Cross-bearing single sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold

Abstract

The invention relates to a cross-bearing single sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold, and belongs to the technical field of rotating machinery fault diagnosis. Aiming at the problems of insufficient global task distribution learning ability, small sample over-fitting, Euclidean modeling limitation and the like of the existing meta learning method in a cross-domain single-sample scene, a cognitive guidance Riemannian meta learning framework is provided. According to the technical scheme, the method comprises the following steps: 1) constructing cognitive prototype learning global task distribution, and guiding a model to extract high-quality general meta-knowledge from multiple tasks; 2) designing a cognitive adaptive factor to dynamically adjust source domain memory, enhancing target domain adaptation and reducing single sample deviation; and 3) introducing a Riemann metric driving strategy, mapping the data to a Grassmann manifold space, and enhancing the non-linear feature discrimination ability by using geodesic distance. According to the method, the average diagnosis accuracy in a cross-bearing single sample task reaches 93.46% and is improved by 10.04% compared with an existing optimal method, and the accuracy and generalization ability under complex working conditions are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotary machinery fault diagnosis, and relates to a cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold. Background Art

[0002] As an important part of rotary mechanical equipment, bearings are widely used in fields such as energy, aviation, and automotive. Due to long-term operation and harsh working environments, bearings are prone to various types of faults, which may lead to economic losses and even casualties. Therefore, fault diagnosis, condition monitoring, and effective maintenance of bearings are of great significance. Deep learning methods have been widely applied in the field of intelligent fault diagnosis with their powerful feature extraction and fault discrimination capabilities, and promising results have been achieved. However, most deep learning-based fault diagnosis methods require sufficient labeled samples, and the data distributions between the source domain and the target domain need to be kept as consistent as possible to obtain better diagnostic performance. These conditions are often difficult to meet in actual engineering. Therefore, methods for cross-bearing fault diagnosis using limited samples have great research potential.

[0003] Meta-learning, as an effective cross-domain few-shot method, learns general meta-knowledge through the training of multiple meta-tasks, enabling it to quickly adapt to new tasks not involved in training in the target domain. In recent years, some studies have focused on exploring the application of meta-learning in the fault diagnosis of rotating machinery. Zhang et al. proposed an adaptive loss-weighted meta-residual network, effectively solving the problem of fault diagnosis with noisy labels. In addition, Zhang et al. designed a momentum prototype method to mitigate the negative impact of prototype oscillation between batches. In 2023, Lin et al. proposed Generalized Model-Agnostic Meta-Learning (GMAML), which uses heterogeneous signals of bearings to drive the model and achieves high-precision cross-domain few-shot fault diagnosis. Wang et al. improved the overfitting problem in traditional few-shot bearing fault diagnosis by adjusting and transferring neuron parameters between different meta-tasks through freezing operations. Li et al. enhanced the meta-learning method using the attention mechanism and demonstrated the effectiveness of this method in fault transfer diagnosis between different bearings. In 2024, Qin et al. proposed an Adaptive Generic Prototype Network (AGPN), which improved the accuracy and training stability of cross-bearing few-shot fault diagnosis by enhancing the prototype representation and its update method. Wang et al. proposed a correlation-weighted Manhattan distance, emphasizing the correlation between different prototypes and correcting the adverse effects of interference values on traditional metric-based classifiers. Shao et al. proposed a method based on task supervision for an almost inner-loop-free model, improving the task adaptability of existing meta-learning models. Wang et al. proposed a brain-inspired meta-learning strategy and successfully applied it to few-shot fault diagnosis of bearings.

