A method and system for diagnosing faults in a rotating machine bearing

Abstract

The application discloses a kind of rotating machinery bearing fault diagnosis method and system, belong to fault diagnosis technical field.Its method includes: obtaining the vibration signal of rotating machinery bearing to be diagnosed;The vibration signal to be diagnosed is input into the bearing fault diagnosis model constructed in advance, and the fault diagnosis result of rotating machinery bearing is output;The bearing fault diagnosis model includes the feature extraction network, GMM feature enhancement module, projection distillation module and dynamic scalable classifier connected in turn.The present application can effectively alleviate catastrophic forgetting, realize the continuous diagnosis of new unexpected failure, in the initial task, the model learns to identify different fault types using the pre-collected fault bearing dataset, providing a basis for subsequent incremental tasks, and the incremental task is composed of several stages, in each stage, the model generates historical class pseudo-features using a Gaussian mixture model to enhance the playback of old knowledge, and uses a projection distillation module to align the feature representations of new and old models.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a fault diagnosis method and system for rotating mechanical bearings. Background Technology

[0002] With the rapid development of industrial intelligence, monitoring the health of equipment operating under harsh conditions has become crucial. This equipment typically needs to withstand environments with high temperatures, high pressures, and strong impacts. Bearings are key components of rotating machinery, widely used in transportation, power generation, aerospace, and many other industries. Their stable operation and reliability directly affect the safety performance and operating costs of the entire system. In harsh environments, bearings operate under heavy pressure for extended periods. Even slight performance degradation can gradually develop into equipment failure. Therefore, developing efficient and reliable bearing fault diagnosis technology is particularly important. This technology helps improve the overall reliability of equipment, providing a fundamental guarantee for safe production.

[0003] Today, rotating machinery is constantly evolving towards greater complexity, larger scale, higher speed, greater intelligence, and greater automation. Diagnosing bearing faults first requires capturing vibration signals from the operating machinery for analysis, then extracting useful bearing information from these signals, and finally identifying the fault using appropriate fault diagnosis methods. Currently widely used, relatively traditional fault diagnosis methods such as Fourier transform, short-time Fourier transform, empirical mode decomposition, and wavelet transform all rely to varying degrees on prior knowledge, and therefore are far from meeting the needs of today's complex systems.

[0004] In recent years, with the continuous maturation of machine learning algorithms, AI-based diagnostic technologies have been widely applied. For example, deep learning-based fault diagnosis technology has achieved considerable success. The deep learning-based fault diagnosis process generally includes three steps: signal acquisition, feature extraction, and prediction / identification. Furthermore, deep learning-based fault diagnosis methods can adaptively extract features from fault data. However, building a deep learning model requires a large amount of fault data. But rotating machinery is mostly in a healthy operating state; if a fault occurs, the equipment stops operating, resulting in a limited amount of available fault data. Therefore, the scarcity of fault samples limits the application of deep learning-based fault diagnosis technology. In addition, in actual diagnostic scenarios, bearings operate in complex environments, inevitably leading to the emergence of new fault types and an imbalanced fault data. The model needs continuous updates to learn new fault types, but most current deep learning-based diagnostic models forget previously learned knowledge when learning new fault types, exhibiting the so-called catastrophic forgetting problem. This results in low detection accuracy for deep learning-based diagnostic models when dealing with different types of faulty bearings. Therefore, how to alleviate catastrophic forgetting is an urgent problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a fault diagnosis method and system for rotating mechanical bearings, which can solve the catastrophic forgetting problem caused by data imbalance and feature distribution drift when the existing deep learning-based diagnostic model continuously learns new fault types.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] On one hand, the present invention provides a fault diagnosis method for rotating machinery bearings, comprising:

[0008] Acquire the vibration signal to be diagnosed in a rotating mechanical bearing;

[0009] The vibration signal to be diagnosed is input into a pre-built bearing fault diagnosis model, and the fault diagnosis result of the rotating machinery bearing is output. The bearing fault diagnosis model includes a feature extraction network, a GMM feature enhancement module, a projective distillation module, and a dynamic scalable classifier connected in sequence.

[0010] The training of the bearing fault diagnosis model includes:

[0011] Obtain historical fault datasets and real-time newly added fault datasets for rotating machinery bearings;

[0012] In the initial task phase, a feature extraction network and a dynamic scalable classifier are trained using a historical fault dataset to obtain the initial feature extraction network and the initial dynamic scalable classifier.

[0013] In the incremental task phase, an initial feature extraction network is trained using newly added fault datasets in real time, resulting in a trained feature extraction network. Historical fault datasets are input into the GMM feature enhancement module to construct Gaussian mixture models for each category of historical fault datasets, generating Gaussian distributions for each category. Controllable perturbation sampling is applied to these Gaussian distributions to generate pseudo-feature data, resulting in a trained GMM feature enhancement module. The historical fault datasets, newly added fault datasets in real time, and pseudo-feature data are input into the projection distillation module for multi-level data projection alignment, generating multi-level features, resulting in a trained projection distillation module. An initial dynamic scalable classifier is trained using these multi-level features, resulting in a trained dynamic scalable classifier.

