A rotating machinery fault diagnosis method based on hierarchical low-rank decoupling and manifold hybrid enhancement

Abstract

This invention discloses a fault diagnosis method for rotating machinery based on hierarchical low-rank decoupling and manifold hybrid enhancement. This invention addresses the problem of poor generalization robustness of models under extreme variable operating conditions due to feature coupling and manifold holes. The method first acquires the original signal and converts it into a time-frequency map; then, it inputs it into a deep network containing hierarchical low-rank normalization units. The shallow layer randomly decouples and amplifies texture, while the deep layer utilizes gradient inversion for adversarial decoupling, actively stripping away specific operating condition styles. Next, manifold interpolation is performed on similar features at the bottleneck layer to fill manifold holes in a physically valid manner and maintain momentum prototypes. Finally, a prototype-based cosine classifier is constructed to immunize amplitude fluctuations, and a sharpness-aware algorithm combined with multiple metrics is used for two-stage smoothing optimization. This invention effectively decouples redundant operating condition features and overcomes minimum value traps, significantly improving zero-shot diagnostic accuracy in unknown variable operating condition scenarios.

Description

Technical Field

[0001] This invention belongs to the field of industrial equipment health management and artificial intelligence technology, specifically relating to a method for high-precision fault diagnosis of rotating machinery (such as rolling bearings, gearboxes, rotor systems, etc.) using deep neural networks under varying operating conditions, strong background noise, and cross-equipment scenarios. Background Technology

[0002] Rotating machinery (such as rolling bearings, gearboxes, and motors) are core power components of modern industrial production lines and major equipment, and their operating status directly affects the safety and reliability of the entire system. In recent years, data-driven fault diagnosis methods based on deep learning, with their superior nonlinear feature mapping capabilities, can automatically extract high-dimensional features containing fault mechanisms and have gradually replaced traditional signal processing methods, becoming the mainstream technology in this field. However, the excellent performance of traditional deep diagnostic models is based on the idealized assumption that training and test data strictly satisfy independent and identically distributed (IID) distributions. In real industrial production environments, existing intelligent diagnostic technologies still face the following serious challenges:

[0003] 1. Domain Shift in Data Distribution and Manifold Holes Conflicting with Mechanisms: Industrial equipment typically operates under frequently changing variable conditions (such as variable speed and load), leading to significant distribution shifts in the vibration signal's amplitude and energy distribution in the time and frequency domains. Furthermore, since actual sampled operating conditions are often discrete, the data manifold exhibits discontinuity in the feature space, containing "holes," which weakens the model's ability to predict "intermediate" data between known operating conditions. Existing cross-class data augmentation methods (such as heterogeneous feature interpolation) severely violate the underlying physical laws of rotating machinery rotor dynamics, easily generating unrealistic "pseudo-feature semantics," which contaminates the original feature manifold space and causes classifiers to confuse unseen operating conditions.

[0004] 2. Feature Coupling and Lack of Targeted Defense Blind Spots under Complex Operating Conditions: The raw vibration signals collected from the field contain a large amount of background noise and domain-specific operating condition features that are highly correlated with specific loads and rotational speeds, leading to severe "feature entanglement" between pure fault features and operating condition features. Existing feature statistical perturbation mechanisms often rely on global blind random amplification, lacking optimization directionality and failing to actively discover and defend against the most destructive operating conditions for the current model; moreover, when applied to shallow networks, they easily destroy the mechanical underlying texture representing weak faults, resulting in the generalization performance lower bound not being effectively guaranteed.

[0005] 3. Classifier decision boundary failure due to drastic fluctuations in feature amplitude: Under varying operating conditions, sudden changes in load and rotational speed can cause drastic fluctuations in the absolute energy of vibration signals, which manifest as sharp changes in the feature vector magnitude in a high-dimensional feature space. Traditional linear classifiers often make decisions based on the inner product of feature vectors, and their decision mechanism is extremely sensitive to the feature magnitude. Simultaneously, in extreme scenarios, class centers calculated based on small batches of data are highly susceptible to interference from local anomalous samples, exhibiting high instability. The combined effect of these two factors leads to severe distortion or even collapse of the classifier's decision boundary during the testing phase.