[0004] These methods provide innovative inspiration for the research on cross-domain few-shot rotating machinery fault diagnosis. However, there are still some limitations and challenges that need to be further explored. 1) In complex cross-domain diagnosis scenarios, existing meta-learning frameworks only adapt to specific meta-tasks through the guidance of support samples, which limits the model's ability to learn general meta-knowledge from the global task distribution, resulting in poor generalization ability for new tasks in the target domain. In addition, when the number of support samples is limited, the error introduced by few-shot samples will have a negative impact on the effective and accurate adaptation of the diagnostic model to the current meta-task. 2) When the training samples are too few, the model is prone to overfitting in the source domain. It may overly memorize the details of each training sample to well adapt to the meta-training task, while ignoring the commonalities between tasks, thus making it difficult to generalize to the meta-test task. Therefore, reducing the task preference in the training domain and making the model applicable to cross-domain tasks pose challenges to meta-learning methods. 3) Most existing meta-learning methods rely on Euclidean modeling. However, in practical applications, data often exhibits highly complex and non-linear structures, and traditional Euclidean structures cannot effectively capture and utilize this geometric information, which limits the model's learning and fault discrimination capabilities. 4) In the face of extreme one-shot scenarios, common few-shot methods often struggle to obtain enough knowledge to adapt to the current task, resulting in a decline in their diagnostic accuracy.

[0005] To address these problems, the present invention proposes a Cognition-guided Riemannian Meta-learning (CGRML) method for cross-domain few-shot fault diagnosis. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a cross-bearing one-shot intelligent diagnosis method based on cognitive guidance and Riemannian manifold.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The cross-bearing one-shot intelligent diagnosis method based on cognitive guidance and Riemannian manifold includes the following steps:

[0009] Step 1: Collect bearing fault signal samples through a rolling bearing test bench;

[0010] Step 2: Divide the collected samples into a source domain and a target domain, and respectively divide support sets and query sets in the source domain and the target domain to generate meta-tasks required for the training and test phases;

[0011] Step 3: Construct a feature extraction module, and initialize the cognitive prototype CP and the cognitive adaptation factor CAF;

[0012] Step 4: In the inner loop, calculate the adaptation loss based on the support set and update the network parameters of the feature extraction module;

[0013] Step 5: In the outer loop, evaluate the similarity between the query sample and each cognitive prototype using the geodesic distance, diagnose the health status of the query sample, calculate the total outer loop loss, and update all model parameters through backpropagation;

[0014] Step 6: Quickly adapt the meta-task of the target domain through the trained CGRML model and diagnose the health status of the target domain samples.

[0015] Furthermore, in the said Step 3, the specific way of initializing the cognitive prototype CP and the cognitive adaptation factor CAF is as follows:

[0016] The cognitive prototype (CP) for each category is randomly sampled from the standard normal distribution, expressed as:

[0017]

[0018] where, is the set of cognitive prototypes for all categories, P n is the prototype of the nth category, ρ i is the learnable parameter, and d is the embedding space dimension;

[0019] The cognitive adaptation factor (CAF) is initialized as:

[0020]

[0021] where K is the number of support samples shot for each category, and the ReLU activation function is used to ensure that τ is positive and greater than the extremely small constant ε.

[0022] Furthermore, in the said Step 4, the calculation formula of the adaptation loss is:

[0023]

[0024] where, is the geodesic distance metric on the Grassmann manifold, is the output of the feature extraction module, τ′τ′ is the CAF value after passing through the ReLU activation, is the support set sample.

[0025] Furthermore, the calculation process of the said geodesic distance metric includes:

[0026] Normalize the cognitive prototype P n :

[0027]

[0028] Map the normalized v n to the one-dimensional subspace basis in the Grassmann manifold space G(1, d);

[0029] Calculate the geodesic distance between the sample feature and the cognitive prototype through the projection mapping:

[0030]

[0031] where Φ is the projection mapping function, is the manifold embedding function.

[0032] Furthermore, in the step 5, the total loss of the outer loop is composed of the classification loss L c and the separability loss L s weighted, and the calculation formula is:

[0033] L t = L c + λL s

[0034] where λ is the weight hyperparameter;

[0035] The classification loss L c is calculated through the cross-entropy loss function:

[0036]

[0037] where, is the true label, and p(y = n) is the probability that the sample belongs to class n;

[0038] The separability loss L s is calculated by constraining the distances between the cognitive prototypes of different classes:

[0039]

[0040] Furthermore, in the step 5, the geodesic distance is used to calculate the similarity between the query sample and each cognitive prototype and generate the classification probability:

[0041]

[0042] where, is the feature extraction module updated by the inner loop, and x n,jQ is the query set sample.

