Rolling bearing small sample open set cross-working-condition fault diagnosis method based on graph learning

Abstract

The invention provides a rolling bearing small sample open set cross-working-condition fault diagnosis method based on graph learning, and relates to the technical field of rotating machine fault diagnosis. According to the invention, through constructing a dynamic adaptive graph convolution multi-classifier fusion network, designing an unknown fault category discrimination joint strategy and adopting an N-way K-shot element learning mode, the model can realize effective identification of known fault types across working conditions and can effectively identify unknown fault types under the constraint of small samples, so that the method has the advantages of high efficiency and high reliability. Therefore, the accuracy of fault diagnosis of the rolling bearing is improved.

Description

Technical Field

[0001] This invention relates to the technical field of rotating machinery fault diagnosis, and more particularly to a method for fault diagnosis of rolling bearings based on graph learning with small sample open sets across operating conditions. Background Technology

[0002] During operation, rotating machinery inevitably experiences varying degrees of damage to critical components such as rolling bearings and gears. Monitoring the operational status and diagnosing faults in rotating machinery is crucial for ensuring production efficiency and personnel safety. However, in actual industrial settings, obtaining fault samples is difficult; real fault data is often limited in quantity and diverse in type, making it difficult to cover all possible fault scenarios. Furthermore, most existing methods are based on the closed-set assumption, assuming that fault types in the testing phase are consistent with those in the training phase. When unknown faults or abnormal states not included in the training set occur, the model is prone to misclassifying them as a known category, affecting the reliability and safety of the diagnostic results.

[0003] Thanks to the development of computer science and artificial intelligence, deep learning-based fault diagnosis methods can effectively analyze large amounts of data and automatically provide diagnostic results, and have been widely studied in rotating machinery fault identification. Compared with traditional machine learning, which usually relies on manual experience to extract and select features, deep learning methods can automatically learn feature representations from raw monitoring signals in an end-to-end manner, and characterize the fault evolution law through multi-layer nonlinear structures, thus possessing strong feature representation capabilities and generalization performance even under fluctuating operating conditions and noise interference. However, research on fault diagnosis for small sample open sets is still relatively limited. This problem requires models to have stronger feature extraction and transfer generalization capabilities under limited sample conditions, and also requires reliable identification and rejection strategies when unknown states not included in the training set occur. Therefore, researching a graph learning-based fault diagnosis method for rolling bearings with small sample open sets across operating conditions, enabling it to effectively identify known faults and reject and detect unknown faults under limited sample conditions, has significant engineering implications. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention provides a graph learning-based method for fault diagnosis of rolling bearings using small-sample open-set data across operating conditions. This invention constructs a dynamic adaptive graph convolutional multi-classifier fusion network, designs a joint strategy for identifying unknown fault categories, and adopts an N-way K-shot meta-learning mode. This enables the model, under small-sample constraints, to not only effectively identify known fault types across operating conditions but also effectively distinguish unknown fault types, thereby improving the accuracy of rolling bearing fault diagnosis.

[0005] The technical means employed in this invention are as follows:

[0006] A graph learning-based method for fault diagnosis of rolling bearings with small sample open sets across operating conditions includes the following steps: S1. Collect raw vibration signals of rolling bearings in different health states under different operating speeds or loads; the different health states include: normal state, inner ring fault state, outer ring fault state, and rolling element fault state. S2. Perform a fast Fourier transform on the original vibration signal to convert it to the frequency domain, and construct a labeled source domain dataset and an unlabeled target domain dataset; the labeled source domain dataset includes: labeled known fault category samples; the unlabeled target domain dataset includes: unlabeled known fault category samples and unknown fault category samples; S3. Construct a dynamic adaptive graph convolutional multi-classifier fusion network model. Based on the training method of N-class K-sample N-way K-shot few-shot meta-learning, train the network model with the labeled source domain dataset to obtain the source domain trained model. S4. Using the model trained in the source domain, perform preliminary classification on the target domain data to obtain the predicted output probability of each sample belonging to a known category; then, select from the prediction results of the known classes with high prediction probabilities... A sample is selected, and then a sample with a low prediction probability is selected. One sample is used as the representative of the unknown class, and the two samples constitute the support set. The remaining samples are used to construct the query set, forming an open set meta-learning test task. The high prediction probability is the predicted output probability being greater than a first threshold. The low prediction probability is the predicted output probability being less than a second threshold. The first threshold is greater than the second threshold. The trained model is rapidly fine-tuned on the support set of the test task to obtain model parameters adapted to the target domain. Finally, the query set samples are input into the fine-tuned model to obtain the predicted output probability. Based on the output probability, a joint decision is made on whether it meets the preset confidence criterion, prediction entropy criterion, and prototype distance criterion to obtain the fault type information of the target domain data. The joint decision is to compare the confidence, prediction entropy, and prototype distance of the target domain data to be predicted with the corresponding preset criteria. If any of the confidence, prediction entropy, and prototype distance meets the judgment condition, the target domain dataset to be predicted is determined to be of unknown fault type. If none of the confidence, prediction entropy, and prototype distance meet the judgment condition of unknown fault type, the current fault type is predicted according to the trained network model. The current fault type is the fault type corresponding to the maximum probability among the known fault types.

