CN-121977835-A - Strong noise and bearing fault diagnosis method under variable working condition based on multi-mode fusion projection

CN121977835ACN 121977835 ACN121977835 ACN 121977835ACN-121977835-A

Abstract

The invention relates to a bearing fault diagnosis method based on multi-mode fusion projection under strong noise and variable working conditions, which belongs to the technical field of bearing fault diagnosis and comprises the following steps of S1, collecting vibration and current signals, carrying out frequency domain transformation, S2, utilizing a dual-path current to guide a vibration characteristic fusion strategy to fuse vibration and current characteristics, S3, introducing a learnable class self-adaptive projection matrix, mapping the fused characteristics into a subspace with unchanged running state, and S4, constructing an encoder-projection coupling loss function, realizing joint optimization of a dual-path encoder and the class self-adaptive learnable projection matrix, so that the model can learn the representation of unchanged running state and class distinction degree under different rotating speeds, and diagnosing and classifying bearing faults.

Inventors

QIN YI
ZHAO LIJUAN
NIU BUZHAO
BAI HOUYI
MAO YONGFANG

Assignees

重庆大学

Dates

Publication Date: 20260505
Application Date: 20260204

Claims (9)

1. A bearing fault diagnosis method based on multi-mode fusion projection under strong noise and variable working conditions is characterized by comprising the following steps: s1, collecting vibration and current signals and performing frequency domain transformation; s2, fusing vibration and current characteristics by utilizing a dual-path current-guided vibration characteristic fusion strategy; s3, introducing a learnable class self-adaptive projection matrix, and mapping the fused features to a subspace with unchanged running state; and S4, constructing an encoder-projection coupling loss function, realizing the joint optimization of the dual-path encoder and the class self-adaptive learning projection matrix, so that the model can learn the representation of unchanged running state and class distinction under different rotating speeds, and diagnosing and classifying bearing faults.
2. The method for diagnosing a bearing fault under a strong noise and variable working condition based on multi-modal fusion projection as set forth in claim 1, wherein the current signal is Containing a fundamental component Harmonic component Fault related component As shown in formula (1): (1) Wherein the fundamental component Is determined by the frequency and the rotating speed of the power supply; The vibration signal Containing fundamental vibration components Random noise component And fault related component As shown in formula (2): (2) random noise component Including ambient noise, sensor noise, and other non-fault disturbances, fault-related components Reflecting the impact or abnormal vibration caused by mechanical failure.
3. The method for diagnosing a bearing fault under strong noise and variable working conditions based on multi-modal fusion projection as set forth in claim 1, wherein the fusing of vibration and current features by using a dual-path current-guided vibration feature fusion strategy in step S2 includes: Inputting vibration and current signals converted into a frequency domain; The vibration branch adopts Morlet wavelet convolution kernel to pre-extract fault frequency band from vibration signal, and the current branch adopts harmonic wavelet convolution kernel to capture the characteristic related to rotation speed; Then, the two branches aggregate multi-scale time information through a plurality of one-dimensional convolution modules and a channel attention module, and the channel characteristics with the most information quantity are highlighted; A current-guided channel-time hybrid gating mechanism is introduced, wherein the rotation speed related features extracted by the current branches are used for generating gating weights, so that the self-adaptive positioning of the vibration features in the channel dimension and the time dimension is realized; Finally, the vibration and current embedding are spliced to form a unified bimodal representation.
4. The method for diagnosing the bearing faults under the strong noise and variable working conditions based on the multi-mode fusion projection as claimed in claim 1 is characterized in that the step S2 comprises the following steps: s21, input vibration and current signals are defined as follows: (3) Wherein the method comprises the steps of In order to be of a batch size, Is the signal length; S22, a learnable Morlet wavelet convolution layer is adopted for the vibration branch, and a kernel function is defined as follows: .(4) learner version importation scale And translation As training parameters: (5) The generation of the first order vibration based on the convolution transform of the Morlet wavelet is represented as follows: (6) the current branches employ a learnable harmonic convolution layer: (7) Wherein the method comprises the steps of And The output of the harmonic convolution is: (8) s23, gradually applying nonlinear transformation to the primary features extracted by the wavelet layer through a multiple convolution module MCB, thereby capturing advanced, rich and discernable depth features: (9) (10) Applying pinch-activated SE-attention mechanisms in both branches to strengthen the discriminating channel and suppress redundant activation, input signature is denoted as The channel compression operation is as follows: (11) s24, learning importance weights of all channels And pass through Channel-level weight allocation is realized: (12) (13) the corresponding vibration and current branches are defined as: (14) S25, introducing a mixed time domain-channel gating mechanism for suppressing environmental noise in vibration characteristics and highlighting fault-related components drifting along with the rotating speed and fault-related components drifting along with the rotating speed: (15) (16) Wherein the method comprises the steps of As a function of the Sigmoid, After gating processing, the adaptive time domain pooling aggregates two types of features into compact embedding: (17) s26, splicing vibration and current embedding to form uniform bimodal representation: (18)。
5. The method for diagnosing the bearing faults under the strong noise and variable working conditions based on the multi-mode fusion projection as claimed in claim 1 is characterized in that the step S3 specifically comprises the following steps: S31 definition of a leachable projection matrix for each class : (19) Wherein the method comprises the steps of Representing the dimensions of the tandem bimodal feature, Representing a low-dimensional subspace dimension of the projection target; S32, projecting the series bimodal features into each subspace: (20) s33 for each category Are all provided with a weight vector capable of learning A logic value for computing a classification in its corresponding low-dimensional subspace: (21) For batch The projection characteristics of the samples are calculated in parallel with the logic value: (22) (23) s34 utilizing Calculating cross entropy loss, enabling each class to define an independent low-dimensional subspace for discrimination projection, thereby realizing alignment of similar features under different operation conditions, and separating out feature and speed dependency change related to faults; s35, obtaining the prediction probability of each category through a softMax function: (24) the predicted class label for each sample is determined by the maximum probability: (25)。
6. The method for diagnosing a bearing fault under a strong noise and variable working condition based on multi-modal fusion projection as set forth in claim 1, wherein the encoder-projection coupling loss function in step S4 includes: Classification loss Adopts standard cross entropy loss as a main supervision target to promote the model to generate higher confidence degree for the real category: (26) Wherein the method comprises the steps of Representing a true category; Within class alignment loss The similar characteristics are aggregated, namely, a pair cosine similarity matrix of all samples in a batch is calculated firstly: (27) Wherein the method comprises the steps of Representing projection features Is a vectorized version of (a); Subsequently, the sample pairs are divided into homogeneous groups according to labels Heterogeneous group The intra-class alignment loss function is defined as follows: (28) projection matrix orthogonality loss Maintaining independence of projection subspace column vectors, ensuring that each category is defined by a learnable projection matrix The defined low-dimensional subspaces remain orthogonal and information independent: (29) Wherein the method comprises the steps of Represents the Fr Luo Beini Us norm, Is a d-dimensional identity matrix; The final joint optimization objective is defined as follows: (30)。
7. an electronic device comprising a memory and a processor; The memory is used for storing a computer program; The processor is configured to implement the method for diagnosing a bearing fault under a strong noise and variable working condition based on multi-modal fusion projection according to any one of claims 1 to 6 when the computer program is executed.
8. A computer readable storage medium, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, the method for diagnosing bearing faults under strong noise and variable working conditions based on multi-mode fusion projection according to any one of claims 1 to 6 is realized.
9. A computer program product comprising a computer program which, when executed by a processor, implements the method for diagnosing a bearing failure under variable conditions and strong noise based on multi-modal fusion projection as set forth in any one of claims 1 to 6.

