Robust fault diagnosis method for rotating machinery based on physical information multi-scale network

Abstract

The application relates to a rotating machinery robust fault diagnosis method and system based on a physical information multi-scale network, a computer device and a storage medium. The method comprises the following steps: acquiring an original pure one-dimensional vibration signal and performing feature processing to obtain multi-scale fusion features; an initial round is based on the features to train a network to establish a benchmark loss; in subsequent rounds, the upper bound of interference is dynamically adjusted according to the loss reduction degree, and after the dynamic pollution signal is synthesized through zero boundary continuous sampling, the training is continued; after the training is completed, the features of the to-be-diagnosed signal are extracted, and then the classification output is obtained. Through deep coupling of physical priori and adaptive curriculum learning, the problem that early weak fault features are submerged under strong compound interference is solved, and high-precision robust diagnosis under a low signal-to-noise ratio condition is realized.

Description

Technical Field

[0001] This application relates to the field of mechanical fault diagnosis technology, and in particular to a robust fault diagnosis method, system, computer equipment and storage medium for rotating machinery based on a multi-scale network of physical information. Background Technology

[0002] With the rapid development of modern industrial systems towards larger scale, greater complexity, and automation, core components of rotating machinery, such as gearboxes and induction motors, operate under harsh conditions including high speed and heavy load for extended periods. Reliable condition monitoring and early fault diagnosis are crucial for preventing sudden equipment downtime, reducing significant economic losses, and avoiding safety accidents. In recent years, deep learning technology, with its powerful end-to-end feature extraction and pattern recognition capabilities, has been widely applied in the field of rotating machinery fault diagnosis. Deep models, particularly one-dimensional convolutional neural networks, have achieved near-perfect diagnostic accuracy on experimental datasets.

[0003] However, the operating environment of rotating machinery in real-world industrial applications is extremely complex. The vibration signals collected by sensors inevitably contain two types of severe composite interference: macroscopic low-frequency mechanical harmonics generated by mechanisms such as rotor imbalance and gear meshing, and environmental Gaussian white noise commonly found in industrial settings. The energy of these two types of interference is often far higher than the weak periodic impact pulses generated by early bearing failures, resulting in extremely low signal-to-interference-plus-noise ratio (SNR) of the collected signals, severely masking the fault pulses. Existing methods typically employ a simple serial splicing architecture between signal processing and deep learning. For example, variational mode decomposition (VMD) and other signal processing techniques are used as the front end to separate the signal frequency bands, and the separation results are then input into a deep learning model for classification. In this loosely coupled architecture, the signal processing module and the deep learning module are designed independently, lacking an effective information transmission and feedback mechanism. This prevents the signal processing process from adaptively adjusting to the needs of subsequent diagnostic tasks. Furthermore, the deep model, acting as a "black box," is highly susceptible to overfitting high-energy interference components, leading to a sharp degradation in diagnostic performance under low SNR conditions. Therefore, how to achieve deep coupling between physical signal priors and deep learning, so that the model can still accurately capture early fault pulse characteristics under extremely low signal-to-interference-plus-noise ratio conditions, is a technical problem that urgently needs to be solved in the field of rotating machinery fault diagnosis. Summary of the Invention

[0004] Therefore, it is necessary to provide a robust fault diagnosis method, system, computer equipment, and storage medium for rotating machinery based on a physical information multi-scale network, which can achieve high-precision and robust diagnosis of early weak faults in rotating machinery under strong composite interference conditions.

[0005] Firstly, this application provides a robust fault diagnosis method for rotating machinery based on a physical information multi-scale network. The method includes: The original pure one-dimensional vibration signal is obtained, and multi-scale fusion features are obtained after performing feature processing on the original pure one-dimensional vibration signal. The feature processing includes, in sequence: standardization, variational mode decomposition, calculation of root mean square value and kurtosis of each channel and concatenation of physical prior vectors, attention weighting, and multi-scale feature extraction. In the initial training rounds, a multi-scale network is trained based on the multi-scale fusion features to establish an initial baseline training loss. In each subsequent training round, a dynamic difficulty factor is calculated based on the decrease in the average training loss of the previous round relative to the initial baseline training loss. The upper bounds of the harmonic interference intensity and Gaussian noise intensity for the current round are adjusted according to the dynamic difficulty factor. In each training batch, the actual harmonic interference intensity and actual Gaussian noise intensity are continuously sampled from a range where the lower bound is zero and the upper bounds are the upper bounds of the harmonic interference intensity and the Gaussian noise intensity, respectively. Based on the actual harmonic interference intensity and the actual Gaussian noise intensity, a dynamic contamination signal is synthesized from the original pure one-dimensional vibration signal of the current batch. The multi-scale network is then trained based on the contamination multi-scale fusion features obtained after performing the feature processing on the dynamic contamination signal. After all training rounds are completed, the vibration signal to be diagnosed is subjected to the aforementioned feature processing, and then classified through the multi-scale network to output the fault diagnosis result.

[0006] In one embodiment, the calculation of the root mean square value and kurtosis of each channel includes: The root mean square value is calculated using the following formula: The kurtosis is calculated using the following formula: in, Let be the root mean square value of the k-th channel. Let be the kurtosis of the k-th channel. This represents the sampled value of the i-th sampling point in the k-th channel. Let i be the sampled value of the i-th sampling point of the k-th intrinsic mode function. For sequence length, Let be the mean of the k-th channel. To prevent division by zero of extremely small constants.

[0007] In one embodiment, calculating the dynamic difficulty factor based on the degree of decrease in the average training loss of the previous round relative to the initial baseline training loss includes: The dynamic difficulty factor is calculated using the following formula: in, Let be the dynamic difficulty factor for the t-th training round. Let be the average training loss for the (t-1)th training round. The initial benchmark training loss is given.

[0008] In one embodiment, adjusting the upper bound of the harmonic interference intensity and the upper bound of the Gaussian noise intensity for the current round according to the dynamic difficulty factor includes: The upper limit of the harmonic interference intensity is adjusted using the following formula: The upper bound of the Gaussian noise intensity is adjusted using the following formula: in, Let this be the upper bound of the harmonic interference intensity in the t-th training round. Let the upper bound of the Gaussian noise intensity in the t-th training round be , To preset the maximum harmonic interference tolerance limit, To preset the maximum Gaussian noise tolerance limit, Let be the dynamic difficulty factor for the t-th training round.

[0009] In one embodiment, the synthesis of dynamic contamination signals from the original pure one-dimensional vibration signals of the current batch includes: The dynamic pollution signal is synthesized using the following formula: in, As a dynamic pollution signal, The actual harmonic interference intensity is... The actual Gaussian noise intensity is... This is the original, pure, one-dimensional vibration signal of the current batch. It is the random fundamental frequency of low-frequency mechanical harmonics. For time vectors, For random phase, It is standard Gaussian white noise.

[0010] In one embodiment, performing the attention weighting includes: The physical prior vector is input into the bottleneck multilayer perceptron, and after being processed sequentially by the ReLU activation function and the Sigmoid activation function, the physical information attention weights are output; the formula for calculating the physical information attention weights is: in, Let be the physical prior vector. The Sigmoid activation function is... The ReLU activation function is... and is the learnable weight matrix of the bottleneck multilayer perceptron.

[0011] In one embodiment, the constrained variational problem of the variational mode decomposition is: The constraints are: in, For time variables in the continuous time domain, Let k be the intrinsic mode function. Let be the center frequency of the k-th intrinsic mode function. It follows the Dirac distribution. For convolution operators, The imaginary unit, For about The partial derivatives, The standardized one-dimensional vibration signal This represents the number of intrinsic mode functions.