[0006] 4. The Sharp Minimum Trap Caused by Empirical Risk Minimization: When faced with non-independent and identically distributed data with severe distributional shifts, conventional optimization strategies (such as SGD and Adam) are often limited to pursuing the current empirical risk minimization (ERM), lacking a global awareness of the topological geometry of the high-dimensional loss surface. This makes the model parameters extremely prone to falling into "sharp local minima" with extremely poor generalization performance. As a result, even a small distributional shift in the test data can cause a sharp increase in the loss function, leading to a precipitous performance degradation of the model under unknown target conditions.

[0007] To address the aforementioned issues, there is currently no systematic method in the technology that can simultaneously solve the problems of data manifold filling, anti-stripping of working condition features, resistance to amplitude fluctuations in decision boundaries, and smoothing of underlying optimization surfaces. Summary of the Invention

[0008] The purpose of this invention is to provide a rotating machinery fault diagnosis method based on anti-low-rank decoupling and manifold hybridization, aiming to solve the technical problems of low diagnostic accuracy and poor generalization robustness caused by extreme variable operating condition domain shift, manifold amplification due to mechanistic conflicts, drastic changes in signal amplitude, and model optimization overfitting in the prior art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for diagnosing rotating machinery faults based on counteracting low-rank decoupling and manifold mixing, comprising the following steps:

[0010] Step 1: Raw Signal Acquisition and Time-Frequency Conversion. Acquire the raw one-dimensional vibration acceleration signal of the rotating machinery continuously operating under multiple known source domain conditions; use an adaptive sliding window to extract signal segments, and utilize Short-Time Fourier Transform (STFT) to convert the one-dimensional time-domain signal into a two-dimensional time-frequency spectrum, which serves as the input data for the model.

[0011] Step 2: Feature extraction based on hierarchical low-rank decoupling. The two-dimensional time-frequency spectrum is input into a deep residual feature extraction network containing hierarchical low-rank normalization units. The hierarchical low-rank normalization units execute a heterogeneous decoupling strategy according to the network depth. That is, random low-rank decoupling is performed in the shallow layers of the network to increase the diversity of physical textures, and adversarial low-rank decoupling controlled by the gradient reversal mechanism is performed in the deep layers of the network. This actively removes the domain-specific style that is highly coupled with specific working conditions, thereby outputting a deep feature tensor with domain invariance.

[0012] Step 3: Feature Projection and Momentum Prototype Maintenance. The depth feature tensor is input into the projection bottleneck layer for dimensionality reduction and nonlinear mapping to obtain low-dimensional projected features. Simultaneously, a set of momentum prototype tensors corresponding to each fault category is established and maintained within the classifier. The momentum prototype tensors are updated using a moving average of the blocking gradient, based on the mean of the current batch of projected features.

[0013] Step 4: Manifold Hybrid Enhancement Based on Similarity. During the training phase, based on the class labels of the samples, a strictly constrained manifold interpolation operation is performed on different feature vectors belonging to the same fault category in the projected feature space to generate similar hybrid features, thereby legally filling the discrete manifold holes in the feature space.

[0014] Step 5: Multi-constraint loss calculation and sharpness-aware optimization. The projected features and similar mixed features are respectively input into a momentum prototype-based cosine classifier to calculate the predicted logical value, constructing an overall objective function that includes source domain classification loss, mixed flow consistency loss, local maximum mean difference alignment loss, center loss, and prototype contrast loss; the sharpness-aware gradient matching optimization algorithm (SAGM) is used to perform two-step gradient updates to sense and cross sharp regions on the loss surface.

[0015] Step Six: Online Equipment Diagnosis. Real-time vibration signals of the rotating machinery under unknown operating conditions are acquired. After preprocessing in Step S1, these signals are input into the trained diagnostic model. A cosine classifier outputs the final fault category based on the spherical cosine similarity between the pure features and the momentum prototypes of each fault.