[0043] Furthermore, the generation method of the meta-task is:

[0044] Each meta-task T i follows the "N-way K-shot" paradigm, where N is the number of classes and K is the number of support samples for each class;

[0045] In the training phase, meta-tasks are randomly sampled from the source domain, and in the testing phase, meta-tasks are sampled from the target domain.

[0046] Furthermore, the feature extraction module is implemented by a multi-scale convolutional neural network CNN, which is used to extract high-dimensional features from the input signal.

[0047] Furthermore, in step 4, the update process of the inner loop is as follows:

[0048]

[0049] where θ is the parameter of the feature extraction module, θ′ represents the updated parameter, α1 is the learning rate of the inner loop, and L a is the adaptation loss.

[0050] Furthermore, in step 5, the update process of the outer loop is as follows:

[0051]

[0052] ρ←ρ - α3▽ ρ ∑L t (ρ)

[0053] τ←τ - α4▽ τ ∑L t (τ′)

[0054] where α2, α3, and α4 are the learning rates of the feature extraction parameter, the cognitive prototype parameter, and the CAF in the outer loop, respectively.

[0055] The beneficial effects of the present invention are as follows:

[0056] (1) A new meta-learning framework, called Cognition-guided Meta-learning (CGML), is proposed to enhance the model's ability to learn high-quality general meta-knowledge from the global task distribution. A Cognitive Prototype (CP) is designed to accurately guide the adaptation process of the inner loop, reduce the model's over-reliance on support samples in specific meta-tasks, and at the same time reduce the negative impact of small-sample introduction errors.

[0057] (2) A Cognition Adaptive Factor (CAF) is proposed and introduced into the loss function to dynamically adjust the fitting strength of the support set, constrain the model's memory of the source domain, and flexibly adapt to target domain tasks.

[0058] (3) The Riemannian Metric-driven Strategy (RMDS) was proposed and implemented in CGML. By imposing Riemannian geometric constraints on the CGML model, the geometric perception update of the model in the Grassmann manifold space was realized. In addition, RMDS enables the model to better capture the geometric structure of the data, thereby enhancing its discriminant classification ability in complex scenarios.

[0059] (4) Based on CGML, CAF, and RMDS, a new fault diagnosis framework called CGRML was constructed. This model was successfully applied to single-sample fault diagnosis under different bearings and working conditions, and the experimental results verified its superiority over other classical and advanced comparison methods.

[0060] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following specification. Brief Description of the Drawings

[0061] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0062] Figure 1 is the diagnostic flowchart of the CGRML model;

[0063] Figure 2 is the schematic diagram of the inner loop of CGML;

[0064] Figure 3 is the schematic diagram of the outer loop of CGML. Detailed Embodiments

[0065] The following specific examples illustrate the embodiments of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention schematically, and the following embodiments and the features in the embodiments can be combined with each other without conflict.

[0066] Among them, the attached drawings are only for illustrative purposes, showing only schematic diagrams rather than actual drawings, and should not be construed as limiting the present invention; in order to better illustrate the embodiments of the present invention, some components in the attached drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the attached drawings may be omitted.

[0067] In the attached drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the attached drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, so the terms describing the positional relationship in the attached drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0068] A cross-bearing single-sample fault diagnosis method based on cognitive-guided Riemannian meta-learning includes the following steps:

[0069] 1) Collect bearing fault signal samples through a rolling bearing test bench;

[0070] 2) Divide the support set and query set in the source domain and target domain, and generate the meta-tasks used in the training and testing phases;

[0071] 3) Build a feature extraction module, and initialize the cognitive prototype and cognitive adaptation factor;

[0072] 4) In the inner loop, calculate the adaptation loss and update the network parameters;

[0073] 5) In the outer loop, use the geodesic distance to evaluate the similarity between the query sample and each class prototype, diagnose the health status of the query sample, and calculate the total outer loss, and update all model parameters through backpropagation;

[0074] 6) Through the trained CGRML model, quickly adapt to the meta-tasks in the target domain and diagnose the health status of the target domain samples.

[0075] Specific flowcharts are as Figure 1 shown.

[0076] Cognitive-guided meta-learning:

[0077] At the beginning of model training, randomly initialize the CPs of each class, and each feature value is sampled from the standard normal distribution:

[0078]

[0079] In the formula is the CPs, P n is the prototype of the nth class, ρ i is the learnable parameter, and d is the hyperparameter representing the dimension of the embedding space.

