Rotating machine bearing intelligent diagnosis method and system for cross-domain data missing

Abstract

The invention discloses a rotating machine bearing intelligent diagnosis method and system oriented to cross-domain data missing. The method comprises the steps that a source domain complete vibration signal and a target domain missing vibration signal of a rotating machine bearing are acquired; complex domain interpolation is carried out on the target domain missing vibration signal, and high-frequency noise is filtered by using an adaptive threshold value to obtain a target domain complete vibration signal after interpolation and denoising; extracting a source domain time-frequency joint feature of the source domain complete vibration signal and a target domain time-frequency joint feature of the target domain complete vibration signal; and performing domain alignment on the target domain time-frequency joint feature and the source domain time-frequency joint feature, decoding the target domain time-frequency joint feature after the domain alignment, and inputting the decoded target domain time-frequency joint feature into a classifier to obtain the fault category probability of the rotating machine bearing. The method is applied to the field of fault diagnosis, and can effectively solve the problem of performance degradation of a traditional method in cross-domain data distribution offset and high-missing-rate scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment fault diagnosis, and in particular to an intelligent diagnosis method and system for rotating machinery bearings facing cross-domain data loss. Background Art

[0002] Traditional bearing fault diagnosis methods face significant technical bottlenecks when dealing with complex industrial scenarios. First, existing data interpolation methods have inherent flaws in maintaining vibration signal integrity. While the KNN algorithm, based on neighboring sample filling, can quickly fill missing values, its local mean strategy disrupts the temporal continuity of the signal, making it difficult to capture transient impact characteristics, especially in scenarios with high missing rate. Generative adversarial networks can improve interpolation quality through data synthesis, but are limited by the generator's ability to model high-frequency details, often resulting in distorted impact response waveforms and, in turn, misjudgment of diagnostic models.

[0003] Secondly, mainstream domain adaptation methods perform poorly when dealing with cross-condition migration tasks. Typical methods, represented by domain adversarial networks and joint adaptation networks, mitigate inter-domain offsets by aligning feature distributions, but their single-domain alignment strategies struggle to handle the problem of joint time-frequency distribution offsets. When equipment speed and load change, aligning only time-domain or frequency-domain features can lead to the loss of critical fault information. More seriously, these methods generally employ a multi-layer neural network stacking structure, with model parameters often reaching millions. When deployed on edge computing devices, they face issues such as excessive memory usage and high inference latency, making it difficult to meet the needs of industrial real-time diagnosis.

[0004] Even more challenging, industrial sites often face the combined challenges of data loss and distribution offset. Existing end-to-end solutions lack breakthroughs in time-frequency feature decoupling and noise robustness. This results in insufficient average diagnostic accuracy in complex operating conditions characterized by variable speeds, high noise levels, and non-uniform data loss, severely hindering the intelligent operation and maintenance of industrial equipment. Summary of the Invention

[0005] In response to the above-mentioned deficiencies in the prior art, the present invention provides an intelligent diagnosis method and system for rotating machinery bearings for cross-domain data missing, which can effectively solve the performance degradation problem of traditional methods in scenarios with cross-domain data distribution offset and high missing rate.

[0006] To achieve the above object, the present invention provides an intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss, comprising the following steps: Step 1: Obtain a complete vibration signal in the source domain and a missing vibration signal in the target domain of a rotating machinery bearing; Step 2: interpolating the missing vibration signal in the target domain in the complex domain and filtering out high-frequency noise using an adaptive threshold to obtain a complete vibration signal in the target domain after interpolation and denoising; Step 3: extracting the source domain time-frequency joint features of the source domain complete vibration signal and the target domain time-frequency joint features of the target domain complete vibration signal; Step 4: align the target domain time-frequency joint features with the source domain time-frequency joint features, decode the aligned target domain time-frequency joint features and input them into the classifier to obtain the fault category probability of the rotating machinery bearing.

[0007] Compared with the prior art, the present invention has the following beneficial technical effects: 1. In the process of interpolating missing vibration signals in the target domain, the present invention significantly improves the signal spectrum resolution and robustness to noise interference through complex domain amplitude-phase fusion and sub-channel phase compensation strategies; 2. This invention uses a domain adaptation mechanism based on joint time-frequency distribution alignment to align source domain features with target domain features, effectively solving the problem of generalization performance degradation of diagnostic models in scenarios with missing cross-domain data. 3. The present invention sets a dual-path decoder in the unsupervised domain adaptation module, thereby taking into account the reconstruction capabilities of the signal periodic components and transient impact characteristics, and improving the interpretability and engineering applicability of the diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0009] Figure 1 This is a flow chart of the intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss in Example 1 of the present invention; Figure 2 This is a flowchart of complex domain adaptive interpolation in embodiment 1 of the present invention; Figure 3 Flowchart of unsupervised domain adaptation in Example 1 of the present invention; Figure 4 This is a structural block diagram of the intelligent diagnosis system for rotating machinery bearings for cross-domain data loss in Example 2 of the present invention.

