CN-122020401-A - Fault data generation method and system based on physical constraint double-domain diffusion model

CN122020401ACN 122020401 ACN122020401 ACN 122020401ACN-122020401-A

Abstract

The invention discloses a fault data generation method and system based on a physical constraint double-domain diffusion model, relates to the technical field of predictive maintenance, and solves the technical problem that the existing data generation method is easy to generate data which are similar in statistics but violate a fault physical mechanism when generating industrial fault data, and cannot meet the actual application requirements; the method comprises the steps of capturing physical characteristics of real fault data in a time domain and a frequency domain, constructing a diffusion model, introducing a FANS mechanism in a forward diffusion process and conditionally injecting the physical characteristics in a reverse denoising process, constructing a total loss function optimization diffusion model based on energy consistency and denoising reconstruction loss to obtain a fault data generation model, and generating simulated fault data by the fault data generation model.

Inventors

XU SHUAI
QI MENGLEI
MA GUANGTONG
Bi Xianning
ZHANG HAN

Assignees

西南交通大学

Dates

Publication Date: 20260512
Application Date: 20260109

Claims (10)

1. The fault data generation method based on the physical constraint double-domain diffusion model is characterized by comprising the following steps of: Step 1, extracting physical characteristics of target real fault data complementary in time domain and frequency domain; step 2, a diffusion model is built, wherein the generation process of the diffusion model comprises a forward diffusion process and a reverse denoising process, a FANS mechanism is introduced in the forward diffusion process, a condition constraint comprising the physical characteristics is injected in the reverse denoising process, and the FANS mechanism is used for adaptively adjusting noise intensity based on spectrum energy distribution; And step3, constructing an energy consistency loss and denoising reconstruction loss to obtain a total loss function, optimizing a diffusion model according to the total loss function to obtain a fault data generation model, and generating target simulation fault data according to the target real fault data by the fault data generation model.
2. The fault data generation method based on the physical constraint two-domain diffusion model according to claim 1, wherein the step 1 comprises: Capturing time sequence features on different scales on a time domain branch by adopting a hierarchical convolution architecture as time domain features; Calculating a logarithmic power spectrum on the frequency domain branch to strengthen weak fault harmonic waves, and then extracting a logarithmic power spectrum characteristic as a frequency domain characteristic; and finally, integrating the refined time domain features and the frequency domain features to obtain physical features based on a physical mechanism.
3. The method for generating fault data based on a physically constrained two-domain diffusion model according to claim 2, wherein the hierarchical convolution architecture comprises a double-layer convolution layer, a first layer of convolution layer is used for capturing coarse time domain features including a fault periodicity and an integral waveform structure, and a second layer of convolution layer is used for extracting fine time domain features including transient impact and envelope modulation.
4. The method of generating fault data based on a physically constrained two-domain diffusion model of claim 2, wherein the logarithmic power spectral features include a characteristic frequency, a family of harmonics, and a resonance band.
5. The fault data generation method based on the physical constraint two-domain diffusion model according to claim 1, wherein the step 2 comprises defining a forward diffusion process as a T-step markov chain, and introducing frequency adaptive noise determined by a fas mechanism therein; The reverse denoising process adopts a denoising network of a U-Net architecture to perform reverse generation, and physical characteristics are explicitly integrated into each level of the U-Net through a multi-scale self-adaptive injection mechanism.
6. The method for generating fault data based on a physically constrained two-domain diffusion model according to claim 1, wherein the FANS mechanism comprises: Defining frequency importance weights for the signal spectrum to quantify the contribution of each frequency, the fault signature bin value trending towards 1, the background noise value trending towards 0; Then, self-adaptive noise modulation is carried out, a frequency selective mask is defined as a function which takes a value on a frequency domain, and the function reduces the noise intensity of a high-energy fault zone by controlling the protection intensity and the frequency importance weight; and finally inverting the expression of the modulation noise in the frequency domain back to the time domain according to the frequency selective mask to obtain the frequency self-adaptive noise.
7. The fault data generation method based on the physical constraint two-domain diffusion model according to claim 5, wherein the multi-scale adaptive injection mechanism comprises judging a region or a frequency band which needs to be guided by physical features according to the feature distribution of the current generated signal through a gating item.
8. The method for generating fault data based on a physically constrained two-domain diffusion model according to claim 1, wherein said constructing an energy consistency loss comprises: Collecting noise signals and predicted noise of each time step of the inverse denoising process by the diffusion model, and inversely deducing and estimating an initial signal predicted value; And respectively calculating time domain energy and frequency domain energy of the initial signal predicted value, and constructing a relative differential loss function as an energy consistency loss according to the calculation result so as to quantify the difference between the two.
9. The method for generating fault data based on a physically constrained two-domain diffusion model according to claim 1, wherein the diffusion model is a denoising diffusion probability model DDPM.
10. The fault data generation system based on the physical constraint double-domain diffusion model is characterized by comprising the following components: The first processing module is used for acquiring target real fault data corresponding to the target object; The second processing module is used for inputting the target real fault data into a fault data generation model, obtaining target simulation fault data output by the target fault data generation model, wherein the fault data generation model is a diffusion model, a FANS mechanism is introduced in a forward diffusion process, physical features are conditionally injected in a reverse denoising process, and finally, a total loss function is built and optimized based on energy consistency and denoising reconstruction loss, the FANS mechanism is used for adaptively adjusting noise intensity based on frequency spectrum energy distribution, and the physical features are a combination of time domain features and frequency domain features extracted from the real fault data.