[0012] Secondly, this application also provides a robust fault diagnosis system for rotating machinery based on a physical information multi-scale network. The system includes: The feature extraction module is used to acquire the original pure one-dimensional vibration signal, and to obtain multi-scale fused features after performing feature processing on the original pure one-dimensional vibration signal. The feature processing includes, in sequence: standardization, variational mode decomposition, calculation of root mean square value and kurtosis of each channel and concatenation of physical prior vectors, attention weighting, and multi-scale feature extraction. The initial training module is used to train a multi-scale network based on the multi-scale fusion features in the initial training rounds and establish an initial baseline training loss. The pollution scheduling training module is used to calculate a dynamic difficulty factor in each subsequent training round based on the decrease in the average training loss of the previous round relative to the initial baseline training loss, and to adjust the upper bound of the harmonic interference intensity and the upper bound of the Gaussian noise intensity for the current round according to the dynamic difficulty factor. In each training batch, the actual harmonic interference intensity and the actual Gaussian noise intensity are continuously sampled from a range of values ​​with a lower bound of zero and upper bounds of the upper bounds of the harmonic interference intensity and the Gaussian noise intensity, respectively. Based on the actual harmonic interference intensity and the actual Gaussian noise intensity, a dynamic pollution signal is synthesized from the original pure one-dimensional vibration signal of the current batch, and the multi-scale network is trained based on the pollution multi-scale fusion features obtained after performing the feature processing on the dynamic pollution signal. The fault diagnosis module is used to perform the feature processing on the vibration signal to be diagnosed after all training rounds are completed, classify it through the multi-scale network, and output the fault diagnosis result.

[0013] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the aforementioned robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information.

[0014] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the aforementioned robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information.

[0015] The aforementioned robust fault diagnosis method, system, computer equipment, and storage medium for rotating machinery based on a physical information multi-scale network firstly achieves robust fault diagnosis of rotating machinery by constructing a physical information multi-scale network. This is achieved by adaptively decoupling the original vibration signal into multiple intrinsic mode functions with clear physical frequency band meanings through variational mode decomposition, thus realizing physical frequency domain decoupling. On this basis, the kurtosis and root mean square value of each channel are jointly evaluated as physical prior descriptors. Physical information attention weights are generated through a bottleneck multilayer perceptron, and the multi-channel tensors are weighted element-wise. This allows the multi-scale network to autonomously highlight fault impact channels with high peak intensity and suppress high-energy, low-impulsivity harmonic interference channels, thereby solving the problem of pulse spike dilution and energy center bias in high-energy harmonic interference due to global average pooling in traditional attention mechanisms. Subsequently, multi-scale feature extraction is performed on the weighted features through a parallel multi-scale one-dimensional convolutional architecture to comprehensively capture multi-frequency component features such as wideband modulation envelope and high-frequency transient spikes. Regarding the training strategy for multi-scale networks, a baseline loss is established based on uncontaminated data in the initial training rounds. In subsequent training rounds, the upper bound of the injected harmonic interference and Gaussian noise intensity is dynamically adjusted according to the degree of decrease in the average loss of the previous round relative to the baseline loss. A continuous uniform sampling mechanism with a strictly zero lower bound is employed to inject actual interference into each batch. This allows the multi-scale network to smoothly transition from clean data to a low signal-to-interference-plus-noise ratio (SNR) composite interference environment, avoiding the catastrophic forgetting problem caused by stage switching in traditional discrete learning. Therefore, through the deep coupling of physical priors and adaptive learning, high-precision and robust diagnosis of early weak faults in rotating machinery under strong composite interference conditions is achieved. Attached Figure Description

[0016] Figure 1 This is an application environment diagram of a robust fault diagnosis method for rotating machinery based on a physical information multi-scale network in one embodiment. Figure 2 This is a flowchart illustrating a robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information in one embodiment. Figure 3 This is a block diagram of a robust fault diagnosis system for rotating machinery based on a physical information multi-scale network in one embodiment. Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Therefore, embodiments of this application provide a robust fault diagnosis method, system, computer device, and storage medium for rotating machinery based on a multi-scale network of physical information.

[0018] The robust fault diagnosis method for rotating machinery based on physical information multi-scale networks provided in this application can be applied to, for example... Figure 1 In the application environment shown, the data acquisition device 102 communicates with the server 104 via a network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated onto the server 104 or placed in the cloud or on another network server. The data acquisition device 102 is deployed on the rotating machinery side to acquire raw vibration signals and send them to the server 104. After acquiring the raw, clean one-dimensional vibration signal, the server 104 performs feature processing to obtain multi-scale fusion features. In the initial training rounds, a multi-scale network is trained based on the multi-scale fusion features to establish an initial benchmark training loss. In each subsequent training round, the upper bound of the interference is dynamically adjusted according to the decrease in the average loss of the previous round relative to the benchmark loss, and a dynamic contamination signal is synthesized before continuing to train the multi-scale network. After all training rounds are completed, the server 104 performs feature processing on the vibration signal to be diagnosed, classifies it through the multi-scale network, and outputs the fault diagnosis result. The data acquisition device 102 can be, but is not limited to, various industrial data acquisition terminals or edge computing devices, such as vibration sensors, data acquisition cards, programmable logic controllers, industrial IoT gateways, edge computing devices, or industrial control computers; the server 104 can be implemented using a standalone server or a server cluster composed of multiple servers.

[0019] Firstly, in rotating machinery fault diagnosis, vibration signals in real industrial scenarios are severely contaminated by multi-source composite interference consisting of macroscopic low-frequency mechanical harmonics and environmental Gaussian white noise, completely drowning out early weak fault pulses. Traditional diagnostic methods either rely on manual feature extraction, which is difficult to adapt to complex operating conditions, or use deep models to directly process the original signal, which is prone to overfitting high-energy interference components, leading to a sharp degradation in diagnostic performance under extremely low signal-to-interference-plus-noise ratio conditions. More critically, in existing methods, the signal processing module and the deep learning module adopt a loosely coupled serial splicing architecture, lacking effective information transmission between them, and the intermediate links do not form a closed-loop feedback, making it impossible for the signal processing process to adaptively adjust according to the needs of subsequent diagnostic tasks. Therefore, in one embodiment, such as Figure 2 As shown, a robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information is provided, which is then applied to... Figure 1Taking the data acquisition device 102 as an example, the following steps are included: Step S1: Obtain the original pure one-dimensional vibration signal, and perform feature processing on the original pure one-dimensional vibration signal to obtain multi-scale fusion features; the feature processing includes: standardization, variational mode decomposition, calculation of root mean square value and kurtosis of each channel and concatenation of physical prior vectors, attention weighting, and multi-scale feature extraction. Step S2: In the initial training rounds, train a multi-scale network based on multi-scale fusion features and establish an initial baseline training loss; Step S3: In each subsequent training round, calculate the dynamic difficulty factor based on the decrease in the average training loss of the previous round relative to the initial baseline training loss, and adjust the upper bound of the harmonic interference intensity and the upper bound of the Gaussian noise intensity for the current round according to the dynamic difficulty factor; in each training batch, continuously sample the actual harmonic interference intensity and the actual Gaussian noise intensity from the range of values ​​with a lower bound of zero and upper bounds of the upper bounds of the harmonic interference intensity and the Gaussian noise intensity, respectively; based on the actual harmonic interference intensity and the actual Gaussian noise intensity, synthesize the dynamic contamination signal from the original pure one-dimensional vibration signal of the current batch, and train the multi-scale network based on the contamination multi-scale fusion features obtained after feature processing of the dynamic contamination signal; Step S4: After all training rounds are completed, feature processing is performed on the vibration signal to be diagnosed, and then classification is performed through a multi-scale network to output the fault diagnosis result.