[0014] Optionally, the group numbers of the Gaussian mixture model for each category of historical fault dataset are represented as follows:

[0015] ;

[0016] in, The group number of the Gaussian mixture model for the historical fault dataset of category c; This represents the number of available samples in the historical fault dataset for category c; , These represent taking the maximum and minimum values, respectively. This indicates rounding to the nearest integer.

[0017] Optionally, the Gaussian distribution of the historical fault datasets for each category is represented as follows:

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] in, This represents the mean of the Gaussian components of the historical fault dataset of class c at time k. Gaussian component variance Gaussian distribution under the following conditions; Represents a diagonal matrix; This represents the mean of the Gaussian components in the historical fault dataset of category c; This represents the Gaussian component variance of the historical fault dataset for category c; Let k represent the perturbation vector at time k; It represents the Hadamardi (or Hadama) stack; The group number of the Gaussian mixture model for the historical fault dataset of category c; Indicates the direction of symbol extraction; This represents the variance adjustment coefficient for a specific component at time k.

[0023] Optionally, the multi-level features are represented as follows:

[0024] ;

[0025] in, Indicates input features Multi-level features obtained by projection alignment; , These represent the weight matrix of the extended layer and the weight matrix of the compressed layer, respectively. Indicates the activation function; Indicates batch normalization; The input features include historical fault datasets, real-time newly added fault datasets, and pseudo-feature data. , Indicates the bias term; This indicates the matrix transpose.

[0026] Optionally, the training of the bearing fault diagnosis model further includes:

[0027] A dynamic weight adjustment mechanism is adopted to dynamically adjust the weights of classification loss and total projective distillation loss to obtain the total loss function for the incremental task stage.

[0028] The bearing fault diagnosis model is trained by using the total loss function in the incremental task phase, resulting in a well-trained bearing fault diagnosis model.

[0029] Optionally, the classification loss includes historical fault classification loss, newly added fault classification loss, and pseudo-feature classification loss;

[0030] The historical fault classification loss is represented as follows:

[0031] ;

[0032] The newly added fault classification loss is represented as follows:

[0033] ;

[0034] The pseudo-feature classification loss is expressed as:

[0035] ;

[0036] in, , , These represent the historical fault classification loss, the newly added fault classification loss, and the pseudo-feature classification loss, respectively. Represents the mathematical expectation; This represents the experience replay buffer; This indicates that new fault datasets are added in real time. This represents a pseudo-feature dataset; Represents the cross-entropy loss function; This represents a dynamically scalable classifier; This represents a feature extraction network; These represent historical fault data and the corresponding labels for historical fault data, respectively. These represent the labels corresponding to newly added fault data in real time and the data itself. These represent pseudo-feature data and the labels corresponding to the pseudo-feature data, respectively.

[0037] Optionally, the total projected distillation loss is expressed as:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] in, This represents the total projected distillation loss; This indicates the projected distillation loss due to historical faults. This indicates the newly added projected distillation loss due to a fault; This represents the loss from pseudo-feature projection distillation. Indicates the projection distillation module; This represents a feature extraction network; Indicates temperature parameter; These represent the feature extraction networks corresponding to historical fault data, real-time newly added fault data, and pseudo-feature data, respectively. Represents the L2 norm; This represents the mean squared error loss; These represent historical fault data, real-time newly added fault data, and pseudo-feature data, respectively.

[0043] Optionally, the weights of the classification loss and the total projective distillation loss are expressed as follows:

[0044] ;

[0045] ;

[0046] in, The weights representing the classification loss; This indicates the total number of fault categories that have been learned; Represents the sigmoid function; The weight representing the total projected distillation loss; The initial weights represent the total projected distillation loss; Indicates the number of the current incremental task; This indicates the total incremental task.

[0047] Optionally, the total loss function for the incremental task phase is expressed as:

[0048] ;

[0049] in, This represents the total loss function for the incremental task phase. , , These represent the historical fault classification loss, the newly added fault classification loss, and the pseudo-feature classification loss, respectively. This represents the total projected distillation loss; The weights representing the classification loss; This represents the weight of the total projected distillation loss.

[0050] On the other hand, the present invention provides a fault diagnosis system for rotating mechanical bearings, comprising:

[0051] The data acquisition module is used to acquire the vibration signal to be diagnosed from the rotating machinery bearing.