[0010] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0011] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0012] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in this field can implement it. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0013] Example 1 like Figure 1 The figure shows an intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss disclosed in this embodiment, which mainly includes the following steps: Step 1: Obtain a complete vibration signal in the source domain and a missing vibration signal in the target domain of a rotating machinery bearing; Step 2: Interpolate the missing vibration signal in the target domain in the complex domain and use an adaptive threshold to filter out high-frequency noise to obtain the complete vibration signal in the target domain after interpolation and denoising; Step 3: extracting the source domain time-frequency joint features of the complete vibration signal in the source domain and the target domain time-frequency joint features of the complete vibration signal in the target domain; Step 4: Align the target domain time-frequency joint features with the source domain time-frequency joint features, decode the aligned target domain time-frequency joint features, and input them into the classifier to obtain the fault category probability of the rotating machinery bearing in the target domain.

[0014] In the specific implementation process of step 2, the missing vibration signal in the target domain is interpolated in the complex domain based on the spectrum-focused Fourier transform, thereby reducing the length and computational complexity of a single Fourier operation and enhancing the spectral resolution to achieve high-precision and lightweight complex domain signal processing. The specific implementation process is as follows: First, the target domain missing vibration signal is processed to form several sub-channels. For example, given the input time domain signal , the time series sampling points can be expressed as , the signal amplitude is , the signal carrier frequency is , the sampling frequency is , the signal phase is , is the sampling point index, is the number of sampling points, and Gaussian white noise is , if the original time domain signal Perform intervals The bit extraction process forms sub-channels, then The time domain signal of each sub-channel is: ; in, is an imaginary unit; Then, the time domain signals of each sub-channel are Perform zero-padding fast Fourier transform operation to obtain the spectrum signal of the sub-channel; Finally, the phase term of the spectrum signal of each sub-channel is compensated Then perform spectrum superposition to obtain the final output spectrum signal ,for: ; ; ;

[0015] in, is the discrete frequency index, is the coherent superposition factor, is the frequency deviation, For the The noise component of each sub-channel.

[0016] When the frequency deviation When the signal component is enhanced due to coherent superposition, the noise is incoherently superposed, which significantly improves the signal-to-noise ratio. Compared with direct FFT, the spectrum focusing in this embodiment compresses the frequency range to , the resolution is determined by Upgrade to , while maintaining the data length Super-resolution analysis is achieved simultaneously.

[0017] In addition, high-frequency noise usually manifests as random fluctuations that deviate from the signal trend and are difficult to analyze. This embodiment proposes an adaptive threshold filtering method that effectively removes high-frequency noise components by dynamically adjusting the filtering strength to adapt to different data characteristics. This method is particularly suitable for non-stationary time series data, whose spectral characteristics may change over time. Figure 2 In this embodiment, the process of using the adaptive threshold to filter out high-frequency noise is specifically as follows: First, the final output spectrum signal Forming a power spectrum , and form a trainable threshold based on the power spectral density , so that the trainable threshold Dynamically distinguish valid signals from noise, i.e. based on trainable thresholds The final output spectrum signal Filter and retain power higher than The frequency components of the filtered spectrum signal are obtained , thereby ensuring that key information is retained while removing noise, and automatically adapting the optimal threshold for different time series data. During the training phase, the threshold can be trained Optimization is performed via gradient descent.

[0018] Then, in order to combine the global periodic pattern with the local detail features, a dynamic fusion weight is constructed based on a learnable complex filter that fuses the global filter and the local filter. ,for: ; in, is a global filter, is a local filter; Then, based on the dynamic fusion weight Spectrum signal Perform feature adjustment to obtain the fused frequency domain features ,for: ; in, Spectrum signal The output after inputting the global filter, Spectrum signal Output after inputting the local filter; Finally, the fused frequency domain features Perform inverse discrete Fourier transform to obtain the denoised and enhanced time domain signal .

[0019] As a preferred implementation, considering the similarity between the input sequence and the global element, two linear layers are introduced as a lightweight self-attention generator for the time domain signal. Perform feature reconstruction and finally obtain the complete vibration signal of the target domain ,for: ; ; Calculate input sequence and global memory weight similarity, and through Normalized attention weight matrix , The input feature is weighted according to the global memory The reconstructed output features, is the transpose of the matrix. Among them, the global memory weight 、 Encode statistical regularities across samples through overall training set optimization.