[0020] Specifically, acquiring the original pure one-dimensional vibration signal refers to the one-dimensional time series signal collected by a piezoelectric accelerometer deployed on the bearing housing or base of rotating machinery. This signal represents the change of vibration acceleration over time in the form of a voltage value sequence, with the sampling unit being mV or g (gravitational acceleration). The signal is temporarily stored in the data buffer before any synthetic interference is applied, serving as the basic data source for all subsequent processing and training operations.

[0021] Feature processing of a raw, pure one-dimensional vibration signal refers to transforming the original time-domain waveform into a feature representation that reflects fault identification information through a series of signal processing and feature extraction operations. Feature processing consists of five sub-steps, which are sequentially connected; that is, the output of one sub-step serves as the input of the next, as detailed below: The first sub-step is standardization, which aims to eliminate differences in signal dimensions and sampling rates under different acquisition devices or operating conditions, laying a unified data foundation for subsequent frequency domain decoupling.

[0022] The second sub-step is variational mode decomposition, which adaptively decouples the standardized one-dimensional time-domain signal into several intrinsic mode functions with sparse characteristics. Each intrinsic mode function is distributed in a narrow band around its own center frequency, thereby achieving explicit separation in the physical frequency domain.

[0023] The third sub-step involves calculating the root mean square (RMS) and kurtosis values ​​for each channel, and concatenating the physical prior vectors. The calculation of the RMS and kurtosis values ​​for each channel refers to calculating the RMS and kurtosis values ​​for each channel in the multi-channel tensor obtained from variational mode decomposition. The RMS value reflects the effective energy level of the signal in that channel, while the kurtosis reflects the impulse strength of the signal. Concatenating the physical prior vectors involves combining the RMS and kurtosis values ​​of each channel into a single physical prior vector of length 2K, following the order of "the first K elements are the RMS values ​​of each channel, and the last K elements are the kurtosis of each channel."

[0024] The fourth sub-step is attention weighting, which generates attention weights based on the physical prior vector and performs element-wise weighting on the multi-channel tensor, so that the fault impact channels with high peak intensity and high impulse intensity receive higher weights, while the high-energy but low-impulsivity harmonic interference channels are suppressed.

[0025] The fifth sub-step is multi-scale feature extraction, which uses convolution kernels of different scales to extract features from the weighted feature tensor, thereby simultaneously capturing multi-frequency component features such as low-frequency modulation envelope and high-frequency transient spikes in a wide frequency range, and finally outputting multi-scale fused features.

[0026] In the initial training epoch (denoted as epoch t=0), the multi-channel tensor obtained from variational mode decomposition is sequentially processed through root mean square and kurtosis calculations, physical prior vector concatenation, attention weighting, and multi-scale feature extraction to obtain multi-scale fused features, which are then input into the multi-scale network. The multi-scale network consists of sequentially connected fully connected layers and a Softmax classifier, outputting the probability distribution of fault categories. The difference between the predicted value and the true label is calculated using the cross-entropy loss function and used as the training loss for this epoch. This loss value is recorded as the initial baseline training loss. In the initial training epoch, no synthetic interference is applied to the input data; the multi-scale network is trained directly on clean data. Therefore, the initial baseline training loss reflects the optimal learning state baseline of the multi-scale network under interference-free conditions.

[0027] In each subsequent training epoch (denoted as epoch t≥1), a dynamic difficulty factor is first calculated based on the decrease in the average training loss of the previous epoch (i.e., epoch t-1) relative to the initial baseline training loss. This dynamic difficulty factor is a continuous scalar value ranging from [0,1], used to quantify the current learning progress of the multi-scale network. Subsequently, this dynamic difficulty factor is multiplied by a preset maximum harmonic interference tolerance limit and a preset maximum Gaussian noise tolerance limit, respectively, to obtain the upper bound of the harmonic interference intensity and the upper bound of the Gaussian noise intensity for the current epoch. These two upper bounds control the maximum intensity of the two types of interference that can be applied in the current epoch. In each training batch, the following operations are performed independently on each sample in the batch: the actual harmonic interference intensity of the sample is obtained by uniformly sampling from an interval with a lower limit of zero and an upper limit of the harmonic interference intensity; the actual Gaussian noise intensity of the sample is obtained by uniformly sampling from an interval with a lower limit of zero and an upper limit of the Gaussian noise intensity; and a dynamic contamination signal is synthesized from the original pure one-dimensional vibration signal of the sample using the above two actual interference intensities. This sampling mechanism ensures that a portion of the samples in each training batch always have interference intensity close to zero, thus maintaining a near-original pure state. The synthesized dynamic contaminated signals undergo the same feature processing as the pure signals, namely, normalization, variational mode decomposition, calculation of root mean square value and kurtosis for each channel, concatenation of physical prior vectors, attention weighting, and multi-scale feature extraction, to obtain contaminated multi-scale fusion features. These contaminated multi-scale fusion features are then input into the multi-scale network to complete the training for that round.

[0028] After all training rounds are completed, from the initial rounds to the preset maximum rounds (e.g., 200 rounds), the diagnostic phase begins. In the diagnostic phase, vibration signals to be diagnosed are acquired under actual operating conditions. These signals undergo the same feature processing as in the training phase (i.e., standardization, variational mode decomposition, calculation of root mean square and kurtosis values ​​for each channel, physical prior vector concatenation, attention weighting, and multi-scale feature extraction) to obtain the multi-scale fusion features to be diagnosed. These features are then input into the trained multi-scale network. The multi-scale network maps the features to a vector of equal length to the number of fault categories through fully connected layers. The vector is then converted into a probability distribution using a Softmax activation function. The category corresponding to the highest probability value is the output fault diagnosis result.

[0029] Based on the above, this embodiment injects physical prior knowledge into the network architecture through the aforementioned feature processing, dynamically adjusts the interference difficulty according to the learning progress of the multi-scale network through closed-loop adaptive interference scheduling training, and employs a continuous training mechanism that trains in parallel with clean and contaminated data. This makes the entire method constitute a complete end-to-end robust diagnostic link from the original signal input to the fault category output. Under extremely low signal-to-interference-plus-noise ratio (SNR) conditions, the physical prior-driven attention mechanism ensures that fault pulse features are highlighted while interference components are suppressed. Closed-loop adaptive training ensures that the multi-scale network can smoothly transition from clean data to an extremely low SNR environment. Continuous training with clean data ensures that the multi-scale network always retains its ability to distinguish clean fault features. The synergistic effect of these three factors achieves high-precision robust diagnosis of early weak faults in rotating machinery under strong composite interference.

[0030] In fault diagnosis, data acquisition devices deployed in different industrial scenarios often employ different sampling rate configurations. For example, some production lines use 48kHz, while others use 51.2kHz. Directly mixing these heterogeneous data sources can lead to inconsistent frequency band division benchmarks in subsequent variational mode decomposition, affecting the accuracy of physical frequency band decoupling. Therefore, in one embodiment, the standardization process involves using multiphase filtering and resampling to uniformly standardize the original one-dimensional vibration signals with different sampling rates to the target sampling rate.

[0031] In practice, polyphase filtering resampling is a sampling rate conversion technique based on a polyphase decomposition structure. Let the sampling rate of the original signal be... The target sampling rate is Then the ratio of the upsampling factor P to the downsampling factor Q satisfies Choose integers P and Q such that the above ratio is equal to the ratio of the target sampling rate to the original sampling rate. For example, when the original sampling rate is 48kHz and the target sampling rate is 24kHz, P=1 and Q=2; when the original sampling rate is 32kHz and the target sampling rate is 24kHz, P=3 and Q=4.