[0052] The fault diagnosis module is used to: input the vibration signal to be diagnosed into a pre-built bearing fault diagnosis model and output the fault diagnosis result of the rotating machinery bearing; the bearing fault diagnosis model includes a feature extraction network, a GMM feature enhancement module, a projective distillation module and a dynamic scalable classifier connected in sequence;

[0053] The model training module is used for: wherein the training of the bearing fault diagnosis model includes:

[0054] Obtain historical fault datasets and real-time newly added fault datasets for rotating machinery bearings;

[0055] In the initial task phase, a feature extraction network and a dynamic scalable classifier are trained using a historical fault dataset to obtain the initial feature extraction network and the initial dynamic scalable classifier.

[0056] In the incremental task phase, an initial feature extraction network is trained using newly added fault datasets in real time, resulting in a trained feature extraction network. Historical fault datasets are input into the GMM feature enhancement module to construct Gaussian mixture models for each category of historical fault datasets, generating Gaussian distributions for each category. Controllable perturbation sampling is applied to these Gaussian distributions to generate pseudo-feature data, resulting in a trained GMM feature enhancement module. The historical fault datasets, newly added fault datasets in real time, and pseudo-feature data are input into the projection distillation module for multi-level data projection alignment, generating multi-level features, resulting in a trained projection distillation module. An initial dynamic scalable classifier is trained using these multi-level features, resulting in a trained dynamic scalable classifier.

[0057] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0058] This invention effectively mitigates catastrophic forgetting and enables continuous diagnosis of newly occurring unexpected faults. The fault training task is divided into an initial task and incremental tasks. In the initial task, the model uses a pre-collected dataset of faulty bearing vibration signals to learn and identify different fault types, obtaining an initial bearing fault diagnosis model that provides the foundation for subsequent incremental tasks. The incremental tasks consist of several stages. In each stage, the model enhances the playback of old knowledge by generating historical category pseudo-features using a Gaussian mixture model and aligns the feature representations of the new and old models using a projective distillation module. The training process employs multi-datastream joint optimization, updating the model together with newly added fault samples, historical playback samples, and generated pseudo-features. This more effectively mitigates the problem of catastrophic forgetting, improves the overall diagnostic accuracy and stability of the model in continuous learning scenarios, and better meets the actual needs of industrial applications. Attached Figure Description

[0059] Figure 1 The diagram shown is a flowchart of one embodiment of the fault diagnosis method for rotating mechanical bearings of the present invention.

[0060] Figure 2 The diagram shown is a schematic diagram of the training process of the bearing fault diagnosis model of the present invention in one embodiment.

[0061] Figure 3 The diagram shown is a schematic diagram of the confusion matrix obtained after completing the incremental task in one embodiment of the present invention;

[0062] Figure 4 The diagram shown is a feature visualization obtained after completing the incremental task in one embodiment of the present invention. Detailed Implementation

[0063] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0064] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0065] Example 1

[0066] To address the catastrophic forgetting issues inherent in imbalanced datasets, which result in low detection accuracy of current deep learning-based diagnostic models when handling faulty bearing types, such as... Figure 1 As shown in the figure, this embodiment introduces a fault diagnosis method for rotating machinery bearings, including the following steps:

[0067] Step 1: Construct a bearing fault diagnosis model, specifically as follows:

[0068] The bearing fault diagnosis model comprises a feature extraction network, a Gaussian Mixture Model (GMM) feature enhancement module, a projective distillation module, and a dynamic scalable classifier, connected in sequence. In this embodiment, an 18-layer weighted deep residual network (ResNet-18) is used as the backbone of the feature extraction network.

[0069] Step 2: Train the bearing fault diagnosis model, specifically:

[0070] During model training, the GMM feature enhancement module and projective distillation module are used to characterize the distribution of each fault sample in the feature space, addressing the imbalance and lack of diversity issues in data replay. This aligns old and new features in the projective subspace, and the dynamically scalable classifier also expands with learning. Figure 2 The following is the training process for the bearing fault diagnosis model:

[0071] Collect and analyze fault data, and acquire historical fault dataset D0 and real-time newly added fault dataset for rotating machinery bearings. The historical fault dataset D0 for rotating machinery bearings is derived from typical examples. Extracted from, real-time updated fault dataset From the classic examples The historical fault dataset D0 extracted from it serves as a buffer for subsequent experience replay. .

[0072] The fault tasks at different stages are divided into initial tasks and incremental tasks, and the number of fault categories in the initial task stage and the incremental task stage is determined, namely the number of fault categories in the historical fault dataset and the number of fault categories in the real-time newly added fault dataset.

[0073] In the initial task phase, a feature extraction network is trained based on the fault samples and corresponding fault categories to obtain the initial feature extraction network. Simultaneously, high-level, abstract feature representations are extracted from the input historical fault dataset to obtain the historical fault features. An initial dynamic scalable classifier is obtained by training a dynamic scalable classifier using historical fault features. .

[0074] The initial task refers to the initial stage of training the bearing fault diagnosis model. It involves training the model using a pre-collected dataset of bearing vibration signals containing various fault types. This dataset typically contains a rich variety of fault vibration signals and their corresponding category labels. During this stage, the model learns to identify different fault types and constructs an initial bearing fault diagnosis model. This model provides the basic feature extraction capabilities and classification architecture for subsequent incremental task stages. After completing the initial task, incremental task learning begins. In these incremental tasks, the model continuously learns new fault types.