[0007] Furthermore, the step S2, which involves performing a fast Fourier transform on the original vibration signal to convert it to the frequency domain, and constructing a labeled source domain dataset and an unlabeled target domain dataset, includes the following steps: S21, Using a window width of... The sliding window segments each original vibration signal, obtaining a length of... The data sample, and The sliding window's movement step size is... ; S22. Perform a Fast Fourier Transform on each data sample to obtain a sample of length [length missing]. Frequency domain signal; S23. Construct a labeled source domain dataset and an unlabeled target domain dataset based on the frequency domain signal; wherein, the labeled source domain dataset... ,in Representing the number of health states contained in the source domain dataset, the i-th A health status dataset ;in, Indicates the first element in the source domain dataset. The frequency domain signal of a data sample, Indicates the first element in the source domain dataset. Labels for each data sample Indicates the number of samples in the source domain dataset; The unlabeled target domain dataset ;in Let represent the number of health states contained in the target domain dataset, then the th A health status dataset ,in, Indicates the first element in the target domain dataset. The frequency domain signal of each data sample; Indicates the number of samples in the target domain dataset; The number of health states contained in the source domain dataset Less than the number of health states contained in the target domain dataset .

[0008] Furthermore, the dynamic adaptive graph convolutional multi-classifier fusion network model includes: a shallow feature convolution extraction unit, a dynamic adaptive graph convolution feature learning unit, a deep feature convolution extraction unit, and a multi-classifier fusion unit; The shallow feature convolution extraction unit includes: multiple cascaded first convolution sub-modules; each of the first convolution sub-modules sequentially includes: a first convolutional layer, a batch normalization layer, a modified linear unit ReLU activation layer, and a max pooling layer; the convolution kernel parameters of each of the first convolution sub-modules are all different; the output of the last first convolution sub-module is also provided with a random deactivation Dropout layer for suppressing overfitting in small sample training; The dynamic adaptive graph convolutional feature learning unit uses the features output by the shallow feature convolutional extraction unit as graph nodes, and dynamically learns the adjacency matrix through scaling dot product attention. And based on the adjacency matrix Graph convolution operations are performed to aggregate neighborhood information of node features. At the same time, the original convolution features and graph convolution features are fused through residual connections, and a dynamic weight adjustment mechanism based on channel attention is introduced to adaptively control the fusion ratio of the two types of features. The deep feature convolution extraction unit includes a second convolutional Conv2 submodule and a third convolutional Conv3 submodule connected in series; both the second convolutional Conv2 submodule and the third convolutional Conv3 submodule include: a one-dimensional convolutional layer stacked in sequence, a batch normalization layer, a modified linear unit ReLU activation layer, and a max pooling layer; the output of the third convolutional Conv3 submodule is also provided with a random deactivation Dropout layer to suppress overfitting in few-shot learning; The multi-classifier fusion unit is used to adaptively learn the output contribution ratio of the two classifiers through a gradient feedback-based dynamic weight learning strategy.

[0009] Furthermore, the adjacency matrix is ​​dynamically learned through scaling dot product attention. The calculation formula is: ; in, and These represent the asymmetric adjacency matrices generated by two convolutional layers with identical structures but independent parameters; C Representing feature dimension, r This represents the scaling factor.

[0010] Furthermore, the adjacency matrix-based The formula for calculating the graph convolution operation is: ; in, This represents the output features of the shallow feature convolution extraction unit; This represents the output features of graph convolution. and These represent the learning weights and biases of the graph convolution, respectively.

[0011] Furthermore, the dynamic weight adjustment mechanism based on channel attention obtains global statistical features of the channel dimension through global average pooling. Then, the weights are learned and fused through a fully connected network. : ; in, and Both represent the weight matrix of the fully connected layer; This represents the Sigmoid activation function.

[0012] Furthermore, the feature fusion is achieved through dynamic residual connections, and the formula for feature fusion is as follows: ; in, This indicates element-wise multiplication.

