Rolling bearing composite fault diagnosis method based on generative zero sample learning

Abstract

The invention discloses a rolling bearing composite fault diagnosis method based on generative zero sample learning. A deep neural network and a generative adversarial network are adopted to construct a zero sample learning model; the zero sample learning model comprises a data preprocessing module, a feature extraction network module, a fault category auxiliary information module, a generative adversarial network module and a classification module; performing data preprocessing on the original vibration signal by adopting continuous wavelet transform, and constructing a wide-kernel convolutional neural network to extract deep fault feature information of a sample; and finally, learning a mapping relationship between the fault feature information and the FCAI by adopting an adversarial training strategy so as to generate high-quality pseudo fault features, and diagnosing a composite fault according to a similarity relationship between the fault features through a distance measurement method. The rolling bearing composite fault diagnosis method based on generative zero sample learning has the advantages that composite faults can be diagnosed through a distance measurement method according to the similarity relation between fault features, and the like.

Description

Technical Field

[0001] This invention relates to a fault diagnosis method, and more particularly to a composite fault diagnosis method for rolling bearings based on generative zero-sample learning. Background Technology

[0002] Modern machinery is rapidly developing towards intelligence in the field of smart manufacturing. In real-world industrial scenarios, equipment operating conditions are highly variable, and timely detection of mechanical equipment faults is difficult. Multiple faults may occur simultaneously during equipment operation, forming complex failure modes. The property losses and safety risks caused by compound faults have attracted widespread attention. Compound faults in rotating machinery involve multiple interrelated individual fault parameters, and the fault characteristics interfere with each other, making the analysis and diagnosis of compound faults extremely challenging. Therefore, conducting research on intelligent diagnosis of compound faults is crucial to ensuring the stability and safety of machinery operation.

[0003] A rolling bearing is a precision mechanical component that transforms the sliding friction between a rotating shaft and its housing into rolling friction, thereby reducing frictional losses. A rolling bearing generally consists of four parts: an inner ring, an outer ring, rolling elements, and a cage. The rolling elements are evenly distributed between the inner and outer rings by the cage, and their shape, size, and number directly affect the performance and lifespan of the rolling bearing. The cage ensures the even distribution of the rolling elements, guides their rotation, and provides lubrication. Rolling bearings operate for extended periods in rotating machinery, and the probability of failure is relatively high during continuous high-speed operation. Moreover, these failures are characterized by randomness, concealment, and concurrency, with multiple failures often coexisting. The complex failures of rolling bearings are not merely a simple linear superposition of multiple individual failure characteristics, but rather involve the mutual coupling and correlation of parameters between individual failures, as well as mutual interference between characteristics, making the analysis and diagnosis of complex bearing failures highly challenging.

[0004] In the field of fault diagnosis, research on traditional composite fault diagnosis methods mainly focuses on vibration signal-based methods, using various time-domain or frequency-domain signal processing methods to acquire various fault characteristics for fault diagnosis. Compared to traditional methods, intelligent diagnosis requires expert knowledge to design powerful algorithms, thereby constructing diagnostic models. It also requires a sufficient number of labeled composite fault samples to train these models. Deep learning and transfer learning methods, with their ability to automatically extract deep fault features and achieve end-to-end fault diagnosis, have attracted considerable attention from researchers.

[0005] With the continuous development of deep learning, deep neural network models with deeper network structures are able to learn more useful features from data. A deep learning-based bearing composite fault diagnosis model utilizes the powerful feature extraction capabilities of deep neural networks to capture key fault features and constructs a multi-label classifier to output multiple fault category labels, enabling intelligent diagnosis of composite faults. However, deep learning neural networks typically require a large amount of data for model training to allow the model to learn effective fault features. Insufficient data can lead to overfitting in the fault diagnosis model.

[0006] Transfer learning leverages knowledge from the source domain to enhance the learning performance of the target domain, reducing the target domain's data requirements and overcoming the limitations of deep learning. Addressing the lack of composite fault data in industrial scenarios, existing technologies have proposed a bearing composite fault decoupling model combining deep neural networks, transfer learning, and multi-label output mechanisms. This model utilizes multi-label learning theory to construct a multi-label decoupling classifier, decoupling composite faults into multiple single fault labels. However, this model typically requires constructing complex models to handle the relationships between multiple labels and setting independent loss functions for each fault class, resulting in large model parameters and increased training time and inference complexity. Furthermore, most current transfer learning methods assume that the source and target domains share the same label space, relying on source domain samples and their label information to predict the labels of target domain data. However, this assumption fails to address the composite fault diagnosis problem under zero-sample conditions, leading to a lack of composite fault training samples in the source domain. This makes it unsuitable for fault diagnosis scenarios where the source and target domains have inconsistent fault types, thus limiting the accuracy of fault diagnosis methods.

[0007] Therefore, existing technologies often suffer from difficulties in collecting composite fault samples, leading to a shortage of training samples. This situation causes class imbalance, which in turn affects the fault diagnosis performance of the model. Currently, how to train a fault diagnosis model and accurately diagnose composite faults in the absence of composite fault samples remains a significant challenge.

[0008] Therefore, there is an urgent need to develop a method for diagnosing composite bearing faults that can solve the problem of a lack of composite fault samples for rolling bearings. Summary of the Invention

[0009] To avoid the shortcomings of the existing technology, the present invention provides a method for diagnosing composite faults in rolling bearings based on generative zero-shot learning, which can use a single fault sample as training data to diagnose composite fault categories.

[0010] The present invention adopts the following technical solution to solve the technical problem.