[0075] Based on the initial feature extraction network, in the incremental task phase, the knowledge of the new and old models is integrated through Gaussian mixture feature playback and projective distillation mechanisms to obtain a trained bearing fault diagnosis model. Here, the new model refers to the model version trained in the current incremental task phase, while the old model refers to the model version trained in the previous phase (including the initial and incremental task phases). Incremental tasks refer to the sequence of tasks that require continuous learning and adaptation of the model during system operation when new fault types or new fault samples appear; these tasks are performed sequentially in stages.

[0076] In each incremental task phase:

[0077] First, by adding fault dataset D in real time t The initial feature extraction network is trained to obtain a well-trained feature extraction network. Simultaneously, high-level, abstract feature representations are extracted from the real-time newly added fault dataset to obtain the new fault features. Taking the first incremental task stage as an example, the initial feature extraction network is trained using the real-time newly added fault dataset D1, which is derived from a typical example. The extracted features are used to obtain a trained feature extraction network. Simultaneously, high-level, abstract feature representations are extracted from the real-time newly added fault dataset to obtain the newly added fault features. .

[0078] Secondly, from the first The incremental task to the 1st When performing incremental tasks, the feature distribution of historical tasks is modeled and enhanced using a Gaussian Mixture Model (GMM). A GMM is constructed for each historical fault category, and the historical fault dataset is input into the GMM feature enhancement module. In this embodiment, the GMM feature enhancement module consists of a two-layer bottleneck multilayer perceptron. The mean of the Gaussian components of the features extracted by the old model for each fault category is used. Generated using class prototypes and perturbation vectors, represented as:

[0079] ;

[0080] ;

[0081] Gaussian component variance Represented as:

[0082] ;

[0083] in, Represents a diagonal matrix; This represents the mean of the Gaussian components in the historical fault dataset of category c; This represents the Gaussian component variance of the historical fault dataset for category c; Let k represent the perturbation vector at time k; It represents the Hadamardi (or Hadama) stack; The group number of the Gaussian mixture model for the historical fault dataset of category c; Indicates the direction of symbol extraction; This represents the variance adjustment coefficient for a specific component at time k;

[0084] Mean Gaussian components of features for each fault category extracted using the old model Gaussian component variance Modeling the feature distribution of historical fault datasets, constructing Gaussian mixture models for each category of historical fault datasets, and determining the component numbers of the Gaussian mixture models for category c historical fault datasets. Represented as:

[0085] ;

[0086] in, This represents the number of available samples in the historical fault dataset for category c; , These represent taking the maximum and minimum values, respectively. This indicates rounding to the nearest integer.

[0087] Gaussian distribution for generating historical fault datasets for each category , This represents the mean of the Gaussian components of the historical fault dataset of class c at time k. Gaussian component variance The Gaussian distribution of the historical fault dataset is used as a basis; the Gaussian distribution of each category is sampled with controlled perturbation to generate pseudo-feature data that is statistically representative and covers a wider feature space, so as to enhance the playback of historical knowledge and obtain a trained GMM feature enhancement module.

[0088] Then, the model introduces a projective distillation module to align the feature representations of the old and new models in a specially constructed subspace to stably transfer existing knowledge. The training process employs multi-datastream joint optimization to extract high-level, abstract feature representations from historical fault datasets. High-level, abstract feature representations extracted from real-time newly added fault datasets Pseudo-feature data The input projection distillation module performs multi-level data projection alignment, generates multi-level features, and jointly updates the model parameters to obtain a trained projection distillation module. The multi-level features are represented as follows:

[0089] ;

[0090] in, Indicates input features Multi-level features obtained by projection alignment; , These represent the weight matrix of the extended layer and the weight matrix of the compressed layer, respectively. Indicates the activation function; Indicates batch normalization; The input features include historical fault datasets, real-time newly added fault datasets, and pseudo-feature data. ; , Indicates the bias term; This indicates the matrix transpose.

[0091] Finally, an initial dynamically scalable classifier is trained using multi-level features to obtain a well-trained dynamically scalable classifier. The classifier is a fully connected layer network whose output dimension equals the total number of fault categories learned so far. The classifier set... It consists of a specific classifier for each task, where, These are incremental tasks. The classifier.

[0092] Furthermore, by adaptively balancing the classification loss and the total projective distillation loss through a dynamic weight adjustment mechanism, the total loss function for the incremental task stage is obtained. This not only effectively learns new fault types but also significantly mitigates catastrophic forgetting, ensuring that the model maintains high diagnostic accuracy and generalization ability throughout the entire task sequence.