[0013] Furthermore, the formula for calculating the predicted output probability of the multi-classifier fusion unit is as follows: ; in, This represents the predicted output probability of the multi-classifier fusion unit; Indicates learnable parameters, Indicates an adjustable weighting factor; This represents the Softmax activation function; This represents the Sigmoid activation function; This represents the predicted output of a standard linear classifier. This is the predicted output of the cosine similarity classifier; Represents the prototype score vector; The prototype score vector is The prototype score vector includes: the first accumulated score vector from the meta-training task. Prototype Center of Each Category and the Within-class variance of each category ;No. The prototype score vectors for each category are: ; in, Indicates the features of the input sample. This represents a preset infinitesimal constant. .

[0014] Compared with the prior art, the present invention has the following advantages: The proposed graph learning-based method for fault diagnosis of rolling bearings across operating conditions with small sample sizes involves several steps. First, a Fast Fourier Transform (FFT) is performed on the collected vibration data, and the resulting spectral signal is used as input to the network model. Second, a dynamic adaptive graph convolutional multi-classifier fusion network model is designed, incorporating an attention-based dynamic residual adjustment mechanism to adaptively fuse original features and graph convolutional features, mitigating the over-smoothing problem caused by graph convolution. Third, a joint decision-making method based on multiple criteria—confidence, prediction entropy, and prototype distance—is designed to distinguish between known and unknown fault types. Finally, an N-way K-shot meta-learning model is used to train and test the constructed network model. The proposed method effectively improves the performance of fault diagnosis for rolling bearings across operating conditions with small sample sizes. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the graph learning-based method for fault diagnosis of rolling bearings with small sample open sets across operating conditions, as proposed in this invention.

[0017] Figure 2 This is a diagram of the dynamic adaptive graph convolutional multi-classifier fusion network structure proposed in this invention.

[0018] Figure 3 This is the confusion matrix of the classification results obtained by the method proposed in this invention.

[0019] Figure 4 This is the confusion matrix of the classification results obtained by the ResNet18-based open-set domain adaptive method.

[0020] Figure 5 This is the confusion matrix of the classification results obtained by the open-set domain adaptive method based on graph neural networks. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] like Figures 1-5 As shown, this invention provides a method for fault diagnosis of rolling bearings with small sample open sets across operating conditions based on graph learning, including the following steps: S1. Collect raw vibration signals of rolling bearings in different health states under different operating speeds or loads; the different health states include: normal state, inner ring fault state, outer ring fault state, and rolling element fault state. S2. Perform a Fast Fourier Transform (FFT) frequency domain transformation on the original vibration signal to construct a labeled source domain dataset and an unlabeled target domain dataset. The labeled source domain dataset includes labeled samples of known fault categories. The unlabeled target domain dataset includes unlabeled samples of known fault categories and samples of unknown fault categories. The FFT frequency domain transformation on the original vibration signal in S2 to construct the labeled source domain dataset and the unlabeled target domain dataset includes the following steps: S21, Using a window width of... The sliding window segments each original vibration signal, obtaining a length of... The data sample, and The sliding window's movement step size is... ; S22. Perform a Fast Fourier Transform on each data sample to obtain a sample of length [length missing]. Frequency domain signal; S23. Construct a labeled source domain dataset and an unlabeled target domain dataset based on the frequency domain signal (in this embodiment, the ratio of labeled source domain dataset to unlabeled target domain dataset is 7:3, and the specific ratio can be set according to the actual situation); the labeled source domain dataset ,in Representing the number of health states contained in the source domain dataset, the i-th A health status dataset ;in, Indicates the first element in the source domain dataset. The frequency domain signal of a data sample, Indicates the first element in the source domain dataset. Labels for each data sample Indicates the number of samples in the source domain dataset; The unlabeled target domain dataset ;in Let represent the number of health states contained in the target domain dataset, then the th A health status dataset ,in, Indicates the first element in the target domain dataset. The frequency domain signal of each data sample; Indicates the number of samples in the target domain dataset; The number of health states contained in the source domain dataset Less than the number of health states contained in the target domain dataset .

[0024] S3. Construct a dynamic adaptive graph convolutional multi-classifier fusion network model. Based on the training method of N-class K-sample N-way K-shot few-shot meta-learning, train the network model using the labeled source domain dataset to obtain the source domain trained model.

