Rotating machinery fault data enhancement method and system based on mechanism calibration residual diffusion

Abstract

The application discloses a rotating machinery fault data enhancement method and system based on mechanism calibration residual diffusion, and belongs to the technical field of fault diagnosis. The method comprises the following steps: obtaining a real vibration sample; calibrating mechanism parameters and generating a mechanism simulation signal; selecting a conditional sample and a target sample, aligning the mechanism signal with the conditional sample, and then constructing inter-domain residual; training a conditional diffusion model with the residual as an object; inversely sampling to synthesize a predicted residual, fusing the predicted residual with the mechanism signal to obtain a candidate enhanced signal; and performing sample-by-sample screening through physical-feature joint gating, and adaptively constructing a high-quality enhanced sample library with consistent capacity of each fault category. The application makes the generated sample effectively approximate the real domain distribution while retaining the fault physical skeleton, reduces the generation freedom and replication risk under the condition of a small sample, and improves the authenticity, diversity and downstream fault diagnosis performance of the enhanced sample.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and more specifically to a method and system for enhancing fault data of rotating machinery based on mechanism-based calibration residual diffusion. Background Technology

[0002] In practical engineering, rotating machinery performs functions such as energy conversion, media transportation, and power transmission. During long-term operation, its gears, bearings, shafts, and transmission components may experience abnormal conditions such as localized damage, surface degradation, material loss, root damage, loosening, or imbalance. These faults alter the periodic impact, shaft frequency order, resonant frequency band, envelope spectrum, and statistical distribution in the vibration signal. Therefore, one-dimensional vibration signals are an important information carrier reflecting the fault state of equipment.

[0003] Deep learning-based fault diagnosis methods rely on sufficient and class-balanced labeled samples. However, in real-world engineering, real-world fault data occurs very infrequently, and dangerous faults are difficult to replicate, resulting in only a small number of samples for each fault category. Directly training the diagnostic model using such a small number of samples can easily lead to insufficient feature learning, unstable class boundaries, and a severe decline in generalization performance.

[0004] To address the problem of insufficient samples, existing technologies employ traditional transformations such as noise addition and scaling, or generative methods such as generative adversarial networks (GANs) and diffusion models. However, traditional transformations can only perform local perturbations and are unlikely to generate new fault patterns; GANs are prone to training instability or pattern collapse under small sample conditions. Although ordinary diffusion models are stable during training, they need to learn complete real waveforms from pure noise. Under extremely small sample conditions, they often generate signals with incorrect impulse periods, inconsistent spectra and envelope spectra, or signals that approximately replicate training samples, making it difficult to stably amplify effective features.

[0005] To address this, some studies have attempted to utilize the prior constraints of rotating machinery mechanisms to generate the process. The shaft frequency order can be obtained from the rotational speed, and local damage manifests as periodic impacts and damped resonance responses. Based on this, category-related mechanism simulation signals can be constructed, providing a physical framework for generation. However, fixed mechanism parameters cannot adapt to the differences in amplitude, damping, resonance bands, and noise caused by different equipment, measurement points, and operating environments, resulting in a significant shift between the mechanism simulation domain and the actual measurement domain. If fixed mechanism signals are directly used to augment the data, the enhanced samples lack true domain characteristics; if mechanism constraints are abandoned and complete waveforms are directly generated, a dilemma of high degrees of freedom and high replication risk arises.

[0006] Existing sample selection methods only evaluate the overall temporal similarity or distribution distance between candidate signals and real data. Retaining only the most similar signals can easily introduce approximate replication, while overemphasizing diversity can cause signals to deviate from the real domain or cross fault category boundaries. Furthermore, temporal similarity alone cannot ensure that the envelope spectrum, weak fault frequency, and diagnostic features still belong to the target fault category, lacking a joint criterion of physical and feature-based judgment.

[0007] Therefore, there is an urgent need for a method that can utilize real small sample inversion mechanism parameters, learn residual correction from simulation domain to real domain, and combine physical-feature multidimensional gating to screen enhanced samples. Summary of the Invention

[0008] To address the aforementioned technical issues, this invention proposes a method and system for enhancing rotating machinery fault data based on mechanism calibration residual diffusion. It utilizes similar real small samples for inversion and constrains mechanism parameters, enabling the mechanism simulation signal to adapt to the real equipment response. It learns the inter-domain residual between the mechanism simulation signal and the real vibration signal, rather than directly generating a complete waveform. Furthermore, it uses real neighborhood, anti-copying mechanisms, mechanism structure, and diagnostic category boundaries to jointly screen candidate signals, constructing a physically reasonable, near-realistic, diverse, and downstream diagnostic enhancement sample library.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, embodiments of the present invention provide a method for enhancing fault data of rotating machinery based on mechanism-based calibration residual diffusion, comprising the following steps: S1. Obtain real vibration samples of the fault category to be enhanced; S2. Based on the real vibration samples, the mechanism parameters of the mechanism simulation model are calibrated, and the calibrated mechanism parameters are used to generate category-related mechanism simulation signals. S3. Select conditional real samples and target real samples from the same fault category respectively, and the correlation coefficient between the selected conditional real samples and the target real samples is within a preset range. Align the mechanism simulation signal with the conditional real samples in the time domain to obtain an aligned mechanism signal, and construct an inter-domain residual using the target real samples and the aligned mechanism signal. S4. Using the inter-domain residual as the diffusion generation object, train the conditional residual diffusion model under the constraints of the alignment mechanism signal and the conditional real sample, and perform multi-objective joint optimization. S5. Backsample and synthesize the prediction residual using the trained conditional residual diffusion model, and fuse the prediction residual with the alignment mechanism signal to obtain the candidate enhancement signal; S6. The candidate enhancement signals are screened sample by sample through physical-feature joint quality gating, and a high-quality enhancement sample library with consistent capacity for each fault category is adaptively constructed according to the fault category.

[0011] Furthermore, in step S1, obtaining real vibration samples of the fault category to be enhanced includes: Read the one-dimensional vibration signal of rotating machinery stored according to state category, remove invalid values ​​and select the target measurement channel; The continuous signal is non-overlappingly sliced ​​according to the preset window length, and mean and standard deviation normalization is performed on each window; K real samples are selected from each fault category to be enhanced to form a small training sample for the generative model, and an interval window is set between the training area and the evaluation area to reduce data leakage.

[0012] Furthermore, in step S2, calibrating the mechanism parameters of the mechanism simulation model based on the real vibration samples includes: Extract time-domain statistical features, spectral features, and envelope spectral features from K real samples of the same fault category; The mechanism parameters are inverted using the aforementioned time-domain statistical features, spectral features, and envelope spectral features; Each mechanism parameter is limited to a preset physical reasonable range.

