CN-121786493-B - Cross-domain generation method of fault data of ball screw pair based on unsupervised condition diffusion

CN121786493BCN 121786493 BCN121786493 BCN 121786493BCN-121786493-B

Abstract

The invention discloses a fault data cross-domain generation method of a ball screw pair based on unsupervised condition diffusion, and belongs to the field of fault diagnosis of ball screw pairs. The method comprises the steps of constructing a source domain data set and a target domain data set, constructing an unsupervised condition diffusion model, extracting a training set according to the source domain data set and the target domain data set, taking the training set as input of the unsupervised condition diffusion model, carrying out unsupervised training on the unsupervised condition diffusion model by adopting a weighted joint loss function, establishing bidirectional mapping of data in the source domain data set and data in the target domain data set to obtain a mapping model, obtaining health state data to be enhanced, and generating enhanced fault state vibration data corresponding to working conditions of the health state data to be enhanced according to the mapping model. The method and the device can input the one-dimensional time sequence vibration data of the unknown working condition in the healthy state and generate the one-dimensional time sequence vibration data corresponding to the fault state of the working condition.

Inventors

LIU CHANG
YANG LEI
HE FEIFEI
WU HAIBO

Assignees

昆明理工大学

Dates

Publication Date: 20260505
Application Date: 20260303

Claims (10)

