Ball screw pair fault data cross-domain generation method based on unsupervised conditional diffusion

Abstract

This invention discloses a cross-domain generation method for ball screw pair fault data based on unsupervised conditional diffusion, belonging to the field of ball screw pair fault diagnosis. The method includes constructing a source domain dataset and a target domain dataset; constructing an unsupervised conditional diffusion model; extracting a training set based on the source and target domain datasets, using the training set as input to the unsupervised conditional diffusion model, and performing unsupervised training on the unsupervised conditional diffusion model using a weighted joint loss function to establish a bidirectional mapping between the data in the source and target domain datasets, obtaining a mapping model; acquiring the health state data to be enhanced; and generating enhanced fault state vibration data corresponding to the operating conditions of the health state data to be enhanced based on the mapping model. This invention can generate one-dimensional time-series vibration data of the corresponding fault state from one-dimensional time-series vibration data of an unknown operating condition in a healthy state.

Description

Technical Field

[0001] This invention relates to a cross-domain generation method for ball screw pair fault data based on unsupervised conditional diffusion, belonging to the field of ball screw pair fault diagnosis. Background Technology

[0002] Ball screw assemblies, as key transmission components in electromechanical equipment, are widely used in high-precision fields such as CNC machine tools, robotics, and aerospace. Fault diagnosis and health monitoring of ball screw assemblies are of great significance, especially against the backdrop of increasing industrial automation and intelligent manufacturing. Traditional ball screw assembly fault diagnosis methods mainly rely on supervised learning models, such as Support Vector Machines (SVM) and Convolutional Neural Networks (CNN). These methods typically rely on large amounts of labeled fault data for training to identify different types of fault modes. However, due to limitations in safety, cost, and equipment availability, it is already difficult to collect fault samples in industrial settings, let alone cover real-world fault data under different loads, speeds, and other operating conditions. Because of data scarcity and lack of diversity, the generalization ability and robustness of the models are often challenged, resulting in unsatisfactory performance in practical applications.

[0003] To address the issues of data scarcity and uneven distribution, data augmentation techniques have been widely adopted as an effective solution. Existing traditional data augmentation methods, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), expand the training dataset by generating synthetic samples. However, these methods have some key limitations. First, GANs may face pattern collapse, meaning the generated samples lack diversity and struggle to cover all aspects of the fault space. While VAE models can generate diverse samples, their generated samples often exhibit cross-domain instability, especially in the case of unpaired data, where using health domain data to generate fault domain data performs poorly. More importantly, traditional data augmentation methods typically rely on large amounts of paired health and fault data (such as training with health and fault data under the same operating condition), but this data requirement is often difficult to meet in real-world industrial environments. Second, existing methods often rely on feature extraction from one-dimensional time-series data followed by feature augmentation, or augmentation of the time-frequency plot of one-dimensional time-series data. These practices often result in the loss of significant original temporal information. Although the extracted features and time-frequency plots can provide some information, the original time-domain information can more directly reflect the dynamic response and instantaneous changes of the equipment, thus providing richer fault diagnosis information.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] This invention provides a cross-domain generation method for ball screw pair fault data based on unsupervised conditional diffusion. Under unsupervised training without the need for paired data (such as training with health data and fault data under the same working condition), a mapping model from healthy state vibration data to fault state vibration data is obtained. It can generate one-dimensional time-series vibration data of the corresponding working condition fault state by inputting one-dimensional time-series vibration data of an unknown working condition (referring to a working condition that has not been used in the training process) in the healthy state.

[0006] The technical solution of this invention is: a method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion, comprising:

[0007] S1. Obtain vibration data of ball screw pairs in different states operating under M working conditions to construct source domain dataset and target domain dataset; where M is a positive integer and M>2.

[0008] S2. Constructing an unsupervised conditional diffusion model: One-dimensional domain translator T AB One-dimensional domain translator T BA Coarse translations are performed on the source domain dataset and the target domain dataset, respectively; a conditional encoder is used to generate conditional vectors; and source domain conditional diffusion denoising networks and target domain conditional diffusion denoising networks are used to achieve cross-domain conditional diffusion and denoising generation under the guidance of conditional vectors.

[0009] S3. Extract training sets based on the source domain dataset and the target domain dataset. Use the training sets as input to the unsupervised conditional diffusion model and use a weighted joint loss function to train the unsupervised conditional diffusion model in an unsupervised manner. Establish a bidirectional mapping between the data in the source domain dataset and the data in the target domain dataset to obtain the mapping model.

[0010] S4. Obtain the health status data to be enhanced; based on the mapping model, generate enhanced fault state vibration data corresponding to the working conditions of the health status data to be enhanced.

[0011] Further, S1 includes:

[0012] S11. Under constant load on the ball screw pair fault simulation test bench, collect vibration data of healthy state and vibration data of fault state under various speed conditions.

[0013] S12. Using a non-overlapping sliding sampling method, select healthy state vibration data segments and fault state vibration data segments with equal data volume from various speed conditions as the source domain sample source and target domain sample source, respectively.

[0014] S13. Calculate the global mean and standard deviation of the source domain sample sources according to the working conditions, and use the global mean and standard deviation of the source domain sample sources to perform working condition-by-working-condition Z-score normalization on the source domain sample sources and the target domain sample sources, so as to construct the source domain dataset and the target domain dataset.