[0032] The basic principle of polyphase filtering resampling is as follows: First, the original signal is upsampled by a factor of P (i.e., P-1 zero points are inserted between every two adjacent sampling points of the original signal). Then, an anti-aliasing filter is applied using a low-pass finite impulse response filter to suppress the mirror spectrum generated by upsampling. Finally, a factor of Q is downsampled (i.e., one sampling point is extracted every Q sampling points) to obtain the reconstructed signal at the target sampling rate. To reduce computational complexity, the above three operations are combined using polyphase decomposition technology: the unit impulse response coefficient of the filter is divided into P groups according to phase, with each group corresponding to a sub-filter for one input sampling point. Each output sampling point only needs to be calculated by performing a weighted summation of the input signal once, thereby reducing the computational complexity from... Reduce to Where L is the length of the input signal. The target sampling rate can be preset (e.g., 24kHz) according to the key fault characteristic frequency range and signal bandwidth requirements of the rotating machinery in the actual application scenario. The Nyquist frequency corresponding to this sampling rate is 12kHz, which is sufficient to cover the high-frequency resonant band of common rotating machinery faults (such as bearing faults) (usually 2kHz to 10kHz), while avoiding the computational redundancy and storage overhead caused by using a higher sampling rate. In practical applications, other sampling rate conversion methods can also be used for multiphase filtering resampling, such as the resampling method based on Lagrange interpolation. This embodiment does not limit this method.

[0033] Based on the above, this embodiment uses multiphase filtering resampling for standardization processing, so that the original signals with different sampling rates have a unified sampling rate and time scale before entering the subsequent variational mode decomposition, ensuring the consistency of physical frequency band division; at the same time, the multiphase decomposition structure reduces the computational complexity of resampling and improves the real-time processing efficiency of the entire diagnostic method.

[0034] When applying variational mode decomposition, the number of decomposition modes directly determines the granularity and accuracy of physical frequency band division. If the number of decomposition modes is too large, the fault characteristics of the same physical frequency band will be excessively split into multiple adjacent modes, weakening the concentrated expression of the fault pulse. If the number of decomposition modes is too small, vibrational components from different physical sources will be forcibly mixed in the same mode, undermining the purpose of physical frequency domain decoupling. Therefore, in one embodiment, the number of decomposition modes in variational mode decomposition is configured to be 4, such that the first intrinsic mode function after decomposition corresponds to low-frequency macroscopic mechanical harmonics, the second intrinsic mode function corresponds to mid-frequency structural resonance, the third intrinsic mode function corresponds to the high-frequency fault pulse resonance band, and the fourth intrinsic mode function corresponds to ultra-high frequency background Gaussian white noise.

[0035] Specifically, based on the physical mechanism of vibration generation in rotating machinery, the number of decomposition modes in variational mode decomposition is fixed at 4. Under this configuration, variational mode decomposition will convert the standardized one-dimensional vibration signal... It is decomposed into 4 discrete intrinsic mode functions. , , , Each intrinsic mode function closely revolves around its respective center frequency. , , , It exhibits a narrow band distribution and satisfies .

[0036] The physical frequency band allocation for the four intrinsic mode functions is as follows: First intrinsic mode function The covered frequency range is mainly below 100Hz, corresponding to low-frequency periodic vibration components such as rotor rotation frequency and its harmonics, gear meshing frequency, etc., i.e., macroscopic mechanical harmonics; the second inherent mode function The covered frequency range is mainly from several hundred hertz to about 1 kHz, corresponding to the mid-frequency natural vibration response of rotating mechanical structural components, i.e., structural resonance; the third natural mode function The covered frequency range is mainly between 2kHz and 10kHz, corresponding to the high-frequency resonant response excited by the periodic impacts caused by early faults such as local bearing damage and tooth surface spalling. This frequency band is the key information carrier for fault diagnosis, namely the high-frequency fault pulse resonance band; the fourth intrinsic mode function The covered frequency range is mainly above 10kHz, corresponding to broadband random noise components in the environment, i.e., ultra-high frequency background Gaussian white noise. The above frequency ranges are typical values. In actual decomposition, the actual center frequencies of each intrinsic mode function are adaptively determined by the variational mode decomposition algorithm based on the input signal and are not limited by the above typical values.

[0037] Based on the above, this embodiment configures the number of decomposition modes to 4, so that the number of output channels of variational mode decomposition matches the physical frequency band structure of the rotating machinery vibration signal. The originally uninterpretable original vibration signal is mapped into four band-limited channels with clear physical meaning, providing a physically interpretable multi-channel tensor basis for the subsequent calculation of channel-level physical prior statistics and the generation of attention weights.

[0038] Variational mode decomposition (VMD), as an adaptive signal decomposition technique, aims to minimize the bandwidth of each mode by solving a constrained variational problem. However, directly solving this constrained variational problem is difficult, requiring the augmented Lagrangian method to transform it into an unconstrained optimization problem and then iteratively solving it using the alternating direction multiplier method. Therefore, in one embodiment, the constrained variational problem of VMD is: The constraints are: in, For continuous time domain, the unit is seconds; Let be the k-th intrinsic mode function, which is related to time. The continuous function represents the k-th band-limited signal component obtained from the decomposition, and its value range is the set of real numbers; Let be the center frequency of the k-th intrinsic mode function, in radians per second, and be the optimization variable to be solved. The Dirac distribution is defined as follows: (when (time) and ; This is the convolution operator, representing the convolution operation between two functions; The imaginary unit satisfies ; For about The first-order partial derivative operator indicates that the function within the subsequent parentheses has respect to... Find the first-order partial derivative; The standardized one-dimensional vibration signal is related to time. Continuous functions; The number of intrinsic mode functions is a positive integer, and in this embodiment, it is set to 4.

[0039] The physical meaning of the above constrained variational problem is: to ensure that the sum of all intrinsic mode functions can accurately reconstruct the original signal. Find K intrinsic mode functions under the constraints. and its corresponding center frequency This ensures that each intrinsic mode function, after undergoing the Hilbert transform, yields an analytic signal, i.e. ,in and The convolution is The Hilbert transform, then multiplied by a complex exponent The spectrum of each mode is shifted to the baseband, and the square of the L2 norm of the gradient is taken as the estimate of the bandwidth of that mode. Minimizing the sum of the estimated bandwidths of each mode ensures the compactness of each mode in the frequency domain, that is, each mode is concentrated as close as possible to the center frequency.

[0040] To solve the aforementioned constrained variational problem, an augmented Lagrange function is constructed by introducing a quadratic penalty factor (a positive real number) and Lagrange multipliers. The alternating direction multiplier method is then used to iteratively update each intrinsic mode function, each center frequency, and the Lagrange multipliers in the frequency domain until convergence. The quadratic penalty factor is used to ensure reconstruction fidelity in the presence of noise, while the Lagrange multipliers are used to strictly enforce the reconstruction constraints. The convergence criterion for the alternating direction multiplier method is that the sum of the changes in the center frequencies of each mode is less than a preset convergence threshold (e.g., 10). −6 The maximum number of iterations can be reached (e.g., 500). The numerical update method for the quadratic penalty factor and Lagrange multipliers can be achieved using the standard alternating direction multiplier method in variational mode decomposition algorithms; this embodiment does not limit this approach.

[0041] Based on the above, this embodiment applies a non-recursive variational mode decomposition algorithm to the physical frequency band adaptive separation of vibration signals by defining the constrained variational problem and solving it using the alternating direction multiplier method. Each mode maintains compact band-limited characteristics in the frequency domain, ensuring the robustness of the decomposition and the accuracy of the reconstruction.