[0093] The bearing fault diagnosis model is trained using the total loss function in the incremental task phase, resulting in a well-trained model. A multi-objective loss function is designed, and projective distillation loss is used to integrate knowledge from the old and new models. The loss function combines historical fault classification loss, new fault classification loss, pseudo-feature classification loss, and total projective distillation loss. In the initial task phase, only cross-entropy loss is used. The projective distillation module uses a dedicated projection head network to map the features extracted from the old and new models to a low-dimensional subspace and align their feature representations within this space, effectively transferring and preserving important knowledge from the old model. Furthermore, a dynamic weight adjustment mechanism is introduced to automatically balance the optimization objectives of learning new knowledge and preserving old knowledge, ultimately resulting in an integrated bearing fault diagnosis model with continuous learning capabilities.

[0094] Here, "new model" and "old model" refer to model versions corresponding to different training stages. The new model is the updated version obtained through joint training in the current incremental learning stage. This model not only inherits the feature representation capabilities for learned fault types from the old model through projective distillation, but also enhances the recognition and classification performance for new fault types through multi-data-stream optimization. The old model refers to the model version that was trained before entering the current incremental stage; its knowledge representation is used as a reference for distillation alignment, providing a stable starting point for feature learning for the new model. Knowledge transfer between the new and old models is mainly achieved through feature alignment in the projective subspace and dynamically weighted loss optimization, thereby effectively maintaining the diagnostic capability for historical tasks while integrating new knowledge.

[0095] The old tasks refer to all fault diagnosis categories that the model has learned and mastered up to the current incremental stage; the new tasks refer to new fault types or new sample sets introduced for the first time in the current stage, requiring the model to learn and adapt. Under the continuous learning framework, the introduction of new tasks requires the model to simultaneously learn new feature patterns and consolidate old knowledge structures under limited historical sample conditions. To this end, this method enhances the feature diversity of old categories through Gaussian mixture feature replay and constrains the drift of the feature space using a projective distillation module, enabling the model to balance the learning objectives of new and old tasks during joint training. The new tasks not only represent the diagnostic scope that the model needs to expand but also reflect the model's key ability to continuously self-evolve in dynamic industrial environments. Through systematic feature replay and knowledge alignment, the model can gradually accumulate diagnostic knowledge, improving its overall generalization performance, stability, and plasticity balance in complex fault scenarios.

[0096] Historical fault classification losses are represented as follows:

[0097] ;

[0098] The loss from the newly added fault classification is represented as follows:

[0099] ;

[0100] The pseudo-feature classification loss is expressed as:

[0101] ;

[0102] in, , , These represent the historical fault classification loss, the newly added fault classification loss, and the pseudo-feature classification loss, respectively. The mathematical expectation refers to the average value of the loss function calculated on a specific data distribution (such as an experience replay buffer, a real-time newly added fault dataset, or a pseudo-feature dataset). This represents the experience replay buffer; This indicates that new fault datasets are added in real time. This represents a pseudo-feature dataset; Represents the cross-entropy loss function; This represents a dynamically scalable classifier; This represents a feature extraction network; These represent historical fault data and the corresponding labels for historical fault data, respectively. These represent the labels corresponding to newly added fault data in real time and the data itself. These represent pseudo-feature data and the labels corresponding to the pseudo-feature data, respectively.

[0103] The total projected distillation loss is expressed as:

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] in, This represents the total projected distillation loss; This indicates the projected distillation loss due to historical faults. This indicates the newly added projected distillation loss due to a fault; This represents the loss from pseudo-feature projection distillation. Indicates the projection distillation module; This represents a feature extraction network; Indicates temperature parameter; These represent the feature extraction networks corresponding to historical fault data, real-time newly added fault data, and pseudo-feature data, respectively. Represents the L2 norm; This represents the mean squared error loss; These represent historical fault data, real-time newly added fault data, and pseudo-feature data, respectively.

[0109] The weights of the classification loss and the total projective distillation loss are expressed as follows:

[0110] ;

[0111] ;

[0112] in, The weights representing the classification loss; This indicates the total number of fault categories that have been learned; Represents the sigmoid function; The weight representing the total projected distillation loss; The initial weights represent the total projected distillation loss; Indicates the number of the current incremental task; This indicates the total incremental task.

[0113] The total loss function for the incremental task phase is expressed as:

[0114] ;

[0115] in, This represents the total loss function for the incremental task phase. , , These represent the historical fault classification loss, the newly added fault classification loss, and the pseudo-feature classification loss, respectively. This represents the total projected distillation loss; The weights representing the classification loss; This represents the weight of the total projected distillation loss.

[0116] Therefore, the parameters of the trained bearing fault diagnosis model include those of the feature extraction network, the GMM feature enhancement module (feature prototype, variance and mixture weights), the projective distillation module, and the dynamic scalable classifier. The model is always trained and updated based on the same backbone network. In the incremental stage, auxiliary modules, namely the GMM feature enhancement module and the projective distillation module, are added to enhance the model's functionality.