[0080] The randomly initialized CPs are iteratively updated through the corresponding loss function in the outer loop, retaining the abstract perception of the task distribution in their parameters and gradually obtaining the general class prototypes that can adapt to various meta-tasks, so that the model can better learn how to learn.

[0081] Meanwhile, initialize the CAF as:

[0082]

[0083] where K represents the number of samples (shot) in the task.

[0084] The CAF is initialized based on the number of samples of each class in the meta-training task and adaptively updated during training. When the number of training samples of a certain class is small, the introduced error is large and the model is more likely to overfit. At this time, τ is given a large initial value to mitigate the negative impact of the introduced error and at the same time adjust the fitting of the model to the source domain. On the contrary, when the samples are sufficient, a small initial value of τ makes the model trust the input training data more.

[0085] Using the ReLU activation function ensures that τ is positive and greater than a very small constant ε to ensure the stability of the model training process:

[0086] τ′ = max(ReLU(τ), ε)

[0087] In CGML, the task distribution is represented as Each meta-task is used as the basic unit for training and testing the model. A multi-scale convolutional neural network is used as the feature extractor of CGML

[0088] In the inner loop of CGML, guides the learning direction of the model. The adaptation loss L of the inner loop a is designed as follows:

[0089]

[0090] where, represents the geodesic distance metric on the Grassmann manifold, which will be discussed later. When the distance between the support sample feature and the class prototype is less than the CAF, L ais set to zero, and at this time, it is considered that the CGML model has completed the adaptation to the current meta-task, and the network parameters will no longer be updated.

[0091] By designing L a , the model adapts to the current meta-task, specifically manifested as bringing the support samples of each category closer to their category prototypes until within the range of τ'. This design ensures that the model learns in a direction beneficial to the current meta-task under its understanding of , rather than determining the learning direction of the inner loop only based on the support set of each specific meta-task. This prevents the model from over-relying on support samples and ignoring the shared meta-knowledge between tasks. In addition, introducing the parameter τ constrains the model to avoid overfitting to a single sample while reducing the negative impact of the error introduced by small samples on the learning process.

[0092] Through gradient descent, the update process for each step in the inner loop is as follows:

[0093]

[0094] where θ represents the feature extractor parameters, θ' represents the updated parameters, and the learning rate α1 is set to 0.1.

[0095] By updating the parameter θ in the inner loop step, the updated feature extractor extracts the sample features that gather in their neighborhoods under the guidance of the CPs of their respective categories, thus completing the adaptation to the specific meta-task.

[0096] As Figure 2 shown, when there is only one sample for each category, using this sample to represent its category will introduce significant bias. In contrast, the CPs learned from different meta-tasks can represent the category prototypes more accurately and effectively. By constraining the samples within the range of τ' close to the CPs, the bias can be effectively reduced, thereby guiding the model to learn in the correct direction and better performing subsequent fault identification. At the same time, the existence of τ' reduces the overfitting of the CGML model to a certain sample.

[0097] In the outer loop of CGML, the health status of the query sample is identified through a classifier based on the geodesic distance metric:

[0098]

[0099] where represents the feature extractor after the inner loop adaptation, and p(y = n) represents the probability that the sample belongs to category n.

[0100] Then, the classification loss L c is calculated as follows:

[0101]

[0102] wherein is the true label of the query sample.

[0103] Meanwhile, CAFτ is used to constrain the clustering boundaries of different classes not to overlap, and then the separability loss L is calculated s as follows:

[0104]

[0105] When the distance between class prototypes is greater than 2τ′, L s is set to zero. At this time, there is no overlap between different classes, and the CGML model pays more attention to learning and updating the positions of class prototypes from the results of the classification meta-task to retain more meta-knowledge.

[0106] The total loss function L of the outer loop t is expressed as:

[0107] L t = L c + λL s

[0108] where λ is the weight hyperparameter, which is set to 0.1 in this study.