[0016] The present invention has the following significant beneficial effects:

[0017] 1. A hierarchical low-rank feature stripping mechanism is proposed to address the challenges of domain-specific style entanglement in deep spatial domains and insufficient diversity of shallow physical features. This invention introduces heterogeneous perturbations at different levels of the backbone network, enhances the model's robustness to weak signal perturbations using shallow random mixing, and strips redundant load and speed styles using deep gradient inversion game theory. This hierarchical decoupling mechanism preserves key fault mechanism textures while improving the high-dimensional generalization performance of features, overcoming the technical bottleneck of blind spots in existing generalization methods.

[0018] 2. A manifold augmentation paradigm based on physical semantic consistency constraints was constructed, overcoming the disruption of mechanical dynamics mechanisms caused by traditional cross-class data augmentation. This invention completely abandons the mainstream cross-class feature mixing mechanism, strictly performing manifold interpolation based on similar samples in the feature dimensionality reduction bottleneck layer, and applying probability distribution consistency regularization constraints between the real and virtual flows. This innovation, while ensuring that the generated virtual features strictly follow the real physical evolution trajectory of rotating machinery, fills the manifold gaps between discrete working conditions with high purity, establishing a new standard for data augmentation that balances feature diversity and physical legitimacy.

[0019] 3. A robust metric topology resistant to amplitude fluctuations has been reconstructed, eliminating the risk of decision boundary failure caused by drastic changes in feature magnitude under extreme conditions. This invention forcibly projects high-dimensional features onto a unit hypersphere for cosine distance measurement, naturally immunizing against absolute energy fluctuations in vibration signals caused by actual load or sudden changes in rotational speed. Simultaneously, a momentum prototype mechanism that blocks gradients is used for smooth class center updates, providing a stable anchor point for model decisions that is unaffected by abnormal fluctuations in local small-batch data, fundamentally ensuring the stability of the target domain decision boundary.

[0020] 4. Achieving smooth optimization through Joint Multiple Metrics and Sharpness Awareness (SAGM) overcomes the sharp minimum trap under empirical risk minimization. This invention establishes a joint metric loss system encompassing macroscopic multi-source alignment (LMMD), microscopic cluster tightening (center and prototype loss), and manifold boundary regularization (hybrid consistency), completely eliminating negative transfer of out-of-class features. Simultaneously, the underlying SAGM algorithm is deeply integrated, matching the empirical risk gradient with the gradient at severe perturbations, forcing the diagnostic model to avoid high-curvature valleys in the parameter space and stably settle into a broad and smooth generalization basin. The synergistic effect of these two approaches endows the model with an exceptionally robust lower bound for diagnosis when facing extreme, invisible, and severe conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2 A schematic diagram of the original one-dimensional time-domain vibration signal waveform of rotating machinery;

[0023] Figure 3 As described in the embodiments of the present invention, by Figure 2 The two-dimensional time-frequency spectrum diagram obtained by converting the one-dimensional time-domain signal into a short-time Fourier transform (STFT) is shown below.

[0024] Figure 4 This is a schematic diagram of the overall network framework and core algorithm structure used in the embodiments of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0026] This invention proposes a fault diagnosis method for rotating machinery based on counteracting low-rank decoupling and manifold mixing, aiming to solve the problems of non-stationary vibration signal distribution and strong noise interference under varying operating conditions. Figure 1 As shown, this method achieves robust end-to-end fault diagnosis by constructing a deep network incorporating hierarchical low-rank decoupling (FLoR) and combining homogeneous manifold hybridization with the SAGM smoothing optimization strategy. The process is mathematically rigorously divided into the following six implementation steps:

[0027] Step 1: Time-frequency spatial mapping of non-stationary signals

[0028] Please see Figure 3 This step aims to establish a mapping from a one-dimensional time-domain physical space to a two-dimensional time-frequency characteristic space, thus eliminating the strong dependence of traditional diagnostics on prior rotational speed.

