Rotating machinery fault diagnosis method and system based on direct gram frequency domain attention

Abstract

The application discloses a rotating machinery fault diagnosis method based on direct gram frequency domain attention, and belongs to the field of rotating machinery state monitoring. The method directly rearranges a fixed-length one-dimensional vibration sample into multiple tokens without word embedding and position encoding, performs real number fast Fourier transform on each token, constructs a gram similarity matrix by the product of a complex frequency domain tensor and the conjugate transpose thereof, obtains an attention matrix through softmax, and directly weights the original complex frequency spectrum without setting Q, K and V learnable projections. After the weighted frequency spectrum is subjected to inverse transformation, residual feedforward, pooling and classification, a fault category is output. The application reduces attention projection parameters, enhances the aggregation of stable fault frequency components under noise conditions, and is suitable for online fault identification of bearings and other rotating machines.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery condition monitoring and intelligent fault diagnosis technology, specifically to a rotating machinery fault diagnosis method and system based on direct Gram frequency domain attention. Background Technology

[0002] Rotating mechanical components such as bearings, gears, and rotors are subjected to cyclic loads, impact loads, and environmental noise for extended periods in industrial equipment. The impact response resulting from localized spalling, cracking, or pitting typically manifests as a non-stationary, low signal-to-noise ratio one-dimensional vibration sequence. Traditional methods relying on manually selecting time-domain statistics, envelope spectra, or characteristic frequencies are easily affected by variations in rotational speed, load, sensor installation location, and noise levels.

[0003] Wide-Deep Convolutional Neural Network (WDCNN) directly processes the original vibration signal through a wider convolutional kernel, exhibiting good in-domain accuracy in bearing fault diagnosis and becoming a commonly used baseline for deep fault diagnosis.

[0004] Existing technologies also propose simultaneously inputting time-domain signals and Fourier spectra into self-attention networks, or setting multiple attention modules in convolutional networks. For example, patent application CN116304861A discloses a fault diagnosis method based on self-attention time-frequency feature fusion; patent application CN110672343A discloses a fault diagnosis method for rotating machinery based on multi-attention convolutional neural networks. These solutions still rely on time-frequency dual-branch systems, convolutional feature extractors, or conventional attention parameterization structures.

[0005] Furthermore, patent application CN120781173A discloses a single-domain generalized rotating machinery fault diagnosis method based on a frequency domain attention diffusion model. This method progressively adds Gaussian noise to the source domain vibration signal and generates enhanced features using an encoder-decoder diffusion model. Attention weights are calculated based on channel spectral energy. The original features are then fused with the real and imaginary parts of multiple diffusion-enhanced features using learnable weighted fusion. Two-stage training is performed using cross-entropy loss and covariance embedding loss.

[0006] The key technical aspect of the aforementioned diffusion enhancement scheme lies in expanding the feature distribution of a single source domain through a generative model, with its frequency domain attention used to form channel weights.

[0007] Conventional Transformer attention typically involves first generating a query Q, key K, and value V using a learnable matrix, and then calculating... For short vibration windows, this projection introduces a large number of parameters and may learn a linear transformation coupled with specific loads or acquisition conditions; meanwhile, word embeddings and positional encoding in text tasks are not necessary conditions for calculating the spectral correlation of vibration signals.

[0008] Therefore, a frequency domain attention method is needed that combines the physical properties of one-dimensional vibration data to reduce unnecessary projection parameters, directly utilize the complex spectrum correlation between different signal blocks to enhance the stable fault frequency components, and improve the classification reliability in noisy environments. Summary of the Invention

[0009] This invention aims to address the problems of large Q / K / V projection parameters, indirect integration of attention similarity and vibration spectrum correlation in existing intelligent fault diagnosis models, and the easy submergence of fault features under additive noise conditions. To solve these problems, this invention does not perform forward diffusion, reverse denoising, enhanced feature generation, or fusion of original and generated features. Instead, it directly rearranges the original vibration samples into multiple vibration tokens and constructs a T×T Gram similarity matrix based on the conjugate inner product of the complex spectra of different tokens. This matrix directly aggregates the complex spectra between tokens. Therefore, the two differ in terms of the technical problems to be solved, the objects of attention calculation, network structure, and training process.

