CN-122020005-A - Interpretable extremely simplified intelligent data enhancement method, system and medium suitable for fault diagnosis of high-end equipment

CN122020005ACN 122020005 ACN122020005 ACN 122020005ACN-122020005-A

Abstract

The invention discloses an interpretable extremely simplified intelligent data enhancement method, a system and a medium suitable for fault diagnosis of high-end equipment, wherein the method comprises the following steps of obtaining vibration signal data; the method comprises the steps of carrying out standardized processing and dividing the model into a training set and a testing set, constructing an interpretable encoder and a decoder, training the interpretable encoder and the decoder, freezing parameters, mapping the training set into a variable set, constructing a greatly simplified data generation model, training the model by using the variable set, sampling to generate a new interpretable variable, carrying out back-translation to obtain a simulated vibration signal, mixing the simulated vibration signal with a sample of the training set, constructing a balance training data set, training a classifier, carrying out fault diagnosis on the testing set by using the classifier, and outputting a diagnosis result. The invention solves the problem that the reliability of the fault diagnosis model is poor due to the fact that the existing intelligent data enhancement method is unexplainable in the generation process, unstable in training and low in sampling efficiency under the conditions of small samples, unbalanced categories and limited engineering calculation force of high-end equipment.

Inventors

XU KUN
XU JIEYI
WANG HUA
LIN KAI
MIAO XIAODONG

Assignees

南京工业大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. The interpretable extremely simplified intelligent data enhancement method suitable for high-end equipment fault diagnosis is characterized by comprising the following steps of: The method comprises the following steps of S1, obtaining vibration signal data corresponding to different fault types of bearings for high-end equipment under different loads or different rotating speeds; S2, constructing an interpretable encoder and an interpretable decoder based on a double-layer residual stacking structure, training the interpretable encoder and the interpretable decoder through a joint objective function based on reconstruction errors and sparse regularities, freezing parameters of the interpretable encoder and the interpretable decoder after training is finished, and mapping vibration signal data in a training set into a low-dimensional, readable and parameterized interpretable variable set through the interpretable encoder; s3, constructing a very simplified data generation model, and training the very simplified data generation model by using an interpretable variable set; S4, inputting a new interpretable variable generated by sampling a few fault types into an interpretable decoder with frozen parameters for back interpretation to obtain a time domain analog vibration signal; And S5, training the balance training data set by using a classifier, performing fault diagnosis on the test set by using the trained classifier, and outputting a diagnosis result.
2. The method for enhancing interpretable and extremely simplified intelligent data applicable to high-end equipment fault diagnosis as recited in claim 1, wherein said interpretable encoder is a double-headed encoder structure, and outputs physical parameter variables with contribution coefficients of each physical component in the vibration signal corresponding to each physical component.
3. The method for enhancing interpretable extremely simplified intelligent data applicable to high-end equipment fault diagnosis according to claim 2, wherein in step S2, the double-layer residual error stacking structure represents a vibration signal as a two-stage fixed residual error mechanism of a steady-state background component and an impact transient component, a first stage fits and reconstructs the background steady-state component in the vibration signal based on a first function packet, a second stage fits and reconstructs the impact transient component in the residual error after the first stage reconstruction based on a second function packet, and the noise intensity in the signal is represented by a final residual error statistic after the second stage reconstruction.
4. The method for enhancing interpretable extremely simplified intelligent data for high-end equipment fault diagnosis as recited in claim 3, wherein said interpretable variable is composed of contribution coefficients of background steady-state components and physical parameters, contribution coefficients of impact transient components and physical parameters, and statistics of noise intensity.
5. The method for enhancing interpretable and extremely simplified intelligent data applicable to high-end equipment fault diagnosis according to claim 1, wherein in step S2, the combined objective function expression based on reconstruction error and sparse regularization is specifically: ; Wherein, the For the joint objective function, x is the original vibration signal, The signal is reconstructed and the signal is then processed, The contribution coefficients for the components are calculated, Is a regularization coefficient.
6. The method for enhancing interpretable extremely simplified intelligent data for high-end equipment fault diagnosis according to claim 1, wherein in step S3, said extremely simplified data generation model generates a backbone network for conditions constructed using a transducer module.
7. The method for enhancing interpretable extremely simplified intelligent data applicable to high-end equipment fault diagnosis as recited in claim 6, wherein training of said extremely simplified data generation model specifically includes adding noise to interpretable variables to construct noisy inputs, the network takes noisy variables, time steps and class conditions as inputs, and adopts a training mode targeting prediction of clean variables to construct supervisory signals and loss functions in the form of flow rates to optimize conditions to generate parameters of the backbone network.
8. The method for enhancing interpretable extremely simplified intelligent data for high-end equipment fault diagnosis of claim 1, wherein said training set is structured as a class imbalance data set comprising a number of health state samples and fault state samples greater than a number of fault state samples, said balance training data set having a balance of health state samples and fault state samples.
9. A system adapted for an interpretable extremely simplified intelligent data enhancement method for high-end equipment fault diagnosis, comprising: The data acquisition and preprocessing module is used for acquiring vibration signal data corresponding to different fault types of each type of bearing for high-end equipment under different loads and different rotating speeds; the system comprises an interpretable space construction module, an interpretable encoder and an interpretable decoder, a storage module and a storage module, wherein the interpretable space construction module is used for constructing the interpretable encoder and the interpretable decoder based on a double-layer residual error stacking structure; the system comprises a very simplified data generation model construction module, a new interpretable variable generation module, a fault model generation module and a fault model generation module, wherein the very simplified data generation model construction module is used for constructing a very simplified data generation model, and training the very simplified data generation model by using a set of interpretable variables; The balance training data set construction module is used for inputting a new interpretable variable generated by sampling a few fault types into an interpretable decoder with frozen parameters for back interpretation to obtain a time domain analog vibration signal; and the fault diagnosis module is used for training the balance training data set by using the classifier, performing fault diagnosis on the test set by using the trained classifier and outputting a diagnosis result.
10. A computer-readable storage medium having stored thereon a computer program, wherein the computer program causes a computer to execute the interpretable extremely simplified intelligent data enhanced method for high-end equipment fault diagnosis as recited in any one of claims 1-8.