[0029] In the physical sampling phase, accelerometers are used to collect raw one-dimensional vibration signal sequences of rotating machinery under multiple source domain conditions. Construct a short-time Fourier transform (STFT) time-frequency mapping operator for a given sliding window function. (In this embodiment, a Hanning window is preferred, with a window length of 2048 and an overlap rate of 50%) and the number of frequency domain sampling points. The signal is at the The first time frame, the first Complex time-frequency response at each frequency point Calculated using the following formula:

[0030] (1)

[0031] In the formula, The sliding step size, The imaginary unit is used. Subsequently, the amplitude is extracted from the complex spectrum and a logarithmic scale is taken to amplify the weak high-frequency impact energy. Bilinear interpolation and Min-Max normalization are then performed to generate a standardized two-dimensional time-frequency feature tensor. (size is) ), which serves as input to deep networks.

[0032] Step 2: Constructing the feature extraction network and decoupling it from the hierarchical low-rank structure.

[0033] This step achieves feature extraction by constructing a deep residual backbone network and utilizes a heterogeneous low-rank normalization strategy to achieve active stripping of working condition styles.

[0034] 2.1 Extraction of statistical measures

[0035] Input Sample After mapping through a deep residual network (such as ResNet-18), the feature tensors of each layer of the network are... Calculate the batch normalized (BN) mean and variance, which include global operating condition statistics, and the instance normalized (IN) mean and variance, which exclude individual style:

[0036] (2)

[0037] (3)

[0038] And obtain normalized features and .

[0039] 2.2 Hierarchical Low-Rank Decoupling Normalization

[0040] Depending on the network depth of the feature map, this invention employs different decoupling strategies, specifically:

[0041] Shallow stochastic low-rank decoupling: In the shallow residual blocks of the network (such as layer1 and layer2), the mixed weights are randomly sampled from a pre-defined probability distribution. Perform random style perturbation fusion:

[0042] (4)

[0043] The aim is to increase the diversity of basic physical textures through randomization operations in the early stages of feature extraction, thereby improving the model's robustness to weak signal perturbations.

[0044] Deep adversarial low-rank decoupling: Adversarial FLoR modules are used in deep residual blocks of the network (such as layer 3 and layer 4) to compute deep features. After obtaining the dual-path normalized features, a learnable channel-level logarithmic parameter is defined. Dynamic blending weights are obtained after Sigmoid activation. Execution of adversarial optimization fusion:

[0045] (5)

[0046] 2.3 Gradient Reversal and Adversarial Optimization

[0047] During the backpropagation phase of training, The gradient inversion layer (GRL) is connected in series in the update path, and its gradient propagation formula is:

[0048] (6)

[0049] In the formula, To counteract the strength factor, this mechanism forces the optimizer to update in the direction that maximizes the classification error. It actively seeks out the "worst combination of statistics" that makes the current classifier most likely to confuse it, forcing the feature extraction network to completely detach itself from the domain style that is highly coupled with specific working conditions.

[0050] Step 3: Bottleneck Projection and Momentum Prototype Space Construction

[0051] 3.1 Feature Dimensionality Reduction Projection

[0052] Will After flattening, the input to the projection bottleneck layer (containing linear layers, BatchNorm1d, and Dropout) is mapped into a compact high-dimensional semantic feature vector. (In this embodiment) ).

[0053] 3.2 Momentum-based central prototype update

[0054] The classifier is built with a size of ( The momentum prototype matrix (where the number of categories is 1) This matrix blocks the backpropagation gradient. During each forward propagation, the true labels within the current batch are extracted. The feature subset. First, calculate the original mean vector of the features of this subset. :

[0055] (7)

[0056] Then the original mean vector is processed. Normalization yields the unit class center vector for the current batch. :

[0057] (8)

[0058] Next, the updated intermediate prototype vector is calculated using an exponential moving average strategy. :

[0059] (9)

[0060] In the formula, Let be the momentum coefficient (set to 0.9). This is the class prototype from the previous iteration. Finally, global normalization is performed on the intermediate prototype vector to obtain the current iteration's class prototype anchored on the unit hypersphere. :

[0061] (10)

[0062] Step 4: Hybridization of similar manifolds based on physical semantic consistency

[0063] To fill the gaps in the data manifold caused by discrete operating conditions, data augmentation with strict physical constraints is performed in the 256-dimensional bottleneck layer feature space.