[0013] Furthermore, the time-domain statistical features include at least one of the following: mean, standard deviation, root mean square, absolute mean, peak value, kurtosis, skewness, peak factor, and impulse factor. The spectral characteristics include at least one of the following: logarithmic amplitude spectrum, mid-to-high frequency band energy ratio, and axis frequency order response; The envelope spectral features include at least one of envelope fluctuations and the main peak of the envelope spectrum; The mechanistic parameters include resonant frequency scale, damping scale, impact amplitude scale, random slip scale, noise scale, amplitude modulation depth scale, and bandwidth scale.

[0014] Furthermore, in step S2, generating category-related mechanism simulation signals using the calibrated mechanism parameters includes: The shaft frequency is obtained based on the rotational speed, and the fundamental periodic components are constructed using the shaft frequency and multiple stable orders. Different impact repetition rate, resonant carrier frequency, damping, noise, and random jitter parameters are set for different fault categories; An exponentially decaying resonance response is superimposed at the impact location, and colored noise is added to form a category-related mechanism simulation signal.

[0015] Furthermore, in step S3, the preset interval is: the correlation coefficient between the conditional real sample and the target real sample is not lower than a first preset threshold and not higher than a second preset threshold. The first preset threshold is used to ensure that the two have the same fault structure, and the second preset threshold is used to prevent the two from being approximately replicated.

[0016] Furthermore, in step S4, the multi-objective joint optimization includes: Random sampling diffusion time steps are performed, and Gaussian noise is added to the inter-domain residuals; The noise-added residual, alignment mechanism signal, conditional real sample, fault category, diffusion time step and similar reference sample set are input into the conditional denoising network. The main loss is the spread noise prediction loss, which is combined with the loss of consistency between the real domain spectrum and envelope spectrum, statistical characteristics and frequency band energy, weak fault frequency and mechanism structure, diagnosis and classification, class center and feature distribution, as well as the loss of real neighborhood, mechanism anchoring, anti-copying and diversity to carry out multi-objective joint optimization.

[0017] Furthermore, in step S6, the physical-feature joint quality gating screening specifically includes: Calculate the maximum temporal correlation coefficient and envelope spectral correlation coefficient between the candidate enhanced signal and all real samples of the same type; If the maximum temporal correlation coefficient is lower than the preset lower limit of the true neighborhood or higher than the preset upper limit of anti-copying, or if the envelope spectrum correlation coefficient is lower than the preset threshold, then the candidate enhancement signal is eliminated. The candidate enhancement signals that pass the initial screening are input into the pre-trained diagnostic network to obtain their diagnostic features, and the first distance between them and the center of the same category, and the second distance between them and the nearest center of the opposite category are calculated. If the difference between the second distance and the first distance is less than a preset interval threshold, the candidate enhancement signal is discarded. The remaining candidate enhancement signals were identified as samples with physical-feature joint quality gating.

[0018] Furthermore, in step S6, adaptively constructing a high-quality enhanced sample library with consistent capacity for each fault category specifically includes: For each fault category to be enhanced, candidate signals are generated in batches. The current batch is then merged with the historical candidate pool, and gating, comprehensive scoring, quality stratification, and diversity sampling are performed again. When the number of high-quality samples for a certain fault category reaches the target capacity, the generation of that category stops. If the maximum number of generation rounds is still insufficient, the sample pool that meets the basic safety threshold is supplemented according to the comprehensive quality score and diversity requirements, and a high-quality augmented sample library with consistent capacity for each fault category is output.

[0019] In a second aspect, embodiments of the present invention provide a system for enhancing rotating machinery fault data based on mechanism-based calibration residual diffusion, applying the method described in the first aspect, the system comprising: The sample acquisition module is used to acquire real vibration samples of the fault category to be enhanced. The mechanism calibration and simulation module is used to calibrate the mechanism parameters of the mechanism simulation model based on the real vibration samples, and to generate category-related mechanism simulation signals using the calibrated mechanism parameters; The alignment and residual construction module is used to select conditional real samples and target real samples from the same fault category, respectively, and the correlation coefficient between the selected conditional real samples and the target real samples is within a preset range. The mechanism simulation signal is time-domain aligned with the conditional real samples to obtain an aligned mechanism signal, and the target real samples and the aligned mechanism signal are used to construct an inter-domain residual. The residual diffusion training module is used to train a conditional residual diffusion model under the constraints of the alignment mechanism signal and the conditional real sample, using the inter-domain residual as the diffusion generation object, and to perform multi-objective joint optimization. The candidate signal synthesis module is used to backsample and synthesize the prediction residual using a trained conditional residual diffusion model, and fuse the prediction residual with the alignment mechanism signal to obtain the candidate enhancement signal. The quality gating screening module is used to screen the candidate enhancement signals sample by sample through physical-feature joint quality gating, and adaptively construct a high-quality enhancement sample library with consistent capacity for each fault category according to the fault category.

[0020] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention focuses on small-sample augmentation of rotating machinery fault data. Instead of using the fault diagnosis model as the main generator, it uses it as a tool to constrain the quality of candidate samples and verify the augmentation effect, thus clearly defining the augmentation objective.

[0021] 2. This invention utilizes the statistical, spectral, and envelope spectral characteristics of similar real small samples to invert and limit the amplitude mechanism parameters, enabling the mechanism simulation signal to adapt to the amplitude, frequency band, damping, and noise characteristics of real equipment.

[0022] 3. This invention changes the diffusion object from the complete vibration waveform to the inter-domain residual between the mechanism simulation signal and the real vibration signal, reducing the generation degree of freedom under extremely small sample conditions, while retaining the fault mechanism skeleton.

[0023] 4. This invention employs decoupling of conditional real samples and target real samples, and performs finite time-shift alignment before residual construction to reduce invalid large residuals caused by input replication and shock phase misalignment.

[0024] 5. This invention utilizes statistics, spectrum, envelope spectrum, category, and similar reference sets to jointly modulate a one-dimensional residual diffusion network, and supplements the global frequency dependence through a low-gain FNO-Lite branch.

[0025] 6. The P2F gating of the present invention simultaneously sets a lower limit for the real neighborhood and an upper limit for preventing duplication, and combines real / mechanistic distance, envelope spectrum, weak fault frequency, diagnostic class center and difficult outlier interval for sample-by-sample screening.

[0026] 7. This invention adaptively generates and constructs a high-quality sample library of fixed capacity in batches according to fault categories, which can alleviate the problem of imbalanced augmented data caused by the different generation difficulties of different fault categories.

[0027] 8. In downstream verification, this invention uses real small samples to anchor weights, so that generated samples are used to expand the fault distribution without replacing the dominant role of real small samples. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0029] Figure 1 This is a flowchart of a rotating machinery fault data enhancement method based on mechanism calibration residual diffusion provided in an embodiment of the present invention.

[0030] Figure 2 This is a schematic diagram of feature extraction and bounded mechanism parameter calibration for similar real small samples provided in the embodiments of the present invention.

[0031] Figure 3 This is a schematic diagram of the simulation signal construction for category-related weak mechanisms provided in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram of the multi-condition residual diffusion denoising network structure provided in an embodiment of the present invention.