[0025] As a preferred implementation, the dynamic adaptive graph convolutional multi-classifier fusion network model includes: a shallow feature convolution extraction unit, a dynamic adaptive graph convolution feature learning unit, a deep feature convolution extraction unit, and a multi-classifier fusion unit; The shallow feature convolution extraction unit includes: multiple cascaded first convolution sub-modules; each of the first convolution sub-modules sequentially includes: a first convolutional layer, a batch normalization layer, a modified linear unit ReLU activation layer, and a max pooling layer; and the convolution kernel parameters of each of the first convolution sub-modules are all different; the output of the last first convolution sub-module is also provided with a random deactivation Dropout layer for suppressing overfitting in small sample training; The dynamic adaptive graph convolutional feature learning unit uses the features output by the shallow feature convolutional extraction unit as graph nodes, and dynamically learns the adjacency matrix through scaling dot product attention. And based on the adjacency matrix Graph convolution operations are performed to aggregate neighborhood information of node features. At the same time, the original convolution features and graph convolution features are fused through residual connections, and a dynamic weight adjustment mechanism based on channel attention is introduced to adaptively control the fusion ratio of the two types of features. The deep feature convolution extraction unit includes a second convolutional Conv2 submodule and a third convolutional Conv3 submodule connected in series; both the second convolutional Conv2 submodule and the third convolutional Conv3 submodule include: a one-dimensional convolutional layer, a batch normalization layer, a modified linear unit ReLU activation layer, and a max pooling layer stacked sequentially; the output of the third convolutional Conv3 submodule is provided with a random deactivation Dropout layer to suppress overfitting in few-shot learning; The multi-classifier fusion unit is used to adaptively learn the output contribution ratio of the two classifiers through a gradient feedback-based dynamic weight learning strategy.

[0026] Among them, the adjacency matrix is ​​dynamically learned by scaling dot product attention. The calculation formula is: ; in, and These represent the asymmetric adjacency matrices generated by two convolutional layers with identical structures but independent parameters; Representing feature dimension, This represents the scaling factor.

[0027] Based on the adjacency matrix The formula for calculating the graph convolution operation is: ; in, This represents the output features of the shallow feature convolution extraction unit; This represents the output features of graph convolution. and These represent the learning weights and biases of the graph convolution, respectively.

[0028] The channel-attention-based dynamic weight adjustment mechanism obtains global statistical features of the channel dimension through global average pooling. Then, the weights are learned and fused through a fully connected network. : ; in, and Both represent the weight matrix of the fully connected layer; This represents the Sigmoid activation function.

[0029] Feature fusion is achieved through dynamic residual connections, and the formula for feature fusion is as follows: ; in, This indicates element-wise multiplication.

[0030] The formula for calculating the predicted output probability of the multi-classifier fusion unit is: ; in, This represents the predicted output probability of the multi-classifier fusion unit; Represents learnable parameters and adjustable weight factors. =0.2; This represents the Softmax activation function; This represents the Sigmoid activation function; This represents the predicted output of a standard linear classifier. This is the predicted output of the cosine similarity classifier; Represents the prototype score vector; The prototype score vector is The prototype score vector includes: the first accumulated score vector from the meta-training task. Prototype Center of Each Category and the Within-class variance of each category ;No. The prototype score vectors for each category are: ; in, Indicates the features of the input sample. This represents a preset infinitesimal constant. .

[0031] S4. Using the model trained in the source domain, perform preliminary classification on the target domain data to obtain the predicted output probability of each sample belonging to a known category; then, select from the prediction results of the known classes with high prediction probabilities... A sample is selected, and then a sample with a low prediction probability is selected. One sample is used as the representative of the unknown class, and the two samples constitute the support set. The remaining samples are used to construct the query set, forming an open set meta-learning test task. The high prediction probability is the predicted output probability being greater than a first threshold. The low prediction probability is the predicted output probability being less than a second threshold. The first threshold is greater than the second threshold. The trained model is rapidly fine-tuned on the support set of the test task to obtain model parameters adapted to the target domain. Finally, the query set samples are input into the fine-tuned model to obtain the predicted output probability. Based on the output probability, a joint decision is made on whether it meets the preset confidence criterion, prediction entropy criterion, and prototype distance criterion to obtain the fault type information of the target domain data. The joint decision is to compare the confidence, prediction entropy, and prototype distance of the target domain data to be predicted with the corresponding preset criteria. If any of the confidence, prediction entropy, and prototype distance meets the judgment condition, the target domain dataset to be predicted is determined to be of unknown fault type. If none of the confidence, prediction entropy, and prototype distance meet the judgment condition of unknown fault type, the current fault type is predicted according to the trained network model. The current fault type is the fault type corresponding to the maximum probability among the known fault types.

[0032] As a preferred implementation, the specific method for joint decision-making using preset confidence criteria, prediction entropy criteria, and prototype distance criteria described in S4 is as follows: During the testing phase, the test sample x is input into the trained dynamic adaptive graph convolutional multi-classifier fusion network model to obtain the predicted output probability of the constructed network model for the test sample. .

[0033] As a preferred implementation, the confidence criterion is: the detection of the maximum predicted output probability. Is it less than the preset confidence threshold? ;if If the condition is met, the sample is determined to be of an unknown fault type; otherwise, the results of the other two criteria are awaited.