[0010] Technical solution:

[0011] This invention first discloses a method for diagnosing rotating machinery faults based on direct Gram frequency domain attention. The method is implemented based on a fault diagnosis model and specifically includes:

[0012] S101. Acquire the one-dimensional vibration signal generated during the operation of rotating machinery and input it into the fault diagnosis model;

[0013] S102. Slice the one-dimensional vibration signal into fixed-length slices to obtain vibration samples of length L;

[0014] S103. Without going through the word embedding layer, position encoding layer and learnable embedding projection, the vibration samples are directly rearranged into T vibration tokens according to the preset feature dimension d, where L=T×d;

[0015] S104. Perform a real-number fast Fourier transform along the characteristic dimension of each vibration token in an orthogonal normalization manner to obtain the complex frequency domain tensor Xf.

[0016] S105. Construct a token Gram similarity matrix G to describe the complex spectral correlation among T vibration tokens. Scale the Gram similarity matrix G according to the number of independent frequency points F and perform softmax processing to obtain the frequency domain attention matrix A. Use the frequency domain attention matrix A to apply the complex frequency domain tensor X... f Weighted aggregation is performed on the token dimension to obtain the weighted complex frequency domain tensor Y. f Among them, in constructing the Gram similarity matrix G and the weighted complex frequency domain tensor Y... f During the process, learnable projections of the query matrix, key matrix, and value matrix are not set;

[0017] S106. Using an orthogonal normalization method corresponding to the real number fast Fourier transform, perform an inverse real number fast Fourier transform on the weighted complex frequency domain tensor to restore the frequency domain features to the time domain features. After residual feedforward processing and classification processing, the fault category of the rotating machinery is output.

[0018] Preferably, the vibration samples are standardized to zero mean on a sample-by-sample basis, and the standardization result is (x-μ) / (σ+ε), where x represents the vibration sample, μ is the mean of the vibration sample, σ is the standard deviation of the vibration sample, and ε is a positive number to prevent division by zero.

[0019] Preferably, the length L is 2048, the number of vibration tokens T is 4, the preset feature dimension d is 512, and the number of independent frequency points F output by the real number fast Fourier transform is 257.

[0020] In the first implementation, for each batch of samples, the complex frequency domain tensors are arranged with the frequency dimension first and the token dimension last, and the Gram similarity matrix G is obtained according to the following formula: .

[0021] Among them, X f H X represents f The conjugate transpose of , Re(·) denotes taking the complex real part, F denotes the number of independent frequency points; the frequency domain attention matrix A=softmax(G), and the weighted complex frequency domain tensor Y f = A · X f .

[0022] As another implementation, for each batch of samples, the Gram similarity matrix G is obtained by arranging the complex frequency domain tensors with the token dimension first and the frequency point dimension last, according to the following formula: .

[0023] Among them, X f H X represents fThe conjugate transpose of , Re(·) denotes taking the complex real part, F denotes the number of independent frequency points; the frequency domain attention matrix A = softmax(G), and the weighted complex frequency domain tensor Y f = A · X f X f The conjugate transpose of Y has dimensions F×T, and the dimensions of G and A are both T×T. f The dimension is T×F; this expression is equivalent to the matrix expression in the first implementation method with the frequency dimension first and the token dimension last, under the transpose storage relationship.

[0024] Preferably, the residual feedforward processing includes: adding the residuals and normalizing the layers of the time-domain features obtained by the inverse transform; sequentially inputting the layer normalization results into the first fully connected layer, the GELU activation layer, the random deactivation layer, and the second fully connected layer; and adding the feedforward output to the residual branch.

[0025] Preferably, a frequency domain attention block is formed by real number fast Fourier transform, direct Gram frequency domain attention, inverse real number fast Fourier transform and residual feedforward processing; N frequency domain attention blocks are set in series, where N is an integer greater than or equal to 1, N represents the number of frequency domain attention blocks, T represents the number of vibration tokens contained in each vibration sample, and N and T are independent of each other.

[0026] Preferably, the classification process includes performing mean pooling on T vibration tokens, inputting the pooling result into at least one fully connected classification layer, and obtaining the class probability corresponding to at least one of the following states: normal state, rolling element fault, inner ring fault, and outer ring fault, through softmax.