[0033] Figure 5 This is a schematic diagram of a lightweight one-dimensional convolutional residual diagnostic network structure provided in an embodiment of the present invention.

[0034] Figure 6 This is a schematic diagram of P2F physical-feature joint quality gating and adaptive database construction provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the construction of the augmented dataset and the verification of downstream augmentation effects provided in this embodiment of the invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Example 1: like Figure 1 As shown, this invention discloses a method for enhancing rotating machinery fault data based on mechanism-based calibration residual diffusion, comprising the following steps: S1. Fault Vibration Data Reading, Windowing, and K-shot Small Sample Construction: Read the one-dimensional vibration signals of rotating machinery stored according to state categories, remove invalid values, and select the target measurement channel; perform non-overlapping windowing on continuous signals according to a preset window length, and perform mean and standard deviation normalization on each window. Select K real samples from each fault category to be enhanced to form a small sample for training the generative model, and set an interval window between the training area and the evaluation area to reduce data leakage caused by adjacent windows. No generative enhancement is performed in the normal state; the enhancement process only applies to fault categories.

[0037] S2. Feature Extraction and Bounded Mechanism Calibration from Small Real Samples of the Same Fault Category: Extract the mean, standard deviation, root mean square, absolute mean, peak value, kurtosis, skewness, peak factor, and impulse factor from K real samples of the same fault category; calculate the logarithmic amplitude spectrum, mid-to-high frequency band energy ratio, axis frequency order response, envelope fluctuation, and envelope spectrum main peak. Utilize the above time-domain statistical features, spectral features, and envelope spectrum features to invert mechanism parameters, including resonant frequency scale, damping scale, impulse amplitude scale, random slip scale, noise scale, amplitude modulation depth scale, and bandwidth scale, and constrain each mechanism parameter within a preset physically reasonable range.

[0038] S3. Generation of Category-Related Mechanism Simulation Signals: The shaft frequency is obtained based on the rotational speed, and a fundamental periodic component is constructed using the shaft frequency and multiple stable orders. Different impact repetition rates, resonant carrier frequencies, damping, noise, and random jitter parameters are set for different fault categories. An exponentially decaying resonant response is superimposed at the impact location, and colored noise is added to form a category-related mechanism simulation signal. When the precise structural parameters of the equipment are unknown, the number of gear teeth or bearing geometric parameters are not forced to be used. Instead, a weak mechanism prior consisting of shaft frequency order, periodic impact, damped resonance, modulation, and noise is adopted.

[0039] S4. Separation and Selection of Conditional Samples and Target Samples, and Alignment of Mechanism Signals: Conditional real samples and target real samples are selected from the same fault category, ensuring they share the same fault characteristics and do not constitute approximate replication. Conditional real samples provide real-domain conditions, while target real samples are used to construct diffusion training targets. A finite-time shift search is performed on the mechanism simulation signal, and the translation result with the highest correlation to the conditional real sample is used as the aligned mechanism signal.

[0040] S5. Residual Construction from Mechanism Domain to Real Domain and Conditional Diffusion Training: The difference between the target real sample and the aligned mechanism signal is standardized at the sample level to obtain the residual from the mechanism domain to the real domain. Diffusion time steps are randomly sampled, and Gaussian noise is added to the residual. The noisy residual, aligned mechanism signal, conditional real sample, fault category, diffusion time step, and a set of similar reference samples are input into the conditional denoising network. The conditional denoising network employs a one-dimensional multi-scale encoder-decoder, fuses conditional information through FiLM conditional modulation and conditionally aware gated jump connections, models global frequency information through spectral and envelope spectrum conditional branches, and performs small-amplitude frequency domain correction on the residual using a low-gain FNO-Lite branch.

[0041] S6. Residual Diffusion Model Optimization under Multi-Objective Constraints: The primary loss is the diffusion noise prediction loss, combined with optimizations based on consistency between the real domain spectrum and envelope spectrum, consistency between statistical features and band energy, weak fault frequencies and mechanistic structures, diagnostic classification and feature constraints, as well as losses for real neighborhood, mechanism anchoring, anti-copying, and diversity. The weights of anti-copying and diversity losses are gradually increased in the latter half of the training phase to prevent the model from deviating too early from the real domain during the initial training phase.

[0042] S7. Residual Backsampling and Candidate Enhancement Signal Synthesis: Based on the trained conditional residual diffusion model, the residual state is initialized with random noise, and backsampling is performed under the constraints of mechanism signal and real sample conditions to obtain the inter-prediction domain residual; the residual is fused with the alignment mechanism signal according to the preset gain and sample-level standardization is performed to obtain the candidate fault enhancement signal.

[0043] S8, P2F Physical-Feature Joint Quality Gating: Calculates the nearest neighbor temporal correlation and envelope spectrum correlation between candidate signals and real samples of the same class, setting a lower limit for the real neighborhood and an upper limit for preventing duplication; calculates the distance ratio of candidate signals relative to real samples and mechanistic signals, eliminating candidate signals that approximately copy real samples, approximately copy mechanistic signals, deviate from the real domain, or lose the mechanistic framework. Weak axis frequency order matching degree is used to verify fault structures, and a pre-trained diagnostic feature network is used to calculate the center distance of the current class, the nearest inter-class center distance, the difficult inter-class center distance, and the class probability interval, eliminating candidate signals located near the boundary of the wrong class.

[0044] S9. Adaptively construct a high-quality augmentation sample library with consistent capacity for each fault category: Generate candidate signals in batches for each fault category to be augmented. Merge the current batch with the historical candidate pool and re-execute hard gating, comprehensive scoring, quality stratification, and diversity sampling. Stop generation when the number of high-quality samples for the current fault category reaches the target capacity; if the maximum number of generation rounds is still insufficient, supplement the sample library from the candidate pool that meets the basic safety threshold according to the comprehensive quality score and diversity requirements, and output a high-quality augmentation sample library with consistent capacity for each fault category.

[0045] S10. Construction of Augmented Dataset and Validation of Augmentation Effect: A fault augmentation dataset is constructed by combining K-shot real fault samples with high-quality generated samples, while the normal category remains unchanged. In the downstream fault diagnosis model training, different training weights are set for real fault samples and generated fault samples to ensure that the small real samples maintain an anchoring effect on the class boundaries. The realism, diversity, and diagnostic gain of the augmented data are evaluated using Real-only, TSTR, TRTS, and Real+Fake methods.

[0046] The following example further illustrates the rotating machinery fault data augmentation method based on mechanism calibration residual diffusion of the present invention: Small-sample enhancement of vibration data from coal mill bearing faults: This embodiment uses the vibration signal of the rolling bearing of a high-speed shaft in a coal mill of a thermal power plant as the object to illustrate the small-sample enhancement method for mechanism calibration residual diffusion described in this invention. The rolling bearing state categories include: normal state, inner ring damage, outer ring damage, rolling element damage, and cage damage. The normal state is not enhanced. The four fault states are respectively trained using K-shot real vibration samples to generate models.