Description

Fault data generation method and system based on physical constraint double-domain diffusion model Technical Field The invention relates to the technical field of predictive maintenance, in particular to a fault data generation method and system based on a physical constraint double-domain diffusion model. Background Accurate and rapid fault diagnosis is critical to ensure operational safety of complex electromechanical equipment systems and to minimize downtime. In recent years, the deep learning method has made remarkable progress in the field of fault diagnosis, but it faces serious data imbalance problems in practical industrial applications. The equipment can continuously run and collect a large amount of data under normal conditions, and the fault conditions often lead to the equipment to be immediately shut down for maintenance, so that fault samples are extremely scarce. Such data imbalance results in a deep learning model that is difficult to fully learn fault characteristics, severely affecting diagnostic accuracy. To solve the data imbalance problem, researchers have attempted to extend the failure samples using data generation techniques. The generation of the countermeasure network (GAN) generates new samples through the countermeasure training of the discriminator and the generator, but the problems of unstable training, pattern collapse, gradient disappearance and the like exist, and the diversity of the generated samples is insufficient. Variational self-encoders (VAEs) often suffer from posterior collapse, and the resulting samples tend to be blurred, making it difficult to preserve the fine features of the fault signal. More importantly, these methods generally lack explicit modeling of the physical characteristics of the mechanical vibration signal, and the resulting data may violate basic physical laws, resulting in reduced performance of downstream diagnostic algorithms. The Denoising Diffusion Probability Model (DDPM) is used as an emerging generation model, and good performance and training stability are shown in the fields of image generation and the like by constructing a reversible noise diffusion process. However, applying it directly to industrial fault diagnosis data generation presents challenges. The mechanical vibration signal has a highly non-uniform frequency domain energy distribution, the fault signature is typically concentrated in a particular frequency band, while the global uniform noise scheduling strategy adopted by standard DDPM can prematurely destroy these weak fault signatures during forward diffusion. For example, bearing faults occur following strict physical mechanisms, and inner ring faults, outer ring faults and rolling body faults correspond to different impact modes and characteristic frequencies respectively, and existing DDPM lacks explicit modeling of these physical constraints and may generate signals against the physical mechanisms. Disclosure of Invention The standard diffusion model, while excellent in the field of image generation, faces many challenges when applied directly to mechanical vibration signals. The frequency domain energy distribution of the mechanical vibration signals is highly non-uniform, the fault characteristics are often concentrated in a specific frequency band, and a standard diffusion model adopts a globally uniform noise scheduling strategy, so that the weak fault characteristics are damaged too early in the forward diffusion process, and are difficult to recover accurately during reverse generation. More importantly, the generation of industrial faults follows strict physical laws, different fault types correspond to different impact modes and characteristic frequencies, and existing generative models generally lack explicit modeling of these physical constraints. Furthermore, when the generation model contains a time-domain and frequency-domain dual-branch structure, if explicit constraints are absent, the two branches can independently optimize the respective reconstruction targets, resulting in the generated signal violating the conservation of energy law, which is physically unrealizable. In order to solve the technical problems, the application provides a fault data generation method and a fault data generation system based on a physical constraint double-domain diffusion model. The fault data generation method based on the physical constraint double-domain diffusion model comprises the following steps: Step 1, extracting physical characteristics of target real fault data complementary in time domain and frequency domain; step 2, a diffusion model is built, wherein the generation process of the diffusion model comprises a forward diffusion process and a reverse denoising process, a FANS mechanism is introduced in the forward diffusion process, a condition constraint comprising the physical characteristics is injected in the reverse denoising process, and the FANS mechanism is used for adaptively adjusting noise intens