Description

Interpretable extremely simplified intelligent data enhancement method, system and medium suitable for fault diagnosis of high-end equipment Technical Field The invention relates to the technical field of high-end equipment data enhancement and intelligent diagnosis, in particular to an interpretable extremely simplified intelligent data enhancement method, an interpretable extremely simplified intelligent data enhancement system and an interpretable extremely simplified intelligent data enhancement medium suitable for high-end equipment fault diagnosis. Background An intelligent fault diagnosis method based on vibration signals is widely used in high-end equipment. In an actual industrial scene, high-end equipment is in a healthy running state for most of time, fault samples are difficult to continuously acquire and systematically accumulate, meanwhile, complex working condition changes such as multiple loads, multiple rotating speeds, noise disturbance and the like exist in equipment, few samples are difficult to cover real distribution, a training data set generally shows unbalanced characteristics of small samples and categories, and therefore judgment boundaries of a diagnosis model on the few categories are unstable, the misclassification rate is increased, and diagnosis reliability is affected. For unbalanced data problems, the prior art typically expands a minority class and builds a balanced dataset by generating simulated fault samples. The patent CN112649198B proposes a fault sample generation thought based on a generation countermeasure network, a few types of samples are directionally expanded in a condition generation mode to improve diagnosis training under unbalanced data conditions, the patent CN119848668A proposes a sample enhancement/diagnosis scheme based on a denoising diffusion probability model, the samples are enhanced through a diffusion generation mechanism to improve diagnosis performance, and in order to improve the interpretability of a generation process, the patent CN120296522A proposes a related scheme of fusion mechanism modeling and interpretable generation constraint, and attempts are made to improve consistency and interpretable clues of the generated samples by introducing mechanism and constraint items. However, the existing scheme still has the defects under the complex working condition of high-end equipment that on one hand, the scheme based on the generation of an anti-network is generally sensitive to training configuration under the conditions of small samples and strong noise, the generation stability and diversity are easily affected, on the other hand, the scheme based on the denoising diffusion model generally relates to a multi-step reverse sampling process, the sampling cost is high, the difficulty of realizing consistent and controllable generation under the condition of multiple categories is high, meanwhile, the existing related scheme of 'interpretable generation' focuses on providing interpretation clues through network constraint or loss constraint, and still is difficult to form low-dimensional, readable, parameterized and traceable generation representation facing to a vibration signal. In addition, from the engineering deployment perspective, the intelligent diagnosis of high-end equipment often faces the constraints of limited computational effort, short iteration cycle, on-site real-time requirements and the like, and the data enhancement module is easy to train, reproduce and deploy. The existing generation method often introduces a complex generation countermeasure structure or a multi-stage denoising sampling process, so that the training process is unstable, a sampling link is long, the super-parameter sensitivity or the reasoning cost is large, and the rapid landing in an actual diagnosis system is not facilitated. Therefore, on the premise of ensuring the consistency of the generation quality and the category, the construction of the extremely simplified data generation framework with simpler structure, more stable training and more efficient sampling is also of great significance. Disclosure of Invention The invention aims to provide an interpretable extremely simplified intelligent data enhancement method, an interpretable extremely simplified intelligent data enhancement system and a medium suitable for fault diagnosis of high-end equipment, so as to solve the problem that the reliability of a fault diagnosis model is poor due to the fact that the existing intelligent data enhancement method is uninterpretable in a generation process, unstable in training and low in sampling efficiency under the conditions of small samples, unbalanced categories and limited engineering calculation force of the high-end equipment. In order to achieve the above purpose, the present invention adopts the following technical scheme: an interpretable extremely simplified intelligent data enhancement method suitable for high-end equipment fault diagnosis, compri