[0109] After iterating over all sets in each training epoch, all parameters of CGML are updated as follows:

[0110]

[0111] where α2 and α3 are set to 0.0001, and α4 is set to 0.01, denotes the meta-task randomly selected from the i-th set. In particular, the update of the feature extractor parameters in the inner loop is not retained.

[0112] The general meta-knowledge learned by the model in different meta-tasks is stored in these parameters to guide it to process subsequent similar meta-tasks. Figure 3 Shows the process of query sample classification in the outer loop, the update of CPs and CAF, and the constraint that the inter-class distance exceeds 2τ′.

[0113] Riemannian metric-driven strategy:

[0114] The intrinsic geometric structure of high-dimensional data can usually be represented in a low-dimensional form, and the embedded subspace can effectively capture this structure. By representing and analyzing the inherent geometric information of data on the manifold space through the embedded subspace, a fault diagnosis model that overcomes the limitations of Euclidean distance-based modeling can be constructed. This method can more accurately identify the health status of samples even when the data shows a complex non-linear distribution.

[0115] When evaluating the similarity between samples, first normalize the class prototype P n as follows:

[0116]

[0117] The normalization process eliminates the influence of vector scale on its position, which is particularly important in cross-domain tasks because data collected by different machines vary. This processing helps enhance the generalization ability of the model and enables it to better adapt to new tasks.

[0118] Then, take v n as the basis vector of the one-dimensional prototype subspace :

[0119]

[0120] The column space of matrix Y n = [v n is used to represent this subspace. It is easy to prove Therefore, P n is mapped to a point in the Grassmann manifold space, denoted as:

[0121]

[0122] can be regarded as a set of undirected axes passing through the origin in the embedding space. Similarly, each sample feature in the embedding space is also mapped to Thus, the embedding space is extended to the Grassmann manifold space, denoted as:

[0123]

[0124] Next, obtain the analytical manifold equivalent to through projection mapping, and use the geodesic distance induced by the Riemannian metric to evaluate the similarity between the query sample feature and P n :

[0125]

[0126] Similarly, all the distance functions used above are based on this equation. Therefore, the Riemannian metric drives the entire learning process of the CGML model.

[0127] By applying RMDS, the meta-learning modeling technique based on Euclidean distance is extended to the Riemannian manifold space, which performs excellently in dealing with complex high-dimensional non-linear data. This geometric learning paradigm provides a mathematically principled method for utilizing limited fault data while maintaining generalization performance, which is crucial for some practical industrial applications where data is extremely scarce.

[0128] 1) The CWRU bearing dataset was collected by Case Western Reserve University (CWRU) and is widely used in the field of fault diagnosis. Four health states in this dataset are used in this paper, including normal condition (NC), inner race fault (IF), ball fault (BF), and outer race fault (OF), with a fault diameter of 7 mils. The present invention only uses the data collected under the conditions of a rotational speed of 1730 rpm and a load of 3 HP.

[0129] 2) The SWJTU bearing dataset was collected by Southwest Jiaotong University (SWJTU). This dataset contains four different health states, including NC, IF, BF, and OF. For this dataset, the present invention only uses the data collected under the conditions of a rotational speed of 896.1 rpm and a load of 3 Nm.

[0130] 3) The bearing vibration signals under time-varying working conditions are collected using a self-built bearing test bench. The self-built test bench mainly consists of a driving motor, a loading motor, a torque measuring instrument, and a gear reducer. During the experiment, vibration signals are collected from four health states (NC, IF, BF, and OF). Each health state is tested under the conditions of a rotational speed range from 500 rpm to 1000 rpm (with a change rate of 300 rpm / s) and a constant load of 2 N. The obtained dataset is called the Bearing Fault Detection (BFD) dataset.

[0131] Table 1 shows the detailed information of the three datasets. Using these datasets, six cross-bearing one-shot tasks are constructed to verify the effectiveness and superiority of the proposed method, as shown in Table 2. It should be noted that only 5 samples are used for training for each health state to simulate an industrial scenario where the number of training samples is extremely scarce, while 200 samples are used for testing for each health state. The sample length is set to 1024.