[0064] Traverse the baseline features within the current batch (Corresponding tags) Forced to retrieve the true label from within the batch heterogeneous sibling samples Interpolation coefficients sampled from the Beta distribution Perform manifold linear interpolation:

[0065] (11)

[0066] This operation, while fully adhering to the dynamics mechanism of rotating machinery rotors, synthesizes virtual features of the same type of fault's minute evolution state, completely eliminating the physical "pseudo-semantics" caused by cross-class interpolation.

[0067] Step 5: Construction of Joint Loss Based on Multiple Metrics and Sharpness-Aware Smoothing Optimization

[0068] To address the shortcomings of traditional optimization algorithms in handling multi-source heterogeneous conditions, where gradients are prone to conflict and the coupling of different loss functions can easily lead to the destruction of physical semantics, this step completely decouples the "local basic diagnostic task" and the "global manifold alignment task" on the computational graph, proposing a two-stage, multi-metric joint optimization strategy. This process is strictly divided into the following six sub-steps:

[0069] 5.1 Cosine Measure and Logical Value Mapping for Resistance to Amplitude Fluctuations

[0070] To eliminate the interference of vibration signal absolute amplitude (modulus) fluctuations caused by drastic changes in operating conditions on the classification boundary, this invention abandons the traditional linear point integrator classifier and adopts a cosine metric mechanism with boundary margin.

[0071] Clean features (or mixed features) )conduct After normalization, they are compared with the various types of momentum prototypes on the unit hypersphere. Calculate cosine similarity :

[0072] (12)

[0073] To increase the safety distance between classes, the true label corresponding to the feature is... Subtract the additive boundary margin from the cosine value of the location. and multiply by the scaling factor. Calculate the logits of the classification with hard decision boundaries:

[0074] (13)

[0075] 5.2 Construction of Local Fundamental Diagnostic Functions and Separation of Source Domain Gradients

[0076] In the first stage of a single training iteration, in order to accurately capture the unique vulnerable features of each physical condition and prevent virtual manifold data and cross-domain alignment operations from interfering with the generation of real physical gradients, the system temporarily shuts down the manifold mixing module (Mixup) and the domain alignment module (LMMD).

[0077] Based on the domain labels to which the samples belong, the current batch of data is divided into... A subset of the source domain. Based on the calculation in step 5.1. The local basic diagnostic loss is computed independently for each subset of the source domain. :

[0078] (14)

[0079] In the formula, These are the currently trainable parameters of the network; For the first The clean feature cross-entropy loss of each source domain is used to provide the basic physical fault classification error; The center loss is used to constrain similar features within a single domain to cluster towards the momentum class center. The prototype contrast loss is used to push away heterogeneous momentum prototypes; To balance the weights.

[0080] For each source domain Perform backpropagation independently to obtain the set of parameter gradient vectors corresponding to each source domain. .

[0081] 5.3 Gradient Conflict Projection and Consensus Gradient Acquisition

[0082] To eliminate the mutual cancellation of optimization directions between different source domain conditions, the obtained gradient set is... Perform conflict-free consensus projection.

[0083] First, the gradient set is randomly shuffled to eliminate order bias. Then, the gradients from any two source domains are iterated over. and ( ), calculate the vector dot product of the two. If This indicates a conflict between the two operating conditions in the direction of parameter update, so the gradient will be... Project to On the normal plane, its projection correction formula is:

[0084] (15)

[0085] like Then, the original gradient remains unchanged. After traversing the projection, all conflict-free gradients are summed as vectors to obtain the conflict-free consensus gradient representing the direction of co-evolution of all source domains. :

[0086]

[0087] 5.4 Dynamic Sharpness Perception and Global Worst Perturbation Generation

[0088] To guide the model to converge to a flat basin while maintaining training efficiency, a neighborhood radius parameter that dynamically decays with the number of training steps is introduced. :

[0089] (16)

[0090] In the formula, The initial maximum disturbance radius is (preferably 0.05 in this embodiment). This represents the current iteration step. This represents the total number of iterations.