[0011] This invention discloses a method for diagnosing complex faults in rolling bearings based on generative zero-shot learning, wherein a zero-shot learning model is constructed using a deep neural network and a generative adversarial network; the zero-shot learning model includes a data preprocessing module, a feature extraction network module, a fault category auxiliary information module, a generative adversarial network module, and a classification module; the method for diagnosing complex faults in rolling bearings includes:

[0012] Step 1: Construct a training dataset D using vibration data from a single fault in a rolling bearing. s The composite fault vibration data was used as the test dataset D. t ;

[0013] Step 2: Based on the training dataset D s and test dataset D t , will the training dataset D s Single-fault vibration data is converted into single-fault image data. s The test dataset D t The composite fault vibration data in the image is converted into composite fault image data. t ;

[0014] Step 3: For single fault image data d s and composite fault image data d t Extract fault feature information v from the two types of fault image data respectively;

[0015] Step 4: Based on the training dataset D s Obtain semantic information of fault categories from the raw vibration data. and manually defined fault description information This allows us to obtain auxiliary information A for a single fault category that has already been identified. s ;

[0016] Step 5: Based on the observed fault characteristic information v s Auxiliary information for known fault categories A s Establish fault characteristic information v and fault category auxiliary information A s The mapping relationship between them is used to obtain pseudo-fault characteristic information. ;

[0017] Step 6: Based on the adversarial training method in Step 5, calculate the fault feature v extracted by the feature extraction network and the pseudo-fault feature information generated by the generator according to the similarity relationship between fault features. The distance d between the samples is used to sort the results from largest to smallest. The category corresponding to the pseudo-fault feature that is closest to the feature of the unseen sample is the diagnosis result of that unseen sample.

[0018] The technical features of the rolling bearing composite fault diagnosis method based on generative zero-shot learning of the present invention also lie in:

[0019] Furthermore, in step 2, the data preprocessing module of the zero-shot learning model uses continuous wavelet transform to convert single-fault vibration data and composite-fault vibration data into single-fault image data d, respectively. s and composite fault image data d t .

[0020] Furthermore, in step 2, the single fault vibration data and the composite fault vibration data are one-dimensional vibration signals.

[0021] Furthermore, in step 3, feature information of the fault image data is extracted. The process includes the following steps:

[0022] Step 31: Extract single fault image data d through the feature extraction network module. s Found fault characteristic information v s And predict the single fault image data d of the seen class through the softmax network layer of the feature extraction network. s Fault labels ;

[0023] Step 32: Calculate the actual fault labels and prediction labels The L1 cross-entropy loss between the classifications is used, and the performance of the feature extraction network is optimized by the Adam optimizer through adaptive learning rate and momentum method to obtain the trained feature extraction network.

[0024] Step 33: Use the trained feature extraction network to extract composite fault feature information of unseen classes. t .

[0025] Furthermore, in step 4, auxiliary information A for observed fault categories is obtained. s The process includes the following steps:

[0026] Step 41: Construct fault category semantic information S, and obtain single fault category semantic information S for the seen classes. s ; composed of semantic information S of a single fault category s Obtain semantic information S of composite fault categories u ;

[0027] Step 42: Construct manually defined fault description information Obtain manually defined single fault description information R s ;

[0028] Step 43: Using semantic information S of a single fault category s And manually defined single fault description information R s Calculate and obtain auxiliary information A for a single fault category of the known type. s ;

[0029] Step 44: Based on the observed single fault category auxiliary information A s Derivation of auxiliary information A for unseen composite fault categories u .

[0030] Furthermore, in step 5, the observed fault characteristic information v s Auxiliary information for known fault categories A s The input is fed into the Generative Adversarial Network (GAN) module, which then obtains pseudo-fault feature information. .

[0031] Furthermore, the generative adversarial network module includes a generator G and a discriminator D.

[0032] Furthermore, auxiliary information on fault categories The pseudo-fault characteristics are obtained by inputting random noise Z into the generator G. .

[0033] The present invention also discloses an electronic device, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the described method for diagnosing composite faults in rolling bearings based on generative zero-sample learning.

[0034] The present invention also discloses a non-transient computer-readable storage medium storing computer instructions for causing the computer to execute the aforementioned method for diagnosing composite faults in rolling bearings based on generative zero-shot learning.

[0035] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0036] This invention discloses a method for diagnosing composite faults in rolling bearings based on generative zero-shot learning. A zero-shot learning model is constructed using deep neural networks and generative adversarial networks. The zero-shot learning model includes a data preprocessing module, a feature extraction network module, a fault category auxiliary information module, a generative adversarial network module, and a classification module. Continuous wavelet transform is used to preprocess the original vibration signals, and a wide-kernel convolutional neural network is constructed to extract deep fault feature information from the samples. Finally, an adversarial training strategy is employed to learn the mapping relationship between fault feature information and FCAI (Fault-Fault Indicator Algorithm), thereby generating high-quality pseudo-fault features. Composite faults are then diagnosed based on the similarity relationship between fault features using a distance metric method.

[0037] The rolling bearing composite fault diagnosis method based on generative zero-sample learning of the present invention has the advantages of being able to diagnose composite faults based on the similarity relationship between fault features through distance metric methods. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a method for diagnosing composite faults in rolling bearings based on generative zero-sample learning, according to the present invention.

[0039] Figure 2 This is a schematic diagram of the zero-sample fault diagnosis model of the present invention.

[0040] Figure 3 This is a schematic diagram illustrating the construction of the FCAI of the present invention;

[0041] Figure 4 This is a schematic diagram of the training model process for known sample types in this invention;

[0042] Figure 5 This is a comparison chart showing the accuracy of composite fault diagnosis between existing methods and the method of this invention in different tasks.