[0117] Step 3: Obtain the vibration signal to be diagnosed from the rotating mechanical bearing.

[0118] Step 4: Obtain the fault diagnosis results of the rotating machinery bearing, specifically:

[0119] The vibration signal to be diagnosed is input into a pre-built bearing fault diagnosis model, and the fault diagnosis results of the rotating machinery bearing are output.

[0120] In actual use of the model for fault diagnosis, the GMM feature enhancement module and projective distillation module are not used. The collected vibration signal to be diagnosed is first preprocessed, then features are extracted through a feature extraction network, and finally a dynamic scalable classifier is used for classification decision to obtain the fault diagnosis result.

[0121] This embodiment can solve the catastrophic forgetting problem caused by data imbalance and feature distribution drift when existing deep learning-based diagnostic models continuously learn new fault types.

[0122] Example 2

[0123] Based on Example 1, this example presents an experimental example of a fault diagnosis method for rotating machinery bearings:

[0124] Vibration signals of bearings with various fault types were collected for the experiment, and a fault diagnosis dataset was constructed. To obtain faulty bearings, wire cutting technology was used to create cracks of different sizes on the inner ring, outer ring, and rollers of the bearings to simulate localized bearing faults in actual production.

[0125] The fault tasks at different stages are divided into initial tasks and incremental tasks, and the number of fault categories for initial tasks and incremental tasks is determined.

[0126] In the experimental parameter settings, the motor speed was set to 800 r / min and the sampling frequency to 10 kHz. The constructed fault dataset contains 16 health states, as shown in Table 1. NO represents the normal state, IF represents the inner ring fault state, OF represents the outer ring fault state, and BF represents the roller fault state. Bearings are divided into two health states: normal and faulty. Faulty bearings include inner ring faults, roller faults, and outer ring faults.

[0127] Table 1 Health Status and Fault Task Setting of Rotating Machinery Bearings

[0128]

[0129] Vibration signals from bearing failures were collected using accelerometers, and a dataset containing various failure types was constructed. The model training task was divided into an initial task phase and an incremental task phase.

[0130] As shown in Table 1, the initial task M0 has four fault types (labels 0-3). M1, M2, and M3 are three incremental tasks, each containing four new fault types (labels 4-7 correspond to M1, labels 8-11 correspond to M2, and labels 12-15 correspond to M3). In this embodiment, the incremental tasks retain fault types from all previous tasks, but the number of retained samples is relatively small, thus creating an imbalanced dataset. Taking M2 as an example, M2 corresponds to labels 8-11, each containing 100 samples. The four types in the initial task M0 and the four types in M1 each contain only 5 to 10 samples.

[0131] In real-world scenarios, the number of bearings in healthy condition far exceeds the number of faulty bearings. These healthy bearings are common, therefore more varied and easier to learn from. Thus, the initial task M0 learns about normal bearings and three types of faulty bearings. To simulate the incremental fault types that occur due to unforeseen events in real-world scenarios, each subsequent incremental task learns fault types including inner ring faults, outer ring faults, roller faults, and combined faults. Each fault type in the initial task M0 has 100 training samples and 100 test samples, and the incremental tasks also have 100 training samples and 100 test samples for each fault type.

[0132] After the incremental task training is completed, i.e., after M3 in Table 1 has been trained, the obtained diagnostic model needs to be tested. The test data includes all learned fault types to verify whether the obtained diagnostic model has the ability to mitigate catastrophic amnesia. The test results are presented in the form of a confusion matrix. Figure 3 The confusion matrix shown demonstrates that this method has a good resistance to forgetting. Figure 3 The confusion matrix intuitively quantifies the forgetting resistance of this method. As clearly seen in the figure, the off-diagonal regions of the matrix represent the parts where old tasks were incorrectly predicted as new tasks, and their values ​​are generally low. Meanwhile, the diagonal blocks representing the categories of each task itself have high values. This directly indicates that the model can still accurately distinguish past categories after sequentially learning multiple stages.

[0133] Figure 4 The effectiveness of this method in classification was further demonstrated through feature visualization. The results show that the trained model obtained by this method can effectively cluster features. Figure 4By using t-Distributed Stochastic Neighbor Embedding (t-SNE) for dimensionality reduction, high-dimensional features are projected onto a two-dimensional plane, providing an intuitive view of the model's internal representation. As shown in the figure, sample points belonging to different stages (labeled 0, 1, 2, 3) each form independent and tightly clustered regions. These clusters are clearly defined with minimal mutual interference. This visualization vividly demonstrates that our method enables the model to learn a feature distribution that exhibits high intra-class aggregation and inter-class separation, thereby ensuring classification robustness during continuous learning.

[0134] This demonstrates that it can effectively alleviate catastrophic amnesia and enable continuous diagnosis of new and unexpected bearing failures.