[0015] Furthermore, the unsupervised conditional diffusion model specifically includes:

[0016] Two one-dimensional domain translators: One-dimensional domain translator T AB One-dimensional domain translator T BA The one-dimensional domain translator T AB One-dimensional domain translator T BA They have the same structure, including a domain translator encoder, a domain translator decoder, and an output layer. The one-dimensional domain translator T... AB Using the source domain dataset as input, a preliminary domain transformation from the source domain to the target domain is achieved; a one-dimensional domain translator T BA Using the target domain dataset as input data, a preliminary domain transformation from the target domain to the source domain is achieved. The structure of the domain translator encoder includes a one-dimensional convolutional block, two stacked "one-dimensional convolutional blocks + first residual blocks", a first residual block, and a one-dimensional convolutional block connected in sequence. Downsampling is introduced into the one-dimensional convolutional block of the two stacked "one-dimensional convolutional blocks + first residual blocks". The structure of the domain translator decoder includes a one-dimensional convolutional block, two stacked "first residual blocks + one-dimensional deconvolutional blocks", and upsampling is introduced into the one-dimensional deconvolutional block of the two stacked "first residual blocks + one-dimensional deconvolutional blocks". The output of the domain translator decoder is used as the input of the output layer, and the output of the output layer is the output of the one-dimensional domain translator.

[0017] The conditional encoder comprises three sequentially connected one-dimensional convolutional blocks, an adaptive average pooling layer, and a fully connected layer, ultimately mapping to a fixed-dimensional conditional vector.

[0018] The source domain conditional diffusion denoising network and the target domain conditional diffusion denoising network have the same structure, both based on a one-dimensional U-Net framework, including a second residual block, a downsampling layer, an upsampling layer, a bottleneck block, and an output layer. The input data passes through two second residual blocks and then enters the first downsampling layer, and then through two more second residual blocks before entering the second downsampling layer, completing the downsampling of the input data. The output features of the second downsampling layer pass through two bottleneck blocks and then enter the first upsampling layer. The output of the first upsampling layer and the input of the second downsampling layer are concatenated along the channel dimension and then connected to two second residual blocks, before entering the second upsampling layer. The output of the second upsampling layer and the input of the first downsampling layer are concatenated along the channel dimension and then passed through two more second residual blocks before finally being output by the output layer.

[0019] Furthermore, the first residual block includes two stacked "one-dimensional convolutional layers + group normalization layers". The first "one-dimensional convolutional layer + group normalization layer" is followed by the SiLU activation function. The output of the second "one-dimensional convolutional layer + group normalization layer" is concatenated with the original input of the first residual block and then input into the Tanh activation function to obtain the output of the first residual block.

[0020] Furthermore, the second residual block includes two blocks of "one-dimensional convolutional layer + group normalization layer + SiLU activation function" and a regularization operation. The output of the first block is concatenated with the output of the time-encoded vector projected by the time embedding projection layer. The concatenated features are regularized and used as the input of the second block. The output of the group normalization layer of the second block is concatenated with the original input of the second residual block and input into the SiLU activation function in the second block to obtain the final output of the second residual block. The time-encoded vector is obtained by embedding the time embedding vector generated by sine-cosine position encoding and adding it to the conditional vector output by the conditional encoder.

[0021] Furthermore, the bottleneck block consists of a second residual block followed by a self-attention mechanism.

[0022] Further, S3 includes:

[0023] S31. From the source domain dataset and the target domain dataset, extract all samples of N working conditions as the training set, and the rest as the test set; where N is a positive integer, and 1 <N<M。

[0024] S32. Training of the unsupervised conditional diffusion model: The training set is used as the input of the unsupervised conditional diffusion model. The AdamW optimizer is used, and a weighted joint loss function is adopted to train the unsupervised conditional diffusion model in an unsupervised manner. A bidirectional mapping between the data in the source domain dataset and the data in the target domain dataset is established to obtain the mapping model. The weighted joint loss function is constructed in a weighted manner based on the noise prediction mean square error loss, the cycle consistency L1 loss, and the multi-resolution STFT log spectrum L1 loss.

[0025] Furthermore, the weighted joint loss function is expressed as follows:

[0026] ;

[0027] in, The mean square error loss for noise prediction is respectively L1 loss for cycle consistency and multi-resolution STFT logarithmic spectrum L1 loss The weighting coefficients.

[0028] Furthermore, the weighting coefficients satisfy the following values: .

[0029] Further, S4 includes:

[0030] S41. Utilize a one-dimensional domain translator T in a mapping model to process enhanced health status data. AB A rough translation, used as input to the source domain conditional diffusion denoising network in the mapping model, is performed according to a preset number of release steps. Perform forward diffusion to generate intermediate noise state signals. ;

[0031] S42. Using the health status data to be enhanced as input to the condition encoder in the mapping model, a condition vector is obtained. In the condition vector Under the guidance of the mapping model, the target domain conditional diffusion denoising network from time step Start with intermediate noise state signal Perform reverse denoising sampling until Gradually generate enhanced fault state vibration data corresponding to the working conditions of the health state data to be enhanced.

[0032] The beneficial effects of this invention are:

[0033] This invention, based on an unsupervised conditional diffusion model trained with multiple loss mechanisms, yields a mapping model capable of generating faulty one-dimensional vibration data from healthy one-dimensional vibration data. This mapping model not only effectively preserves background information related to the operating conditions and processing environment but also enables cross-domain generation from healthy vibration data under unknown operating conditions to faulty vibration data. The method of this invention directly enhances one-dimensional time-series data, preserving more useful time-series information and avoiding the data details that may be lost during time-frequency diagram conversion or feature extraction in traditional methods, thereby improving the quality and diversity of the generated data. Furthermore, the unsupervised conditional diffusion model can be trained without paired data, reducing reliance on large amounts of labeled data and demonstrating strong potential for industrial applications. Attached Figure Description

[0034] Figure 1 This is a flowchart of the present invention.