[0043] The present invention will be further described below through specific embodiments and in conjunction with the accompanying drawings. Detailed Implementation

[0044] See Figures 1-5 This invention discloses a method for diagnosing complex faults in rolling bearings based on generative zero-shot learning, wherein a zero-shot learning model is constructed using a deep neural network and a generative adversarial network; the zero-shot learning model includes a data preprocessing module, a feature extraction network module, a fault category auxiliary information module, a generative adversarial network module, and a classification module; the method for diagnosing complex faults in rolling bearings includes:

[0045] Step 1: Construct a training dataset D using vibration data from a single fault in a rolling bearing. s The composite fault vibration data was used as the test dataset D.t ;

[0046] The deep neural network and generative adversarial network in this invention are both conventional models. In specific implementation, raw fault vibration data (Data) of the rolling bearing is collected and categorized into single-fault vibration data and composite-fault vibration data based on their type. The single-fault vibration data is used as the training dataset (D). s The composite fault vibration data was used as the test dataset D. t Satisfying D s ∩D t =Ø; where the training dataset D s The single-fault vibration data sample in the dataset belongs to the known class, and the test dataset D t The composite fault vibration data samples in the data belong to the unseen class.

[0047] Step 2: Based on the training dataset D s and test dataset D t , will the training dataset D s Single-fault vibration data is converted into single-fault image data. s The test dataset D t The composite fault vibration data in the image is converted into composite fault image data. t ;

[0048] Training dataset D s The single-fault vibration data in the dataset consists of observed single-fault vibration signals. The test dataset D... t The composite fault vibration data in the dataset consists of composite fault vibration signals of unseen classes, both of which are one-dimensional vibration signals. This is based on the training dataset D partitioned in step 1. s and test dataset D t The one-dimensional vibration signals of single and combined faults in rolling bearings are converted into single fault image data d. s and composite fault image data d t ;

[0049] Step 3: For single fault image data d s and composite fault image data d t Extract fault feature information v from the two types of fault image data respectively;

[0050] like Figure 1 Based on the two types of fault image data obtained in step 2, this invention constructs a deep neural network to extract feature information v from the two types of fault image data; wherein, single fault image data d s The extracted feature information is a single fault characteristic of the known class. s Composite fault image data d tThe extracted composite fault feature information is of the unseen class. t ; to transfer the single fault feature information v s and composite fault characteristic information v t Input to the Generative Adversarial Network (GAN) module.

[0051] Step 4: Based on the training dataset D s Obtain semantic information of fault categories from the raw vibration data. and manually defined fault description information This allows us to obtain auxiliary information A for a single fault category that has already been identified. s ;

[0052] From the training dataset D s Semantic information S of a single fault category is obtained from the original single-fault vibration data. s And manually defined single fault description information R s Based on the semantic information S of a single fault category of the known classes s And manually defined single fault description information R s Construct auxiliary information A for a single fault category of known types. s Then, based on the already seen single fault category auxiliary information A s Derivation of Auxiliary Information A for Unseen Composite Fault Categories u .

[0053] Step 5: Based on the observed fault characteristic information v s Auxiliary information for known fault categories A s Establish fault characteristic information v and fault category auxiliary information A s The mapping relationship between them is used to obtain pseudo-fault characteristic information. ;

[0054] Step 6: Based on the adversarial training method in Step 5, calculate the fault feature v extracted by the feature extraction network and the pseudo-fault feature information generated by the generator according to the similarity relationship between fault features. The distance d between the samples is used to sort the results from largest to smallest. The category corresponding to the pseudo-fault feature that is closest to the feature of the unseen sample is the diagnosis result of that unseen sample.

[0055] like Figure 1This is a flowchart of a rolling bearing composite fault diagnosis method based on generative zero-shot learning according to the present invention. Fault semantic information and manually defined fault description information are combined to construct Fault Category Auxiliary Information (FCAI), which is used to learn the correlation between single faults and composite faults. Furthermore, continuous wavelet transform is used to preprocess the original vibration signal, and a wide-kernel convolutional neural network is constructed to extract deep fault feature information from the samples. Finally, an adversarial training strategy is used to learn the mapping relationship between fault feature information and FCAI, thereby generating high-quality pseudo-fault features. Composite faults are then diagnosed based on the similarity relationship between fault features using a distance metric method.

[0056] In specific implementation, in step 2, the data preprocessing module of the zero-shot learning model uses continuous wavelet transform to convert single-fault vibration data and composite-fault vibration data into single-fault image data d, respectively. s and composite fault image data d t .

[0057] like Figure 2 and Figure 3 As shown, the zero-shot learning model comprises five parts: a data preprocessing module, a feature extraction network module, a fault category auxiliary information module, a generative adversarial network module, and a classification module.

[0058] The data preprocessing module performs time-frequency domain transformation on the raw fault vibration data (Data) of the rolling bearing. In this invention, the time-frequency domain transformation is a continuous wavelet transform method, which converts the one-dimensional time-series fault vibration data of both seen and unseen classes into fault image data. During the continuous wavelet transform, the window size and sliding step size are set to convert the single fault vibration data of the seen class into single fault image data (d). s The vibration data of unseen composite faults are converted into image data of unseen composite faults. t The specific calculation process is shown in the following formula (1).