[0135] In summary, compared with traditional deep learning methods, this invention can alleviate the problem of catastrophic forgetting and is more in line with the actual scenarios of industrial applications.

[0136] Example 3

[0137] This embodiment introduces a fault diagnosis system for rotating machinery bearings, including:

[0138] The data acquisition module is used to acquire the vibration signal to be diagnosed from the rotating machinery bearing.

[0139] The fault diagnosis module is used to: input the vibration signal to be diagnosed into a pre-built bearing fault diagnosis model and output the fault diagnosis result of the rotating machinery bearing; the bearing fault diagnosis model includes a feature extraction network, a GMM feature enhancement module, a projective distillation module and a dynamic scalable classifier connected in sequence;

[0140] The model training module is used for: wherein the training of the bearing fault diagnosis model includes:

[0141] Obtain historical fault datasets and real-time newly added fault datasets for rotating machinery bearings;

[0142] In the initial task phase, a feature extraction network is trained using a historical fault dataset to obtain the initial feature extraction network;

[0143] In the incremental task phase, an initial feature extraction network is trained using newly added fault datasets in real time, resulting in a trained feature extraction network. Historical fault datasets are input into the GMM feature enhancement module to construct Gaussian mixture models for each category of historical fault datasets, generating Gaussian distributions for each category. Controllable perturbation sampling is applied to these Gaussian distributions to generate pseudo-feature data, resulting in a trained GMM feature enhancement module. The historical fault datasets, newly added fault datasets in real time, and pseudo-feature data are input into the projection distillation module for multi-level data projection alignment, generating multi-level features, resulting in a trained projection distillation module. Finally, a dynamically scalable classifier is trained using these multi-level features, resulting in a trained dynamically scalable classifier.

[0144] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0145] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0146] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0147] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0148] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0149] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for diagnosing faults in rotating machinery bearings, characterized in that, include: Acquire the vibration signal to be diagnosed in a rotating mechanical bearing; The vibration signal to be diagnosed is input into a pre-built bearing fault diagnosis model, and the fault diagnosis result of the rotating machinery bearing is output. The bearing fault diagnosis model includes a feature extraction network, a GMM feature enhancement module, a projective distillation module, and a dynamic scalable classifier connected in sequence. The training of the bearing fault diagnosis model includes: Obtain historical fault datasets and real-time newly added fault datasets for rotating machinery bearings; In the initial task phase, a feature extraction network and a dynamic scalable classifier are trained using a historical fault dataset to obtain the initial feature extraction network and the initial dynamic scalable classifier. In the incremental task phase, an initial feature extraction network is trained using newly added fault datasets in real time, resulting in a trained feature extraction network. Historical fault datasets are input into the GMM feature enhancement module to construct Gaussian mixture models for each category of historical fault datasets, generating Gaussian distributions for each category. Controllable perturbation sampling is applied to these Gaussian distributions to generate pseudo-feature data, resulting in a trained GMM feature enhancement module. The historical fault datasets, newly added fault datasets in real time, and pseudo-feature data are input into the projection distillation module for multi-level data projection alignment, generating multi-level features, resulting in a trained projection distillation module. An initial dynamic scalable classifier is trained using these multi-level features, resulting in a trained dynamic scalable classifier. The training process of the projection distillation module adopts multi-data stream joint optimization. High-level and abstract features extracted from the historical fault dataset, high-level and abstract features extracted from the fault dataset are added in real time, and pseudo-feature data are input into the projection distillation module to perform multi-level data projection alignment, generate multi-level features, and jointly update the model parameters to obtain the trained projection distillation module.

2. The fault diagnosis method for rotating machinery bearings according to claim 1, characterized in that, The component numbers of the Gaussian mixture model for each category of historical fault dataset are represented as follows: ； in, The group number of the Gaussian mixture model for the historical fault dataset of category c; This represents the number of available samples in the historical fault dataset for category c; , These represent taking the maximum and minimum values, respectively. This indicates rounding to the nearest integer.

3. The fault diagnosis method for rotating mechanical bearings according to claim 1, characterized in that, The Gaussian distribution of the historical fault datasets for each category is represented as follows: ； ； ； ； in, This represents the mean of the Gaussian components of the historical fault dataset of class c at time k. Gaussian component variance Gaussian distribution under the following conditions; Represents a diagonal matrix; This represents the mean of the Gaussian components in the historical fault dataset of category c; This represents the Gaussian component variance of the historical fault dataset for category c; Let k represent the perturbation vector at time k; It represents the Hadamardi (or Hadama) stack; The group number of the Gaussian mixture model for the historical fault dataset of category c; Indicates the direction of symbol extraction; This represents the variance adjustment coefficient for a specific component at time k.

4. The fault diagnosis method for rotating mechanical bearings according to claim 1, characterized in that, The multi-level features are represented as follows: ； in, Indicates input features Multi-level features obtained by projection alignment; , These represent the weight matrix of the extended layer and the weight matrix of the compressed layer, respectively. Indicates the activation function; Indicates batch normalization; The input features include historical fault datasets, real-time newly added fault datasets, and pseudo-feature data. , Indicates the bias term; This indicates the matrix transpose.

