shows the classification results for CDDG and DDGFN, while the second row, from left to right, shows the classification results for IEDGNet and the proposed method. This demonstrates that the proposed method achieves superior classification results. These experimental results confirm that the DSFIMDG method, through its innovative design including mechanism simulation data supplementation, dynamic matching feature extraction, and domain-sensitive feature suppression, effectively improves the model's performance in cross-condition fault diagnosis tasks.

[0160] The embodiment of the present application provides a rolling bearing fault diagnosis method based on digital analog drive. The method collects the vibration signals of the rolling bearing under different working conditions, and inputs them into the trained multi-source domain fault diagnosis model after fast Fourier transform and sliding window segmentation preprocessing to obtain the rolling bearing fault diagnosis results. The training process of the multi-source domain fault diagnosis model includes: constructing a rolling bearing local defect dynamics model to generate simulation data, which together with the measured data constitutes a multi-source domain data set; extracting multi-scale features through a multi-source domain dynamic matching feature extraction module, and adopting an adaptive parameter allocation strategy to achieve dynamic feature matching; using a domain-sensitive feature suppression strategy to identify and discard specific domain-sensitive channels to reduce domain-related feature interference; introducing a classification aggregation loss function to achieve cross-domain class condition-invariant representation learning at the feature space level; performing loss backpropagation based on dynamic matching feature loss, domain-sensitive feature suppression loss, and classification aggregation loss to optimize model parameters, thereby improving the accuracy and reliability of the rolling bearing fault diagnosis results.

[0161] Figure 9 This is a schematic diagram of the structure of the rolling bearing fault diagnosis device based on digital-analog drive provided in the embodiment of the present application. Figure 9As shown, the rolling bearing fault diagnosis device 200 based on digital-analog drive provided in this embodiment includes:

[0162] The first processing module 201 is used to collect vibration signals of the rolling bearing under different working conditions;

[0163] The second processing module 201 is used to input the collected vibration signal into the trained multi-source domain fault diagnosis model after fast Fourier transform and sliding window segmentation for analysis and processing to obtain the rolling bearing fault diagnosis result.

[0164] The rolling bearing fault diagnosis device based on digital-analog drive provided in this embodiment can execute the rolling bearing fault diagnosis method based on digital-analog drive provided in the above method embodiment. Its implementation principle and technical effects are similar, and are not described in detail in this embodiment.

[0165] Figure 10 This is a schematic diagram of the structure of the rolling bearing fault diagnosis device based on digital analog drive provided in the embodiment of the present application. Figure 10 As shown, the rolling bearing fault diagnosis device based on digital-analog drive provided in an embodiment of the present application, the rolling bearing fault diagnosis device based on digital-analog drive 300 includes: a receiver 301, a transmitter 302, a processor 303 and a memory 304.

[0166] Receiver 301, for receiving instructions and data;

[0167] Transmitter 302, used to send instructions and data;

[0168] Memory 304, for storing computer-executable instructions;

[0169] The processor 303 is configured to execute the computer-executable instructions stored in the memory 304 to implement the various steps of the rolling bearing fault diagnosis method based on digital-analog drive in the above embodiment. For details, please refer to the relevant description in the above embodiment of the rolling bearing fault diagnosis method based on digital-analog drive.

[0170] Optionally, the memory 304 may be independent or integrated with the processor 303 .

[0171] When the memory 304 is independently provided, the electronic device further includes a bus for connecting the memory 304 and the processor 303 .

[0172] An embodiment of the present application also provides a computer storage medium, which stores computer execution instructions. When the processor executes the computer execution instructions, it implements the rolling bearing fault diagnosis method based on digital analog drive as performed by the above-mentioned rolling bearing fault diagnosis device based on digital analog drive.

[0173] An embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the above-mentioned rolling bearing fault diagnosis method based on digital-analog drive.

[0174] It will be appreciated by those skilled in the art that all or some of the steps, systems, and functional modules / units in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. In hardware implementations, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or temporary medium). As is well known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable, and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those skilled in the art that communication media generally embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0175] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of the present application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0176] It should be understood that the present application is not limited to the exact structure described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A rolling bearing fault diagnosis method based on digital analog drive, characterized in that: The method comprises: The vibration signals of rolling bearings under different working conditions are collected. After fast Fourier transform and sliding window segmentation, the collected vibration signals are input into the trained multi-source domain fault diagnosis model for analysis and processing to obtain the fault diagnosis results of the rolling bearings. The training process of the multi-source domain generalized fault diagnosis model includes: Obtain real fault data; Generate simulation fault data based on rolling bearing dynamics model; Perform fast Fourier transform and sliding window segmentation processing on the real fault data and the simulated fault data to obtain the pre-processed real fault data and the simulated fault data; A multi-source domain dataset is constructed based on the pre-processed real fault data and simulated fault data. The multi-source domain dataset includes: a source domain and a target domain. Each source domain includes real fault data and simulated fault data, and the target domain is the vibration signal to be diagnosed. Perform multi-scale feature extraction on multi-source domain datasets to obtain source domain features and target domain features; The domain discriminator is used to identify domain-sensitive channels in domain features, and the probabilistic channel discarding mechanism is used to shield domain-sensitive channels. By introducing the classification aggregation loss function, the model is trained across domain classes at the feature space level based on the extracted source domain features. Use the fault classifier to classify and predict the target domain features based on the cross-domain class learning results; The total loss is determined according to the feature extraction loss, domain adaptation loss, classification aggregation loss and classification prediction loss. The parameters of the multi-source domain generalized fault diagnosis model are adjusted through the back propagation algorithm until the total loss is less than the preset loss value, and the trained multi-source domain generalized fault diagnosis model is obtained.