[0042] Traditional attention mechanisms (such as compressed excitation networks) rely on global average pooling to extract global contextual features. This operation averages over time, which severely dilutes the highly localized transient spikes of early fault impacts. This results in channels containing sparse fault impacts exhibiting low average energy despite high impulsivity, while channels containing continuous pure harmonic interference receive the highest pooling energy, leading to an energy center bias. Therefore, in one embodiment, the root mean square value and kurtosis of each channel are calculated, including: The root mean square value is calculated using the following formula: Kurtosis is calculated using the following formula: in, is the root mean square value of the k-th channel; Let be the kurtosis of the k-th channel; This represents the sampled value of the i-th sampling point in the k-th channel. This represents the sampled value at the i-th sampling point of the k-th intrinsic mode function; The sequence length is the total number of sampling points per channel, and it is a positive integer. This is the mean of the k-th channel; To prevent division by zero of extremely small constants (e.g., taking 10). −8 The subscript i is the index of the sampling point, and its value range is... , where is a positive integer.

[0043] Specifically, the tensor formed by the K intrinsic mode functions obtained from variational mode decomposition... Perform the following calculations independently for each channel: For the k-th channel, extract a one-dimensional sequence of length L. First, calculate the mean of the sequence. The mean is a scalar, having the same dimensions as the sampled values. Then, the root mean square value of this channel is calculated using the root mean square formula. Calculate the sum of squares of the amplitudes at each sampling point, divide by the sequence length, and add a minimal constant. To prevent the square root from being zero in subsequent calculations, the square root is taken again to obtain the root mean square (RMS) value. The RMS value has the same dimensions as the sampled value and reflects the effective energy level of the channel signal (the larger the signal amplitude or the longer the duration, the larger the RMS value). Simultaneously, the kurtosis of the channel is calculated using the kurtosis formula. Calculate the sum of the fourth powers of the deviations from the mean at each sampling point, divided by the average sequence length, as the numerator; calculate the sum of the squares of the deviations from the mean at each sampling point, divided by the sequence length, and add a minimal constant. The result is used as the base of the denominator, and the square of this base is used as the denominator; the numerator is divided by the denominator to obtain the kurtosis value. Kurtosis is a dimensionless statistic that is highly sensitive to transient impacts (the more prominent the peaks and the thicker the tails of the signal, the larger the kurtosis value); the kurtosis of a pure harmonic signal is about 1.5, the kurtosis of Gaussian white noise is about 3, while the kurtosis of a signal containing fault impacts is usually much greater than 3.

[0044] Based on the above, this embodiment constructs a complete descriptor describing the two-dimensional physical characteristics of each channel's "energy-pulsivity" by jointly calculating the root mean square (RMS) value and kurtosis of each channel. The RMS value reflects the energy level of the channel, while the kurtosis reflects the strength of the channel's pulsivity. The two are complementary: the high-energy but low-pulsivity harmonic interference channel has a high RMS value but a low kurtosis (close to 1.5), while the high-energy but high-pulsivity fault pulse channel has a low RMS value but a high kurtosis (much greater than 3). The combination of the two achieves complete differentiation between fault channels and interference channels, fundamentally correcting the centering bias caused by traditional global average pooling relying solely on the single dimension of "energy".

[0045] After obtaining the physical prior vector, which is composed of the root mean square value and kurtosis of each channel, it needs to be transformed into attention weights that can be used to weight the multi-channel tensor. If the physical prior vector is directly used as the weights, it lacks learnable adaptive capabilities and cannot be optimized according to the specific diagnostic task. Therefore, in one embodiment, attention weighting is performed, including: The physical prior vector is input into the bottleneck multilayer perceptron, and after being processed by the ReLU activation function and the Sigmoid activation function in sequence, the physical information attention weights are output; the formula for calculating the physical information attention weights is: in, For physical prior vectors, It is the Sigmoid activation function. It is the ReLU activation function. and This is the learnable weight matrix for the bottleneck multilayer perceptron.

[0046] Specifically, physical prior vectors The dimension is 2K, and its construction method is as follows: the first K elements are the root mean square values ​​of each channel. The last K elements are the kurtosis values ​​of each channel, in order. ,Right now superscript This indicates transpose.

[0047] The bottleneck multilayer perceptron consists of a dimensionality reduction layer and a dimensionality increase layer. The dimensionality reduction layer uses a learnable weight matrix. input vector A linear mapping from dimension 2K to dimension ,Right now The reduction ratio A pre-defined positive integer, for example, 8, indicates that the dimension of the intermediate layer is equal to the input dimension. This creates an information bottleneck structure, forcing multi-scale networks to learn the most critical feature combinations from physical priors. The output of the dimensionality reduction layer is then activated by the ReLU function. Perform nonlinear transformation: This means performing an operation independently on each element of the vector, setting negative values ​​to zero and leaving positive values ​​unchanged. The dimensionality-upgrading layer uses a learnable weight matrix. From the dimension A linear mapping back to K is performed. Finally, the Sigmoid activation function is applied. Element-wise compression is performed on the output of the dimension-upgrading layer: The value of each element is compressed to the (0,1) interval to form the final physical information attention weights. ,in This represents the weight value of the k-th channel. Learnable weight matrix. and The initial values ​​of the elements are initialized using a uniform distribution of Xavier, and updated during the training of the multi-scale network through backpropagation algorithm and gradient descent optimizer (such as Adam optimizer), so that the generation of attention weights can be adaptively adjusted according to the loss function of the classification task.

[0048] After obtaining the physical information attention weights, they are multiplied element-wise with the original multichannel tensor: for all L sampling points of the k-th channel, they are multiplied by the corresponding weight value. The purification feature tensor is obtained. The i-th element of its k-th channel is This soft mask operation makes the peak channel (i.e., the weight) Channels close to 1 are preserved and enhanced, while low-kurtosis, high-energy channels (i.e., weighted channels) are preserved and enhanced. Channels close to 0 were effectively suppressed.

[0049] After obtaining the cleanup feature tensor, feature extraction needs to be performed on this tensor to capture fault discrimination information. If a single-scale convolution kernel is used, it is difficult to simultaneously capture two types of fault features with very different characteristics: wideband low-frequency modulation envelope and sharp high-frequency transient spikes. Therefore, in one embodiment, multi-scale feature extraction adopts a parallel three-way one-dimensional convolution architecture, with one-dimensional convolution kernel sizes of 3, 7, and 15 for the three branches, respectively; each branch performs a one-dimensional convolution operation on the cleanup feature tensor and outputs a feature map of the corresponding scale after nonlinear mapping by the ReLU activation function; the feature maps output by the three branches are concatenated along the channel dimension to obtain multi-scale fused features.

[0050] Specifically, three parallel convolutional branches are set up within each convolutional block, and each branch independently processes the cleaned feature tensor of the input. Perform a one-dimensional convolution operation, where K is the number of input channels (i.e., the number of intrinsic mode functions, fixed at 4), and L is the time series length. The kernel sizes for the three branches are 3, 7, and 15, respectively, and the number of kernels in each branch is [missing information]. (For example, set to 64), meaning each branch outputs... Each feature map.

[0051] The branch with a kernel size of 3: The convolution kernel has a time window length of 3 sampling points, corresponding to a time span of approximately 0.125 milliseconds at a sampling rate of 24 kHz. This branch has the smallest receptive field, is sensitive to local detail changes in the signal, and excels at extracting fine time-domain waveform features of high-frequency transient impulses, such as the rising edge and peak shape of fault pulses. The branch with a kernel size of 7: The convolution kernel has a time window length of 7 sampling points, corresponding to a time span of approximately 0.292 milliseconds. It has a moderate receptive field and is suitable for extracting medium-scale time-domain waveform patterns, such as the envelope shape of pulse sequences. The branch with a kernel size of 15: The convolution kernel has a time window length of 15 sampling points, corresponding to a time span of approximately 0.625 milliseconds. It has the largest receptive field, is sensitive to the global trend of the signal, and excels at extracting low-frequency modulation envelope features over a wide frequency range, such as the overall envelope of periodic impulse sequences generated by faults. The above time span calculation is based on a 24kHz sampling rate (1 / 24000 seconds × kernel size). For other standardized target sampling rates, the corresponding time span is converted in the same way.