[0027] Preferably, the construction and training of the fault diagnosis model includes: acquiring vibration records of rotating machinery labeled with fault categories, performing fixed-length slicing and sample-by-sample standardization on the vibration records, and dividing the original vibration records into training, validation, and test sets; constructing a fault diagnosis model that sequentially includes a token construction unit, N cascaded frequency domain attention blocks, a token pooling unit, and a fully connected classification unit, wherein each frequency domain attention block performs residual feedforward processing, and the token pooling unit and the fully connected classification unit perform classification processing; setting the initial learning rate, weight decay coefficient, batch size, and number of training rounds, and using the cross-entropy loss function and gradient-based optimization algorithm to perform mini-batch iterative training on the fault diagnosis model, and selecting model parameters based on the validation set loss; during training, optionally adding additive noise with random signal-to-noise ratio and / or performing random cyclic translation to the training vibration samples.

[0028] This invention also discloses a rotating machinery fault diagnosis system based on direct Gram frequency domain attention, comprising:

[0029] The signal acquisition module is used to acquire one-dimensional vibration signals of rotating machinery;

[0030] The preprocessing module is used to standardize and slice one-dimensional vibration signals into fixed-length slices.

[0031] The token construction module is used to directly rearrange vibration samples into T vibration tokens without setting a word embedding layer, a position encoding layer, and a learnable embedding projection.

[0032] The frequency domain transformation module is used to perform a real-number fast Fourier transform on each vibration token in an orthogonal normalization manner to obtain a complex frequency domain tensor.

[0033] The direct Gram frequency domain attention module is used to construct a token Gram similarity matrix that describes the complex spectral correlation between T vibration tokens. Without setting a learnable projection of the query matrix, key matrix, and value matrix, it uses the frequency domain attention matrix obtained from the Gram similarity matrix to perform weighted aggregation of complex frequency domain tensors in the token dimension.

[0034] The classification module is used to perform an inverse real-number fast Fourier transform on the weighted complex frequency domain tensor in an orthogonal normalization manner corresponding to the real-number fast Fourier transform, and to perform residual feedforward processing, token pooling and classification processing to output the rotating machinery fault category.

[0035] Beneficial effects:

[0036] This invention combines algorithmic features with techniques for acquiring vibration signals from rotating machinery, performing spectral analysis, and classifying faults. The input data corresponds to the one-dimensional vibration amplitude actually collected by the sensors, the output data corresponds to the fault state of the mechanical components, and the results are used for equipment condition monitoring and maintenance decisions.

[0037] Compared to conventional Q / K / V attention, this invention directly utilizes or By forming frequency domain attention, the parameter matrices required for Q, K, V, and attention output projection are eliminated. When d=512 and the conventional structure includes four d×d matrices for Q, K, V, and output, each attention block can avoid approximately 4d²=1,048,576 projection parameters, thereby reducing the number of model parameters and the risk of overfitting.

[0038] The complex conjugate Gram matrix utilizes both spectral amplitude and phase relationships to assign high correlation weights to fault frequency components that consistently appear in different vibration tokens, and suppresses broadband noise that only appears randomly in local blocks.

[0039] Orthogonal normalization transformation using RFFT / IRFFT is employed to maintain numerical stability and invertibility before and after the transformation; direct block partitioning avoids uncontrollable transformation of the original sampled information caused by additional embedding projection.

[0040] Unlike schemes that require a diffusion model to perform multi-stage noise addition, denoising, and enhanced feature fusion, this invention employs a deterministic single-pass forward diagnostic process. It eliminates the need to train an independent diffusion encoder / decoder or calculate the covariance embedding loss between generated and reference features, thereby reducing additional computational links during training and deployment. This effect stems from the diffusion-free generation, Q / K / V projection-free, and direct Gram-domain aggregation structure defined in this invention.

[0041] In additive Gaussian noise experiments using CWRU bearing data, the present invention outperformed the WDCNN baseline without target distribution adaptation at eight signal-to-noise ratio (SNR) points ranging from -4dB to 10dB. The average accuracy at these eight SNR points increased from 68.82% to 82.66%, an average improvement of 13.84 percentage points. This effect is limited by the data partitioning and training conditions of the embodiments described in this specification and does not constitute an unconditional guarantee for all equipment and operating conditions. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the overall process of the rotating machinery fault diagnosis method of the present invention.

[0043] Figure 2 This is a schematic diagram of the structure of the four-token direct frequency domain attention diagnostic network of the present invention.

[0044] Figure 3 This is a schematic diagram of the direct Gram frequency domain attention calculation without Q / K / V projection according to the present invention.