2. The multi-domain generalized fault diagnosis method based on digital-analog fusion drive according to claim 1 is characterized in that: Generate simulation fault data based on rolling bearing dynamics model, including: The dynamic models of local defects of the inner ring, outer ring and rolling element of the rolling bearing are established respectively; Based on Hertz contact theory, calculate the time-varying displacement excitation generated when the rolling element contacts the inner and outer rings of the rolling bearing; Based on the established dynamic model and the calculated time-varying displacement excitation, the nonlinear dynamic differential equation is constructed; The constructed nonlinear dynamic differential equation is solved by the fourth-order Runge-Kutta numerical integration method to obtain the simulated fault data.

3. The multi-domain generalized fault diagnosis method based on digital-analog fusion drive according to claim 1 is characterized in that: Perform multi-scale feature extraction on multi-source domain datasets to obtain source domain features and target domain features, including: Construct a multi-source domain dynamic matching feature extraction module, which includes: a convolutional feature extraction module and a dynamic feature extraction module. The convolutional feature extraction module consists of multiple convolutional layers, and the dynamic feature extraction module consists of a multi-scale block, a first cascade layer, an adaptive average pooling layer, a convolution layer, a softmax layer, multiple identical independent convolutional layers, and a second cascade layer. Performing feature extraction on the source domain data and the target domain data through a convolutional feature extraction module to obtain a first source domain feature and a first target domain feature; Extract multi-scale features of source domain data and target domain data through the multi-scale blocks of the dynamic feature extraction module; The multi-scale features extracted by the multi-scale blocks are processed in sequence by the first cascade layer, the adaptive average pooling layer, and the convolution layer to obtain multiple dynamic matching coefficients, which are then normalized using the Softmax layer. Assigning each dynamic matching coefficient to the corresponding convolutional layer to obtain the second source domain feature and the second target domain feature; A source domain feature and a target domain feature are determined according to the first source domain feature, the first target domain feature, the second source domain feature, and the second target domain feature.

4. The multi-domain generalized fault diagnosis method based on digital-analog fusion drive according to claim 1 is characterized in that: The domain discriminator is used to identify domain-sensitive channels in domain features, and a probabilistic channel discarding mechanism is used to shield domain-sensitive channels, including: The domain classification prediction of each source domain feature is performed through the domain discriminator to obtain the domain sensitivity score; Calculate the contribution of each feature channel to domain classification and generate channel importance weight; According to the domain sensitivity score and channel importance weight, a probabilistic discarding mechanism is used to dynamically shield high domain sensitive feature channels.

5. The multi-domain generalized fault diagnosis method based on digital-analog fusion drive according to claim 1 is characterized in that: By introducing the classification aggregation loss function, the model is trained on cross-domain classes at the feature space level based on the extracted source domain features, including: Maximize the mutual information between samples of the same category in different source domains through classification aggregation loss function; The loss is calculated using the standard form of contrastive learning, where pairs of samples of the same category from different source domains are considered positive pairs, and pairs of samples of different categories are considered negative pairs. By minimizing the contrastive learning loss, the model learns class-conditional invariant representations across multiple source domains.

6. The multi-domain generalized fault diagnosis method based on digital-analog fusion drive according to claim 1 is characterized in that: The total loss is determined based on the feature extraction loss, domain adaptation loss, classification aggregation loss, and classification prediction loss. The parameters of the multi-source domain generalized fault diagnosis model are adjusted through the backpropagation algorithm, including: Determine a classification prediction loss, where the classification prediction loss indicates the difference between the prediction result and the true label; determining a feature extraction loss, wherein the feature extraction loss is used to constrain the distribution characteristics of the generated source domain features and target domain features; determining a domain adaptation loss, wherein the domain adaptation loss is used to reduce the feature distribution difference between the source domain and the target domain; Determine a classification aggregation loss that maximizes the mutual information of features of similar categories in different domains and suppresses environmental variation information that is irrelevant to the task; Construct a total loss function based on classification prediction loss, feature extraction loss, domain adaptation loss, and classification aggregation loss; Through the back propagation algorithm, the parameters of the multi-source domain generalized fault diagnosis model are optimized based on the total loss function to realize the training of the multi-source domain generalized fault diagnosis model.

7. A multi-domain generalized fault diagnosis device based on digital-analog fusion drive, characterized in that: The device comprises: The first processing module is used to collect vibration signals of the rolling bearing under different working conditions; The second processing module is used to input the collected vibration signal into the trained multi-source domain fault diagnosis model for analysis and processing after fast Fourier transform and sliding window segmentation to obtain the rolling bearing fault diagnosis result.

8. A multi-domain generalized fault diagnosis device based on digital-analog fusion drive, characterized in that: The device comprises: Memory; processor; wherein the memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the multi-domain generalized fault diagnosis method based on digital-analog fusion drive according to any one of claims 1 to 6.

9. A computer storage medium, characterized in that The computer storage medium stores computer execution instructions, which, when executed by a processor, are used to implement the multi-domain generalized fault diagnosis method based on digital-analog fusion drive as described in any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which is used to implement the multi-domain generalized fault diagnosis method based on digital-analog fusion drive as described in any one of claims 1 to 6 when executed by a processor.

Images