[0059] (1);

[0060] In formula (1), X s and Y s Let represent the fault sample space and label space of the observed class, respectively. Let _i_ be the index of a single fault image data point of the observed class, where _i_i represents the i-th training sample, i = 1, 2, ..., N, and N be the training dataset D. s The total number of training samples; For the i-th known single fault sample, For this fault sample Corresponding label; X t and Y t Let D be the fault sample space and label space for the unseen class, respectively. Let j be the index of the composite fault image data for the unseen class, and j represent the j-th test sample, j=1,2,…,M, where M is the test dataset D. t The total number of test samples; For the j-th unseen composite fault sample, For this fault sample The corresponding label; x(t) is the original fault input signal, which is a function of time t; () is the complex conjugate function of the wavelet function; It is the scaling parameter of the continuous wavelet transform. is the translation parameter of the continuous wavelet transform; W(a, b) is the formula for the continuous wavelet transform.

[0061] In specific implementation, in step 2, the single fault vibration data and the composite fault vibration data are one-dimensional vibration signals.

[0062] In specific implementation, step 3 involves extracting feature information from the fault image data. The process includes the following steps:

[0063] Step 31: Extract single fault image data d through the feature extraction network module. s Found fault characteristic information v s And predict the single fault image data d of the seen class through the softmax network layer of the feature extraction network. s Fault labels ;

[0064] Step 32: Calculate the actual fault labels and prediction labels The L1 cross-entropy loss between the classifications is used, and the performance of the feature extraction network is optimized by the Adam optimizer through adaptive learning rate and momentum method to obtain the trained feature extraction network.

[0065] The calculation process of the classification cross-entropy loss L1 is shown in the following formula (2);

[0066] (2);

[0067] In formula (2), i is the sequence number of the single fault image data of the observed class, i=1,2,…,N; c is the sequence number of the single fault category of the observed class, c=1,2,…,q; q is the total number of single fault categories. The i-th fault label for a single fault in class c of the known classes; Indicates the first A single fault sample Corresponding fault category labels; It is a single fault sample predicted by the feature extraction network. Fault category labels;

[0068] Step 33: Use the trained feature extraction network to extract composite fault feature information of unseen classes. t .

[0069] The feature extraction network is trained using observed single-fault vibration data, thereby extracting observed fault feature information v from the single-fault vibration data. s And predict the fault label of the fault image. After training, the trained feature extraction network is used to extract features from the composite fault image data d. t Obtain unseen composite fault feature information from unseen composite fault vibration data v t Composite fault characteristic information v t The generator is used to calculate similarity and is optimized through adversarial training to generate pseudo-fault feature information for various composite faults.

[0070] Next, the fault features extracted by the feature extraction network are calculated using the following formula (10). Pseudo-fault features generated by the generator Distance between By distance The similarity is identified, and the distances are sorted in descending order. The category corresponding to the pseudo-composite fault feature that is closest to the composite fault feature is the diagnostic result of that composite fault sample.

[0071] In specific implementation, in step 4, auxiliary information A of the observed fault categories is obtained. s The process includes the following steps:

[0072] Step 41: Construct fault category semantic information S, and obtain single fault category semantic information S for the seen classes. s ; composed of semantic information S of a single fault category s Obtain semantic information S of composite fault categories u ;

[0073] In this invention, fault category auxiliary information A is obtained by combining fault category semantic information and manually defined fault descriptions. sAs shown in Table 1, the semantic information of each type of single fault vibration signal is calculated using time-domain and frequency-domain formulas to obtain a C×1 dimension semantic information vector, where C is the number of time-domain and frequency-domain formulas (in one embodiment of the present invention, Table 1 below contains 20 conventional time-domain and frequency-domain formulas, so the value of C should be 20); the process of calculating the semantic information of the kth type of single fault is shown in the following formula (3).

[0074] (3)

[0075] In formula (3), C represents a C-dimensional vector. A C-dimensional vector representing the semantic information of a single fault category in the k-th class. This represents the semantic information of the first fault category of the k-th single fault. This represents the semantic information of the Cth fault category of the kth type of single fault.

[0076] Table 1: Formulas for Statistical Characteristics in the Time and Frequency Domains

[0077]

[0078] Table 1 contains classic time-domain and frequency-domain signal formulas in this field, which will not be explained in detail here.

[0079] To facilitate the processing of fault category semantic information data, the calculation results of time-domain and frequency-domain feature formulas are obtained. Perform one-hot encoding, i.e. Replace with 0 or 1; set the threshold φ, when hour, The value is 1; conversely, when hour, The value is set to 0 to obtain the C-dimensional fault category semantic information for each type of single fault. However, this setup has some shortcomings in fault diagnosis tasks. On the one hand, the calculation results of the time-domain and frequency-domain formulas... The significant differences make it difficult to set a suitable threshold for the one-hot coding method; on the other hand, some time-domain and frequency-domain statistical characteristic formulas yield different calculation results when calculating different fault categories. All are far greater than or much smaller (In specific implementation, The threshold set by yourself is generally 0.5 (with a range of 0 to 1). This means that the C-dimensional vectors obtained by using one-hot encoding will be highly similar and cannot serve as unique auxiliary information for each type of fault.

[0080] For the reasons mentioned above, this invention proposes a relative threshold method:

[0081] First, time-domain and frequency-domain formulas are used to calculate q types of single fault samples, with each formula yielding q calculation results. And compare the sizes, and calculate the maximum value. For a type of fault, output 1, and calculate the minimum value. A single type of fault outputs 0; q represents the total number of fault categories.

[0082] Secondly, set a relative threshold. (In specific implementation, The threshold you set for yourself is generally 0.5 (ranging from 0 to 1), excluding the maximum value of the above calculation result. and the minimum value of the calculation result Intermediate value calculation results Greater than the relative threshold The output is 0.6, and the intermediate value is calculated as follows. Less than the relative threshold The output is 0.3; finally, the C-dimensional category semantic information S of each type of single fault is obtained. s See formula (4) below; formula (4) includes q C-dimensional category semantic information S s ;

[0083] (4)

[0084] In formula (4), This represents the semantic information of a single fault category, representing the k-th type. This represents the semantic information of the i-th fault category of the k-th single fault. The q-th single fault samples are calculated using time-domain and frequency-domain formulas, with each formula yielding q calculation results. And compare their sizes. The result of the formula for the characteristic is used as a substitute. and There are q calculation results respectively. The minimum and maximum values ​​in the formula. Each formula, after calculation, yields q values. , For the kth .

