CN-122020540-A - Rotary mechanical multi-mode fault diagnosis method based on multi-level cross attention fusion network

Abstract

The invention belongs to the technical field of fault diagnosis of deep learning rotary machines, and particularly relates to a rotary machine multi-mode fault diagnosis method based on a multi-level cross attention fusion network, which comprises the steps of adopting a kurtosis screening and constant Q non-steady gabor time-frequency conversion method based on variation modal decomposition to respectively convert original signals from different sensors into time-frequency diagrams with self-adaptive resolution; the method comprises the steps of constructing a fusion self-adaptive multi-branch attention module, comprehensively and finely extracting fault characteristics through a parallel multi-scale structure and a dual attention mechanism, constructing a multi-level fusion strategy and a final fusion module based on cross attention, carrying out dynamic self-adaptive characteristic fusion and refining on the extracted fault characteristics from shallow layers to deep layers, and outputting final characteristic representation with high discrimination. The method obviously improves the accuracy and reliability of fault diagnosis by effectively fusing the complementary information from different sensors.

Inventors

• FENG ZHIGANG
• WANG NAN

Assignees

• 沈阳航空航天大学

Dates

Publication Date

20260512

Application Date

20260129

Claims (5)

1. 1. The rotary mechanical multi-mode fault diagnosis method based on the multi-level cross attention fusion network is characterized by comprising the following steps of: Adopting a kurtosis screening and constant Q non-stationary gabor time-frequency conversion method based on variation modal decomposition to respectively convert original signals from different sensors into a time-frequency diagram with self-adaptive resolution; Constructing a fusion self-adaptive multi-branch attention module, and comprehensively and finely extracting fault characteristics through a parallel multi-scale structure and a dual attention mechanism; And constructing a multi-level fusion strategy and a final fusion module based on cross attention, and dynamically and adaptively fusing and refining the extracted fault characteristics from shallow layers to deep layers to output final characteristic representation with high discriminant.
2. 2. The rotary mechanical multi-mode fault diagnosis method based on the multi-level cross attention fusion network according to claim 1 is characterized in that the kurtosis screening and constant Q non-stationary gabor time-frequency transformation method based on variation mode decomposition comprises the steps of selecting a plurality of IMF components to be overlapped through variation mode decomposition of an original signal, completing reconstruction of the signal, and carrying out constant Q non-stationary gabor transformation on the reconstructed signal.
3. 3. The rotary mechanical multi-mode fault diagnosis method based on the multi-level cross attention fusion network according to claim 2, wherein the method is characterized in that the reconstruction of the signals is completed by performing variation mode decomposition on an original signal, selecting a plurality of IMF components to be overlapped, and performing constant Q non-stationary gabor transformation on the reconstructed signals, and specifically comprises the following steps: using VMD to convert original input signals Adaptively decomposed into K eigenmode functions IMF with a specific center frequency and limited bandwidth: Wherein the method comprises the steps of Is a set of K modal components to be solved; is the set of center frequencies corresponding to each modality, Is a dirac function; Representation pair Performing Hilbert transform to obtain an analytic signal; Represents the partial derivative with respect to time; Representing the square of the L2 norm for estimating the bandwidth of the signal; calculating kurtosis value of each IMF Each modal component is quantized The intensity of the impact component contained in the matrix for a discrete time series of length N The calculation formula is as follows: Wherein the method comprises the steps of Is a mode of Is a mean value of (c). Selecting a plurality of IMF components with highest kurtosis values for superposition to complete the reconstruction of signals and obtain reconstructed signals ; Reconstructed signal Performing constant Q non-stationary Gabor transformation to generate a two-dimensional time-frequency representation with adaptive resolution; defining a logarithmic frequency axis, giving the lowest frequency Highest frequency And the number of frequency points B in each octave, the th Center frequency of each frequency point Can be determined by the following formula: wherein K is as follows And thus, Factor is derived from the frequency point number B According to a constant Definition of the factor, in terms of sample points, specific frequency The required time window length The calculation formula is as follows: Wherein, the Is the sampling frequency of the signal; Using gaussian windows as window functions For the first Frequency channels, discretized window functions thereof The definition is as follows: Wherein the method comprises the steps of Is a parameter controlling window width, for discrete input signals CQ-NSGT coefficient thereof Representing the index of the signal in the time window And frequency channel index Time-frequency energy at the location; the inner product of the signal and the corresponding Gabor atom is calculated to obtain: Wherein the method comprises the steps of Is a time step, Is the first A start sampling point of each time window, Is a window function Complex conjugate of (C), Is of central frequency Is a complex sinusoidal basis function of (2); Taking the mould length, i.e The elements constituting the time-frequency representation matrix.
4. 4. The rotary mechanical multi-mode fault diagnosis method based on multi-level cross attention fusion network according to claim 1, wherein the constructing the fusion self-adaptive multi-branch attention module comprehensively and finely extracts fault characteristics through a parallel multi-scale structure and a dual attention mechanism, and comprises the following steps: the time-frequency chart obtained by VMD-KS-CQNSGT is recorded as 、 Inputting two time-frequency images into two independent two-dimensional convolution layers with the same kernel size in parallel for preliminary feature extraction to obtain corresponding shallow feature images And ; Two shallow feature maps at multi-layer fusion stage And Interaction is carried out through a cross attention fusion module, and a fused shallow feature map is generated Simultaneously, the two shallow feature graphs are respectively sent into two parallel feature learning branches, and each branch is formed by stacking fusion self-adaptive multi-branch attention modules: And Respectively through 2 fusion self-adaptive multi-branch attention modules, the two modules are converted into a middle-level characteristic diagram with more discriminant power And In the middle layer fusion stage, And Also fused via cross-attention mechanism to middle layer fusion features ; The two middle-level feature graphs are further refined into deep-level feature graphs through 2 fusion adaptive multi-branch attention modules And And fused into deep fusion features by cross-attention in the deep fusion stage 。
5. 5. The rotary mechanical multi-mode fault diagnosis method based on multi-level cross attention fusion network of claim 1, wherein the final fusion stage is to fuse features of shallow, medium and deep three levels 、 、 Performing self-adaptive average pooling, and splicing the obtained three feature graphs with consistent sizes on the channel dimension to form a composite feature of multi-scale and multi-level information; And in the final refining stage, the feature is subjected to dimension reduction, information refining and deep optimization through a polymerization bottleneck and a dynamic gating mechanism, and a final feature representation with high discriminant is output.