[0020] The time domain features of a time series (such as mean, variance, and trend) reflect the local dynamic changes of the signal, while the frequency domain features (extracted through Fourier transform or wavelet transform) reveal the global periodic structure of the signal. The periodicity of the signal is manifested in the frequency domain as a specific frequency component with concentrated energy, which has stronger cross-domain invariance. Therefore, in step 3, this embodiment captures the time domain dynamics and frequency domain structure simultaneously through time-frequency joint coding to form a more comprehensive feature representation, that is, the process of extracting the source domain time-frequency joint features and the target domain time-frequency joint features is as follows: for the complete vibration signal in the target domain or the complete vibration signal in the source domain, its time domain features are extracted based on the time domain encoder, and its frequency features are extracted based on the frequency domain encoder. The extracted time domain features are combined with the frequency features to obtain the source domain time-frequency joint features or the target domain time-frequency joint features.

[0021] Frequency feature shift can be regarded as the manifestation of covariate shift in the frequency domain. At this time, although the frequency domain-label mapping relationship learned in the source domain is effective, direct application to the target domain will degrade performance due to differences in frequency domain distribution. In addition, the performance of domain adaptation is limited by the distribution differences between the source domain and the target domain. When only the time domain features are aligned, if the time domain shift is large, it is difficult for the model to narrow the inter-domain differences. Frequency domain features often contain more stable cross-domain invariance. By jointly aligning time-frequency features, the overall inter-domain differences can be reduced and the theoretical performance upper limit of domain alignment can be improved. In this embodiment, the frequency domain encoder uses the above-mentioned spectrum-focused Fourier transform to extract frequency features, and the time domain encoder uses a time domain multi-scale decomposition network to extract time domain features. Reference Figure 3 , the time domain encoder in this embodiment is: the original time series is , decompose it into trend components through linear filtering or smoothing operations and the residual component ,satisfy , ,in For parameters Filter operator.

[0022] In the specific implementation process of step 4, the Sinkhorn divergence domain alignment method is used to align the target domain time-frequency joint features with the source domain time-frequency joint features. The Sinkhorn divergence domain alignment method is an entropy regularization-based domain alignment method that aims to balance the geometric properties of optimal transmission with the sample efficiency of maximum mean difference. Given two discrete probability measures: , ;in, 、 is a non-negative probability vector satisfying , the Sinkhorn divergence is obtained by optimizing the coupling matrix and the cost matrix, where To satisfy the row and , column and A set of non-negative matrices. In this embodiment, 、 They can represent the probability distribution of source domain samples and target domain samples respectively.

[0023] In the specific implementation of step 4, the domain-aligned target domain time-frequency joint features undergo dual-path decoding, including a frequency reconstruction path for recovering the signal's periodic components through inverse Fourier transforms, and a time-domain reconstruction path for reconstructing transient impulse features using specific time-domain methods. This approach balances the ability to reconstruct both the signal's periodic components and transient impulse features, improving the interpretability and engineering applicability of diagnostic results.

[0024] It is worth noting that although this embodiment Figure 1 The steps in the diagram are shown in the order indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0025] Example 2 Based on the intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss in Example 1, this embodiment discloses an intelligent diagnosis system for rotating machinery bearings facing cross-domain data loss, referring to Figure 4 The intelligent diagnosis of rotating machinery bearings for cross-domain data missing includes a complex domain adaptive interpolation module and an unsupervised domain adaptation module. The complex domain adaptive interpolation module consists of a spectrum focusing Fourier transform submodule, an adaptive threshold filter submodule and a lightweight self-attention generator, and the unsupervised domain adaptation module consists of a time-frequency joint encoder, a domain alignment classifier and a dual-path decoder.

[0026] In this embodiment, the specific working process and working principle of the spectrum focusing Fourier transform submodule, the adaptive threshold filter submodule, the lightweight self-attention generator, the time-frequency joint encoder, the domain alignment classifier and the dual-path decoder are the same as those in Example 1, so they will not be described in detail in this embodiment. Each unit module can be implemented in whole or in part by software, hardware and a combination thereof. Each unit module can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above each unit module.

[0027] The above description is only a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. All equivalent structural transformations made by using the contents of the present invention description and drawings under the inventive concept of the present invention, or direct / indirect application in other related technical fields are included in the scope of protection of the present invention.