[0085] Composite faults consist of two or more individual faults. Furthermore, based on the correlation between the composite fault and the individual faults, semantic information of the individual fault categories is used to determine the composite fault. Derivation of semantic information of composite fault categories Semantic information of composite fault categories The dimension is also C×1, and the composite fault semantic information of the r-th class is calculated. The formula is shown in formula (5) below.

[0086] (5)

[0087] In formula (5), C represents a C-dimensional vector (in one embodiment of the present invention, Table 1 below contains 20 conventional time-domain and frequency-domain formulas, so the value of C should be 20). The semantic information of the r-th type of composite fault category is obtained by performing a logical OR operation on the semantic information of multiple single fault categories that constitute the composite fault. , ,…, The semantic information of a single fault constitutes a composite fault, ultimately yielding the semantic information of p-class composite fault categories. Where 1≤r≤p, and p represents the semantic information of the composite fault category. The total number;

[0088] Step 42: Construct manually defined fault description information Obtain manually defined single fault description information R s ;

[0089] Manually defined single fault description information R s Combined fault description information R defined by humans u The acquisition process is as follows: Manual labeling is performed using a one-hot encoding method (0-1). A 1 is assigned if the attribute is present, and a 0 is assigned otherwise. Each type of fault is labeled using multiple attributes, such as fault location, fault severity, and fault cause, to obtain R. s and R u .

[0090] Artificially designed fault attributes, used as auxiliary information A, can be used to establish relationships between known and unseen fault classes. Fault attributes allow for the sharing of fault features, such as fault location, fault severity, and fault cause. These features are easily annotated and converted into readable vector forms through one-hot encoding. Multiple fine-grained attributes are specified as descriptive information for each type of fault, enabling attribute transfer from training faults to target faults, thereby diagnosing unseen fault classes.

[0091] Relative attributes represent the degree to which a category possesses a certain attribute. They are usually normalized to [0,1]. They not only include the functionality of binary attributes but also describe the degree of a certain attribute more precisely. For example, the attribute "active" can be represented by the numbers "0.9", "0.6", and "0.3" in relative attributes, which respectively express "very active", "somewhat active", and "not very active". In contrast, binary attributes can only be divided by 0 and 1. Relative attributes not only include the functionality of binary attributes but also describe the degree of a certain attribute more precisely.

[0092] Therefore, for multiple fine-grained attributes of fault location, operating parameters, and fault severity, a relative attribute approach is used to construct manually defined fault description information R. The fine-grained attributes of fault location include three types: "inner ring," "outer ring," and "rolling element." When a certain attribute is present, the attribute value is set to 1; otherwise, it is 0. The fine-grained attributes of operating parameters include three types: "low speed," "medium speed," and "high speed," with attribute values ​​represented by "0.33," "0.66," and "0.99," respectively. The fine-grained attributes of fault severity include two types: "moderate loss" and "severe loss," with attribute values ​​represented by "0.5" and "0.8," respectively. Manually defined faults are relatively simple. For example, for inner ring faults, the attribute is set to 1, 0, 0; for outer ring faults, it is set to 0, 1, 0; for rolling element faults, it is set to 0, 0, 1; then the combined inner and outer ring faults are set to 1, 1, 0, and other attributes are similar. For example, inner ring faults are expressed as (Ng1, Ng2, Ng3), outer ring faults as (Wg1, Wg2, Wg3), and rolling element faults as (Gg1, Gg2, Gg3). Then, the combined inner and outer ring faults can be represented as (Ng1∪Wg1, Ng2∪Wg2, Ng3∪Wg3), and the combined inner, outer ring, and rolling element faults can be represented as (Ng1∪Wg1∪Gg1, Ng2∪Wg2∪Gg2, Ng3∪Wg3∪Gg3).

[0093] Step 43: Using semantic information S of a single fault category s And manually defined single fault description information R s Calculate and obtain auxiliary information A for a single fault category of the known type. s ;

[0094] In the field of fault diagnosis, manually defined fault feature information is usually limited and cannot fully describe the characteristic attributes of various faults, thus affecting the accuracy of fault diagnosis. Considering that the original fault vibration signal contains rich time-domain and frequency-domain features, fusing manually defined fault attributes and fault semantics defined by feature formulas can provide more complete and accurate fault category auxiliary information A. This method helps to classify target faults more accurately, thereby improving the accuracy of fault diagnosis. In this invention, single fault category auxiliary information A... s The calculation formula is shown in the following formula (6).

[0095] (6)

[0096] In formula (6), λ1 represents the semantic information S of a single fault category. s The weighting parameter, λ2, is the manually defined single fault description information R. s The weight parameters.

[0097] Step 44: Based on the observed single fault category auxiliary information A s Derivation of auxiliary information A for unseen composite fault categories u .

[0098] The composite fault, which serves as the target fault, is composed of multiple individual faults, and its auxiliary information A is derived from the correlation between the individual faults and the composite fault. Therefore, according to the above formula (6), the auxiliary information A of the composite fault category is obtained. u The calculation formula is shown in the following formula (7).

[0099] (7).

[0100] In formula (7), the values ​​of λ1 and λ2 are the same as in formula (6); S u For composite fault category semantic information, R u Define composite fault description information manually.