Description

Strong noise and bearing fault diagnosis method under variable working condition based on multi-mode fusion projection Technical Field The invention belongs to the technical field of bearing fault diagnosis, and relates to a method for diagnosing bearing faults under strong noise and variable working conditions based on multi-mode fusion projection. Background Bearings are critical components of rotary machines, the operating conditions of which directly affect the safety and stability of the equipment. The bearing is easy to generate various faults due to long-term bearing alternating load, friction and abrasion and assembly errors. Timely and accurate fault identification can prevent sudden failure of equipment, reduce economic loss and ensure continuous and reliable operation of a mechanical system. Therefore, bearing fault diagnosis has important significance in the fields of intelligent manufacturing and equipment health monitoring. However, industrial equipment often operates under non-stationary conditions (e.g., variable speed or variable load), and conventional diagnostic methods designed for constant conditions have limited effectiveness. In addition, the vibration signals collected under the complex working condition are often severely interfered by strong noise, so that the difficulty of feature extraction and fault identification is multiplied. Therefore, the method has great significance in diagnosing bearing faults in variable speed and noise environments. Single mode signals tend to be difficult to fully capture the complexity of bearing failure. Reliance on a single sensor type may result in missed detection and exacerbate sensitivity to external disturbances and noise, thereby reducing fault diagnosis accuracy. In recent years, a multi-mode fusion technology has been attracting attention in the field of intelligent fault diagnosis. The multimode fusion can more comprehensively present the running state of the equipment by integrating the complementary information of different types of signals such as vibration, acoustics, current, temperature and the like. Compared with a single-mode signal, the multi-mode data fusion can enhance the discrimination capability and the robustness of the extracted features. The different modes respectively reflect multiple dimensions of the mechanical fault, and the combined analysis of the method can not only improve the stability and reliability of the diagnosis result, but also relieve the influence of sensor abnormality or information loss. Therefore, the bearing fault diagnosis method based on multi-mode fusion has remarkable advantages in the aspect of realizing intelligent state monitoring with high precision and strong robustness. For example Choudhary and the like propose a reliable fault diagnosis method for an electric automobile motor based on wavelet synchronous extrusion transformation and a multi-input fusion network, which fuses vibration and current signals, while Wang and the like combine rotor rotation speed and stator current signals to realize bearing fault diagnosis of a permanent magnet synchronous motor under a variable speed working condition. Grandson et al developed a multi-modal Convolutional Neural Network (CNN) that fuses vibration and acoustic signals to capture complementary fault related information. Particularly, a deep learning framework based on fusion constraint of acoustic-vibration physical information is provided, and diagnosis precision and generalization capability are improved through weighting multiple physical information and embedding a model training process. Vibration signals are typically used in conjunction with other signal types as the core modality for multi-modality fusion bearing fault diagnosis. The main reason for this is that vibration signals are sensitive to mechanical shock and contain rich equipment operating state information. Under variable speed conditions, the rotational speed variation significantly affects the vibration signal characteristics. In the time domain, rotational speed variations change the periodicity and amplitude of the impulse response, resulting in failure-induced pulses exhibiting non-uniform and time-varying waveforms. In the frequency domain, the harmonic frequency related to the fault signature drifts with the rotational speed, making it difficult for conventional methods to accurately capture fault related information. The current signal under the variable speed working condition mainly reflects the influence of load and torque change on the motor current. The frequency characteristic is closely related to the rotating speed, and the running state of the equipment can be reflected. In addition, the current signal is indirectly influenced by the mechanical transmission path, so that the noise immunity is higher, and the additional information can be provided for the fault state. Complementary information is realized by fusing vibration and current signals, so that diagnosi