[0047] The sampling frequency, window length, number of training rounds, and gating threshold in this embodiment are preferred implementation parameters, which can be adjusted according to the coal mill speed, bearing structure, measuring point location, and vibration signal distribution.

[0048] S1. Data acquisition, windowing, and K-shot small sample construction of coal mill bearing vibration: An accelerometer is installed at the bearing housing of the coal mill to collect the radial vibration signal of the bearing and record the rotational speed simultaneously. Using the one-dimensional vibration signal from the selected measuring point as input, data files are created for normal condition, inner ring damage, outer ring damage, rolling element damage, and cage damage, respectively.

[0049] After reading the original vibration signal, remove null values, infinite values, missing acquisition segments, and signal segments with obviously abnormal amplitudes, and select the target vibration channel.

[0050] In this embodiment, the continuous signal is non-overlappingly windowed according to 1024 sampling points, and the windowing step size is also set to 1024 sampling points. The duration of each vibration window is 0.2s, which includes approximately 4 axis rotation cycles.

[0051] For each vibration window, normalize by removing the mean and standard deviation:

[0052] Where Norm(·) is the sample-level standardization operator. This is the original vibration window. and These are the mean and standard deviation of the window, respectively. A positive number set to prevent the denominator from being zero.

[0053] To reduce information leakage between consecutive adjacent windows, the data is divided into a training region, an interval region, and an evaluation region, and 100 vibration windows are reserved as an isolation zone between the training region and the evaluation region.

[0054] K real vibration samples are selected from the training region for each bearing fault category to form a small set of real samples for the corresponding fault category:

[0055] Among them, X c Let x be the true small sample set of fault category c. c,k Let K be the k-th real vibration sample in category c, where K is the number of real samples selected for each fault category. This embodiment uses both 5-shot and 10-shot settings, meaning that only 5 or 10 real samples are used to train the generation model for each fault category.

[0056] S2. Similar real small sample feature extraction and bounded mechanism calibration: For each bearing fault category, extract time-domain statistical features, spectral features, and envelope spectral features from K real small samples.

[0057] Time-domain statistical characteristics include mean, standard deviation, root mean square (RMS), absolute mean, peak value, kurtosis, skewness, peak factor, and impulse factor. Among them, the RMS reflects the bearing vibration energy, kurtosis and impulse factor reflect the impact degree of local faults, and the peak factor reflects the change of transient peak value relative to the overall vibration energy.

[0058] Perform a fast Fourier transform on the real vibration sample to obtain the logarithmic amplitude spectrum, and calculate the energy ratio of the low-frequency, mid-frequency and high-frequency bands, the shaft frequency and its harmonic response and the main spectral peaks.

[0059] The Hilbert transform is then used to obtain the envelope of the vibration signal, and the envelope fluctuation, the position of the main peak of the envelope spectrum, the amplitude of the main peak, and the energy near the repetition frequency of the fault impact are calculated.

[0060] When the number of bearing rolling elements, pitch circle diameter, rolling element diameter, and contact angle are available, the theoretical failure frequencies of the inner ring, outer ring, rolling elements, and cage can be calculated based on the bearing geometric parameters. When the bearing structural parameters cannot be accurately obtained, the shaft frequency order and the stable main peak in the envelope spectrum of similar real small samples are used as the prior for weak failure frequencies.

[0061] The simulation parameters of the mechanism are calibrated based on the above-mentioned real characteristics. The set of mechanism parameters is represented as follows:

[0062] in, The resonant frequency scale, For damping scale, To measure the impact amplitude, For random sliding scales, For noise scale, For amplitude modulation depth scale, It is a frequency bandwidth scale.

[0063] Specifically, the average of the logarithmic amplitude spectra of K real samples of the same type is calculated, and the intermediate frequency energy ratio R is calculated based on the average spectral amplitude of the entire frequency range. M,c High-frequency energy ratio R H,c Simultaneously calculate the average kurtosis K. c Average peak factor C c σ_envelope e,c and fault category severity coefficient v c .

[0064] Each calibration scale is calculated using the following bounded nonlinear mapping:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071] Where tanh(·) is the hyperbolic tangent function; , , , , and These represent, in order, the average peak factor, average kurtosis, high-frequency energy ratio, mid-frequency energy ratio, envelope standard deviation, and preset fault severity coefficient for category c; , , , , , and The meaning is the same as before.

[0072] , and The arithmetic mean was obtained by calculating the corresponding features of each of the K real samples of the same type one by one. and The fault severity coefficient is obtained by integrating the average logarithmic amplitude spectrum of K real samples of the same type across preset mid-frequency and high-frequency bands and normalizing it relative to the average energy of the entire frequency range. Based on on-site maintenance records, fault level, or known damage extent, the severity is predetermined and normalized to [0,1]; when only a fault category label is available without a clear severity classification, ... Take 0.5. In the formula, 3.0, 0.90, and 0.65 are the reference centers for the corresponding features, and 2.0, 4.0, 5.0, 0.35, 0.30, 0.45, and 0.25 are the sensitivity scales for the feature deviation. The coefficients before each tanh term are used to limit the maximum correction amplitude of the corresponding mechanistic scale relative to the baseline value of 1. The above reference centers, sensitivity scales, and correction coefficients are predetermined on real samples in the training area and used in conjunction with subsequent physical reasonable interval limits. When applied to other rotating machinery, they can be recalibrated according to the center position, dispersion, and allowable variation range of mechanistic parameters of similar real samples in the training area.

[0073] To prevent a small number of outlier samples from distorting the mechanistic parameters, range limiting is applied to each parameter:

[0074] This yields mechanistic calibration parameters that are compatible with the actual vibration characteristics of current coal mill bearing failures. The specific calibration process for similar small-sample feature extraction, parameter inversion, and bounded amplitude calibration is as follows: Figure 2 As shown.

[0075] S3. Generation of simulation signals related to bearing category: Map the mechanism calibration parameters to the baseline simulation parameters of category c. Let... , , , , , and Given the uncalibrated resonant carrier frequency, attenuation coefficient, impulse amplitude, jitter standard deviation at impulse moment, noise standard deviation, amplitude modulation depth, and random bandwidth, the mapping rule is as follows:

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] in, Sampling frequency, The resonant frequency scale, For damping scale, To measure the impact amplitude, For random sliding scales, For noise scale, For amplitude modulation depth scale, represents the bandwidth scale; represents the category simulation parameters before and after calibration, respectively. 0.45 This is used to ensure the resonant carrier frequency is below the Nyquist frequency while preserving the transition band. During simulation, [the following is used:] Controlling the zero-mean Gaussian random slip at the moment of impact, in order to Control the amplitude of colored noise, in order to Control the random bandpass bandwidth.

[0083] Calculate shaft frequency based on high-speed shaft speed:

[0084] in, This refers to the shaft frequency, measured in Hz. Rotational speed, in r / min; in this embodiment r / min, therefore shaft frequency Hz.