[0052] The one-dimensional convolution operation in each branch can adopt any existing one-dimensional convolution implementation method. By sliding the convolution kernel along the time dimension, the input values ​​within each local window are weighted, summed, and a bias term is added to obtain the output feature map. The initial weights of the convolution kernels in the three branches are initialized using a uniform He distribution, and the initial value of the bias term is set to zero. It is updated through backpropagation during the multi-scale network training process.

[0053] The convolutional output of each branch undergoes a non-linear mapping using the ReLU activation function, that is, the ReLU activation function is applied to each element y of the output feature map. By setting negative values ​​to zero and keeping positive values ​​unchanged, non-linear expressive power is introduced.

[0054] Finally, the feature maps output from the three branches are concatenated along the channel dimension: Let the output feature maps of the three branches be respectively... , , , The sequence length of the output feature map in the time dimension after a one-dimensional convolution operation is obtained by concatenating it along the channel dimension. Its channel count is the sum of the output channel counts of the three branches ( The time dimension remains unchanged. This concatenation operation allows subsequent layers to access feature information extracted at three different scales simultaneously. The above multi-scale feature extraction process can be repeated multiple times in a multi-scale network (e.g., stacked 2 or 3 times), with downsampling performed after each stack using a pooling layer or convolutional layer with a stride of 2 to gradually deepen the feature abstraction level. The connection relationships between layers can adopt any inter-layer connection method in existing convolutional neural networks; this embodiment does not limit this.

[0055] Based on the above, this embodiment uses a parallel three-way one-dimensional convolutional architecture, where each of the three branches extracts local features at different receptive field scales. The fused features after splicing contain detailed information of high-frequency transient impacts, waveform pattern information at medium scales, and global trend information of low-frequency modulation envelopes, comprehensively covering the different frequency component features of rotating machinery faults.

[0056] Traditional learning methods employ discrete, hard-coded phase divisions (e.g., no noise for the first 50 rounds, a small amount of noise for the middle 50 rounds, and a large amount of noise for the last 50 rounds). This open-loop paradigm requires manually pre-setting phase switching points and the interference intensity for each phase, introducing cumbersome hyperparameters and causing peaks in training loss during phase switching, triggering catastrophic forgetting. Therefore, in one embodiment, a dynamic difficulty factor is calculated based on the degree of decrease in the average training loss of the previous round relative to the initial baseline training loss, including: The dynamic difficulty factor is calculated using the following formula: in, is the dynamic difficulty factor for the t-th training round, which is a continuous scalar with a value range of [0,1]; t is the training round index, which is a non-negative integer. When t=0, it represents the initial training round, and when t≥1, it represents the subsequent training rounds. Let be the average training loss in the (t-1)th training round. It is a scalar, and its unit is the dimension of the loss function (cross-entropy loss is dimensionless). The initial baseline training loss is a scalar, and its unit is the dimension of the loss function (cross-entropy loss is dimensionless). This indicates taking the larger value between the number inside the parentheses and 0, to ensure that the dynamic difficulty factor is non-negative.

[0057] Specifically, in the initial round (denoted as round t=0), the model completes a full round of training on completely uncontaminated, clean data, that is, iterates through all training samples once and calculates the average loss of each batch as the initial baseline training loss. This value is then fixed and saved as a constant reference for all subsequent rounds. Starting from the first round, before the start of each training round, the average loss value of the previous round over the entire training set is read. .

[0058] Dynamic difficulty factor The calculation logic is as follows: ratio This reflects the proportion of the previous round's loss relative to the initial baseline, when When the ratio is close to 1, it indicates that the multi-scale network has not yet made any learning progress, and the dynamic difficulty factor is 0, meaning that the current round should maintain the lowest possible interference difficulty; when When, the ratio A dynamic difficulty factor of 0.5 indicates that the multi-scale network has achieved a 50% loss reduction, and the current difficulty can be increased to an intermediate level; when When the ratio approaches 0, the dynamic difficulty factor approaches 1, indicating that the multi-scale network has learned sufficiently and the current difficulty can be increased to the highest level. The function ensures that in extreme cases (such as...) (That is, when the model performance degrades compared to the initial baseline) the dynamic difficulty factor is set to 0 instead of negative, to prevent the interference intensity from becoming negative.

[0059] Based on the above, this embodiment continuously calculates the dynamic difficulty factor by using the actual training loss from the previous round as a feedback signal. This allows the adjustment of training difficulty to be driven entirely by the model's own learning state in real time, without the need for manual pre-setting of stage divisions and switching points. In this closed-loop feedback mechanism, the loss serves as a negative feedback signal (lower loss indicates better learning and higher difficulty), achieving automatic matching between training difficulty and network learning ability.

[0060] After obtaining the dynamic difficulty factor, it needs to be converted into specific intensity upper bounds for harmonic interference and Gaussian noise, respectively, to control the injection levels of the two different types of interference. The two types of interference have different physical sources (harmonic interference originates from deterministic mechanical vibrations such as rotor imbalance and gear meshing, while Gaussian noise originates from random environmental noise in industrial settings), and their energy levels and ways of affecting diagnosis differ, requiring the setting of independent tolerance limits for each. Therefore, in one embodiment, the upper bounds of the harmonic interference intensity and the Gaussian noise intensity are adjusted according to the dynamic difficulty factor for the current round, including: The upper limit of harmonic interference intensity can be adjusted using the following formula: The upper bound of the Gaussian noise intensity is adjusted using the following formula: in, Let be the upper bound of the harmonic interference intensity in the t-th training round. is a scalar, a dimensionless amplitude coefficient, representing the maximum amplitude of the harmonic interference allowed to be applied in the current round as a multiple of the original signal amplitude. is the upper bound of the Gaussian noise intensity in the t-th training round, and is a scalar with a dimensionless amplitude coefficient, representing the maximum standard deviation of Gaussian white noise allowed to be applied in the current round as a multiple of the original signal amplitude; The maximum harmonic interference tolerance limit is a scalar, dimensionless amplitude coefficient. For example, a value of 3.0 means that the maximum allowable harmonic interference amplitude is 3.0 times the original signal amplitude. The maximum tolerance limit for Gaussian noise is preset. It is a scalar, and the unit is a dimensionless amplitude coefficient. For example, a value of 1.5 means that the maximum allowable standard deviation of Gaussian white noise is 1.5 times the amplitude of the original signal. Let be the dynamic difficulty factor for the t-th training round.

[0061] Specifically, a maximum harmonic interference tolerance limit is preset. and preset maximum Gaussian noise tolerance limit These are two independently configurable hyperparameters, pre-set based on the maximum interference level that may be encountered in the target industrial scenario before actual application. The degree of harmonic interference... A value of 3.0 indicates that under the strongest interference, the amplitude of the synthesized harmonic can reach three times the amplitude of the original signal, corresponding to the situation in actual industrial scenarios where the amplitude of macroscopic mechanical vibration is much larger than the amplitude of weak fault pulses; Gaussian noise A value of 1.5 indicates that under the strongest noise conditions, the standard deviation of the synthesized noise can reach 1.5 times the amplitude of the original signal, corresponding to a harsh operating condition with an extremely low signal-to-interference-plus-noise ratio.

[0062] In each subsequent round, the dynamic difficulty factor will be... Multiply by two preset maximum tolerance limits respectively: upper limit of harmonic interference intensity along with Incrementing smoothly from 0 to 1. Upper bound of Gaussian noise intensity along with Incrementing smoothly from 0 to 1. .because Since the values ​​are continuous, the upper bounds of the two types of interference also change continuously during the training process, without any jumps or abrupt changes. This fundamentally avoids the loss spikes and catastrophic forgetting that are common when switching between discrete stages.