[0035] Figure 2 This is a diagram of a one-dimensional domain translator.

[0036] Figure 3 This is a diagram of the source domain conditional diffusion denoising network and the target domain conditional diffusion denoising network.

[0037] Figure 4 This is the structure diagram of the second residual block.

[0038] Figure 5 This is a flowchart of cross-domain data enhancement based on the mapping model of the present invention.

[0039] Figure 6 It is a time-domain waveform diagram of the enhanced fault state vibration data of the lead screw pitting fault provided in Example 2.

[0040] Figure 7 It is a time-domain waveform diagram based on the vibration data of the actual fault state provided in Example 2.

[0041] Figure 8 The vibration data spectrum diagram of the enhanced fault state of the lead screw pitting fault provided in Example 2 is based on the example 2.

[0042] Figure 9 It is a spectrum diagram of vibration data under real fault conditions provided in Example 2.

[0043] Figure 10 It is a comparison chart of the amplitude distribution of the enhanced fault state vibration data and the actual fault state vibration data of the lead screw pitting fault provided in Example 2. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0045] Example 1: As Figures 1-5 As shown, a cross-domain generation method for ball screw pair fault data based on unsupervised conditional diffusion includes:

[0046] S1. Obtain vibration data of ball screw pairs in different states operating under M working conditions to construct source domain dataset and target domain dataset; where M is a positive integer and M>2.

[0047] S2. Constructing an unsupervised conditional diffusion model: One-dimensional domain translator T AB One-dimensional domain translator T BA Coarse translations are performed on the source domain dataset and the target domain dataset, respectively; a conditional encoder is used to generate conditional vectors; and source domain conditional diffusion denoising networks and target domain conditional diffusion denoising networks are used to achieve cross-domain conditional diffusion and denoising generation under the guidance of conditional vectors.

[0048] S3. Extract training sets based on the source domain dataset and the target domain dataset. Use the training sets as input to the unsupervised conditional diffusion model. Use the AdamW optimizer and a weighted joint loss function to train the unsupervised conditional diffusion model in an unsupervised manner. Establish a bidirectional mapping between the data in the source domain dataset and the data in the target domain dataset to obtain the mapping model.

[0049] S4. Obtain the health status data to be enhanced; based on the mapping model, generate enhanced fault state vibration data corresponding to the working conditions of the health status data to be enhanced.

[0050] Further, S1 includes:

[0051] S11. Under constant load on the ball screw pair fault simulation test bench, collect vibration data of healthy state and vibration data of fault state under various speed conditions.

[0052] S12. Using a non-overlapping sliding sampling method, select healthy state vibration data segments and fault state vibration data segments with equal data volume from various speed conditions as the source domain sample source and target domain sample source, respectively.

[0053] S13. Calculate the global mean and standard deviation of the source domain sample sources according to the working conditions, and use the global mean and standard deviation of the source domain sample sources to perform working condition-by-working-condition Z-score normalization on the source domain sample sources and the target domain sample sources, so as to construct the source domain dataset and the target domain dataset.

[0054] Furthermore, the unsupervised conditional diffusion model specifically includes:

[0055] Two one-dimensional domain translators: One-dimensional domain translator T AB One-dimensional domain translator T BA The one-dimensional domain translator T AB One-dimensional domain translator T BA They have the same structure, including a domain translator encoder, a domain translator decoder, and an output layer. The one-dimensional domain translator T... AB Using the source domain dataset as input, a preliminary domain transformation from the source domain to the target domain is achieved; a one-dimensional domain translator T BAUsing the target domain dataset as input data, a preliminary domain transformation from the target domain to the source domain is achieved. The encoder structure of the domain translator includes a sequentially connected one-dimensional convolutional block, two stacked "one-dimensional convolutional blocks + first residual blocks", a first residual block, and a one-dimensional convolutional block. Downsampling operations are introduced in the one-dimensional convolutional block of the two stacked "one-dimensional convolutional blocks + first residual blocks" to extract features from the input data by downsampling layer by layer during the convolution operation. The decoder structure of the domain translator includes a sequentially connected one-dimensional convolutional block, two stacked "first residual blocks + one-dimensional deconvolutional blocks" to introduce an upsampling operation in the one-dimensional deconvolutional block of the two stacked "first residual blocks + one-dimensional deconvolutional blocks" to recover features by progressively upsampling during the deconvolution operation. According to the temporal resolution; the output layer adopts a one-dimensional convolutional layer combined with the Tanh activation function, and uses the output of the domain translator decoder as the input of the output layer, and the output of the output layer is the output of the one-dimensional domain translator; the one-dimensional convolutional block is composed of a one-dimensional convolutional layer, a group normalization layer, and a SiLU activation function connected in sequence; the one-dimensional deconvolutional block is composed of a one-dimensional deconvolutional layer, a group normalization layer, and a SiLU activation function connected in sequence; the first residual block includes two stacked "one-dimensional convolutional layer + group normalization layer", the first "one-dimensional convolutional layer + group normalization layer" is followed by the SiLU activation function, and the output of the second "one-dimensional convolutional layer + group normalization layer" is concatenated with the original input of the first residual block and then input into the Tanh activation function to obtain the output of the first residual block.