1. The fault data cross-domain generation method of the ball screw pair based on the unsupervised condition diffusion is characterized by comprising the following steps of: s1, vibration data of ball screw pairs in different states, which run under M working conditions, are obtained and used for constructing a source domain data set and a target domain data set, wherein M is a positive integer and M is more than 2; S2, an unsupervised condition diffusion model is built, wherein a one-dimensional domain translator T AB and a one-dimensional domain translator T BA respectively carry out rough translation on a source domain data set and a target domain data set, and a condition encoder is used for generating a condition vector; S3, extracting a training set according to the source domain data set and the target domain data set, taking the training set as input of an unsupervised conditional diffusion model, carrying out unsupervised training on the unsupervised conditional diffusion model by adopting a weighted joint loss function, and establishing bidirectional mapping of data in the source domain data set and data in the target domain data set to obtain a mapping model; S4, acquiring health state data to be enhanced, and generating enhanced fault state vibration data corresponding to the working condition of the health state data to be enhanced according to the mapping model; the unsupervised conditional diffusion model specifically includes: The system comprises two one-dimensional domain translators, namely a one-dimensional domain translator T AB and a one-dimensional domain translator T BA , wherein the one-dimensional domain translator T AB and the one-dimensional domain translator T BA have the same structure and comprise a domain translator encoder, a domain translator decoder and an output layer, the one-dimensional domain translator T AB takes a source domain data set as input data to realize primary domain conversion from a source domain to a target domain, the one-dimensional domain translator T BA takes the target domain data set as input data to realize primary domain conversion from the target domain to the source domain, the structure of the domain translator encoder comprises a one-dimensional convolution block, two stacked ' one-dimensional convolution blocks+first residual blocks ', a first residual block, a one-dimensional convolution block and one-dimensional convolution block, wherein the one-dimensional convolution blocks in the two stacked ' one-dimensional convolution blocks+first residual blocks ' are sequentially connected, the structure of the domain translator decoder comprises a one-dimensional convolution block, two stacked ' first residual blocks+one-dimensional deconvolution blocks ', the two stacked ' one-dimensional convolution blocks ' and the one-dimensional convolution blocks ' are sequentially introduced into the one-dimensional convolution blocks, and the output layer of the one-dimensional translation blocks is translated into the output layer; the condition encoder comprises three one-dimensional convolution blocks, a self-adaptive average pooling layer and a full-connection layer which are sequentially connected, and finally maps to a condition vector with a fixed dimension; the source domain condition diffusion denoising network and the target domain condition diffusion denoising network have the same structure, and all take one-dimensional U-Net as a frame, and comprise a second residual block, a downsampling layer, an upsampling layer, a bottleneck block and an output layer.
2. The method for generating fault data of a ball screw pair based on unsupervised condition diffusion according to claim 1, wherein S1 comprises: S11, under the constant load of the ball screw pair fault simulation test bed, respectively acquiring corresponding health state vibration data and fault state vibration data under various rotating speed working conditions; s12, selecting healthy state vibration data fragments and fault state vibration data fragments with equal data quantity from a plurality of rotating speed working conditions in a non-overlapping sliding sampling mode to serve as a source domain sample source and a target domain sample source respectively; S13, calculating global mean and standard deviation of the source domain sample sources according to working conditions, and carrying out working condition-by-working condition Z-score normalization on the source domain sample sources and the target domain sample sources according to the global mean and standard deviation of the source domain sample sources so as to construct a source domain data set and a target domain data set.
3. The method for generating the fault data of the ball screw pair based on the unsupervised condition diffusion according to claim 1 is characterized in that input data enters a first downsampling layer after passing through two second residual blocks, then enters the second downsampling layer after passing through the two second residual blocks, and downsampling of the input data is completed, output characteristics of the second downsampling layer enter a first upsampling layer after passing through two bottleneck blocks, the output of the first upsampling layer and the input of the second downsampling layer are connected with each other after being spliced in a channel dimension, then enter the second upsampling layer, the output of the second upsampling layer and the input of the first downsampling layer pass through the two second residual blocks after being spliced in the channel dimension, and finally are output by an output layer.
4. The method for generating the cross-domain fault data of the ball screw pair based on the unsupervised condition diffusion according to claim 1, wherein the first residual block comprises two stacked one-dimensional convolution layers and group normalization layers, a SiLU activation function is introduced after the first one-dimensional convolution layer and group normalization layer, and the output of the second one-dimensional convolution layer and group normalization layer is spliced with the original input of the first residual block and then is input with a Tanh activation function, so that the output of the first residual block is obtained.
5. The method for generating fault data of the ball screw pair based on unsupervised condition diffusion according to claim 1, wherein the second residual block comprises two blocks of 'one-dimensional convolution layer + group normalization layer + SiLU activation function', regularization operation, wherein the output of a first block is spliced with the output of a time coding vector projected by a time embedded projection layer, the spliced characteristic is used as the input of the second block after regularization operation, the output of the group normalization layer of the second block is spliced with the original input of the second residual block and is input into SiLU activation function in the second block to obtain the final output of the second residual block, and the time coding vector is obtained by adding a time embedding vector generated by sine-cosine position coding and a conditional vector output by a conditional encoder.
6. The method for generating fault data of the ball screw pair based on the unsupervised condition diffusion according to claim 1, wherein the bottleneck block is composed of a second residual block and a self-attention mechanism.
7. The method for generating fault data of a ball screw pair based on unsupervised condition diffusion according to claim 1, wherein S3 comprises: s31, extracting all samples of N working conditions from a source domain data set and a target domain data set to serve as a training set and the rest as a test set, wherein N is a positive integer and 1< N < M; And S32, training an unsupervised conditional diffusion model, namely taking a training set as input of the unsupervised conditional diffusion model, using AdamW optimizers, carrying out unsupervised training on the unsupervised conditional diffusion model by adopting a weighted joint loss function, and establishing bidirectional mapping of data in a source domain data set and data in a target domain data set to obtain a mapping model, wherein the weighted joint loss function is constructed in a weighted mode according to noise prediction mean square error loss, cyclic consistency L1 loss and multi-resolution STFT log spectrum L1 loss.
8. The method for generating fault data of the ball screw pair based on the unsupervised condition diffusion according to claim 7, wherein the weighted joint loss function is expressed as: ; Wherein, the Predicting mean square error loss for noise respectively Loop consistency L1 penalty Multi-resolution STFT log spectrum L1 penalty Weight coefficient of (c) in the above-mentioned formula (c).
9. The method for generating fault data of a ball screw pair across domains based on unsupervised condition diffusion according to claim 8, wherein the weight coefficient value satisfies the following condition 。
10. The method for generating fault data of a ball screw pair based on unsupervised condition diffusion according to claim 1, wherein S4 comprises: S41, roughly translating the health state data to be enhanced by using a one-dimensional domain translator T AB in a mapping model, taking the one-dimensional domain translator T AB as input of a source domain condition diffusion denoising network in the mapping model, and according to the preset release step number Forward diffusion is carried out to generate an intermediate noise state signal ; S42, taking the health state data to be enhanced as the input of a condition encoder in the mapping model to obtain a condition vector In the condition vector Under the guidance of (1), the target domain condition in the mapping model spreads the denoising network from time step Start to intermediate noise state signal Performing inverse denoising sampling until And gradually generating the enhanced fault state vibration data corresponding to the working condition of the health state data to be enhanced.