[0063] Based on the above, this embodiment determines the upper bound of the intensity of the two types of interference by using an independent preset maximum tolerance limit and a unified dynamic difficulty factor. This ensures that the injection intensity of the two types of interference is synchronously and adaptively adjusted with the network learning progress, and also adapts to the different energy levels and physical characteristics of the two types of interference in actual industrial scenarios, thereby achieving independent and precise control of the two types of interference with different properties.

[0064] After obtaining the upper bounds of the harmonic interference intensity and Gaussian noise intensity for the current round, a specific dynamic contamination signal needs to be synthesized for each sample in each training batch. If a fixed interference intensity is directly used, the multi-scale network will overfit to the specific interference pattern, reducing its generalization ability; if each sample is independently and randomly sampled, the specific sampling method needs to be clearly defined. Therefore, in one embodiment, the synthesis of a dynamic contamination signal from the original pure one-dimensional vibration signal of the current batch includes: Dynamic pollution signals are synthesized using the following formula: in, The dynamic pollution signal is a one-dimensional vector with dimension L, which is the same length as the original signal. The actual harmonic interference intensity is a scalar quantity, with units of dimensionless amplitude coefficients, and its value range is... ; The actual Gaussian noise intensity is a scalar quantity, with units of a dimensionless amplitude coefficient, and its value range is... ; The original, pure one-dimensional vibration signal of the current batch is a one-dimensional vector with dimension L; It is the random fundamental frequency of low-frequency mechanical harmonics, a scalar quantity, and its unit is Hertz (Hz), with a value range of 10Hz to 50Hz; Let L be the time vector, a one-dimensional vector of dimension L, and let its i-th element be L. ,in The sampling interval (seconds) is the sampling interval. The sampling rate (Hz) is given by the subscript i, which is the index of the sampling point and has a range of values. , where is a positive integer; It is a random phase, a scalar, and its unit is radians (rad), with a value range from 0 to 2π. The noise is standard Gaussian white noise, a one-dimensional vector of dimension L, whose elements independently follow a standard normal distribution. Mid-sampling, i.e. .

[0065] Specifically, during the loading process of each training batch, the following synthesis operation is performed independently for each sample in the batch: The first step is to determine the actual harmonic interference intensity. From continuous uniform distribution A random value is sampled from the data. The lower bound of this distribution is 0, and the upper bound is the upper limit of the harmonic interference intensity of the current round. Each sample is sampled independently, meaning different samples are... The values ​​are all different. Similarly, the actual Gaussian noise intensity From continuous uniform distribution A random value is independently sampled from the sample.

[0066] The second step is to determine the random fundamental frequency of low-frequency mechanical harmonics. From continuous uniform distribution A random value is sampled, in Hertz, covering the common rotational frequencies of rotating machinery (e.g., 20Hz to 50Hz) and their lower harmonics; random phase. From continuous uniform distribution A random value is sampled from each sample, in radians. Both parameters are also sampled independently for each sample.

[0067] Step 3: Standard Gaussian white noise From the standard normal distribution A random vector of length L is generated, and the sampling of each element is performed independently.

[0068] Fourth step, according to the formula Calculate the dynamic pollution signal. Among them, A frequency of was generated Phase is The pure harmonic vector, whose amplitude is 1, multiplied by The amplitude was then obtained. Harmonic interference components; To multiply each element of standard Gaussian white noise by The standard deviation was obtained as The Gaussian noise interference component. The above synthesis process is completed before variational mode decomposition, that is, the original pure one-dimensional vibration signal is first processed. Synthetic dynamic pollution signals The vibration signal to be processed is obtained, and then standardization and variational mode decomposition are performed on the vibration signal to be processed.

[0069] Based on the above, this embodiment expands the diversity of training data by independently and randomly sampling the actual interference intensity, random fundamental frequency, random phase, and random noise among each sample, ensuring that each sample in each batch has a unique interference pattern (different harmonic amplitudes, different harmonic frequencies, different harmonic phases, different noise intensities, and different noise samples). Simultaneously, it always... and The sampling lower limit is anchored to zero, ensuring that a certain proportion of samples in each batch always have zero or close to zero interference intensity. This allows the multi-scale network to access pure or near-pure samples in each training round, continuously rehearsing pure physical fault mechanisms and fundamentally preventing catastrophic forgetting.

[0070] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0071] Secondly, based on the same inventive concept, this application also provides a system for implementing the robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information as described above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more system embodiments provided below can be found in the limitations of the robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information described above, and will not be repeated here.

[0072] In one embodiment, such as Figure 3 As shown, the system includes: a feature extraction module, an initial training module, a pollution scheduling training module, and a fault diagnosis module. Among them: The feature extraction module is used to acquire the original pure one-dimensional vibration signal. After performing feature processing on the original pure one-dimensional vibration signal, multi-scale fused features are obtained. The feature processing includes: standardization, variational mode decomposition, calculation of root mean square value and kurtosis of each channel and concatenation of physical prior vectors, attention weighting, and multi-scale feature extraction. The initial training module is used to train a multi-scale network based on multi-scale fusion features in the initial training rounds and establish the initial baseline training loss. The pollution scheduling training module is used to calculate a dynamic difficulty factor in each subsequent training round based on the decrease in the average training loss of the previous round relative to the initial baseline training loss, and to adjust the upper bound of the harmonic interference intensity and the upper bound of the Gaussian noise intensity for the current round according to the dynamic difficulty factor. In each training batch, the actual harmonic interference intensity and the actual Gaussian noise intensity are continuously sampled from a range of values ​​with a lower bound of zero and upper bounds of the upper bounds of the harmonic interference intensity and the Gaussian noise intensity, respectively. Based on the actual harmonic interference intensity and the actual Gaussian noise intensity, a dynamic pollution signal is synthesized from the original pure one-dimensional vibration signal of the current batch, and a multi-scale network is trained based on the pollution multi-scale fusion features obtained after feature processing of the dynamic pollution signal. The fault diagnosis module is used to perform feature processing on the vibration signal to be diagnosed after all training rounds are completed, classify it through a multi-scale network, and output the fault diagnosis results.

[0073] In one embodiment, the feature extraction module includes a numerical calculation unit. The numerical calculation unit is used to calculate the root mean square value using the following formula: The numerical calculation unit is used to calculate kurtosis using the following formula: in, Let be the root mean square value of the k-th channel. Let be the kurtosis of the k-th channel. The sampled value of the i-th sampling point in the k-th channel. Let i be the sampled value of the i-th sampling point of the k-th intrinsic mode function. For sequence length, Let be the mean of the k-th channel. To prevent division by zero of extremely small constants.

[0074] In one embodiment, the pollution scheduling training module includes a difficulty factor calculation unit. This unit calculates the dynamic difficulty factor using the following formula: in, Let be the dynamic difficulty factor for the t-th training round. Let be the average training loss for the (t-1)th training round. The initial baseline training loss is used.

[0075] In one embodiment, the pollution scheduling training module includes a boundary adjustment unit. This boundary adjustment unit is used to adjust the upper bound of the harmonic interference intensity using the following formula: The boundary adjustment unit is used to adjust the upper bound of the Gaussian noise intensity using the following formula: in, Let this be the upper bound of the harmonic interference intensity in the t-th training round. Let the upper bound of the Gaussian noise intensity in the t-th training round be , To preset the maximum harmonic interference tolerance limit, To preset the maximum Gaussian noise tolerance limit, Let be the dynamic difficulty factor for the t-th training round.