[0056] The conditional encoder comprises three sequentially connected one-dimensional convolutional blocks, an adaptive average pooling layer, and a fully connected layer, ultimately mapping to a fixed-dimensional conditional vector. Each one-dimensional convolutional block consists of a one-dimensional convolutional layer, a group normalization layer, and a SiLU activation function connected sequentially. The one-dimensional convolutional block progressively compresses the temporal dimension using different kernel sizes and strides. The group normalization layer and SiLU activation function are placed after the convolutional layers of the convolutional block to enhance feature stability and non-linear expression. The adaptive average pooling layer takes the output of the one-dimensional convolutional block as input, compresses it to one dimension, flattens it, and then uses it as input to the fully connected layer, ultimately mapping to a fixed-dimensional conditional vector.

[0057] The source domain conditional diffusion denoising network and the target domain conditional diffusion denoising network. The source domain conditional diffusion denoising network and the target domain conditional diffusion denoising network have the same structure. Both are based on a one-dimensional U-Net framework and include second residual blocks, downsampling layers, upsampling layers, bottleneck blocks, and output layers. The input data passes through two second residual blocks and then enters the first downsampling layer. After passing through two more second residual blocks, it enters the second downsampling layer to complete the downsampling of the input data. The output features of the second downsampling layer pass through two bottleneck blocks and then enter the first upsampling layer to start upsampling. The output of the first upsampling is concatenated with the input of the second downsampling layer in the channel dimension and then connected to two second residual blocks. Then it enters the second upsampling layer. The output of the second upsampling layer is concatenated with the input of the first downsampling layer in the channel dimension and then passes through two more second residual blocks. Finally, the noise prediction is output by the output layer. The second residual block includes two "one-dimensional convolutional layer + group normalization layer + SiLU activation function" blocks and a regularization operation. The output of the first block is concatenated with the output of the time encoding vector projected by the time embedding projection layer and then passed through the regularization operation as the input of the second block. The output of the group normalization layer of the second block is concatenated with the original input of the second residual block and input into the SiLU activation function in the second block to obtain the final output of the second residual block. Among them, the time encoding vector is obtained by adding the time embedding vector generated by the sine-cosine position encoding and the conditional vector output by the conditional encoder. The time embedding projection layer consists of a SiLU activation function and a fully connected layer. The downsampling layer is composed of a one-dimensional convolutional layer. The upsampling layer is composed of a one-dimensional transposed convolutional layer. The bottleneck block consists of a second residual block followed by a self-attention mechanism. The output layer is composed of a group normalization layer, a SiLU activation function, and a one-dimensional convolutional layer connected in sequence.

[0058] Further, the S3 includes:

[0059] S31. In the source domain dataset and the target domain dataset, all samples of N working conditions are extracted as the training set, and the rest are used as the test set. Here, N is a positive integer, and 1 < N < M. Exemplarily, if the source domain dataset and the target domain dataset involve 3 working conditions, all samples of 2 working conditions can be extracted as the training set. If the source domain dataset and the target domain dataset involve 4 working conditions, all samples of 2 or 3 working conditions can be extracted as the training set.

[0060] S32. Training of the unsupervised conditional diffusion model: The training set is used as the input of the unsupervised conditional diffusion model. The AdamW optimizer is used, and a weighted joint loss function is adopted to train the unsupervised conditional diffusion model in an unsupervised manner. A bidirectional mapping between the data in the source domain dataset and the data in the target domain dataset is established to obtain the mapping model. The weighted joint loss function is constructed in a weighted manner based on the noise prediction mean square error loss, the cycle consistency L1 loss, and the multi-resolution STFT log spectrum L1 loss.

[0061] Furthermore, the weighted joint loss function is expressed as follows:

[0062] ;

[0063] in, The mean square error loss for noise prediction is respectively L1 loss for cycle consistency and multi-resolution STFT logarithmic spectrum L1 loss The weighting coefficients.

[0064] Furthermore, the weighting coefficients satisfy the following values: .

[0065] The noise prediction mean square error loss is calculated independently by the source domain conditional diffusion denoising network and the target domain conditional diffusion denoising network during their respective diffusion processes, and then summed, as shown in the following formula:

[0066] ;

[0067] in, and The original vibration data input to the source domain conditional diffusion denoising network and the target domain conditional diffusion denoising network are respectively (i.e., To train health status samples from the source domain dataset, (For fault state samples from the target domain dataset in the training set); For the diffusion time step of random sampling; Standard Gaussian noise; and They represent and At time step The noise-added state; This indicates that the source domain conditional diffusion denoising network, given a conditional vector, is in operation. Below The noise prediction output, This indicates that the target domain conditional diffusion denoising network, given a conditional vector, is in operation. Below The noise prediction output; and These are cross-domain conditional vectors extracted by the conditional encoder from pseudo-health data and pseudo-fault data generated by the one-dimensional domain translator, respectively. Represents the L2 norm; This represents the joint expectation of the source domain data distribution, the target domain data distribution, the uniform distribution of the diffusion time step, and the standard Gaussian noise distribution.

[0068] The cycle consistency L1 loss is calculated separately by the two one-dimensional domain translators and then summed, as shown in the following formula:

[0069] ;

[0070] in, Represents a one-dimensional domain translator T AB right The positive translation result; Represents a one-dimensional domain translator T BA right The positive translation result; express The reconstructed signal after forward translation followed by backward translation; express The reconstructed signal after forward translation followed by backward translation; Represents the L1 norm; For one-dimensional domain translator T AB Calculate the expectation of the cyclic reconstruction error. For one-dimensional domain translator T BA Calculate the expected value of the cyclic reconstruction error.