[0085] First, the fundamental periodic components of the shaft system are constructed using the shaft frequency and its multiple stable orders:

[0086] in, For axis frequency order components, For time, For the set of axial frequency orders, For order numbering, For shaft frequency, and These represent the amplitude and initial phase for the corresponding order, respectively.

[0087] Subsequently, different impact repetition rates were set for different bearing failure categories. The stable failure peaks in the corresponding real small sample envelope spectra were used as the impact repetition frequencies for inner ring damage, outer ring damage, rolling element damage, and cage damage.

[0088] The ideal impact time is determined according to the stated impact repetition frequency, and random slip is added to the impact position to simulate the impact position deviation caused by speed fluctuations, rolling element slippage, and load changes. Each impact excitation exhibits exponentially decaying resonance.

[0089] in, The number of impacts within the current window. To impact the serial number, Let k be the amplitude of the impact. For the impact moment that includes random slip, The attenuation coefficient is... For resonant carrier frequency, For the initial phase, To ensure No earlier than The unit step function.

[0090] By superimposing shaft frequency order, periodic impact, damped resonance, amplitude modulation, and colored noise, simulation signals related to bearing categories are obtained:

[0091] Where, x sim The original mechanism simulation signal, For impact attenuation resonance components, Colored noise corresponding to the bearing failure category.

[0092] Different impact repetition rates, resonant frequencies, damping, and noise parameters are used for inner ring, outer ring, rolling element, and cage faults to ensure that the mechanistic signals of different fault categories have distinguishable impact periods and envelope spectrum responses. The combined construction process of shaft frequency order, category-related impact repetition rate, random slip and amplitude modulation, exponentially decaying resonance, and colored noise is as follows: Figure 3 As shown.

[0093] S4. Separation and selection of similar conditional samples and target samples, and alignment of mechanism signals: For any bearing failure category, select conditional real samples and target real samples from the K-shot real sample set for that category.

[0094] Conditional true samples are used to provide the residual diffusion model with the true amplitude, frequency band, statistics, and envelope spectrum conditions for the current bearing fault category; target true samples are used to construct the diffusion training target. Both belong to the same bearing fault category but use different vibration windows.

[0095] Calculate the Pearson correlation coefficient between the conditional real sample and other real samples of the same type, and preferentially select the target real sample that satisfies the following formula:

[0096] Where PCC is the Pearson correlation coefficient. For conditional true samples, For the target real sample, the absolute value is used to eliminate the effect of phase reversal. The lower correlation limit of 0.35 is used to ensure that the conditional sample and the target sample have the same bearing fault structure; the upper correlation limit of 0.955 is used to prevent the two from being approximately replicated.

[0097] Since the initial impact position in the mechanism simulation signal is random, a finite time-shift search is performed on the mechanism simulation signal before constructing the inter-domain residuals:

[0098] in, The candidate cyclic time shift is expressed in units of sampling points. This is the optimal cyclic time shift. For time-shifted search sets, For cyclic translation operations, This indicates the mechanism simulation signal. Cyclic translation One sampling point, For the conditional true sample, PCC is the Pearson correlation coefficient.

[0099] The translation result with the highest correlation is taken as the alignment mechanism signal:

[0100] Where Δopt is the optimal cyclic translation amount. This is a cyclic translation operation.

[0101] In this embodiment, the time-shift search range is set to 160 to 160 sampling points were used, with a search step size of 8 sampling points to reduce the phase deviation between the mechanistic impact location and the actual bearing impact location.

[0102] S5. Residual construction from the mechanistic domain to the real domain and conditional diffusion training: The difference between the target real bearing vibration sample and the alignment mechanism signal is standardized to obtain the residual from the mechanism domain to the real domain:

[0103] in, The standardized residual from the mechanistic domain to the real domain. For the target real bearing vibration sample, Norm is a sample-level normalization operator used to align the mechanism signal.

[0104] The residuals include the amplitude difference, resonance frequency band difference, transmission path difference, load modulation difference, and background noise difference between the mechanism simulation signal and the actual bearing vibration signal.

[0105] Random sampling diffusion time step Add standard Gaussian noise to the inter-domain residuals:

[0106] in, For the first The noise-added residual at each diffusion moment For noise-free inter-domain residuals, For cumulative noise scheduling parameters, To and Noise of the same dimension and following a standard Gaussian distribution N(0,I), This represents the diffusion time step. In this embodiment, the total number of diffusion steps is set to 50, and the noise scheduling initial value is set to... The termination value is set to 0.035.

[0107] The noise-added residual, alignment mechanism signal, conditional true sample, bearing fault category, diffusion time step, and similar reference sample set are all input into the conditional denoising network.

[0108] The conditional denoising network employs a one-dimensional multi-scale encoder-decoder. The inputs include the denoised residual, the aligned mechanism signal, the conditional true sample, and the difference between the mechanism signal and the conditional sample. The network channels are configured as follows:

[0109] Statistical features, spectral features, envelope spectral features, bearing fault categories, and diffusion time steps together form a conditional vector, which is then applied to residual blocks at various scales using FiLM modulation.

[0110]

[0111] in, and The first Features before and after layer modulation It is a conditional vector formed by fusing diffusion time step, fault category, mechanism / real statistical characteristics, and spectrum-envelope spectrum characteristics; and For the first The layer can learn linear mapping parameters, and the output of the linear mapping is evenly divided along the channel dimension. and ; coefficient 0.25 and Together, they limit the scaling and bias amplitudes to prevent conditional modulation from disrupting the mechanistic framework.

[0112] The decoding end employs conditionally aware gated skip connections to pass filtered encoded features to the corresponding decoding layer. Spectrum and envelope spectrum conditional branches are used to maintain the bearing fault frequency, impact resonance band, and envelope spectrum response.

[0113] In addition, a lightweight Fourier Neural Operator (FNO-Lite) global frequency domain residual branch is set with a width of 48, retaining 24 frequency modes, containing 2 layers of spectral convolution, and its output is injected into the main network output with a gain of 0.10. The above-mentioned multi-condition residual diffusion denoising network structure, driven by the inter-noise domain residual, the alignment mechanism simulation signal, the real samples under similar conditions, the set of similar reference samples, the fault category, and the diffusion time step, is as follows: Figure 4 As shown.

[0114] S6. Optimization of residual diffusion model under multi-objective constraints: Based on the loss prediction of diffused noise, the bearing residual diffusion model is jointly optimized:

[0115] Where L is the total loss after joint optimization. To predict the loss for diffused noise, To compensate for the loss of consistency between the real-domain spectrum and the envelope spectrum, as well as the loss of consistency between statistical characteristics and band energy, For weak fault frequency and mechanism structure loss, To diagnose classification, class center, and feature distribution losses, For real-world neighborhoods, mechanism anchoring, copy prevention, and diversity loss; , , , and For the corresponding non-negative weights.