Description

Rotary mechanical multi-mode fault diagnosis method based on multi-level cross attention fusion network Technical Field The invention belongs to the technical field of fault diagnosis of deep learning rotary machines, and particularly relates to a rotary machine multi-mode fault diagnosis method based on a multi-level cross attention fusion network. Background With the continuous improvement of industrial modernization and intelligent manufacturing level, the rotating machinery is used as a key power device, and the health state of the rotating machinery is critical to guaranteeing the stable operation and production safety of the whole industrial system. However, in practical industrial situations, the core components of rotating machinery (e.g., rolling bearings, gears, etc.) are extremely susceptible to damage and failure due to long-term exposure to high strength, heavy loads, and harsh environments. More complicated, signals such as vibration, acoustics and the like generated when faults occur usually show typical non-stable and nonlinear characteristics, weak early fault characteristics are easily submerged by strong background noise, and these factors bring serious challenges to achieving accurate and reliable fault diagnosis. Therefore, the novel intelligent diagnosis technology capable of effectively resisting noise interference and fully excavating fault information is researched, and has great research significance and application value for preventing catastrophic accidents, reducing maintenance cost and improving production benefits. Conventional fault diagnosis methods often rely on sensor information of a single modality (e.g., using only vibration signals). However, the information provided by a single sensor has a limited dimension, and it is difficult to fully characterize the operational state of complex devices. In contrast, the multi-sensor information fusion technology can provide richer and more complementary device status information by integrating data from different sources (such as vibration, sound, current, etc.), thereby obtaining higher diagnostic reliability and accuracy, and has become a research hotspot in recent years. And processing and fusing the multi-sensor data by adopting a short-time Fourier transform (SFTF) technology, and then classifying motor faults by utilizing a Support Vector Machine (SVM). The vibration acceleration signals of the permanent magnet synchronous motor at different positions are utilized, then the sensor signals from three different mounting positions are combined into an image by utilizing a multi-signal Grignard angle difference field (MGADF) method, and finally the multi-texture features are fused to extract the features of the image. In addition, a series of methods such as deep random forest fusion, ensemble Empirical Mode Decomposition (EEMD) entropy feature fusion based classifier fusion with K-fold cross validation, KNN combination improvement of D-S evidence theory, K-nearest neighbor (KNN) classifier and advanced sensor fusion, VMD decomposition extraction MPE modeling learning and the like have all made beneficial progress in the field of multi-sensor fault diagnosis. Although the above methods show the potential of multi-sensor fusion, most of them rely heavily on artificial feature extraction and field prior knowledge, and their "self-learning" ability is limited, model generalization and robustness are often insufficient, and it is difficult to cope with complex failure modes and interference of strong noise. In recent years, the multi-sensor fusion technology based on deep learning has shown great potential to overcome the above limitations due to its strong end-to-end feature learning capability. For example, a Variational Mode Decomposition (VMD) is combined with a residual network containing a attentive mechanism to enable fault diagnosis of various components of the hydraulic system. A method for extracting time series information and quasi-static information based on expert experience knowledge. Diagnostic framework based on glas Mi Anjiao field imaging and connected neural network. ResTransUnet model, through integrating the residual connection of the transducer layer and the improvement, the overall situation and multistage feature modeling capability of the model is effectively enhanced. In addition, deep learning frameworks such as modal feature enhanced fusion networks, full convolution networks and deep feature fusion, and failure frequency prior fusion. In particular, some studies explore methods of converting one-dimensional signals into two-dimensional images (e.g., RGB images, relative angle matrices) and processing using convolutional neural networks. However, in spite of the existing fusion method based on deep learning, the method still has the limitations of (1) single fusion level, wherein most methods only fuse deep features at the end of a network, neglect abundant details and structural information contained i