[0071] The multi-resolution STFT logarithmic spectrum L1 loss is determined by the one-dimensional domain translator T. AB One-dimensional domain translator T BA The translation results are weighted and summed with the actual signals, as shown in the following formula:

[0072] ;

[0073] in, Number of resolutions; For the first STFT window length; For the corresponding weights; Indicates the window length to use Short-time Fourier transform (e.g.) Indicates the window length to use right Find the short-time Fourier transform (and so on for others); Indicates taking the complex amplitude spectrum; It is the numerical stability constant; Represents the L1 norm; The expression represents the statistical average of the weighted L1 norm of the logarithmic amplitude spectrum difference between the translation result and the real signal at different STFT window scales over all training samples, i.e., the mathematical expectation. This indicates a logarithmic transformation of the STFT amplitude spectrum.

[0074] Further, S4 includes:

[0075] S41. Treatment of Enhanced Health Status Data Using a one-dimensional domain translator T in a mapping model AB Rough translation As the input to the source domain conditional diffusion denoising network in the mapping model, it is released according to a preset number of steps. Perform forward diffusion to generate intermediate noise state signals. The calculation formula is:

[0076] ;

[0077] in, This represents the total number of diffusion steps; This is the cumulative noise scheduling coefficient, with a value ranging from (0,1); Represents a normal distribution; This represents random noise sampled from a standard normal distribution.

[0078] For example, health status samples from the source domain dataset in the test set can be used as health status data to be enhanced; or health status data collected under any unknown working conditions can be used as health status data to be enhanced.

[0079] S42. Using the health status data to be enhanced as input to the condition encoder in the mapping model, a condition vector is obtained. In the condition vector Under the guidance of the mapping model, the target domain conditional diffusion denoising network from time step Start with intermediate noise state signal Perform reverse denoising sampling until Gradually generate enhanced fault state vibration data corresponding to the operating conditions of the health state data to be enhanced. The calculation formula is:

[0080] ;

[0081] in, Indicates the use of time step Begin, proceed in sequence ,common The process of gradually reversing noise reduction; Health status data to be enhanced At time step The noise-added state; This indicates that the target domain conditional diffusion denoising network, given a conditional vector, is in operation. Below The noise prediction output.

[0082] S43. Using the maximum mean difference (MMD) and Fréchet Inception distance (FID), the vibration data under enhanced fault conditions are analyzed. Vibration data compared to actual fault conditions The similarity of data distributions is assessed using the following formulas:

[0083] ;

[0084] in, The first vibration data derived from actual fault conditions represents the... Each feature sample The first [data] is derived from the enhanced fault condition vibration data. One feature sample; and The number of samples in each of the two categories; This represents the kernel mapping function, used to map the original features to a high-dimensional kernel feature space; and These represent the mean embeddings of the vibration data under the actual fault state and the vibration data under the enhanced fault state in the kernel feature space, respectively. This represents the L2 norm.

[0085] ;

[0086] in, , These are the characteristic mean values ​​of vibration data under real fault conditions and the characteristic mean values ​​of vibration data under enhanced fault conditions, respectively. , These are the feature covariances of real fault data and the feature covariances of augmented fault data, respectively. This indicates summing the elements along the main diagonal of the matrix.

[0087] S43. Extract statistical features, time-domain features, frequency-domain features, and time-frequency-domain features from enhanced fault-state vibration data and real fault-state vibration data. Calculate similarity using the KS test statistic, Wasserstein distance, and Jensen-Shannon divergence, respectively. Then, sum the results of the three methods with equal weights to obtain the similarity index. Specifically:

[0088] The first of the enhanced fault state vibration data The first characteristic and the vibration data of the actual fault state Calculate similarity scores for each feature. for:

[0089] ;

[0090] in: Let KS test statistic be used. Indicates the first Any value that a feature can take. Represents all possible eigenvalues Take the maximum value above. Indicates enhanced vibration data under fault conditions In the The empirical cumulative distribution function over a feature is equal to the value of that feature taking no more than a certain value. proportion, Vibration data representing the actual fault state In the The empirical cumulative distribution function over a feature is equal to the value of that feature taking no more than a certain value. The proportion; For the first Wasserstein distance of each feature For the first A range of eigenvalues ​​is used for normalization; For the first The feature is based on the Jensen-Shannon divergence of equal-width bins. .

[0091] Similarity score for all features The arithmetic mean is taken to obtain the global feature average similarity. .

[0092] Example 2: Figures 1-10 As shown, the present invention will be specifically described below in conjunction with experiments:

[0093] I. A ball screw pair fault simulation test bench was used to obtain vibration data of the ball screw pair under healthy and pitting fault conditions. The screw motion was reciprocating, and the load was 25 kg. The vibration sensor for collecting vibration data was installed on the nut seat. The relevant experimental parameters and specifications are shown in Table 1.

[0094] Table 1 shows the specifications of experimental parameters.

[0095]

[0096] After acquiring the time-series vibration data of the ball screw pair under four speed conditions—healthy state and screw pitting fault state—a non-overlapping sliding sampling method with a fixed window size of 2560 sampling points and a step size of 2560 was used, 300 data segments were selected from each speed condition to form a total of 2400 samples. Among the 2400 samples, 1200 vibration data segments under healthy state were used as the source domain sample source, and 1200 vibration data segments under fault state were used as the target domain sample source. The global mean and standard deviation of the source domain samples are calculated separately for each operating condition. Then, Z-score normalization is performed on both the source and target domain samples using the global mean and standard deviation of the source domain samples (for example, at 600 rpm, Z-score normalization is performed on the source domain samples using the global mean and standard deviation, and on the target domain samples using the same global mean and standard deviation; this process is repeated for subsequent operating conditions to obtain normalized data for four operating conditions). This constructs the source domain dataset. and target domain dataset (Source domain dataset) That is, the normalized data of the source domain samples under four operating conditions, and the target domain dataset. That is, the normalized data of the target domain sample source under four operating conditions.