[0132] Table 1 Detailed information of the three datasets

[0133] Name Dataset Rotational speed Load A CWRU 1730 rpm 3 HP B SWJTU 896.1 rpm 3 Nm C BFD 500 to 1000 rpm (300 rpm / s) 2 Nm

[0134] Table 2 Specific descriptions of different tasks

[0135] Task Source domain Target domain Number of classes Number of samples <![CDATA[T1]]> A B 4-way 1-shot <![CDATA[T2]]> B A 4-way 1-shot <![CDATA[T3]]> A C 4-way 1-shot <![CDATA[T4]]> C A 4-way 1-shot <![CDATA[T5]]> B C 4-way 1-shot <![CDATA[T6]]> C B 4-way 1-shot

[0136] To comprehensively verify the effectiveness and superiority of the proposed CGRML method, several classical and advanced fault diagnosis methods were used as comparison methods, including Relation Network (RelaNet), Prototype Network (ProtoNet), Model-agnostic Meta-learning (MAML), GMAML, and AGPN. In the meta-training stage, 100 epochs were set, and each round contained 30 sets corresponding to 30 meta-tasks. It should be noted that each meta-task was in the "4-way 1-shot" paradigm to simulate extreme single-sample scenarios. The number of inner-loop steps was set to 16. In the test stage, 200 meta-tasks were input into the trained model, and the average accuracy and standard deviation (SD) were recorded. To reduce the influence of randomness, all experiments were repeated 10 times.

[0137] In the six cross-bearing single-sample tasks, the experimental results are shown in Table 3. Compared with the classical meta-learning methods RelaNet, ProtoNet, and MAML, the average accuracy of the proposed CGRML increased by 59.57%, 32.76%, and 19.69% respectively in the six tasks. Compared with the advanced cross-domain few-shot fault diagnosis methods GMAML and AGPN, the average accuracy of CGRML increased by 16.92% and 10.04% respectively. These results highlight the effectiveness and superiority of the proposed CGRML in the six cross-bearing single-sample tasks. It should be noted that in the case where the target domain involves time-varying working conditions (especially T3 and T5), due to the increased intra-class variability and more complex data distribution, cross-bearing single-sample fault diagnosis becomes more challenging. Even in this case, the proposed CGRML still achieved diagnostic accuracies of 88.52% and 90.51%, which were 22.85% and 17.65% better than the sub-optimal methods respectively. In addition, the average SD of the proposed method was the smallest, indicating higher stability in terms of diagnostic accuracy.

[0138] Table 3 Experimental results of six tasks

[0139]

[0140]

[0141] To demonstrate the positive impact of the proposed CP, CAF, and RMDS on the performance of the fault diagnosis model, the present invention will conduct ablation experiments taking T1 as an example.

[0142] Five different model configurations were applied in Task T1, including ProtoNet, CGRML model without CP, CGRML model without CAF, CGRML model based on Euclidean Distance-driving Strategy (EDDS), and the proposed CGRML model. Specifically, the "CGRML model without CP" configuration means generating class prototypes by averaging the support samples, that is, directly using the support samples as class prototypes in one-shot tasks instead of using the learnable CP.

[0143] Table 4 shows the average accuracy and SD obtained by these five model configurations. Compared with the classic ProtoNet, all the proposed improvements significantly improved the diagnostic accuracy. When CP was missing, the average accuracy of the proposed method decreased by 8.81%, indicating that CP is a more optimal classification prototype as a whole in cross-domain one-shot tasks and is more effective than directly using the support samples as class prototypes, enhancing the model's ability to learn and retain high-quality general meta-knowledge from the task distribution. By guiding the learning direction of the inner loop to adapt to the current meta-task and retain meta-knowledge (i.e., the cognition of the overall meta-task distribution), CP effectively improved the diagnostic accuracy of the CGRM model. When CAF was missing, the average accuracy decreased by 5.44%, indicating that CAF alleviated the overfitting of the model to the source domain, enabling it to diagnose target domain data more efficiently and accurately. When the proposed model was based on EDDS, the average accuracy decreased by 12.42%, indicating that RMDS better optimized the model parameters by imposing Riemannian geometric constraints on the model and more accurately measured the similarity between samples by capturing the complex structure of the data, thus enhancing the diagnostic ability of the CGRML model in challenging scenarios. In addition, the SD of CGRML was significantly smaller than that of the CGRML model based on EDDS, indicating that the model constructed based on RMDS has higher stability.