Claims

1. An intelligent diagnosis method for rotating machinery bearings with cross-domain data loss, characterized in that: The steps include: Step 1: Obtain a complete vibration signal in the source domain and a missing vibration signal in the target domain of a rotating machinery bearing; Step 2: interpolating the missing vibration signal in the target domain in the complex domain and filtering out high-frequency noise using an adaptive threshold to obtain a complete vibration signal in the target domain after interpolation and denoising; Step 3: extracting the source domain time-frequency joint features of the source domain complete vibration signal and the target domain time-frequency joint features of the target domain complete vibration signal; Step 4: align the target domain time-frequency joint features with the source domain time-frequency joint features, decode the aligned target domain time-frequency joint features, and input them into the classifier to obtain the fault category probability of the rotating machinery bearing.

2. The intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss according to claim 1 is characterized in that: In step 2, the missing vibration signal in the target domain is interpolated in the complex domain based on the spectrum-focused Fourier transform, specifically: Performing value extraction processing on the missing vibration signal of the target domain to form a plurality of sub-channels; Performing a zero-padding fast Fourier transform operation on the time domain signal of each sub-channel to obtain a spectrum signal of the sub-channel; The spectrum signals of the sub-channels are compensated for the phase terms and then spectrum superimposed to obtain a final output spectrum signal.

3. The intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss according to claim 2 is characterized in that: The final output spectrum signal is: in, is the final output spectrum signal, is the discrete frequency index, represents the signal amplitude, is the imaginary unit, is the signal phase, is the coherent superposition factor, is the number of sub-channels, is the frequency deviation, For the The noise component of the sub-channel, is the length of the missing vibration signal in the target domain, is the zero-padding multiple, is the signal carrier frequency, is the sampling frequency.

4. The intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss according to claim 1, 2 or 3, characterized in that: In step 2, the use of an adaptive threshold to filter out high-frequency noise specifically includes: Setting the trainable threshold , and based on the trainable threshold The final output spectrum signal Filter and retain power higher than The frequency components of the filtered spectrum signal are obtained ; Constructing dynamic fusion weights based on learnable complex filters that fuse global filters and local filters ,for: in, is a global filter, is a local filter; Based on dynamic fusion weight Spectrum signal Perform feature adjustment to obtain the fused frequency domain features ,for: in, Spectrum signal The output after inputting the global filter, Spectrum signal Output after inputting the local filter; The fused frequency domain features Perform inverse discrete Fourier transform to obtain the denoised and enhanced time domain signal .

5. The intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss according to claim 4 is characterized in that: After using adaptive threshold to filter out high-frequency noise, two linear layers are introduced as lightweight self-attention generators for time domain signals. Perform feature reconstruction to obtain the complete vibration signal of the target domain, which is: in, is the complete vibration signal in the target domain, is the attention weight matrix, 、 Global memory weights optimized for trainability, is the transpose of the matrix.

6. The intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss according to claim 1, 2 or 3, characterized in that: The process of extracting the source domain time-frequency joint features and the target domain time-frequency joint features is as follows: For the target domain complete vibration signal or the source domain complete vibration signal, extract its time domain features based on a time domain encoder, and extract its frequency features based on a frequency domain encoder, Combining the extracted time domain features with the frequency features to obtain the source domain time-frequency joint features or the target domain time-frequency joint features; The time domain encoder uses a time domain multi-scale decomposition network to extract time domain features, and the frequency domain encoder uses a spectrum focused Fourier transform to extract frequency features.

7. The intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss according to claim 1, 2 or 3, characterized in that: In step 4, dual-path decoding is performed on the domain-aligned target domain time-frequency joint features, including: Frequency reconstruction path, used to recover the periodic component of the signal through inverse Fourier transform; The time domain reconstruction path is used to reconstruct transient impact characteristics using a specific time domain method.

8. The intelligent diagnosis method for rotating machinery bearings facing cross-domain data loss according to claim 1, 2 or 3, characterized in that: In step 4, the Sinkhorn divergence domain alignment method is used to perform domain alignment of the target domain time-frequency joint features and the source domain time-frequency joint features.

9. An intelligent diagnosis system for rotating machinery bearings facing cross-domain data loss, characterized in that: According to the method described in any one of claims 1 to 8, the intelligent diagnostic system for rotating machinery bearings comprises: The complex domain adaptive interpolation module consists of a spectrum-focused Fourier transform submodule, an adaptive threshold filter submodule, and a lightweight self-attention generator; The unsupervised domain adaptation module consists of a joint time-frequency encoder, a domain alignment classifier, and a dual-path decoder.

Images