[0076] In one embodiment, the pollution scheduling training module includes a pollution signal synthesis unit. This pollution signal synthesis unit synthesizes a dynamic pollution signal using the following formula: in, As a dynamic pollution signal, This represents the actual harmonic interference intensity. This represents the actual Gaussian noise intensity. This represents the original, pure, one-dimensional vibration signal of the current batch. It is the random fundamental frequency of low-frequency mechanical harmonics. For time vectors, For random phase, It is standard Gaussian white noise.

[0077] In one embodiment, the feature extraction module includes an attention-weighted execution unit. This attention-weighted execution unit inputs the physical prior vector into the bottleneck multilayer perceptron, processes it sequentially using a ReLU activation function and a Sigmoid activation function, and outputs physical information attention weights. The formula for calculating the physical information attention weights is: in, For physical prior vectors, It is the Sigmoid activation function. It is the ReLU activation function. and This is the learnable weight matrix for the bottleneck multilayer perceptron.

[0078] In one embodiment, the feature extraction module includes a variational mode decomposition execution unit. The variational mode decomposition execution unit is used to configure the constrained variational problem of the variational mode decomposition as follows: The constraints are: in, For time variables in the continuous time domain, Let k be the intrinsic mode function. Let be the center frequency of the k-th intrinsic mode function. It follows the Dirac distribution. For convolution operators, The imaginary unit, For about The partial derivatives, This is a standardized one-dimensional vibration signal. This represents the number of intrinsic mode functions.

[0079] Each module in the aforementioned robust fault diagnosis system for rotating machinery based on physical information multi-scale networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware within or independently of the processor in a computer device, or stored in software within the memory of a computer device, allowing the processor to invoke and execute the corresponding operations of each module.

[0080] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores all data required for system operation. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information.

[0081] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0082] Thirdly, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described in the above method embodiments.

[0083] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of any of the methods described in the above method embodiments.

[0084] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0085] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0086] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information, characterized in that, The method includes: The original pure one-dimensional vibration signal is obtained, and multi-scale fusion features are obtained after performing feature processing on the original pure one-dimensional vibration signal. The feature processing includes, in sequence: standardization, variational mode decomposition, calculation of root mean square value and kurtosis of each channel and concatenation of physical prior vectors, attention weighting, and multi-scale feature extraction. In the initial training rounds, a multi-scale network is trained based on the multi-scale fusion features to establish an initial baseline training loss. In each subsequent training round, a dynamic difficulty factor is calculated based on the decrease in the average training loss of the previous round relative to the initial baseline training loss. The upper bounds of the harmonic interference intensity and Gaussian noise intensity for the current round are adjusted according to the dynamic difficulty factor. In each training batch, the actual harmonic interference intensity and actual Gaussian noise intensity are continuously sampled from a range where the lower bound is zero and the upper bounds are the upper bounds of the harmonic interference intensity and the Gaussian noise intensity, respectively. Based on the actual harmonic interference intensity and the actual Gaussian noise intensity, a dynamic contamination signal is synthesized from the original pure one-dimensional vibration signal of the current batch. The multi-scale network is then trained based on the contamination multi-scale fusion features obtained after performing the feature processing on the dynamic contamination signal. After all training rounds are completed, the vibration signal to be diagnosed is subjected to the aforementioned feature processing, and then classified through the multi-scale network to output the fault diagnosis result.

2. The method according to claim 1, characterized in that, The calculation of the root mean square value and kurtosis of each channel includes: The root mean square value is calculated using the following formula: The kurtosis is calculated using the following formula: in, Let be the root mean square value of the k-th channel. Let be the kurtosis of the k-th channel. This represents the sampled value of the i-th sampling point in the k-th channel. Let i be the sampled value of the i-th sampling point of the k-th intrinsic mode function. For sequence length, Let be the mean of the k-th channel. To prevent division by zero of extremely small constants.

3. The method according to claim 1, characterized in that, The calculation of the dynamic difficulty factor based on the degree of decrease in the average training loss of the previous round relative to the initial baseline training loss includes: The dynamic difficulty factor is calculated using the following formula: in, Let be the dynamic difficulty factor for the t-th training round. Let be the average training loss for the (t-1)th training round. The initial benchmark training loss is given.

4. The method according to claim 1 or 3, characterized in that, The step of adjusting the upper bound of harmonic interference intensity and the upper bound of Gaussian noise intensity for the current round according to the dynamic difficulty factor includes: The upper limit of the harmonic interference intensity is adjusted using the following formula: The upper bound of the Gaussian noise intensity is adjusted using the following formula: in, Let this be the upper bound of the harmonic interference intensity in the t-th training round. Let the upper bound of the Gaussian noise intensity in the t-th training round be , To preset the maximum harmonic interference tolerance limit, To preset the maximum Gaussian noise tolerance limit, Let be the dynamic difficulty factor for the t-th training round.

5. The method according to claim 1, characterized in that, The synthesis of dynamic contamination signals from the original pure one-dimensional vibration signals of the current batch includes: The dynamic pollution signal is synthesized using the following formula: in, As a dynamic pollution signal, The actual harmonic interference intensity is... The actual Gaussian noise intensity is... This is the original, pure, one-dimensional vibration signal of the current batch. It is the random fundamental frequency of low-frequency mechanical harmonics. For time vectors, For random phase, It is standard Gaussian white noise.

6. The method according to claim 1, characterized in that, Performing the attention weighting includes: The physical prior vector is input into the bottleneck multilayer perceptron, and after being processed sequentially by the ReLU activation function and the Sigmoid activation function, the physical information attention weights are output; the formula for calculating the physical information attention weights is: in, Let be the physical prior vector. The Sigmoid activation function is... The ReLU activation function is... and is the learnable weight matrix of the bottleneck multilayer perceptron.

7. The method according to claim 1, characterized in that, The constrained variational problem of the variational mode decomposition is: The constraints are: in, For time variables in the continuous time domain, Let k be the intrinsic mode function. Let be the center frequency of the k-th intrinsic mode function. It follows the Dirac distribution. For convolution operators, The imaginary unit, For about The partial derivatives, The standardized one-dimensional vibration signal This represents the number of intrinsic mode functions.

8. A robust fault diagnosis system for rotating machinery based on a physical information multi-scale network, characterized in that, The system includes: The feature extraction module is used to acquire the original pure one-dimensional vibration signal, and to obtain multi-scale fused features after performing feature processing on the original pure one-dimensional vibration signal. The feature processing includes, in sequence: standardization, variational mode decomposition, calculation of root mean square value and kurtosis of each channel and concatenation of physical prior vectors, attention weighting, and multi-scale feature extraction. The initial training module is used to train a multi-scale network based on the multi-scale fusion features in the initial training rounds and establish an initial baseline training loss. The pollution scheduling training module is used to calculate a dynamic difficulty factor in each subsequent training round based on the decrease in the average training loss of the previous round relative to the initial baseline training loss, and to adjust the upper bound of the harmonic interference intensity and the upper bound of the Gaussian noise intensity for the current round according to the dynamic difficulty factor. In each training batch, the actual harmonic interference intensity and the actual Gaussian noise intensity are continuously sampled from a range of values ​​with a lower bound of zero and upper bounds of the upper bounds of the harmonic interference intensity and the Gaussian noise intensity, respectively. Based on the actual harmonic interference intensity and the actual Gaussian noise intensity, a dynamic pollution signal is synthesized from the original pure one-dimensional vibration signal of the current batch, and the multi-scale network is trained based on the pollution multi-scale fusion features obtained after performing the feature processing on the dynamic pollution signal. The fault diagnosis module is used to perform the feature processing on the vibration signal to be diagnosed after all training rounds are completed, classify it through the multi-scale network, and output the fault diagnosis result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the robust fault diagnosis method for rotating machinery based on a multi-scale network of physical information as described in any one of claims 1 to 7.

Images