[0116] The core losses are as follows:

[0117] Where B is the number of samples in the current training batch, and b is the sample sequence number within the batch; This refers to the standard Gaussian noise added to the residual between the b-th noiseless domains; For a conditional denoising network with parameter θ; The diffusion time step for the random sampling of the b-th sample; For the b-th sample at diffusion time step The noise-added residual; This represents the fusion condition corresponding to the b-th sample.

[0118]

[0119] in, The candidate signals for the training phase are synthesized based on the predicted noiseless residuals and alignment mechanism signals. For the target real sample, For the logarithmic magnitude spectrum operator, For envelope spectrum operators, To smooth the L1 loss, statistical characteristics, frequency band energy, and a low-weighted spectral-envelope spectrum consistency term for conditionally true samples are incorporated. .

[0120]

[0121] in, To align the mechanism signal, This refers to the energy retention loss in the neighborhood of the shaft frequency order or the repetition frequency of the fault impact. The weights are set as follows: 0.7 and 0.3 are the combined weights of the envelope spectrum structure and the amplitude spectrum structure, respectively.

[0122]

[0123] in, For a frozen diagnostic classifier, CE represents the cross-entropy loss. To diagnose the 128-dimensional feature mapping of the network, For the true feature center of category c, To generate a distribution consistency loss between the generated features and the true features' mean and standard deviation, and For the corresponding weights.

[0124]

[0125]

[0126] in, Only when the correlation between the candidate signal and the conditional true sample exceeds the anti-copying threshold. Punishment will be imposed at that time; and These represent the candidate signal and conditional true sample for the training phase corresponding to the b-th training sample, respectively. ReLU(a) = max(0,a); Minimize the average of the pairwise absolute correlation coefficients of different generated samples within the same fault category batch. To improve intra-class diversity; i and j are sample indices, c i and c j For the corresponding category. It also includes distance ratio constraints that prevent the generated signal from deviating excessively from the real neighborhood and from deviating excessively from the mechanistic signal.

[0127] In the early stages of training, the main focus is on optimizing noise prediction, temporal reconstruction, and spectral envelope features to ensure that the generated results reflect the real bearing fault distribution. In the latter half of the training, the weights for anti-copying and diversity loss are gradually increased to prevent the generated signals from repeating real K-shot samples.

[0128] In this embodiment, the residual diffusion model is trained for 220 rounds, with 70 effective training steps performed in each round.

[0129] S7. Residual backsampling and candidate enhancement signal synthesis: After model training is complete, DDIM backsampling is performed starting from standard Gaussian noise to gradually obtain inter-domain residuals. To improve sample generation efficiency, this embodiment uses 8-step or 16-step DDIM sampling.

[0130] The candidate sample signal is obtained by superimposing the predicted residual and the alignment mechanism signal:

[0131] in, To align the mechanism signal, To predict inter-domain residuals, For residual gain, Norm(·) is the sample-level standardization operator. This enhances the signal for candidate bearing faults.

[0132] In this embodiment, the residual gain is set to 1.05. The mechanistic signal provides the shaft frequency order, bearing fault impact, and attenuated resonance framework, while the predicted residual supplements the transmission path, load fluctuation, and noise characteristics in the actual coal mill bearing vibration.

[0133] S8, P2F Physical-Feature Joint Mass Gating: For each candidate bearing fault enhancement signal, a joint screening process is performed based on the real neighborhood, anti-copying criteria, envelope spectrum, mechanistic structure, and diagnostic boundary.

[0134] First, calculate the nearest neighbor correlation between the candidate signal and similar K-shot real samples:

[0135] And satisfy:

[0136] Where, ρ NN (xg) is the candidate signal x g K real samples x of category c c,j The maximum absolute Pearson correlation coefficient between them, where j is the true sample number. This is the lower bound of the true neighborhood. To prevent copying from exceeding the limit.

[0137] In this embodiment, the lower limit of the true neighborhood Set to 0.62, maximum anti-copying limit. Set it to 0.965.

[0138] Subsequently, the envelope spectrum correlation between the candidate signal and the real sample of the same type is calculated, and the envelope spectrum correlation is required to be no less than 0.14; the shaft frequency order and bearing fault impact frequency retention are calculated, and the weak fault frequency retention is required to be no less than 0.20.

[0139] The candidate signals are further input into a pre-trained bearing fault diagnosis feature network. The network employs, for example... Figure 5 The lightweight one-dimensional convolutional residual diagnostic network DiagnosisFeatureNet, as shown, first performs Z-score normalization on the candidate enhancement signal and increases the channel dimension. Then, it sequentially passes through downsampled one-dimensional convolutional blocks (1→32 and 32→64), a residual block with 64 channels and a dilation rate of 1, a downsampled convolutional block with 64→96 channels, a residual block with 96 channels and a dilation rate of 2, a downsampled convolutional block with 96→128 channels, and a residual block with 128 channels and a dilation rate of 4. The output of the convolutional backbone is then adaptively global average pooled, and 128-dimensional diagnostic features z are obtained through Linear(128→128), LayerNorm, LeakyReLU, and Dropout. These features are then processed through Linear(128→5) and Softmax to obtain the class probabilities p of five bearing states. The 128-dimensional diagnostic features are used to calculate the distances from the candidate signal to the target fault class center, the nearest out-of-class center, and the difficult out-of-class center.

[0140]

[0141]

[0142] in, The nearest dissimilar category interval, For difficult heterogeneous intervals, The probability interval for the target category; The Euclidean distance from the 128-dimensional diagnostic features of the candidate signal to the center of the target fault category is given. The minimum Euclidean distance from this diagnostic feature to all heterogeneous class centers. The minimum Euclidean distance from this diagnostic feature to the center of the difficult heterogeneous class; The predicted probability of the target fault category c output by the pre-trained diagnostic network. Let be the predicted probability of the out-of-class category k, where k is the category index and k ≠ c. Difficult out-of-class categories are determined based on the confusion relationships of the pre-trained diagnostic network on real samples in the training region, prioritizing the out-of-class category most easily confused with the target fault category. This embodiment requires the nearest out-of-class interval. Not less than 0.03, difficult inter-class interval Not less than 0.01, and the probability interval of the target category. Not less than 0, meaning the probability of the target fault category is not lower than the probability of the highest out-of-category fault.

[0143] Calculate the overall quality score for candidate signals that pass the hard threshold:

[0144] in, The overall quality score of the candidate enhanced signal; The nearest neighbor temporal correlation between candidate signals and real K-shot samples of the same type. This refers to the envelope spectrum correlation between candidate signals and similar real samples. For weak fault frequency retention, , and The meanings are the same as before; D is the diagnostic feature diversity score, used to characterize the degree of difference between the candidate signal and the currently retained similar samples in the diagnostic feature space; O is the real domain-mechanism coverage deviation penalty, used to penalize the candidate signal for being too far away from the similar real samples or too far away from the aligned mechanism signal; C is the approximate replication penalty, used to penalize the case where the candidate signal is too highly correlated with any K-shot real sample or the retained similar sample; w1~w9 are the non-negative weights of each scoring item. Each positive evaluation item is normalized to [0,1] before scoring, and O and C are deducted from the total score as penalty items, resulting in a comprehensive quality score. The larger the value, the higher the overall quality of the candidate signal.