[0034] The prediction entropy criterion is as follows: The uncertainty of the test sample is measured by calculating the entropy value of the predicted output probability distribution; the closer the entropy value is to 1, the greater the uncertainty of the prediction result; the calculation formula is: ; in, Represents the entropy value; log represents the logarithm to the base 2; if the entropy value is normalized... Greater than the preset threshold ,Right now ,in If the value of is 0.75, the sample is determined to be of the unknown fault type; otherwise, the results of the other two criteria are awaited. The prototype distance criterion is: the source domain data features output by the deep feature convolution extraction unit in the constructed network model. ,in This indicates the c-th fault type. Features of each sample data This represents the number of samples for the c-th fault type, and the prototype centers of the N known fault categories are calculated. Intra-class average distance and intra-class variance , serving as the geometric reference for the prototype distance criterion.

[0035] Therefore, in summary, the prototype center for the c-th fault type... The calculation formula is as follows: ; The intra-class average distance of the c-th fault type The calculation formula is as follows: ; The intraclass variance of the c-th fault type The calculation formula is as follows: ; Then, the Euclidean distance between the test sample x and each known fault category is calculated. ,in, , This represents the data features of the test sample x output by the deep feature convolution extraction unit of the constructed network model; Then, based on the principle of minimum distance, the prototype of the known fault category that is closest to the test sample x is determined. : ; Finally, based on The principle sets a distance threshold, which is the distance between the test sample x and the nearest known fault category prototype. The distance satisfies ,in and Representing category c If the intra-class mean distance and intra-class variance are within a certain range, the fault type is determined to be unknown; otherwise, the results of the other two criteria are awaited.

[0036] Example: In one embodiment, S1, a vibration acceleration sensor is used to collect operating data of the rolling bearing under different operating conditions; the sampling frequency is set to 12.8 kHz, and the health status of the rolling bearing under each operating condition includes four types: normal, inner ring fault, outer ring fault, and rolling element fault. Operating condition 1—speed is 2600 r / min, load is 0; operating condition 2—speed is 3500 r / min, load is 0.

[0037] S2. Each original vibration signal is segmented using a sliding window with a window width of L=1024 to obtain data samples with a length of L=1024. The moving step size of the sliding window is 512. A fast Fourier transform is performed on each data sample with a length of 1024 to obtain a frequency domain signal with a length of 512. Based on the frequency domain signals of each operating condition (operating condition 1, operating condition 2), a labeled source domain dataset and an unlabeled target domain dataset are constructed.

[0038] For operating condition 1, three health states (i.e., normal - label 0, rolling element fault - label 1, inner race fault - label 2) are selected from the frequency domain signal to construct the labeled source domain dataset. Then, four health states (normal, rolling element fault, inner race fault, unknown fault) are selected from the frequency domain signal after selecting the source domain data to construct the unlabeled target domain dataset, ensuring that the ratio of the number of source domain data to the number of target domain data is 7:3.

[0039] For operating condition 2, data signals of three health states (normal - label 0, rolling element fault - label 1, inner race fault - label 2) are selected from the frequency domain signals to construct the labeled source domain dataset. Then, data signals of four health states (normal, rolling element fault, inner race fault, unknown fault) are selected from the frequency domain signals after the source domain data is selected to construct the unlabeled target domain dataset, ensuring that the ratio of the number of source domain data to the number of target domain data is 7:3.

[0040] In this implementation case, in order to achieve cross-condition fault diagnosis, the source domain dataset of condition 2 and the target domain dataset of condition 1 are selected for analysis.

[0041] S3. Construct a dynamic adaptive graph convolutional multi-classifier fusion network; like Figure 2 As shown, the shallow feature convolution extraction unit consists of two one-dimensional convolutional sub-modules connected in series. Each one-dimensional convolutional sub-module includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer in sequence. The convolutional kernel sizes in the two modules are 5×1 and 3×1, respectively. A Dropout layer is set at the output of the last one-dimensional convolutional sub-module to suppress overfitting in small sample training.

[0042] The dynamic adaptive graph convolutional feature learning unit first uses the 128-dimensional features output by the shallow feature extraction unit as graph nodes, and generates a QK matrix through two structurally identical but parameter-independent 1×1 convolutional layers, according to the formula... Generate adjacency matrix Where C=128 (feature dimension). (Scaling factor) This is used to avoid gradient vanishing due to excessively large inner products. Furthermore, it is based on the normalized adjacency matrix. According to the formula Perform graph convolution operations, where The learnable weight matrix is ​​128×128. A 128-dimensional bias vector is used to aggregate neighborhood information of node features. Then, global average pooling is applied to process the shallow features X to obtain the channel-dimensional global statistical features g, which are then learned and fused using a fully connected network. Fully connected layers W1 and W2, according to the formula Calculation, where The activation function is Sigmoid. Feature fusion is achieved through dynamic residual connections, according to the formula... The calculation, where ⊙ represents element-wise multiplication, adaptively controls the fusion ratio of the original convolutional features and the graph convolutional features, alleviating the problem of over-smoothing of graph convolutional features.

