CN-121540426-B - Fault diagnosis method and system for rotary mechanical bearing

CN121540426BCN 121540426 BCN121540426 BCN 121540426BCN-121540426-B

Abstract

The invention discloses a fault diagnosis method and system for a rotary mechanical bearing, and belongs to the technical field of fault diagnosis. The method comprises the steps of obtaining a vibration signal to be diagnosed of a rotary mechanical bearing, inputting the vibration signal to be diagnosed into a pre-constructed bearing fault diagnosis model, and outputting a fault diagnosis result of the rotary mechanical bearing, wherein the bearing fault diagnosis model comprises a feature extraction network, a GMM feature enhancement module, a projection distillation module and a dynamic expandable classifier which are sequentially connected. The invention can effectively relieve disastrous forgetting, realize continuous diagnosis of new unexpected faults, in an initial task, a model learns and identifies different fault types by utilizing a pre-collected fault bearing data set, a foundation is provided for a subsequent incremental task, the incremental task is composed of a plurality of stages, in each stage, the model enhances playback of old knowledge through historical category pseudo-features generated by a Gaussian mixture model, and a projection distillation module is utilized to align feature representation of the new model and the old model.

Inventors

SHEN CHANGQING
HE YI
CHEN LIANG
YE XIAOFEN
SHI JUANJUAN
HUANG WEIGUO
ZHU ZHONGKUI

Assignees

苏州大学

Dates

Publication Date: 20260512
Application Date: 20260120

Claims (10)