[0097] It should be noted that, considering that healthy vibration data (i.e., the source domain) is usually more stable and easier to obtain, while faulty vibration data (the target domain) often contains abnormal impacts, leading to distribution shifts (such as higher amplitudes or variances), this invention utilizes the global mean and standard deviation of the source domain samples to perform Z-score normalization on the target domain samples. This can "force mapping" the faulty vibration data to the scale of the healthy vibration data, reducing the inter-domain gap and helping the subsequent unsupervised conditional diffusion model focus on fault mode differences (such as impact pulses) rather than absolute numerical deviations, avoiding interference from fault outliers during training and improving the quality of cross-domain generation. At the same time, considering that vibration data before normalization often have different scales (such as the amplitude of vibration signals varying with rotational speed or load), this invention uses Z-score normalization. Z-score normalization makes features on a common scale, avoiding high-variance features dominating the unsupervised conditional diffusion model. Furthermore, considering the significant differences in signal distribution under different operating speeds (such as 600 rpm, 800 rpm, 900 rpm, and 1200 rpm), statistics (i.e., global mean and standard deviation) are calculated independently for each operating condition. This preserves the relative characteristics (such as periodicity) of each operating condition, avoids confusion caused by global normalization, and ensures that the subsequent unsupervised conditional diffusion model can generalize and adapt under unknown operating conditions.

[0098] II. Constructing an unsupervised conditional diffusion model, specifically:

[0099] One-dimensional domain translator T in unsupervised conditional diffusion model AB One-dimensional domain translator T BA Structure such as Figure 2As shown, specifically: First, for the input source domain data or target domain data, preliminary feature extraction is performed in the domain translator encoder through the first one-dimensional convolutional block Conv1dBlock. The one-dimensional convolutional block contains a one-dimensional convolutional layer with a kernel size of 7, a group normalization layer, and a SiLU activation function, keeping the temporal resolution of the input data unchanged and mapping the number of input channels to the number of basic channels, used to capture the local time-frequency features of the signal; then, the second one-dimensional convolutional block performs a 2x downsampling of the features. The one-dimensional convolutional layer in the second one-dimensional convolutional block has a kernel size of 5 and a stride of 2, expanding the number of feature channels to twice the number of basic channels, used to extract mesoscale feature patterns within a larger receptive field; then, nonlinear transformation and feature enhancement are performed through the first residual block ResBlock. The first residual block includes two one-dimensional convolutional layers with a kernel size of 3. The system first passes through layers and residual connections, maintaining the number of channels at twice the base number to mitigate gradient vanishing and enhance feature learning capabilities. Next, a third one-dimensional convolutional block performs a 2x downsampling, with a kernel size of 3 and a stride of 2, expanding the feature channels to four times the base number for extracting higher-level semantic features. Then, two first residual blocks perform a deep nonlinear transformation on the high-dimensional features, with each first residual block having the same structure as the first one, maintaining the number of channels at four times the base number, for learning more complex feature representations. The output then passes through a fourth one-dimensional convolutional block as the encoder output of the domain translator, with a kernel size of 3 for the one-dimensional convolutional layer, expanding the feature channels to eight times the base number and generating a highly compressed bottleneck feature representation.The domain translator decoder employs a structure symmetrical to the domain translator encoder for feature reconstruction. Specifically: First, a one-dimensional convolutional block (Conv1dBlock) reduces the channel count of the bottleneck features output by the domain translator encoder to 4 times. The kernel size of the one-dimensional convolutional layer in this block is 3, reducing the number of feature channels from 8 times the base number to 4 times, thus initiating the decoding process. Then, a first residual block (ResBlock) performs nonlinear transformations and feature refinement on the dimensionality-reduced features, maintaining the channel count at 4 times the base number. Next, a one-dimensional deconvolutional block (ConvT1dBlock) performs a 2x upsampling, with a kernel size of 3 and a stride of 2 in the one-dimensional deconvolutional layer, reducing the number of feature channels to twice the base number and restoring some temporal resolution. Then, a first residual block (ResBlock) further refines the features. The ResBlock block further refines the upsampled features while maintaining the number of channels at twice the base number. Next, a one-dimensional deconvolution block, ConvT1dBlock, performs a 2x upsampling, reducing the feature channel count to the base number and restoring the temporal resolution to near the original input scale. Finally, the output layer maps the features back to the original input channel count. This output layer contains a one-dimensional convolutional layer with a kernel size of 7 and a Tanh activation function, outputting the final translation result. The output maintains the same time length as the input data, achieving a coarse translation between the source and target domains. This provides a preliminary starting point for inter-domain mapping. The coarse translation result contains the basic feature patterns of the source or target domain data, but has not yet undergone refinement.

[0100] The conditional encoder structure in the unsupervised conditional diffusion model is as follows: First, a one-dimensional convolutional block (Conv1dBlock) performs preliminary feature extraction on the input data. This one-dimensional convolutional block contains a one-dimensional convolutional layer with a kernel size of 7, a group normalization layer, and a SiLU activation function, mapping the number of input channels to the base number of channels while maintaining the temporal resolution. Then, a second one-dimensional convolutional block performs downsampling by a factor of 2. The kernel size of the one-dimensional convolutional layer in this second block is 5, and the stride is 2, expanding the number of feature channels to twice the base number and compressing the temporal dimension. Next, a third one-dimensional convolutional block performs further downsampling by a factor of 2. The kernel size of the one-dimensional convolutional layer in this third block is 3, and the stride is 2, expanding the number of feature channels to four times the base number, further compressing the temporal dimension and extracting high-level semantic features. Finally, an adaptive average pooling layer is used... Pooling compresses the features into a fixed length over time to obtain a global feature representation. Then, the multidimensional features are flattened into a one-dimensional vector through a flattening operation. Finally, the feature vector is mapped to a condition vector of a preset dimension through a fully connected layer and the SiLU activation function. This condition vector contains the operating condition features of the input data (such as rotational speed, load, etc.).