[0045] Figure 4 This is a comparison chart of the diagnostic accuracy of the present invention embodiment with the WDCNN baseline at different signal-to-noise ratios. Detailed Implementation

[0046] The present invention will be further described below with reference to embodiments, but the scope of protection of the present invention is not limited thereto:

[0047] Data Acquisition and Sample Construction

[0048] like Figure 1 As shown, an accelerometer is installed at the drive end of the bearing housing to be monitored, and vibration signals are collected at a sampling frequency of 12kHz. The sensor output is converted into a one-dimensional digital vibration sequence by a data acquisition card. In practical applications, the sensor can be replaced with a velocity sensor, displacement sensor, or acoustic emission sensor, and the sampling frequency is set according to the characteristic frequency of the equipment and the Nyquist condition.

[0049] The continuous vibration sequence was divided into vibration samples of length L=2048. Overlapping sliding windows were used to increase the number of training samples, while non-overlapping windows were used for test samples to reduce the risk of information leakage. Zero-mean normalization was performed on each window separately to avoid overall amplitude shifts caused by different sensor gains and loads.

[0050] This embodiment sets ten state categories: normal state; three types of rolling element faults with fault diameters of 0.007 inches, 0.014 inches, and 0.021 inches (approximately 0.178 mm, 0.356 mm, and 0.533 mm); three types of inner ring faults with fault diameters of 0.007 inches, 0.014 inches, and 0.021 inches (approximately 0.178 mm, 0.356 mm, and 0.533 mm); and three types of outer ring faults with fault diameters of 0.007 inches, 0.014 inches, and 0.021 inches (approximately 0.178 mm, 0.356 mm, and 0.533 mm). The number of categories and fault sizes are not limited by this invention, and the output dimensions of the classifier can be adjusted according to the actual equipment state set.

[0051] Direct frequency domain attention neural network

[0052] like Figure 2 As shown, the input tensor shape is [B, 1, 2048]. After removing the channel dimension, without going through the embedding layer, position encoding layer, or any learnable embedding projection, it is directly rearranged into [B, 4, 512], i.e., T=4 and d=512. The four tokens correspond to four consecutive 512-point intervals in the original vibration window, so no additional learnable linear mapping is performed on the original sampling points.

[0053] Each frequency domain attention block first performs layer normalization on the time domain token, and then calls RFFT using an orthogonal normalization convention to obtain a complex frequency domain tensor of [B, 4, 257]. Compared with manually constructing the Fourier matrix, pairwise RFFT / IRFFT operations using the same orthogonal normalization convention can maintain the numerical scale stability and invertibility before and after the transformation, and avoid information loss caused by zero columns or non-strict inverse matrices of the transpose matrix.

[0054] like Figure 3 As shown, the complex frequency domain tensor Stored as [B,257,4], and computed using batch matrix multiplication. and The product of the two tokens is used. Since the product is generally a complex number, its real part is taken as the similarity between tokens, and multiplied by a scaling factor to control the scale of the softmax input. The attention matrix has a shape of [B,4,4], where each row represents the aggregate weight of a target token on the spectral information of the four source tokens.

[0055] The frequency domain weighting result adopts the orthogonal normalization convention corresponding to RFFT, and is recovered as a real tensor of [B,4,512] through IRFFT, forming a residual connection with the frequency domain attention block input. The intermediate dimension of the feedforward network is set to 1024, the activation function is GELU, and the random inactivation rate is 0.1. Four frequency domain attention blocks are cascaded.

[0056] Finally, the average of the four tokens is calculated to obtain the global representation of [B, 512]. This representation is then passed through a 512×128 fully connected layer, a GELU activation layer, a random deactivation layer, and a 128×10 fully connected layer to output ten types of fault logits.

[0057] Training methods

[0058] Obtaining the training dataset involves: collecting or reading one-dimensional vibration records labeled with normal state and each fault state category, performing fixed-length slicing according to length L, and performing zero-mean standardization on each slice; dividing the training set, validation set, and test set into units based on the original continuous vibration records, so that adjacent windows generated from the same original vibration record are not distributed across datasets, in order to avoid information leakage caused by window overlap.

[0059] The fault diagnosis model consists of an input unit, a token construction unit, N cascaded frequency domain attention blocks, a token mean pooling unit, and a fully connected classifier. In this embodiment, the input shape is [B, 1, 2048], T=4, d=512, and N=4. Each frequency domain attention block sequentially performs layer normalization, orthogonal normalized RFFT, direct Gram frequency domain attention, the corresponding orthogonal normalized IRFFT, residual connection, and feedforward transform. The fully connected classifier sequentially includes a 512×128 fully connected layer, a GELU activation layer, a random deactivation layer, and a 128×10 fully connected layer.