[0145] The final selection of enhanced bearing fault samples was based on comprehensive quality score, quality level, and diversity of diagnostic features. The complete process, including the above-mentioned real neighborhood screening, anti-copying screening, real domain and mechanism coverage ratio screening, envelope spectrum and weak fault frequency preservation screening, class center distance and inter-class interval discrimination, and subsequent adaptive library construction by category, is as follows: Figure 6 As shown.

[0146] S9. Adaptively build a high-quality enhanced sample library based on bearing failure categories: Candidate sample pools and high-quality enhanced sample libraries were established for inner ring damage, outer ring damage, rolling element damage, and cage damage, respectively.

[0147] In each round, 512 candidate vibration signals are generated for the current fault category. The newly generated candidate signals are merged with the historical candidate pool, and P2F hard gating, comprehensive scoring, quality stratification and diversity sampling are re-executed.

[0148] The target sample library capacity for each bearing fault category is set to 500 samples, and the maximum number of generated rounds is set to 24. When the number of high-quality samples for a fault category that passes the gating reaches 500, generation for that category stops; fault categories that have not yet reached the target capacity continue to be generated.

[0149] If the number of samples is still insufficient after reaching the maximum number of generation rounds, the samples are supplemented from the candidate pool that meets the basic thresholds such as true neighborhood, anti-copying, envelope spectrum, and class consistency, based on the comprehensive quality score and diversity, and finally 500 high-quality enhanced samples are obtained for each bearing fault category.

[0150] S10. Construction and effect verification of augmented dataset for coal mill bearing faults: From 500 high-quality generated samples for each bearing fault category, 400 samples were selected as enhanced training samples, and the remaining 100 samples were used as generated test samples.

[0151] Augmented training data consists of the following two parts:

[0152] in, For each category, there are K real bearing fault samples. High-quality samples are generated for the corresponding category. The normal category uses only real normal vibration samples and does not generate normal samples.

[0153] In the training of the downstream coal mill bearing fault diagnosis model, higher training weights are assigned to K-shot real fault samples, while lower auxiliary weights are assigned to generated samples. This ensures that real samples maintain the anchoring effect of the fault category decision boundary. The process for constructing the augmented dataset of small K-shot real fault samples, high-quality P2F generated samples, and normal category real samples, the sample weight allocation, and the downstream augmentation effect verification are as follows: Figure 7 As shown.

[0154] The enhancement effect can be evaluated in the following ways: Real-only: Training is performed using only K-shot real bearing failure samples and evaluation is conducted on a real test set; TSTR: Train using generated samples and evaluate on a real test set; TRTS: Train using real samples and evaluate on a generated test set; Real+Fake: Jointly trains using real samples and generated samples and evaluates on a real test set.

[0155] Evaluation metrics include accuracy, precision, recall, weighted F1, Kappa, and MCC, and are combined with temporal correlation, envelope spectrum correlation, nearest neighbor replication rate, and intraclass diversity to comprehensively evaluate the authenticity, non-replicability, and diagnostic gain of generated samples.

[0156] To evaluate the actual contribution of generated samples to fault diagnosis with small sample sizes, this embodiment compares two training strategies, Real-only and Real+Fake, under the same data partitioning, downstream classifier, training parameters, and test set conditions. Under 5-shot conditions, when training with only real samples, the classifier's F1 score was 84.86%; after adding MR-FiLM-RD generated samples, the F1 score increased to 98.19%, an absolute improvement of 13.33 percentage points. Meanwhile, Real+Fake achieved an accuracy, precision, and recall of 98.20%, with a Kappa and MCC of 0.9775, indicating that generated samples effectively alleviated the problems of insufficient intra-class distribution coverage and unstable classification boundaries caused by a very small number of real samples.

[0157] The quality results further demonstrate that the temporal correlation and envelope spectrum correlation between the generated samples and the real samples are 0.9697 and 0.9696, respectively, effectively preserving the amplitude variations, periodic impulses, and envelope modulation characteristics of the real fault signals. The nearest neighbor replication rate is 0, and the intra-class diversity is 0.7271, indicating that the improvement in diagnostic performance does not primarily stem from direct replication of the training samples, but rather from the fact that the generated samples supplement effective intra-class variations while maintaining fault authenticity. Therefore, the significant improvement achieved by Real+Fake compared to Real-only proves that the generated samples possess practical training value and diagnostic gain.

[0158] This embodiment calibrates bearing failure mechanism parameters using similar real small samples, constructs a failure skeleton using shaft frequency order, periodic impact, and damped resonance, uses a residual diffusion model to learn the correction amount from the mechanism simulation domain to the real coal mill bearing vibration domain, and constructs a class-balanced high-quality fault enhancement sample library through P2F joint quality gating, thereby alleviating the problem of insufficient real fault data for coal mill bearings.

[0159] Example 2: Based on the same inventive concept, embodiments of the present invention also provide a system for enhancing rotating machinery fault data based on mechanism-based calibration residual diffusion, the system comprising: The sample acquisition module is used to acquire real vibration samples of the fault category to be enhanced; The mechanism calibration and simulation module is used to calibrate the mechanism parameters of the mechanism simulation model based on the real vibration samples, and to generate category-related mechanism simulation signals using the calibrated mechanism parameters; The alignment and residual construction module is used to select conditional real samples and target real samples from the same fault category, respectively, and the correlation coefficient between the selected conditional real samples and the target real samples is within a preset range. The mechanism simulation signal is time-domain aligned with the conditional real samples to obtain an aligned mechanism signal, and the target real samples and the aligned mechanism signal are used to construct an inter-domain residual. The residual diffusion training module is used to train a conditional residual diffusion model under the constraints of the alignment mechanism signal and the conditional real sample, using the inter-domain residual as the diffusion generation object, and to perform multi-objective joint optimization. The candidate signal synthesis module is used to backsample and synthesize the prediction residual using a trained conditional residual diffusion model, and fuse the prediction residual with the alignment mechanism signal to obtain the candidate enhancement signal. The quality gating screening module is used to screen the candidate enhancement signals sample by sample through physical-feature joint quality gating, and adaptively construct a high-quality enhancement sample library with consistent capacity for each fault category according to the fault category.

[0160] Since these systems and the principles underlying the problems they address are similar to the aforementioned method for enhancing rotating machinery fault data based on mechanism-based calibration residual diffusion, the implementation of this system can be found in the implementation of the aforementioned method, and the repetitions will not be repeated.