[0091] Based on the consensus gradient With dynamic radius In the current model parameters Within the neighborhood of , find the worst-case perturbation that causes the local diagnostic errors of each operating condition to increase the fastest in a coordinated manner. :

[0092] (17)

[0093] Subsequently, this perturbation is virtually applied to the model weights to obtain the parameter points against adverse conditions. In this state, the network parameters are in their "most vulnerable space," where they are highly susceptible to misjudgment under all operating conditions.

[0094] 5.5 Construction of Global Objective Function and Multi-Constraint Coordination

[0095] In the second phase, the preparation period for parameter meta-updating begins. This involves acquiring adversarial parameter points. Afterward, the system will re-input the complete batch data containing all multi-source domains into the network and fully activate the similar manifold mixing mechanism and multi-source distribution alignment module.

[0096] In this worst-case parameter space, the system forward propagates and computes the global objective function that includes all six physical and regularization constraints. :

[0097] (18)

[0098] In the formula, to These are the balance coefficients of each sub-loss function. At this point, the newly activated loss function plays a central role in the adversarial space:

[0099] The local maximum mean difference loss based on multi-kernel Gaussian kernel is used to force cross-domain edge distribution alignment under the worst feature distribution shift, thus completely eliminating global operating condition differences.

[0100] and These are the mixed-flow classification loss and the dual-flow consistency KL divergence loss, respectively. Their function is to force the prediction logic of the virtual interpolation features to be consistent with the real flow in the perturbation space, so as to legally repair and widen the decision manifold void in extreme environments.

[0101] 5.6 Generalized Basin Parameter Element Update

[0102] The global objective function calculated based on the above steps Perform backpropagation to compute the meta-gradient at the current adversarial parameter point, i.e. .

[0103] Ultimately, the system discards the virtual adversarial state and restores the model's original parameters. And by utilizing the meta-gradient obtained in the adversarial space, combined with the learning rate Perform actual parameter iteration updates:

[0104] (19)

[0105] In practical engineering implementation, this parameter update step smooths the meta-gradient by combining an adaptive variance correction algorithm. This two-stage joint optimization mechanism steadily drives the model parameters to converge to a broad and highly smooth generalization minimum basin by decoupling the two processes of "finding adversarial weaknesses" and "global regularization patching".

[0106] Step Six: Online Equipment Fault Diagnosis

[0107] After completing the multi-objective joint optimization training of the network, the online deployment phase begins. At this point, the model is switched to Evaluation Mode, all learnable parameters in the network are fixed, the feature manifold mixing module is removed, and the momentum prototype update is stopped.

[0108] Vibration signals of the device under unknown target conditions are acquired in real time by sensors and converted into a spectrum in step S1 before being input into the network. In the backbone feature extraction stage, the adversarial low-rank normalized unit (FLoR) no longer performs adversarial optimization but directly extracts the adversarial parameters converged during the training phase. After Sigmoid activation, fixed feature mixing weights are obtained. :

[0109] (20)

[0110] And the batch normalized features of the test samples are obtained using this fixed mixture weight. Instance-normalized features Perform stable weighted fusion:

[0111] (twenty one)

[0112] The fused features are passed sequentially through the backbone network and the bottleneck layer to extract pure 256-dimensional projection features. and carry out Normalization. The cosine classifier calculates the normalized feature and its relationship to the prototype of each frozen category. The cosine similarity between them. During diagnostic decision-making, the additive margin from the training phase is removed, retaining only the scaling factor. Directly calculate the evaluation logits:

[0113] (twenty two)

[0114] Finally, the category with the largest logical value (i.e., the closest spherical cosine distance) is selected as the diagnostic result output:

[0115] (twenty three)

[0116] The online diagnostic process fully preserves the domain-invariant decoupling weights learned during the training phase and makes decisions based on the hyperspherical orientation of the feature vector, completely eliminating energy drift across different operating conditions and achieving highly reliable zero-shot domain fault identification.