[0101] In unsupervised conditional diffusion models, the source domain conditional diffusion denoising network and the target domain conditional diffusion denoising network have the same structure, both using a one-dimensional U-Net framework, such as... Figure 3 , Figure 4As shown, the input of the source domain conditional diffusion denoising network (target domain conditional diffusion denoising network) undergoes preliminary feature extraction and temporal information fusion through two second residual blocks, maintaining the base number of feature channels. It then enters the first downsampling layer, which performs a 2x downsampling operation using a convolution with a stride of 2, expanding the number of channels to twice the base number and compressing the temporal dimension. Next, two second residual blocks perform nonlinear transformations and feature enhancement on the downsampled features. The second second residual block expands the number of channels to four times the base number. The second downsampling layer then performs a 2x downsampling operation, maintaining the base number of channels at four times, completing the encoding downsampling of the input data. The output features of the second downsampling layer serve as the input to the bottleneck block, expanding the number of channels to eight times the base number after passing through two bottleneck blocks. A self-attention mechanism is introduced in each bottleneck block to capture global temporal dependencies. Finally, the first upsampling layer begins decoding. This first upsampling layer performs a 2x upsampling operation using a one-dimensional deconvolution layer with a stride of 2. The first upsampling layer samples and reduces the number of channels to four times the base number. The features output from the first upsampling layer are concatenated with the features input from the second downsampling layer along the channel dimension to form a skip connection to preserve detailed information. After the first upsampling layer, two second residual blocks are connected. The number of channels of the concatenated features is reduced to four times and two times the base number, respectively, before entering the second upsampling layer. The second upsampling layer performs a 2x upsampling and reduces the number of channels to the base number. The features output from the second upsampling layer are concatenated with the features input from the first downsampling layer along the channel dimension and then passed through two second residual blocks to reduce the number of channels of the concatenated features to the base number and perform final feature refinement. Finally, the output layer maps the features back to the original number of input channels to obtain the output. In each second residual block, a time-encoded vector is formed by embedding a time embedding vector generated by sine-cosine position encoding and a conditional vector generated by a conditional encoder. This allows the denoising prediction process to retain source domain operating condition feature information (such as speed, load, etc.) while adapting to different diffusion time steps, achieving operating condition adaptive cross-domain noise prediction.

[0102] 3. Based on the source domain dataset and the target domain dataset, a training set is extracted. The training set is used as the input of the unsupervised conditional diffusion model. The AdamW optimizer is used, and a weighted joint loss function is adopted to train the unsupervised conditional diffusion model in an unsupervised manner. A bidirectional mapping between the data in the source domain dataset and the data in the target domain dataset is established to obtain the mapping model.

[0103] Specifically, all data from the source and target domain datasets for three operating conditions (600 rpm, 900 rpm, and 1200 rpm) are randomly shuffled within their respective domains and used as the training input for the unsupervised conditional diffusion model (data from the 800 rpm operating condition is used as the test set). During training, health status samples from the source domain dataset are used... After passing through the one-dimensional domain translator T AB A rough translation sample from health data to fault data was then obtained. Fault state samples from the target domain dataset After passing through the one-dimensional domain translator T BA A rough translation sample of fault data into health data was then obtained. The condition encoder then processes the obtained data. and Perform condition-independent feature extraction and encode the extracted features into a condition vector of length 256. and In the source domain conditional diffusion denoising network and the target domain conditional diffusion denoising network: the source domain conditional diffusion denoising network uses health status samples from the source domain dataset. As input, in the condition vector With the embedding of time-encoded vectors, forward noise addition and reverse denoising are performed (noise is linearly scheduled, with a total diffusion step count T=3000); the target domain conditional diffusion denoising network is trained on fault state samples from the target domain dataset. As input, in the condition vector With the embedding of time-encoded vectors, forward noise addition and reverse noise reduction are performed. The entire training process is completed under the joint optimization of AdamW optimizer and weighted joint loss function, with a learning rate of 0.0002, 400 training rounds, and a batch size of 32. The weighted joint loss function is constructed by weighting the noise prediction mean square error loss, cycle consistency L1 loss, and multi-resolution STFT log spectrum L1 loss, with weights of 1.0, 2.0, and 0.5, respectively. After training, a mapping model between the source domain and the target domain is obtained.