5. The fault diagnosis method for rotating machinery bearings according to claim 1, characterized in that, The training of the bearing fault diagnosis model also includes: A dynamic weight adjustment mechanism is adopted to dynamically adjust the weights of classification loss and total projective distillation loss to obtain the total loss function for the incremental task stage. The bearing fault diagnosis model is trained by using the total loss function in the incremental task phase, resulting in a well-trained bearing fault diagnosis model.

6. The fault diagnosis method for rotating machinery bearings according to claim 5, characterized in that, The classification loss includes historical fault classification loss, newly added fault classification loss, and pseudo-feature classification loss; The historical fault classification loss is represented as follows: ； The newly added fault classification loss is represented as follows: ； The pseudo-feature classification loss is expressed as: ； in, , , These represent the historical fault classification loss, the newly added fault classification loss, and the pseudo-feature classification loss, respectively. Represents the mathematical expectation; This represents the experience replay buffer; This indicates that new fault datasets are added in real time. This represents a pseudo-feature dataset; Represents the cross-entropy loss function; This represents a dynamically scalable classifier. This represents a feature extraction network; These represent historical fault data and the corresponding labels for historical fault data, respectively. These represent the labels corresponding to newly added fault data in real time and the data itself. These represent pseudo-feature data and the labels corresponding to the pseudo-feature data, respectively.

7. The fault diagnosis method for rotating machinery bearings according to claim 5, characterized in that, The total projected distillation loss is expressed as: ； ； ； ； in, This represents the total projected distillation loss; This indicates the projected distillation loss due to historical faults. This indicates the newly added projected distillation loss due to a fault; This represents the loss from pseudo-feature projection distillation. Indicates the projection distillation module; This represents a feature extraction network; Indicates temperature parameter; These represent the feature extraction networks corresponding to historical fault data, real-time newly added fault data, and pseudo-feature data, respectively. Represents the L2 norm; This represents the mean squared error loss; These represent historical fault data, real-time newly added fault data, and pseudo-feature data, respectively.

8. The fault diagnosis method for rotating machinery bearings according to claim 5, characterized in that, The weights of the classification loss and the total projective distillation loss are expressed as follows: ； ； in, The weights representing the classification loss; This indicates the total number of fault categories that have been learned; Represents the sigmoid function; The weight representing the total projected distillation loss; The initial weights represent the total projected distillation loss; Indicates the number of the current incremental task; This indicates the total incremental tasks.

9. The fault diagnosis method for rotating machinery bearings according to claim 5, characterized in that, The total loss function for the incremental task phase is expressed as: ； in, This represents the total loss function for the incremental task phase. , , These represent the historical fault classification loss, the newly added fault classification loss, and the pseudo-feature classification loss, respectively. This represents the total projected distillation loss; The weights representing the classification loss; This represents the weight of the total projected distillation loss.

10. A fault diagnosis system for rotating machinery bearings, characterized in that, include: The data acquisition module is used to acquire the vibration signal to be diagnosed from the rotating machinery bearing. The fault diagnosis module is used to: input the vibration signal to be diagnosed into a pre-built bearing fault diagnosis model and output the fault diagnosis result of the rotating machinery bearing; the bearing fault diagnosis model includes a feature extraction network, a GMM feature enhancement module, a projective distillation module and a dynamic scalable classifier connected in sequence; The model training module is used for: wherein the training of the bearing fault diagnosis model includes: Obtain historical fault datasets and real-time newly added fault datasets for rotating machinery bearings; In the initial task phase, a feature extraction network and a dynamic scalable classifier are trained using a historical fault dataset to obtain the initial feature extraction network and the initial dynamic scalable classifier. In the incremental task phase, an initial feature extraction network is trained using newly added fault datasets in real time, resulting in a trained feature extraction network. Historical fault datasets are input into the GMM feature enhancement module to construct Gaussian mixture models for each category of historical fault datasets, generating Gaussian distributions for each category. Controllable perturbation sampling is applied to these Gaussian distributions to generate pseudo-feature data, resulting in a trained GMM feature enhancement module. The historical fault datasets, newly added fault datasets in real time, and pseudo-feature data are input into the projection distillation module for multi-level data projection alignment, generating multi-level features, resulting in a trained projection distillation module. An initial dynamic scalable classifier is trained using these multi-level features, resulting in a trained dynamic scalable classifier. The training process of the projection distillation module adopts multi-data stream joint optimization. High-level and abstract features extracted from the historical fault dataset, high-level and abstract features extracted from the fault dataset are added in real time, and pseudo-feature data are input into the projection distillation module to perform multi-level data projection alignment, generate multi-level features, and jointly update the model parameters to obtain the trained projection distillation module.

Images