[0101] In specific implementation, in step 5, the observed fault characteristic information v s And auxiliary information on known fault categories The input is fed into the Generative Adversarial Network (GAN) module, which then obtains pseudo-fault feature information. .

[0102] In specific implementation, the generative adversarial network module includes a generator G and a discriminator D.

[0103] In practice, the fault category auxiliary information will be used. The pseudo-fault characteristics are obtained by inputting random noise Z into the generator G. .

[0104] The present invention also discloses an electronic device, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the described method for diagnosing composite faults in rolling bearings based on generative zero-sample learning.

[0105] The present invention also discloses a non-transient computer-readable storage medium storing computer instructions for causing the computer to execute the aforementioned method for diagnosing composite faults in rolling bearings based on generative zero-shot learning.

[0106] like Figure 2 As shown, based on the fault feature information obtained from the image data in step 3 and the fault category auxiliary information obtained in step 4, the fault feature information v and the fault category auxiliary information A are learned through adversarial training. sThe mapping relationship between them is used to obtain high-quality pseudo-fault feature information. .

[0107] like Figure 3 and Figure 4 As shown, the generative adversarial network learns auxiliary information about fault features and fault categories through the generator G and the discriminator D. Mapping capabilities between them; fault category auxiliary information Random noise Z is used as conditional information input to the generator G to generate pseudo-fault features. Discriminator D is used to identify generated pseudo-fault features. and extracted fault features The differences between them; the loss functions of the generator and discriminator are calculated as shown in the following formula (8).

[0108] (8).

[0109] In formula (8), Let D represent the expected value, D represent the variance, and x represent the sample data. The data is randomly distributed and is in the generation phase. and real data The data points are uniformly sampled between the distributions, as shown in the third formula of formula (8); Pseudo-fault features generated by the generator A represents auxiliary information for a single, known fault category. s Z represents Gaussian noise; This represents the fault features extracted by the deep neural network; λ is the gradient penalty coefficient. For gradient penalty, The parameters for gradient penalty; random numbers. The value is [0, 1].

[0110] The adversarial training method for the generative adversarial network is as follows:

[0111] (1) The generator uses a single fault category to assist information. and Gaussian noise As a condition, the above formula (8) is used to generate pseudo-fault features for a single fault. The discriminator identifies the generated pseudo-fault features. Fault features obtained by feature extraction network Determine whether it is true or false, and then calculate. and Loss of difference between them;

[0112] (2) After training is completed, the generator uses single fault category auxiliary information. and Gaussian noise As a condition, generate single fault characteristic information. composite fault category auxiliary information and Gaussian noise As a condition, generate composite fault feature information. .

[0113] Based on the idea of ​​zero-shot learning, feature information generated solely from random noise is often insufficient for generating features for unknown categories. Therefore, by fusing FCAI (Fault-Copyright-Induced Asynchronous Learning), the mapping relationship between fault features and auxiliary information is learned, thereby improving the quality of generated pseudo-fault features and the diagnostic performance of the model. The objective function of this generative adversarial network is... Represented as formula (9).

[0114] (9)

[0115] In formula (9), z represents random noise. Let D represent the expected value, D represent the variance, and x represent the sample data. The data is randomly distributed and is in the generation phase. and real data The data points are uniformly sampled from the distribution, as shown in the third formula of formula (8). Pseudo-fault features generated by the generator A represents auxiliary information for a single, known fault category. s ; This represents the fault features extracted by the deep neural network; λ is the gradient penalty coefficient. This is a gradient penalty.

[0116] The distance metric method trains a model using a generative adversarial network to generate pseudo-fault feature information for various faults. The fault features extracted by the feature extraction network are calculated using the following formula (10). Pseudo-fault features generated by the generator Distance between The results are output in descending order. The category corresponding to the pseudo-fault feature that is closest to the fault feature is the diagnosis result of the fault sample. The specific calculation process is shown in the following formula (10).

[0118] (10).

[0119] Formula (10), v represents the composite fault feature information extracted by the feature extraction network. t , This represents the pseudo-fault characteristics generated by the generator; p represents the classification accuracy, and p is the semantic information of the composite fault category S.u Total number This represents the number of faulty samples in the test set, where h represents the h-th faulty sample in the test set. The sample label correctly predicted by the model Quantity, This is the predicted label for the h-th fault sample in the test set. After calculating the distance between features, the distances are sorted in descending order. The category corresponding to the pseudo-composite fault feature that is closest to the composite fault feature is the diagnostic result for that composite fault sample.

[0120] p represents the semantic information of the composite fault category S. u The total number of faults, such as the fourth type of compound fault, is p, which is 4.

[0121] By calculating the Euclidean distance between the fault features of each sample and the features of various pseudo-faults, the calculation results are sorted by size and encoded for output. The composite fault category corresponding to the minimum value is selected as the prediction result for the composite fault image. The prediction result is compared with the true label to verify the correctness of the composite fault diagnosis. This invention uses only single fault samples of known classes to train the model, and then diagnoses composite fault samples to verify the effectiveness of the proposed zero-sample fault diagnosis method.

[0122] Specifically, to verify the performance of the proposed generative zero-shot learning-based composite fault diagnosis method for rolling bearings, this method was applied to a fault dataset collected on an experimental platform. On the platform, accelerometers were installed at the 3 o'clock, 9 o'clock, and 12 o'clock positions of the bearing housing to acquire three-channel vibration signals at a sampling frequency of 12800 Hz. The acquired data underwent data fusion processing. This rolling bearing fault dataset includes three single fault types (inner ring fault IF, outer ring fault OF, rolling element fault BF) and four composite fault types (inner and outer ring composite fault IF&OF, inner ring and rolling element composite fault IF&BF, outer ring and rolling element composite fault OF&BF, and inner and outer ring and rolling element composite fault IF&OF&BF). The three single faults were considered as known faults. The effectiveness of this invention was evaluated by measuring the diagnostic accuracy for the four unseen composite fault types, as shown in Table 2.