Claims

1. A method for fault diagnosis of rotating machinery based on hierarchical low-rank decoupling and manifold hybrid enhancement, characterized in that, Includes the following steps: Step 1: Raw Signal Acquisition and Time-Frequency Conversion. Acquire the raw one-dimensional vibration acceleration signals of the rotating machinery under multiple known source domain conditions, and convert the one-dimensional time-domain signal into a two-dimensional time-frequency spectrum using short-time Fourier transform. Step 2: Feature extraction based on hierarchical low-rank decoupling. The two-dimensional time-frequency spectrum is input into a deep residual feature extraction network containing hierarchical low-rank normalization units. Random low-rank decoupling is performed in the shallow layers of the network to amplify the diversity of physical textures, while adversarial low-rank decoupling controlled by a gradient inversion mechanism is performed in the deep layers to actively remove work-specific styles, outputting a deep feature tensor with domain invariance. Step 3: Feature Projection and Momentum Prototype Maintenance. The depth feature tensor is input into the projection bottleneck layer for dimensionality reduction mapping to obtain low-dimensional projected features. Simultaneously, a set of momentum class prototypes corresponding to each physical fault category is established and maintained within the classifier. The momentum class prototypes are updated using a moving average of the blocking gradients based on the low-dimensional projected features of the current batch. Step 4: Manifold Hybrid Enhancement Based on Similarity. During the training phase, based on the actual fault category labels of the samples, restricted manifold interpolation is performed on different feature vectors belonging to the same fault category in the low-dimensional projected feature space to generate similar hybrid features. Step 5: Multi-constraint Loss Calculation and Sharpness-Aware Optimization. The low-dimensional projected features and similar mixed features are input into a cosine classifier based on momentum class prototypes to calculate the predicted logical value, constructing a multi-constraint overall objective function; a sharpness-aware gradient matching optimization algorithm based on multi-source domain conflict projection is used to perform parameter optimization. Step Six: Online Equipment Diagnosis. Real-time vibration signals of the rotating machinery to be diagnosed under unknown target operating conditions are acquired, preprocessed, and input into the trained and converged diagnostic model. A cosine classifier, based on the pure projection features and the spherical cosine similarity between the momentum class prototypes in each frozen state, outputs the final equipment fault diagnosis category.

2. The method according to claim 1, characterized in that... The feature extraction process in step two includes a hierarchical low-rank decoupling mechanism, specifically comprising the following steps: Step 2 (a) Parallel extraction of statistics: extracting feature tensors from each layer of the network. Calculate the batch normalized statistic and the instance normalized statistic separately: （1）； （2）； Batch normalized features are obtained based on the aforementioned statistics. Instance-normalized features , Step 2 (II) Shallow Random Low-Rank Decoupling: Targeting the shallow feature tensors of the network Randomly sampled from a preset probability distribution to generate stochastic mixed weights that block the gradient. Execution style perturbation fusion: （3）； Step 2 (3), Deep Adversarial Low-Rank Decoupling: Targeting the deep feature tensor of the network Define learnable channel-level logarithmic parameters The dynamic blending weights are obtained after processing with the Sigmoid activation function. Calculate adversarial fusion features: （4）； Step 2 (iv), Gradient Inversion Style Stripping: During the backpropagation phase, the channel-level logarithmic parameters... The gradient reversal layer is connected in series in the update path, and its gradient propagation formula is: （5）； In the formula, The overall objective function is... For classifying losses, To counter the strength factor.

3. The method according to claim 1, characterized in that... The process of updating the momentum class prototype in step three includes the following steps: Step 3 (a) Extract the real labels within the current batch. Given a low-dimensional projected feature subset, calculate the original mean vector of the features of this subset. : （6）； Step 3 (II): Process the original mean vector. Normalization yields the unit class center vector. : （7）； Step 3 (iii): Using the exponential moving average strategy, calculate the updated intermediate prototype vector. : （8）； Step 3 (4): Perform a global operation on the intermediate prototype vector. Normalization yields the prototype of the current round class. : （9）； In the formula These are low-dimensional projection features within a subset. The number of samples in the subset. The momentum coefficient, This is the class prototype from the previous iteration.