[0043] Furthermore, a deep feature extraction unit (each module consisting of a cascaded Conv2 and Conv3 module, each composed of a one-dimensional convolutional layer, a batch normalization (BN) layer, a ReLU activation layer, and a max pooling layer stacked sequentially, with kernel sizes of 5×1 and 3×1 in the two modules, is used. A Dropout layer is added to the output of the Conv3 module to further suppress the overfitting risk of few-shot learning, thus completing the deep feature extraction. A combination of a standard linear classifier and a cosine similarity classifier is employed. The cosine similarity classifier calculates the cosine similarity between the input features and the prototype vectors of each category, according to the formula... Calculate the predicted output probability ,in , For learnable parameters, This represents the Softmax activation function; This represents the Sigmoid activation function. The prototype score vector s is derived from the prototype centers of the three categories accumulated during the meta-training task. and within-class variance The calculation yields the following formula: ,in =1e-6 (to avoid the denominator being zero).

[0044] Based on the N-way K-shot few-shot meta-learning training mode, the network model parameters are trained and updated using the source domain dataset to obtain the model trained in the source domain, enabling accurate identification and classification of known fault categories in the source domain. The N-way K-shot few-shot meta-learning training mode is adopted, setting N=3 (corresponding to the 3 health states in the source domain) and K=1 (the number of samples for each health state). One sample is selected from each health state to construct the support set, and the remaining samples are used as the query set. Meta-training is implemented based on the Model-Agnostic Meta-Learning algorithm, with a batch size of 100, 50 training rounds, an inner loop learning rate of 0.01, and an outer loop learning rate of 0.001. Stochastic gradient descent is used for optimization, and the loss function is cross-entropy loss. For each meta-training task... Utilizing support sets Perform parameter adaptation according to the formula Update the model parameters ,in This is the cross-entropy loss function. Further utilization of the query set... Calculate the elemental loss according to the formula. Update initial parameters ,in =10 (each training round contains 10 meta-tasks), and cross-task knowledge sharing is achieved through interactive updates in inner and outer loops, ultimately obtaining meta-parameters with rapid adaptation capabilities for small samples. .

[0045] S4. N-way K-shot small-sample meta-learning testing mode; set N=3 (3 known health states), K=1 (number of samples for each health state), P=1 (number of unknown fault categories). Specifically, the model trained in the source domain is used to initially classify the target domain data, obtaining the predicted output probability of each sample belonging to a known category. Then, one sample is selected from each of the known class predictions with high prediction probabilities (predicted output probability greater than 0.8), and one sample is selected from each of the unknown class predictions with low prediction probabilities (predicted output probability less than 0.3). These two samples constitute the support set, and the remaining samples construct the query set, forming an open-set meta-learning testing task. The trained model is quickly fine-tuned on the support set of the testing task to obtain model parameters adapted to the target domain. Finally, the query set samples are input into the fine-tuned model to obtain the predicted output probabilities. ( 1,2,3).

[0046] By jointly making decisions using preset confidence criteria, prediction entropy criteria, and prototype distance criteria, the system can accurately identify and classify known and unknown fault types in the target domain.

[0047] For the confidence criterion: a pre-set confidence threshold. =0.5, calculate the maximum value of the predicted output probability. ;like If so, the sample is determined to belong to an unknown fault type; For the prediction entropy criterion: according to the formula... Calculate the predicted entropy value, by passing through Perform normalization processing. Preset normalization entropy threshold. =0.75; if If so, the sample is determined to belong to an unknown fault type; For the prototype distance criterion: based on the prototype centers of the three known fault categories in the source domain learned during meta-training. Intra-class average distance and intra-class variance Calculate the Euclidean distance between the test sample x and the center of each known class prototype, according to the formula. Determine the most recent prototype category .based on Principles set distance thresholds ,like If so, the sample is determined to belong to an unknown fault type.

[0048] The three criteria are combined by logical OR operation. If any criterion satisfies the discrimination condition for an unknown fault type, the test sample is finally determined to be of an unknown fault type. If none of the three criteria are satisfied, the maximum value of the predicted probability from the model is used for output. This is to determine the known fault types of the test samples.

[0049] This embodiment evaluates the fault diagnosis performance of the model using a confusion matrix, and compares and verifies it with two comparative methods (an open-set domain adaptive method based on ResNet18 and an open-set domain adaptive method based on graph neural networks).