[0144] Table 4 Results of ablation experiments

[0145] Model configuration Result (%) ProtoNet 40.32±4.06 CGRML model without CP 84.45±3.51 CGRML model without CAF 87.82±3.47 CGRML model based on EDDS 80.84±7.86 CGRML model 93.26±2.47

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold, characterized in that: The following steps are involved: Step 1: Collect bearing fault signal samples through a rolling bearing test bench; Step 2: Divide the collected samples into source domain and target domain, and divide the support set and query set in the source domain and target domain respectively to generate the meta-tasks required for training and testing phases; Step 3: Construct a feature extraction module and initialize the cognitive prototype CP and cognitive adaptive factor CAF; Step 4: In the inner loop, the adaptation loss is calculated based on the support set and the network parameters of the feature extraction module are updated; Step 5: In the outer loop, the geodesic distance is used to evaluate the similarity between the query sample and each cognitive prototype, the health status of the query sample is diagnosed, the total loss of the outer loop is calculated, and all model parameters are updated through back propagation; Step 6: Use the trained CGRML model to quickly adapt to the meta-task of the target domain and diagnose the health status of the target domain samples.

2. The cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold according to claim 1, wherein: In step 3, the specific method of initializing the cognitive prototype CP and the cognitive adaptive factor CAF is: The cognitive prototype CP of each category is randomly sampled from a standard normal distribution and is expressed as: Among them, is the set of cognitive prototypes for all categories, and P n is the prototype of the nth category, ρ i is the learnable parameter, and d is the dimension of the embedding space; The cognitive adaptive factor CAF is initialized as: Among them, K is the number of supporting samples shot for each category, and the ReLU activation function is used to ensure that τ is positive and greater than the minimum constant ε.

3. The cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold according to claim 1, characterized in that: In step 4, the calculation formula of the adaptation loss is: Among them, is the geodesic distance metric on the Grassmann manifold, is the output of the feature extraction module, τ′τ′ is the CAF value after ReLU activation, is the support set sample.

4. The cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold according to claim 3, characterized in that: The geodesic distance metric The calculation process includes: Normalize the cognitive prototype P n Perform normalization processing: Map the normalized v n to the basis of one-dimensional subspaces in the Grassmann manifold space G(1, d); The geodesic distance between sample features and cognitive prototypes is calculated by projection mapping: where Φ is the projection mapping function, is the manifold embedding function.

5. The cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold according to claim 1, wherein: In the step 5, the total outer loop loss is composed of the classification loss L c and the separability loss L s weighted, and the calculation formula is: L t = L c + λL s Among them, λ is the weight hyperparameter; Classification loss L c Calculated by the cross-entropy loss function: Among them, is the true label, and p(y=n) is the probability that the sample belongs to class n; Divisibility loss L s Calculated by constraining the distance between cognitive prototypes of different categories:

6. The cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold according to claim 1, characterized in that: In step 5, the geodesic distance is used to calculate the similarity between the query sample and each cognitive prototype, and generate the classification probability: Among them, is the feature extraction module after inner loop update, and x n,jQ is the query set sample.

7. The cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold according to claim 1, characterized in that: The meta-task is generated as follows: Each meta-task T i follows the "N-way K-shot" paradigm, where N is the number of classes and K is the number of support samples for each class; In the training phase, meta-tasks are randomly sampled from the source domain, and in the testing phase, meta-tasks are sampled from the target domain.

8. The cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold according to claim 1, characterized in that: The feature extraction module is implemented using a multi-scale convolutional neural network (CNN) and is used to extract high-dimensional features from input signals.

9. The cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold according to claim 1, characterized in that: In step 4, the updating process of the inner loop is: Among them, θ is the parameter of the feature extraction module, θ′ represents the updated parameter, α1 is the inner loop learning rate, and L a is the adaptation loss.

10. The cross-bearing single-sample intelligent diagnosis method based on cognitive guidance and Riemannian manifold according to claim 1, characterized in that: In step 5, the update process of the outer loop is: Among them, α2, α3 and α4 are the learning rates of feature extraction parameters, cognitive prototype parameters and CAF in the outer loop respectively.

Images