[0060] The network is trained using cross-entropy loss, with AdamW as the optimizer and an initial learning rate of 3×10⁻⁶. -4 The weight decay coefficient is 1×10 -4 The batch size is 128, and the number of training epochs is 30. Cosine annealing is used to adjust the learning rate, and the gradient norm is clipped to 1.0. After each training epoch, the loss is calculated on the validation set, and the model parameters with the lowest validation loss are saved. Optionally, additive noise with random signal-to-noise ratio and / or random cyclic shifting are added to the training window, provided that only the parameters are generated using the training set.

[0061] In the CWRU implementation, the three load domains of 1 hp, 2 hp, and 3 hp contain a total of 19,800 training windows and 750 test windows; training is repeated using three random seeds of 0, 1, and 2, and the results are compared with the WDCNN baseline under the same data partitioning, test noise, and random seed conditions.

[0062] Noise reduction effect

[0063] Zero-mean Gaussian white noise is added to the standardized clean test windows to achieve a preset signal-to-noise ratio for each window; the noisy composite windows are then normalized again using the same window-by-window method as the clean samples. The WDCNN baseline uses the exact same training set, test set, noisy samples, and random seed as this invention.

[0064] like Figure 4 As shown in Table 1, across all test points from -4dB to 10dB, this invention outperforms the unadapted WDCNN. Specifically, it improves performance by 26.36 and 26.93 percentage points at 0dB and 2dB, respectively, indicating that the direct complex spectrum correlation weighting has a significant suppression effect on moderate-intensity noise.

[0065] On a noise-free test set, the accuracy of this invention is 99.60% ± 0.19%, while WDCNN achieves 99.78% ± 0.06%, with both showing similar performance in the same domain classification. Therefore, the main proven advantage of this invention is its diagnostic stability under noisy conditions, rather than its absolute accuracy under noise-free conditions.

[0066] Table 1 Comparison of test accuracy of the present invention and the WDCNN method at different test points.

[0067]

[0068] Deployment method

[0069] The trained model can be deployed on industrial computers, edge computing boxes, coprocessor units of programmable logic controllers, or cloud servers. During online diagnostics, the acquisition module continuously receives vibration signals. After caching 2048 points, it performs standardization, four-token rearrangement, frequency domain attention, and classification, and outputs the category number, category probability, and timestamp to the status monitoring platform.

[0070] When the maximum category probability is lower than a preset threshold, an unknown state can be output or manual review can be triggered; when multiple consecutive windows are judged as faulty, a maintenance alarm can be generated. The threshold judgment and continuous window rules are engineering applications of the diagnostic results and do not change the core calculation of the direct Gram frequency domain attention of this invention.

[0071] The above embodiments are for illustrative purposes only and are not intended to limit the present invention. The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them (e.g., changing the window length, number of tokens, feedforward dimension, number of frequency domain attention blocks, and number of fault categories), but without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for diagnosing rotating machinery faults based on direct Gram frequency domain attention, characterized in that, The method is based on a fault diagnosis model and specifically includes the following steps: S101. Acquire the one-dimensional vibration signal generated during the operation of rotating machinery; S102. Slice the one-dimensional vibration signal into fixed-length slices to obtain vibration samples of length L; S103. Without going through the word embedding layer, position encoding layer and learnable embedding projection, the vibration samples are directly rearranged into T vibration tokens according to the preset feature dimension d, where L=T×d; S104. Perform a real-number fast Fourier transform along the characteristic dimension of each vibration token in an orthogonal normalized manner to obtain the complex frequency domain tensor X. f ; S105. Construct a token Gram similarity matrix G to describe the complex spectral correlation among T vibration tokens. Scale the Gram similarity matrix G according to the number of independent frequency points F and perform softmax processing to obtain the frequency domain attention matrix A. Use the frequency domain attention matrix A to apply the complex frequency domain tensor X... f Weighted aggregation is performed on the token dimension to obtain the weighted complex frequency domain tensor Y. f Among them, in constructing the Gram similarity matrix G and the weighted complex frequency domain tensor Y... f During the process, learnable projections of the query matrix, key matrix, and value matrix are not set; S106. Using an orthogonal normalization method corresponding to the real number fast Fourier transform, perform an inverse real number fast Fourier transform on the weighted complex frequency domain tensor to restore the frequency domain features to the time domain features. After residual feedforward processing and classification processing, the fault category of the rotating machinery is output.