4. The method according to claim 1, characterized in that... The manifold interpolation operation in step four includes the following steps: Step 4 (a): In the low-dimensional projection feature space, anchor any low-dimensional projection feature within the current batch as the benchmark evolution feature characterizing the basic operating state. And lock its corresponding fault category label. , Step 4 (II) Constructing labels for the aforementioned fault categories A strictly aligned local homogeneous feature search domain is used to force the extraction of homogeneous fluctuation features that differ from the baseline evolutionary features from the search domain. and from Manifold interpolation coefficients are generated by sampling from the distribution to simulate small transient states of the operating condition. , Step 4 (3): Based on the manifold interpolation coefficients, perform a restricted feature space affine combination on the baseline evolution feature and similar fluctuation features to generate similar hybrid features. : （10）。 5. The method according to claim 1, characterized in that... The cosine classifier in step five includes the following steps when calculating the predicted logistic value: Step 5 (a) Perform feature analysis After normalization, it is compared with the momentum prototypes of various categories. Calculate cosine similarity : （11）； Step 5 (II) Find the true label corresponding to the feature. Subtract the additive boundary margin from the cosine value of the location. and multiply by the scaling factor. Calculate the logarithmic odds : （12）。 6. The method according to claim 5, characterized in that... Step five, which involves constructing a multi-constraint overall objective function and performing separation and conflict projection of local basic diagnostic gradients, specifically includes the following steps: Step 5 (3): Temporarily disable the manifold blending module and the neighborhood alignment module, and divide the current batch of data into... A subset of source domains Step 5 (iv): Independently compute the local basic diagnostic loss, which includes source domain classification, central clustering, and prototype repulsion, for each source domain subset. : （13）； Step 5 (v): Perform backpropagation on the local basic diagnostic loss to obtain the gradient set. , Step 5 (6): Traverse any two source domain gradients, if the vector inner product... Then Project to On the normal plane, its projection correction formula is: （14）； Step 5 (7): Summate all the conflict-free gradients by vector summation to obtain the conflict-free consensus gradient. : （15）。 7. The method according to claim 6, characterized in that... Step five involves obtaining the conflict-free consensus gradient. Then, generating the global worst-case perturbation includes the following steps: Step 5 (8): Calculate the parameter neighborhood radius that dynamically decays with the number of training steps. : （16）； Step 5 (9): Based on the conflict-free consensus gradient, calculate the worst-case perturbation that causes the local diagnostic errors of each operating condition to increase the fastest in a coordinated manner. : （17）； Step 5 (10): Virtually apply the worst-case perturbation to the model weights to obtain the parameter points that withstand adverse conditions. : （18）。 8. The method according to claim 7, characterized in that The final update of the parameters in step five includes the following steps: Step 5 (XI): Obtain the parameter points under the adverse conditions described above. Afterwards, the similar manifold mixing mechanism and multi-source distribution alignment module were fully restored and activated. Step 5 (12): Calculate the overall objective function containing global multiple constraints. : （19）； Step 5 (13): Perform backpropagation and calculate the smoothed meta-gradient at the current adversarial parameter point. , Step 5 (XIV) Restore the original parameters of the model And utilize the smooth meta-gradient combined with the learning rate Perform real-time parameter downhill iterative updates: （20）。 9. The method according to claim 8, characterized in that... In the overall objective function: This is a loss function based on the local maximum mean difference of a multi-kernel Gaussian kernel. Cross-entropy loss for mixed stream classification, The KL divergence loss for consistent two-stream prediction distribution.

10. The method according to claim 1, characterized in that... The online device diagnostic process in step six includes the following steps: Step 6 (a): Freeze all learnable parameters of the network, remove manifold interpolation operations, and stop updating the momentum class prototype. Step 6 (II): The adversarial low-rank normalized unit directly extracts the adversarial parameters that converged during the training phase, obtaining fixed feature mixing weights. : （21）； Step 6 (3): Extracting Pure Projection Features Then normalize it and compare it with the momentum class prototypes in each frozen state. Calculate cosine similarity. Step 6 (iv): Cancel the additive boundary margin and calculate the online evaluation logic value. : （22）； Step 6 (5): Select the category with the largest online evaluation logic value as the final fault diagnosis result of the device and output it.

Images