[0123] Table 2: Laboratory Bearing Failure Categories

[0124]

[0125] The fault diagnosis tasks are divided as shown in Table 3. Each fault sample consists of 1024 continuous sampling data points. In tasks A1 and A2, the number of samples per fault class is 500, and in tasks A3 and A4, the number of samples per fault class is 1000. A comparative analysis was conducted with three other methods, as follows:

[0126] (1) f-CLSWGAN (Y. Xian, T. Lorenz, B. Schiele and Z. Akata, "FeatureGenerating Networks for Zero-Shot Learning," 2018 IEEE / CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 2018, pp.5542-5551);

[0127] (2) f SAE (E. Kodirov, T. Xiang and S. Gong, "Semantic Autoencoder for Zero-Shot Learning," 2017 IEEE Conference on Computer Vision and PatternRecognition (CVPR), Honolulu, HI, USA, 2017, pp. 4447-4456)

[0128] (3) f GZLCFDM (J. Xu, K. Li, Y. Fan and X. Yuan, “A label informationvector generative zero-shot model for the diagnosis of compound faults,” Expert Syst.Appl., vol.233, Dec.2023, Art.no.120875).

[0129] Under the same dataset and training method, five repeated experiments were conducted for each task, and the average of the experimental results was used as the final fault identification accuracy. The fault diagnosis accuracy (%) of each method on different tasks is shown below. Figure 5 As shown, tasks A2 and A4 introduce a composite fault consisting of three types of faults: inner race, outer race, and rolling element. This increases the complexity compared to tasks A1 and A3, leading to a decrease in fault diagnosis performance. Experimental results show that the proposed method achieves optimal results in all fault diagnosis tasks. Specifically, the average fault diagnosis accuracy of the proposed method reaches 75.98%, which is 15.21% higher than the second-best comparison method, demonstrating the effectiveness of using auxiliary information to generate pseudo-fault features for fault diagnosis under zero-sample conditions.

[0130] Table 3 Bearing Fault Diagnosis Tasks

[0131]

[0132] Based on the above, to address the problem of insufficient training samples for compound faults in imbalanced scenarios, this invention proposes a method for training a model using only single fault samples to diagnose compound fault samples. The method comprises five parts: data preprocessing, a feature extraction network, fault category auxiliary information, a generative adversarial network, and a classification module, forming a zero-shot learning fault diagnosis model. First, continuous wavelet transform is performed on the original vibration signal to obtain time-frequency domain image data. Then, a wide-kernel convolutional neural network is used to extract fault feature information. Simultaneously, fault category semantic information and manually defined fault descriptions are used as auxiliary information to learn the correlation between single and compound faults. An adversarial training method is employed to learn the mapping relationship between fault feature information and fault category auxiliary information, obtaining high-quality pseudo-fault features. Finally, a distance metric method is used to diagnose compound faults based on the similarity relationship between fault features. Experiments using a laboratory bearing dataset demonstrate the effectiveness of the proposed method.

[0133] The rolling bearing composite fault diagnosis method based on generative zero-sample learning of the present invention has the following technical advantages and features.

[0134] (1) In view of the scarcity of composite fault samples, this invention proposes a zero-shot learning fault diagnosis model for diagnosing composite faults. The model consists of a feature extraction network, a fault category auxiliary information construction module, a pseudo-fault feature generation and classification module. It uses single fault samples and labels, as well as auxiliary information shared by single faults and composite faults, to diagnose the category of composite fault samples, thereby alleviating the problem of insufficient composite fault sample quantity.

[0135] (2) To establish the connection between single faults and compound faults, this invention proposes a method for designing fault semantic information using relative thresholds. Vibration signals for each type of fault are calculated using time-domain and frequency-domain formulas, while a threshold value is set to encode the semantic information of each type of fault. To enhance the connection between single faults and compound faults, a method for defining auxiliary information for fault categories is proposed. By combining fault semantic information and fault description information, a new auxiliary information is obtained. The semantic information of each type of single fault vibration signal is calculated using time-domain and frequency-domain formulas, and a relative threshold method is used to encode the new fault category semantic information. Simultaneously, artificially designed fault attributes are used as auxiliary information to establish the connection between single faults and compound faults. For multiple fine-grained attributes such as fault location, operating parameters, and fault severity, a relative attribute approach is used to construct fault description information. The auxiliary information for compound faults is derived from the correlation between single faults and compound faults; therefore, compound fault samples are not required, and the auxiliary information for compound fault categories can be obtained using only single fault samples.

[0136] (3) To address the problem of training a fault diagnosis model and accurately diagnosing compound faults in the absence of compound fault samples, and to alleviate the class imbalance problem, this invention proposes a generative zero-shot learning method for intelligent diagnosis of compound faults in rolling bearings. This method considers that single fault samples are easier to obtain than compound fault samples, and aims to diagnose unknown compound faults using single fault samples and FCAI (Fail-Free Learning). The model can identify compound fault categories using only single fault samples for training. Experimental results show that the model has good intelligent diagnostic performance for compound faults.