2. The method according to claim 1, characterized in that, The vibration samples are standardized to zero mean on a sample-by-sample basis. The standardized result is (x-μ) / (σ+ε), where x represents the vibration sample, μ is the mean of the vibration sample, σ is the standard deviation of the vibration sample, and ε is a positive number to prevent division by zero.

3. The method according to claim 1, characterized in that, The length L is 2048, the number of vibration tokens T is 4, the preset feature dimension d is 512, and the number of independent frequency points F output by the real fast Fourier transform is 257. Both the real fast Fourier transform and the inverse real fast Fourier transform adopt the orthogonal normalization convention.

4. The method according to claim 1, characterized in that, The complex frequency domain tensors are arranged with the frequency dimension first and the token dimension second. The Gram similarity matrix G is obtained according to the following formula: Among them, X f H X represents f The conjugate transpose of , Re(·) denotes taking the complex real part, F denotes the number of independent frequency points; the frequency domain attention matrix A=softmax(G), and the weighted complex frequency domain tensor Y f = A · X f .

5. The method according to claim 1, characterized in that, The complex frequency domain tensors are arranged with the token dimension first and the frequency point dimension second. The Gram similarity matrix G is obtained according to the following formula: Among them, X f H X represents f The conjugate transpose of , Re(·) denotes taking the complex real part, F denotes the number of independent frequency points; the frequency domain attention matrix A = softmax(G), and the weighted complex frequency domain tensor Y f = A · X f .

6. The method according to claim 1, characterized in that, The residual feedforward processing includes: summing the residuals and normalizing the layers of the time-domain features obtained by the inverse transform; inputting the normalized layer results into the first fully connected layer, the GELU activation layer, the random deactivation layer, and the second fully connected layer in sequence; and adding the feedforward output to the residual branch.

7. The method according to claim 6, characterized in that, A frequency domain attention block is formed by real-number fast Fourier transform, direct Gram frequency domain attention, inverse real-number fast Fourier transform, and residual feedforward processing. N frequency domain attention blocks are set in series, where N is an integer greater than or equal to 1. The number of frequency domain attention blocks N is independent of the number of vibration tokens T contained in each vibration sample.

8. The method according to claim 1, characterized in that, The classification process includes mean pooling of T vibration tokens, inputting the pooling result into at least one fully connected classification layer, and obtaining the class probability corresponding to at least one of the following states: normal state, rolling element fault, inner ring fault, and outer ring fault through softmax.

9. The method according to claim 1, characterized in that, The construction and training of the fault diagnosis model includes: acquiring vibration records of rotating machinery labeled with fault categories, slicing the vibration records into fixed-length slices and standardizing them sample by sample, and dividing the original vibration records into training, validation, and test sets; constructing a fault diagnosis model consisting of a token constructing unit, N cascaded frequency domain attention blocks, a token pooling unit, and a fully connected classification unit, wherein each frequency domain attention block performs residual feedforward processing, and the token pooling unit and the fully connected classification unit perform classification processing; setting the initial learning rate, weight decay coefficient, batch size, and number of training rounds, and using the cross-entropy loss function and gradient-based optimization algorithm to perform mini-batch iterative training on the fault diagnosis model, selecting model parameters based on the validation set loss; adding additive noise with random signal-to-noise ratio and / or performing random cyclic translation to the training vibration samples during training.

10. A rotating machinery fault diagnosis system based on direct Gram frequency domain attention, characterized in that, include: The signal acquisition module is used to acquire one-dimensional vibration signals of rotating machinery; The preprocessing module is used to standardize and slice one-dimensional vibration signals into fixed-length slices. The token construction module is used to directly rearrange vibration samples into T vibration tokens without setting a word embedding layer, a position encoding layer, and a learnable embedding projection. The frequency domain transformation module is used to perform a real-number fast Fourier transform on each vibration token in an orthogonal normalization manner to obtain a complex frequency domain tensor. The direct Gram frequency domain attention module is used to construct a token Gram similarity matrix that describes the complex spectral correlation between T vibration tokens. Without setting a learnable projection of the query matrix, key matrix, and value matrix, it uses the frequency domain attention matrix obtained from the Gram similarity matrix to perform weighted aggregation of complex frequency domain tensors in the token dimension. The classification module is used to perform an inverse real-number fast Fourier transform on the weighted complex frequency domain tensor in an orthogonal normalization manner corresponding to the real-number fast Fourier transform, and to perform residual feedforward processing, token pooling and classification processing to output the rotating machinery fault category.

Images