1. A fault diagnosis method of a rotary machine bearing, comprising: Acquiring a vibration signal to be diagnosed of a rotary mechanical bearing; Inputting the vibration signal to be diagnosed into a pre-constructed bearing fault diagnosis model, and outputting a fault diagnosis result of a rotary mechanical bearing, wherein the bearing fault diagnosis model comprises a feature extraction network, a GMM feature enhancement module, a projection distillation module and a dynamic expandable classifier which are sequentially connected; the training of the bearing fault diagnosis model comprises the following steps: acquiring a historical fault data set and a real-time newly-added fault data set of a rotary mechanical bearing; in an initial task stage, training a feature extraction network and a dynamic expandable classifier through a historical fault data set to obtain an initial feature extraction network and an initial dynamic expandable classifier; In the incremental task stage, an initial feature extraction network is trained through a real-time newly-added fault data set to obtain a trained feature extraction network, a history fault data set is input into a GMM feature enhancement module, a Gaussian mixture model of each type of history fault data set is constructed, gaussian distribution of each type of history fault data set is generated, controllable disturbance sampling is carried out on the Gaussian distribution of each type of history fault data set to generate pseudo feature data, and the trained GMM feature enhancement module is obtained; The training process of the projection distillation module adopts multi-data flow joint optimization, the high-level and abstract features extracted in the historical fault data set are added in real time, the high-level and abstract features extracted in the fault data set are input into the projection distillation module, multi-level data projection alignment is carried out, multi-level features are generated, and model parameters are updated together, so that the trained projection distillation module is obtained.
2. The method of claim 1, wherein the components of the gaussian mixture model of each of the historical fault datasets are expressed as: ; Wherein, the Component numbers of the gaussian mixture model representing the class c historical fault dataset; representing the number of available samples of the class c historical fault dataset; 、 respectively representing taking the maximum value and the minimum value; Representing rounding.
3. The method of claim 1, wherein the gaussian distribution of each of the historical fault datasets is represented as: ; ; ; ; Wherein, the Gaussian component mean representing class c historical fault dataset at time k Variance of Gaussian components The gaussian distribution below; Representing a diagonal matrix; A Gaussian component mean value representing a class c historical fault dataset; Representing gaussian component variance of the class c historical fault dataset; A disturbance vector representing the moment k; representing the Hadamard product; component numbers of the gaussian mixture model representing the class c historical fault dataset; Representing the extracted symbol direction; a variance adjustment coefficient representing a specific component at time k.
4. The method of claim 1, wherein the multi-level feature is expressed as: ; Wherein, the Representing input features Projecting and aligning the obtained multi-level characteristics; 、 respectively representing an expansion layer weight matrix and a compression layer weight matrix; Representing an activation function; Representing batch normalization; Representing input features including a historical fault data set, a real-time newly-added fault data set and pseudo feature data; 、 Representing a bias term; representing the matrix transpose.
5. The method of claim 1, wherein the training of the bearing failure diagnosis model further comprises: dynamically adjusting weights of the classification loss and the total projection distillation loss by adopting a dynamic weight adjustment mechanism to obtain a total loss function of an incremental task stage; And training a bearing fault diagnosis model through the total loss function of the incremental task stage to obtain a trained bearing fault diagnosis model.
6. The method of diagnosing a fault in a rotating machine bearing according to claim 5, wherein the classification loss includes a historical fault classification loss, a newly added fault classification loss, and a pseudo feature classification loss; the historical fault classification loss is expressed as: ; The added fault classification loss is expressed as: ; the pseudo-feature classification loss is expressed as: ; Wherein, the 、、 Respectively representing historical fault classification loss, newly added fault classification loss and pseudo feature classification loss; representing mathematical expectations; representing an experience playback buffer; Representing a real-time newly-added fault dataset; Representing a pseudo-feature data set; representing a cross entropy loss function; representing a dynamic scalable classifier; Representing a feature extraction network; Respectively representing historical fault data and labels corresponding to the historical fault data; Respectively representing the labels corresponding to the real-time newly-added fault data; the pseudo-feature data and the labels corresponding to the pseudo-feature data are respectively represented.
7. The method of claim 5, wherein the total projected distillation loss is expressed as: ; ; ; ; Wherein, the Representing total projected distillation loss; representing historical fault projected distillation losses; representing newly increased fault projected distillation loss; representing pseudo-feature projected distillation losses; representing a projective distillation module; Representing a feature extraction network; representing a temperature parameter; the characteristic extraction network respectively represents the corresponding characteristics of the historical fault data, the real-time newly-added fault data and the pseudo characteristic data; Represents an L2 norm; represents the mean square error loss; respectively representing historical fault data, real-time newly-added fault data and pseudo characteristic data.
8. The fault diagnosis method of a rotary machine bearing according to claim 5, wherein the weights of the classification loss and the total projected distillation loss are expressed as: ; ; Wherein, the A weight representing a classification loss; representing a total number of learned fault categories; representing a sigmoid function; Weights representing total projected distillation loss; initial weights representing total projected distillation loss; A number representing the current incremental task; Representing the total incremental task.
9. The method of claim 5, wherein the total loss function of the incremental mission phase is expressed as: ; Wherein, the Representing the total loss function of the incremental task phase; 、、 Respectively representing historical fault classification loss, newly added fault classification loss and pseudo feature classification loss; Representing total projected distillation loss; A weight representing a classification loss; The weight of the total projected distillation loss is represented.
10. A fault diagnosis system for a rotary machine bearing, comprising: the data acquisition module is used for acquiring vibration signals to be diagnosed of the rotary mechanical bearing; The fault diagnosis module is used for inputting the vibration signal to be diagnosed into a pre-constructed bearing fault diagnosis model and outputting a fault diagnosis result of the rotary mechanical bearing, wherein the bearing fault diagnosis model comprises a feature extraction network, a GMM feature enhancement module, a projection distillation module and a dynamic expandable classifier which are connected in sequence; the model training module is used for training the bearing fault diagnosis model, and comprises the following steps: acquiring a historical fault data set and a real-time newly-added fault data set of a rotary mechanical bearing; in an initial task stage, training a feature extraction network and a dynamic expandable classifier through a historical fault data set to obtain an initial feature extraction network and an initial dynamic expandable classifier; In the incremental task stage, an initial feature extraction network is trained through a real-time newly-added fault data set to obtain a trained feature extraction network, a history fault data set is input into a GMM feature enhancement module, a Gaussian mixture model of each type of history fault data set is constructed, gaussian distribution of each type of history fault data set is generated, controllable disturbance sampling is carried out on the Gaussian distribution of each type of history fault data set to generate pseudo feature data, and the trained GMM feature enhancement module is obtained; The training process of the projection distillation module adopts multi-data flow joint optimization, the high-level and abstract features extracted in the historical fault data set are added in real time, the high-level and abstract features extracted in the fault data set are input into the projection distillation module, multi-level data projection alignment is carried out, multi-level features are generated, and model parameters are updated together, so that the trained projection distillation module is obtained.