Description

Cross-domain generation method of fault data of ball screw pair based on unsupervised condition diffusion Technical Field The invention relates to a fault data cross-domain generation method of a ball screw pair based on unsupervised condition diffusion, and belongs to the field of fault diagnosis of ball screw pairs. Background The ball screw pair is used as a key transmission component in electromechanical equipment and is widely applied to the high-precision fields of numerical control machine tools, robots, aerospace and the like. The method has important significance for fault diagnosis and health monitoring of the ball screw pair, and is especially in the background of the increasing development of industrial automation and intelligent manufacturing. The conventional fault diagnosis method of the ball screw pair mainly relies on supervised learning models such as Support Vector Machines (SVMs) and Convolutional Neural Networks (CNNs). These methods typically rely on a large amount of labeled fault data to train to identify different types of fault patterns. However, due to safety, cost and equipment availability limitations, it is inherently difficult to collect fault samples on an industrial site, and even more so, it is almost impossible to cover real fault data of different load, rotation speed and other working conditions. The generalization ability and robustness of the model are often challenged due to data scarcity and data lack of diversification, resulting in unsatisfactory results in practical applications. In order to cope with the problems of data scarcity and maldistribution, etc., data enhancement techniques are widely used as an effective solution. Existing conventional data enhancement methods, such as generation of a countermeasure network (GAN) and a variational self-encoder (VAE), augment the training dataset by generating synthetic samples. However, these methods have some key limitations. First, generating a countermeasure network (GAN) may face a pattern collapse problem, that is, the generated samples lack diversity and are difficult to cover various aspects of the fault space, while the VAE model is capable of generating various samples, but the generated samples often have a problem of cross-domain instability, and particularly, in the case of unpaired data, the effect of generating fault domain data by using healthy domain data is poor. More importantly, conventional data enhancement methods typically rely on a large number of paired health and fault data (as trained on health and fault data under the same conditions), but such data requirements are often difficult to meet in a practical industrial environment. Secondly, the existing methods generally rely on feature extraction of one-dimensional time series data and then feature enhancement, or enhancement of a time-frequency diagram of the one-dimensional time series data, which often lose a large amount of original time series information. Although the extracted features and the time-frequency diagram can provide certain information, the original time-domain information can reflect the dynamic response and the instantaneous change of the equipment more directly, so that richer fault diagnosis information is provided. In view of this, the present invention has been made. Disclosure of Invention The invention provides a fault data cross-domain generation method of a ball screw pair based on unsupervised condition diffusion, which obtains a mapping model from healthy state vibration data to fault state vibration data under unsupervised training without pairing data (as the training of healthy data and fault data under the same working condition), and can realize the input of one-dimensional time sequence vibration data of unknown working conditions (which refer to the working conditions of the model which are not used in the training process) in the healthy state to generate one-dimensional time sequence vibration data corresponding to the fault state of the working condition. The technical scheme of the invention is that the cross-domain generation method of the fault data of the ball screw pair based on unsupervised condition diffusion comprises the following steps: s1, vibration data of ball screw pairs in different states, which run under M working conditions, are obtained and used for constructing a source domain data set and a target domain data set, wherein M is a positive integer and M is more than 2. S2, an unsupervised condition diffusion model is built, wherein a one-dimensional domain translator T AB and a one-dimensional domain translator T BA are used for respectively carrying out rough translation on a source domain data set and a target domain data set, a condition encoder is used for generating a condition vector, and a source domain condition diffusion denoising network and a target domain condition diffusion denoising network are used for realizing cross-domain condition diffusion and denoising generation u