[0050] like Figure 3 As shown in the diagram, the confusion matrix of the diagnostic results of the method of the present invention on the target domain data shows that the recognition accuracy of known fault types (labels 0, 1, and 2) is 99.23%, 99.07%, and 98.54%, respectively, with an average accuracy of 98.95%; the recognition accuracy of unknown fault types (label 3) is 99.33%, with only a small number of samples being misclassified as known categories; the overall diagnostic accuracy is as high as 99.04%, indicating that the model has excellent fault recognition capabilities in small sample, cross-condition, and open set scenarios.

[0051] like Figure 4 As shown, its average accuracy rate for identifying known fault types is 52.54% (i.e., Figure 4The average accuracy rates for identifying known faults (labels 0, 1, and 2) were 52.76%, 56.54%, and 48.32%, respectively; for unknown faults (i.e.... Figure 4 The accuracy rate for identifying the Chinese label 3) was 83.10%, while the overall accuracy rate was only 60.18%; an open-set adaptive method based on graph neural networks, such as... Figure 5 As shown, its average accuracy rate for identifying known fault types is 66.67% (i.e., Figure 5 The average accuracy rates for identifying known faults (labels 0, 1, and 2) were 95.50%, 53.50%, and 51.00%, respectively; for unknown faults (i.e.... Figure 5 The accuracy rate of the middle label 3) is 100%, and the overall accuracy rate is 75.00%. Although it performs well in the identification of unknown faults, the accuracy rate of the identification of known faults is low.

[0052] Comparative experimental results show that the method of this invention alleviates the problem of feature oversmoothing by using a dynamic adaptive graph convolutional multi-classifier fusion network, thus improving feature extraction and transfer generalization capabilities under small sample conditions. Simultaneously, the multi-criteria joint decision-making strategy achieves accurate identification and rejection of unknown faults, significantly outperforming existing comparative methods. The method of this invention significantly improves fault diagnosis generalization ability and identification accuracy in small sample, cross-operating condition, and open-set scenarios, providing a reliable technical solution for rotating machinery fault diagnosis.

[0053] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the above embodiments of the present invention, the descriptions of each embodiment have their own emphasis; parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. It should be understood that the disclosed technical content in the several embodiments provided by the present invention can be implemented in other ways.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fault diagnosis of rolling bearings with small sample open sets across operating conditions based on graph learning, characterized in that, Includes the following steps: S1. Collect the original vibration signals of rolling bearings in different health states under different operating speeds or loads during operation; The different health states include: normal state, inner ring fault state, outer ring fault state, and rolling element fault state; S2. Perform a fast Fourier transform on the original vibration signal to convert it to the frequency domain, and construct a labeled source domain dataset and an unlabeled target domain dataset; the labeled source domain dataset includes: labeled known fault category samples; the unlabeled target domain dataset includes: unlabeled known fault category samples and unknown fault category samples; S3. Construct a dynamic adaptive graph convolutional multi-classifier fusion network model. Based on the training method of N-class K-sample N-way K-shot few-shot meta-learning, train the network model with the labeled source domain dataset to obtain the source domain trained model. S4. Using the model trained in the source domain, perform preliminary classification on the target domain data to obtain the predicted output probability of each sample belonging to a known category; then, select from the prediction results of the known classes with high prediction probabilities... A sample is selected, and then a sample with a low prediction probability is selected. One sample is used as the representative of the unknown class, and the two samples constitute the support set. The remaining samples are used to construct the query set, forming an open set meta-learning test task. The high prediction probability is the predicted output probability being greater than a first threshold. The low prediction probability is the predicted output probability being less than a second threshold. The first threshold is greater than the second threshold. The trained model is rapidly fine-tuned on the support set of the test task to obtain model parameters adapted to the target domain. Finally, the query set samples are input into the fine-tuned model to obtain the predicted output probability. Based on the output probability, a joint decision is made on whether it meets the preset confidence criterion, prediction entropy criterion, and prototype distance criterion to obtain the fault type information of the target domain data. The joint decision is to compare the confidence, prediction entropy, and prototype distance of the target domain data to be predicted with the corresponding preset criteria. If any of the confidence, prediction entropy, and prototype distance meets the judgment condition, the target domain dataset to be predicted is determined to be of unknown fault type. If none of the confidence, prediction entropy, and prototype distance meet the judgment condition of unknown fault type, the current fault type is predicted according to the trained network model. The current fault type is the fault type corresponding to the maximum probability among the known fault types.