Description

Fault diagnosis method and system for rotary mechanical bearing Technical Field The invention relates to the technical field of fault diagnosis, in particular to a fault diagnosis method and system for a rotary mechanical bearing. Background With the rapid development of industrial intelligence, it is becoming very important to monitor the health of operating equipment under severe conditions. These devices are often required to cope with environments of high temperature, high pressure, strong impact, etc. Bearings are critical components of rotating machinery, and are widely used in transportation, power generation, aerospace, and many other industries. Their stable operation and reliability directly affect the safety performance and operation cost of the whole system. In severe environments, bearings operate under heavy pressure for long periods of time. Even a slight decrease in performance may develop into equipment failure. Therefore, it is important to develop a bearing fault diagnosis technology with high efficiency and reliability. The technology is helpful for improving the overall reliability of equipment, and provides a fundamental guarantee for safe production. Today, equipment based on rotary machinery is continuously developed towards complicating, enlarging, accelerating, intelligentizing and automatizing. The diagnosis of a bearing fault requires firstly capturing vibration signals for analysis in the operating mechanical equipment, then extracting useful bearing information in the vibration signals, and identifying the fault by combining a corresponding fault diagnosis method. The prior knowledge of the current widely applied Fourier transformation, short-time Fourier transformation, empirical mode decomposition, wavelet transformation and other fault diagnosis methods is more or less dependent on the prior knowledge, so that the requirements of the complex system at present can not be met. In recent years, as machine learning is becoming more and more algorithmically mature, diagnostic techniques based on artificial intelligence are widely used. Such as fault diagnosis techniques based on deep learning, have achieved much better results. The fault diagnosis process based on deep learning generally comprises three steps of signal acquisition, feature extraction and predictive recognition. And the fault diagnosis method based on deep learning can adaptively extract the characteristics of fault data. A large amount of fault data is required as support in the process of establishing the deep learning model, but the rotating machinery is in a healthy operation state under most conditions, because if the equipment fails, the equipment stops operating, so the fault data which can be collected is quite limited, and therefore, the defect of fault samples limits the application of fault diagnosis technology based on deep learning. In addition, in an actual diagnosis scenario, the bearing is in a complex working environment, so that new fault types inevitably occur, and the fault data is in an unbalanced state, and at this time, the model needs to be continuously updated to learn the new fault types, but when most of the deep learning-based diagnosis models forget old knowledge learned before learning the new fault types, so-called catastrophic forgetting problems occur, which makes the deep learning-based diagnosis models lower in detection accuracy when facing the problem of processing the fault bearing types. Thus, how to alleviate the catastrophic forgetfulness is a problem that is currently urgently needed to be solved. Disclosure of Invention The invention aims to overcome the defects in the prior art, and provides a fault diagnosis method and system for a rotary mechanical bearing, which can solve the problem of disastrous forgetting caused by unbalanced data and characteristic distribution drift when a deep learning-based diagnosis model continuously learns a new fault type in the prior art. In order to achieve the above purpose, the invention is realized by adopting the following technical scheme: in one aspect, the present invention provides a fault diagnosis method for a rotary machine bearing, including: Acquiring a vibration signal to be diagnosed of a rotary mechanical bearing; Inputting the vibration signal to be diagnosed into a pre-constructed bearing fault diagnosis model, and outputting a fault diagnosis result of a rotary mechanical bearing, wherein the bearing fault diagnosis model comprises a feature extraction network, a GMM feature enhancement module, a projection distillation module and a dynamic expandable classifier which are sequentially connected; the training of the bearing fault diagnosis model comprises the following steps: acquiring a historical fault data set and a real-time newly-added fault data set of a rotary mechanical bearing; in an initial task stage, training a feature extraction network and a dynamic expandable classifier through a historical fault data set