[0161] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0162] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for enhancing rotating machinery fault data based on mechanism-based calibration residual diffusion, characterized in that, Includes the following steps: S1. Obtain real vibration samples of the fault category to be enhanced; S2. Based on the real vibration samples, the mechanism parameters of the mechanism simulation model are calibrated, and the calibrated mechanism parameters are used to generate category-related mechanism simulation signals. S3. Select conditional real samples and target real samples from the same fault category respectively, and the correlation coefficient between the selected conditional real samples and the target real samples is within a preset range. Align the mechanism simulation signal with the conditional real samples in the time domain to obtain an aligned mechanism signal, and construct an inter-domain residual using the target real samples and the aligned mechanism signal. S4. Using the inter-domain residual as the diffusion generation object, train the conditional residual diffusion model under the constraints of the alignment mechanism signal and the conditional real sample, and perform multi-objective joint optimization. S5. Backsample and synthesize the prediction residual using the trained conditional residual diffusion model, and fuse the prediction residual with the alignment mechanism signal to obtain the candidate enhancement signal; S6. The candidate enhancement signals are screened sample by sample through physical-feature joint quality gating, and a high-quality enhancement sample library with consistent capacity for each fault category is adaptively constructed according to the fault category.

2. The method as described in claim 1, characterized in that, In step S1, obtaining real vibration samples of the fault category to be enhanced includes: Read the one-dimensional vibration signal of rotating machinery stored according to state category, remove invalid values ​​and select the target measurement channel; The continuous signal is non-overlappingly sliced ​​according to the preset window length, and mean and standard deviation normalization is performed on each window; K real samples are selected from each fault category to be enhanced to form a small training sample for the generative model, and an interval window is set between the training area and the evaluation area to reduce data leakage.

3. The method as described in claim 1, characterized in that, In step S2, calibrating the mechanism parameters of the mechanism simulation model based on the real vibration samples includes: Extract time-domain statistical features, spectral features, and envelope spectral features from K real samples of the same fault category; The mechanism parameters are inverted using the aforementioned time-domain statistical features, spectral features, and envelope spectral features; Each mechanism parameter is limited to a preset physical reasonable range.

4. The method as described in claim 3, characterized in that, The time-domain statistical features include at least one of the following: mean, standard deviation, root mean square, absolute mean, peak value, kurtosis, skewness, peak factor, and impulse factor. The spectral characteristics include at least one of the following: logarithmic amplitude spectrum, mid-to-high frequency band energy ratio, and axis frequency order response; The envelope spectral features include at least one of envelope fluctuations and the main peak of the envelope spectrum; The mechanistic parameters include resonant frequency scale, damping scale, impact amplitude scale, random slip scale, noise scale, amplitude modulation depth scale, and bandwidth scale.

5. The method as described in claim 1, characterized in that, In step S2, generating category-related mechanism simulation signals using the calibrated mechanism parameters includes: The shaft frequency is obtained based on the rotational speed, and the fundamental periodic components are constructed using the shaft frequency and multiple stable orders. Different impact repetition rate, resonant carrier frequency, damping, noise, and random jitter parameters are set for different fault categories; An exponentially decaying resonance response is superimposed at the impact location, and colored noise is added to form a category-related mechanism simulation signal.

6. The method as described in claim 1, characterized in that, In step S3, the preset interval is: the correlation coefficient between the conditional real sample and the target real sample is not lower than a first preset threshold and not higher than a second preset threshold. The first preset threshold is used to ensure that the two have the same fault structure, and the second preset threshold is used to prevent the two from being approximately replicated.

7. The method as described in claim 1, characterized in that, In step S4, the multi-objective joint optimization includes: Random sampling diffusion time steps are performed, and Gaussian noise is added to the inter-domain residuals; The noise-added residual, alignment mechanism signal, conditional real sample, fault category, diffusion time step and similar reference sample set are input into the conditional denoising network. The main loss is the spread noise prediction loss, which is combined with the loss of consistency between the real domain spectrum and envelope spectrum, statistical characteristics and frequency band energy, weak fault frequency and mechanism structure, diagnosis and classification, class center and feature distribution, as well as the loss of real neighborhood, mechanism anchoring, anti-copying and diversity to carry out multi-objective joint optimization.

8. The method as described in claim 1, characterized in that, In step S6, the physical-feature joint quality gating screening specifically includes: Calculate the maximum temporal correlation coefficient and envelope spectral correlation coefficient between the candidate enhanced signal and all real samples of the same type; If the maximum temporal correlation coefficient is lower than the preset lower limit of the true neighborhood or higher than the preset upper limit of anti-copying, or if the envelope spectrum correlation coefficient is lower than the preset threshold, then the candidate enhancement signal is eliminated. The candidate enhancement signals that pass the initial screening are input into the pre-trained diagnostic network to obtain their diagnostic features, and the first distance between them and the center of the same category, and the second distance between them and the nearest center of the opposite category are calculated. If the difference between the second distance and the first distance is less than a preset interval threshold, the candidate enhancement signal is discarded. The remaining candidate enhancement signals were identified as samples with physical-feature joint quality gating.

9. The method as described in claim 1, characterized in that, In step S6, the adaptive construction of a high-quality enhanced sample library with consistent capacity for each fault category specifically includes: For each fault category to be enhanced, candidate signals are generated in batches. The current batch is then merged with the historical candidate pool, and gating, comprehensive scoring, quality stratification, and diversity sampling are performed again. When the number of high-quality samples for a certain fault category reaches the target capacity, the generation of that category stops. If the maximum number of generation rounds is still insufficient, the sample pool that meets the basic safety threshold is supplemented according to the comprehensive quality score and diversity requirements, and a high-quality augmented sample library with consistent capacity for each fault category is output.

10. A system for enhancing rotating machinery fault data based on mechanism-based calibration residual diffusion, characterized in that, The system comprises: The sample acquisition module is used to acquire real vibration samples of the fault category to be enhanced; The mechanism calibration and simulation module is used to calibrate the mechanism parameters of the mechanism simulation model based on the real vibration samples, and to generate category-related mechanism simulation signals using the calibrated mechanism parameters; The alignment and residual construction module is used to select conditional real samples and target real samples from the same fault category, respectively, and the correlation coefficient between the selected conditional real samples and the target real samples is within a preset range. The mechanism simulation signal is time-domain aligned with the conditional real samples to obtain an aligned mechanism signal, and the target real samples and the aligned mechanism signal are used to construct an inter-domain residual. The residual diffusion training module is used to train a conditional residual diffusion model under the constraints of the alignment mechanism signal and the conditional real sample, using the inter-domain residual as the diffusion generation object, and to perform multi-objective joint optimization. The candidate signal synthesis module is used to backsample and synthesize the prediction residual using a trained conditional residual diffusion model, and fuse the prediction residual with the alignment mechanism signal to obtain the candidate enhancement signal. The quality gating screening module is used to screen the candidate enhancement signals sample by sample through physical-feature joint quality gating, and adaptively construct a high-quality enhancement sample library with consistent capacity for each fault category according to the fault category.

Images