[0137] This invention presents a method for diagnosing composite faults in rolling bearings based on generative zero-shot learning. It employs deep neural networks and generative adversarial networks to construct a zero-shot learning fault diagnosis model, using only known fault samples to diagnose unknown target faults. This invention provides a novel fault attribute description method, combining fault semantic information with manually defined fault description information to construct Fault Category Auxiliary Information (FCAI), used to learn the correlation between single and composite faults. Furthermore, continuous wavelet transform is used to preprocess the original vibration signals, and a wide-kernel convolutional neural network is constructed to extract deep fault feature information from the samples. Finally, an adversarial training strategy is employed to learn the mapping relationship between fault feature information and FCAI, thereby generating high-quality pseudo-fault features. Composite faults are then diagnosed based on the similarity relationship between fault features using a distance metric method.

[0138] This invention addresses the problem of scarce composite fault samples by proposing a zero-shot learning fault diagnosis model for diagnosing composite faults. This model can identify composite fault categories using only a single fault sample for training.

[0139] This invention proposes a composite fault diagnosis method based on generative zero-shot learning. It designs a fault category auxiliary information to learn the correlation between single and composite faults, thereby assisting the generator in generating fault features and alleviating the problem of limited model diagnostic performance due to a lack of target samples. The model is trained using only single fault samples from known classes, and then composite fault samples are diagnosed to verify the effectiveness of the proposed zero-shot fault diagnosis method.

[0140] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0141] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A rolling bearing composite fault diagnosis method based on generative zero-shot learning, characterized in that, The zero-shot learning model is constructed by using a deep neural network and a generative adversarial network; the zero-shot learning model comprises a data preprocessing module, a feature extraction network module, a fault category auxiliary information module, a generative adversarial network module and a classification module; and the rolling bearing compound fault diagnosis method comprises: Step 1: Constructing training dataset D using single fault vibration data of rolling bearings s and using composite fault vibration data as test dataset D t ; Step 2: Based on the training dataset D s and test dataset D t , will the training dataset D s Single-fault vibration data is converted into single-fault image data. s The test dataset D t The composite fault vibration data in the image is converted into composite fault image data. t ; Step 3: Extracting the fault feature information v of the two kinds of fault image data d and d respectively s and d t ​ Step 4: Based on the training dataset D s Obtain semantic information of fault categories from the raw vibration data. and manually defined fault description information This allows us to obtain auxiliary information A for a single fault category that has already been identified. s ; Step 5: based on the seen similar fault feature information v s and the seen similar fault category auxiliary information A s , a mapping relationship between the fault feature information v and the fault category auxiliary information A s is established, and then the pseudo fault feature information v' is obtained ; Step 6: Based on the adversarial training method in Step 5, calculate the fault feature v extracted by the feature extraction network and the pseudo-fault feature information generated by the generator according to the similarity relationship between fault features. The distance d between the samples is used to sort the results from largest to smallest. The category corresponding to the pseudo-fault feature that is closest to the feature of the unseen sample is the diagnosis result of that unseen sample.

2. The rolling bearing composite fault diagnosis method based on generative zero-shot learning according to claim 1, characterized in that, In step 2, in the data preprocessing module of the zero sample learning model, the single fault vibration data and the compound fault vibration data are respectively converted into single fault image data d s and compound fault image data d t .

3. The rolling bearing composite fault diagnosis method based on generative zero-shot learning according to claim 1, characterized in that, In the step 2, the single fault vibration data and the compound fault vibration data are one-dimensional vibration signals.

4. The rolling bearing composite fault diagnosis method based on generative zero-shot learning according to claim 1, characterized in that, In the step 3, the feature information of the fault image data is extracted The process includes the following steps: Step 31: extracting single fault image data d by a feature extraction network of a feature extraction network module s of the seen class fault feature information v s , and predicting the single fault image data d of the seen class by a softmax network layer of the feature extraction network s fault label ; Step 32: Calculate the classification cross-entropy loss L1 between the real fault label and the predicted label and the predicted label , and optimize the performance of the feature extraction network using the adam optimizer with adaptive learning rate and momentum method to obtain the trained feature extraction network. Step 33: use the trained feature extraction network to extract the composite fault feature information v of the unseen class t .

5. The rolling bearing composite fault diagnosis method based on generative zero-shot learning according to claim 1, characterized in that, In the step 4, the seen fault category auxiliary information A is obtained s The process includes the following steps: Step 41 : constructing the failure category semantic information S, obtaining the single failure category semantic information S of the seen class s ; obtaining the compound failure category semantic information S from the single failure category semantic information S s u ;​ Step 42: Constructing the artificially defined fault description information , obtaining the artificially defined single fault description information R s ; Step 43: obtaining the single fault category auxiliary information A by the single fault category semantic information S s and the artificial definition single fault description information R s Step 44: calculating the seen single fault category auxiliary information A s ; Step 44: Derive seen-class single fault category auxiliary information A s , derive unseen-class composite fault category auxiliary information A u .

6. The rolling bearing composite fault diagnosis method based on generative zero-shot learning according to claim 1, characterized in that, The step 5, the seen similar fault feature information v s And the seen similar fault category auxiliary information A s Input into the generative adversarial network module, obtain the pseudo fault feature information through the generative adversarial network module .

7. The rolling bearing composite fault diagnosis method based on generative zero-shot learning according to claim 6, characterized in that, The generative adversarial network module comprises a generator G and a discriminator D.

8. The rolling bearing composite fault diagnosis method based on generative zero-shot learning according to claim 7, characterized in that, assistance information for a fault category and a random noise z input generator g obtain pseudo fault features .

9. An electronic device comprising at least one processor and a memory connected in communication with the at least one processor; characterized in that, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the rolling bearing compound fault diagnosis method based on generative zero-shot learning according to any one of claims 1-8.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to enable the computer to perform the rolling bearing compound fault diagnosis method based on generative zero-shot learning according to any one of claims 1-8.

Images