2. The method for fault diagnosis of rolling bearings based on graph learning with small sample open sets across operating conditions as described in claim 1, characterized in that, The step S2, which involves performing a fast Fourier transform on the original vibration signal to convert it to the frequency domain and constructing labeled source domain datasets and unlabeled target domain datasets, includes the following steps: S21, Using a window width of... The sliding window segments each original vibration signal, obtaining a length of... The data sample, and The sliding window's movement step size is... ; S22. Perform a Fast Fourier Transform on each data sample to obtain a sample of length [length missing]. Frequency domain signal; S23. Construct a labeled source domain dataset and an unlabeled target domain dataset based on the frequency domain signal; wherein, the labeled source domain dataset... ,in Representing the number of health states contained in the source domain dataset, the i-th A health status dataset ;in, Indicates the first element in the source domain dataset. The frequency domain signal of a data sample, Indicates the first element in the source domain dataset. Labels for each data sample Indicates the number of samples in the source domain dataset; The unlabeled target domain dataset ;in Let represent the number of health states contained in the target domain dataset, then the th A health status dataset ,in, Indicates the first element in the target domain dataset. The frequency domain signal of each data sample; Indicates the number of samples in the target domain dataset; The number of health states contained in the source domain dataset Less than the number of health states contained in the target domain dataset .

3. The method for fault diagnosis of rolling bearings based on graph learning with small sample open sets across operating conditions as described in claim 1, characterized in that: The network model includes: a shallow feature convolution extraction unit, a dynamic adaptive graph convolution feature learning unit, a deep feature convolution extraction unit, and a multi-classifier fusion unit; The shallow feature convolution extraction unit includes: multiple cascaded first convolution sub-modules; each of the first convolution sub-modules sequentially includes: a first convolutional layer, a batch normalization layer, a modified linear unit ReLU activation layer, and a max pooling layer; the convolution kernel parameters of each of the first convolution sub-modules are all different; the output of the last first convolution sub-module is also provided with a random deactivation Dropout layer for suppressing overfitting in small sample training; The dynamic adaptive graph convolutional feature learning unit uses the features output by the shallow feature convolutional extraction unit as graph nodes, and dynamically learns the adjacency matrix through scaling dot product attention. And based on the adjacency matrix Graph convolution operations are performed to aggregate neighborhood information of node features. At the same time, the original convolution features and graph convolution features are fused through residual connections, and a dynamic weight adjustment mechanism based on channel attention is introduced to adaptively control the fusion ratio of the two types of features. The deep feature convolution extraction unit includes a second convolutional Conv2 submodule and a third convolutional Conv3 submodule connected in series; both the second convolutional Conv2 submodule and the third convolutional Conv3 submodule include: a one-dimensional convolutional layer, a batch normalization layer, a modified linear unit ReLU activation layer, and a max pooling layer stacked sequentially; the output of the third convolutional Conv3 submodule is provided with a random deactivation Dropout layer to suppress overfitting in few-shot learning; The multi-classifier fusion unit is used to adaptively learn the output contribution ratio of the two classifiers through a gradient feedback-based dynamic weight learning strategy.

4. The method for fault diagnosis of rolling bearings based on graph learning with small sample open sets across operating conditions as described in claim 3, characterized in that, The adjacency matrix is ​​dynamically learned through scaling dot product attention. The calculation formula is: ； in, and These represent the asymmetric adjacency matrices generated by two convolutional layers with identical structures but independent parameters; Representing feature dimension, This represents the scaling factor.

5. The method for fault diagnosis of rolling bearings based on graph learning with small sample open sets across operating conditions as described in claim 3, characterized in that, The adjacency matrix The formula for calculating the graph convolution operation is: ； in, This represents the output features of the shallow feature convolution extraction unit; This represents the output features of graph convolution. and These represent the learning weights and biases of the graph convolution, respectively.

6. The method for fault diagnosis of rolling bearings based on graph learning with small sample open sets across operating conditions as described in claim 5, characterized in that, The dynamic weight adjustment mechanism based on channel attention is as follows: global statistical features of the channel dimension are obtained through global average pooling. Then, the weights are learned and fused through a fully connected network. : ； in, and Both represent the weight matrix of the fully connected layer; This represents the Sigmoid activation function.

7. The method for fault diagnosis of rolling bearings based on graph learning with small sample open sets across operating conditions as described in claim 6, characterized in that, Feature fusion is achieved through dynamic residual connections, and the formula for feature fusion is as follows: ； in, This indicates element-wise multiplication.

8. The method for fault diagnosis of rolling bearings based on graph learning with small sample open sets across operating conditions as described in claim 3, characterized in that, The formula for calculating the predicted output probability of the multi-classifier fusion unit is as follows: ； in, This represents the predicted output probability of the multi-classifier fusion unit; Indicates learnable parameters, Indicates an adjustable weighting factor; This represents the Softmax activation function; This represents the Sigmoid activation function; This represents the predicted output of a standard linear classifier. This is the predicted output of the cosine similarity classifier; Represents the prototype score vector; The prototype score vector is The prototype score vector includes: the first accumulated score vector from the meta-training task. Prototype Center of Each Category and the Within-class variance of each category ;No. The prototype score vectors for each category are: ； in, Indicates the features of the input sample. This represents a preset small constant. .

Images