[0104] It should be noted that, considering the core task of the diffusion model is to accurately predict noise distribution, this invention uses the noise prediction mean square error loss as the main loss function, with a weight of 1.0, to supervise the model's learning to accurately predict noise from noisy data, ensuring the diffusion model's denoising capability. Simultaneously, considering the crucial importance of reversibility and consistency in domain translation—that is, the loop translation from source domain data to target domain data and back to source domain data should be able to restore the original data, and the loop translation from target domain data to source domain data and back to target domain data should also be able to restore the original data—this invention introduces a loop consistency L1 loss, with a weight of 2.0 (higher than the main loss). By constraining the loop reconstruction error, it prevents the one-dimensional domain translator from losing key information during the translation process, ensuring that the source domain operating characteristics (such as speed, load, etc.) are accurately predicted. The cross-domain generation process preserves the data, preventing pattern collapse or information degradation. Furthermore, considering the crucial role of frequency domain characteristics of vibration signals (such as fault characteristic frequencies and harmonic components) in fault diagnosis, simple time-domain loss may not adequately constrain frequency domain characteristics. Therefore, this invention introduces a multi-resolution STFT logarithmic spectrum L1 loss with a weight of 0.5. By performing a short-time Fourier transform (STFT) on the time-domain signal and calculating the logarithmic spectrum error, the frequency domain consistency between the generated data and the real data is constrained across multiple time-frequency resolution scales. This ensures that the generated fault data is not only similar in time-domain waveforms but also accurately reproduces the spectral structure of the real fault (such as impulse modulation and sidebands) in the frequency domain, avoiding frequency domain distortion or missing frequency components, thereby improving the effectiveness and reliability of the generated data in practical fault diagnosis applications. Through the weighted combination of the above three loss functions, this invention achieves multi-dimensional constraints on time-domain, frequency-domain, and cross-domain consistency. The source-domain and target-domain mapping model obtained after training can generate high-quality, high-fidelity unknown operating condition fault vibration data while preserving the operating condition characteristics.

[0105] IV. Cross-Domain Data Enhancement and Evaluation: Utilizing a mapping model and a one-dimensional domain translator T AB A coarse translation is performed on health status samples from the source domain dataset in the test set. The coarse translation result is used as the input to the source domain conditional diffusion denoising network, where incomplete forward diffusion is performed to obtain intermediate noise state signals. These intermediate noise state signals are then used as the input to the target domain conditional diffusion denoising network, where denoising is progressively guided by the conditional vectors generated by the conditional encoder. This achieves cross-domain generation of health data from unknown operating conditions to corresponding fault data. Data distribution similarity and feature similarity are used to evaluate the generated fault data samples and verify the quality of the generated data.

[0106] Cross-domain data enhancement specifically refers to: in the source domain dataset In this process, data with a rotational speed of 800 rpm is selected and cross-domain data enhancement is performed based on the mapping model of this invention to generate enhanced fault state vibration data, and the target domain dataset is selected. The data from 800 rpm was compared with the enhanced fault-state vibration data to evaluate the quality of the generated enhanced fault-state vibration data; the cross-domain data enhancement process is as follows: Figure 5 As shown, the trained mapping model is loaded, and a one-dimensional domain translator T is used. AB Input Rough translation As the input to the source domain conditional diffusion denoising network, the conditional encoder simultaneously... Feature extraction is performed and encoded into a 256-dimensional conditional vector. ,Then In the target domain conditional diffusion denoising network Positive noise addition to the step, Gradually adding noise to obtain intermediate noise state signals Then As input to the target domain conditional diffusion denoising network, in the conditional vector Guided by this, reverse denoising prediction is performed, from... Step-by-step denoising =0 steps, ultimately achieving cross-domain generation of unknown working condition data, obtaining enhanced fault state vibration data of lead screw pitting fault at 800 rpm. The time-domain waveforms of the generated enhanced fault-state vibration data and the actual fault-state vibration data are shown below. Figure 6 , Figure 7 As shown in the figure, the time-domain waveforms reveal that the generated enhanced fault-state vibration data is highly similar to the real fault-state vibration data in terms of vibration amplitude, waveform morphology, and impact characteristics. The spectrum diagrams of the generated enhanced fault-state vibration data and the real fault-state vibration data (obtained by performing a Fast Fourier Transform (FFT) on the time-domain vibration signals) are shown below. Figure 8 , Figure 9 As shown in the spectrum, the generated enhanced fault-state vibration data and the real fault-state vibration data are highly consistent in frequency domain distribution. The main frequency components, amplitude distribution, and concentrated spectral energy regions are basically the same, verifying the effectiveness of the method of the present invention in frequency domain feature fidelity preservation. A comparison of the amplitude distributions of the generated enhanced fault-state vibration data and the real fault-state vibration data is shown in the figure below. Figure 10As shown in the amplitude distribution diagram, the statistical characteristics of the generated enhanced fault state vibration data are highly consistent with those of the real fault state vibration data. The amplitude probability density distribution curves basically overlap, indicating that the generated enhanced fault state vibration data is not only similar to the real fault state vibration data in waveform morphology, but also maintains a high degree of consistency in statistical characteristics. This fully verifies the quality and reliability of the data generated by the method of this invention, and can be effectively used for fault data enhancement under unknown working conditions.

[0107] The cross-domain data evaluation specifically involves: under the 800 rpm operating condition, 200 samples are randomly selected from both the generated vibration data of the enhanced pitting fault state of the leadscrew and the actual vibration data of the pitting fault state of the leadscrew for data distribution similarity evaluation and feature similarity evaluation. In the data distribution similarity evaluation, nine statistical features are extracted (mean, root mean square, skewness, kurtosis, peak value, peak-to-peak value, impulse index, peak index, and margin index); the maximum mean difference (MMD) and Fréchet index are used. Using the Inception distance (FID) as a metric, we obtained MMD=0.078 and FID=3.798. For feature distribution similarity evaluation, we extracted statistical features (mean, standard deviation, variance, skewness, kurtosis, RMS, maximum value, minimum value, peak-to-peak value, waveform factor, peak factor, impulse factor), time-domain features (mean absolute difference, difference standard deviation, energy, mean absolute value, zero-crossing rate), frequency-domain features (dominant frequency, spectral centroid, spectral roll-off, spectral spread, spectral flatness, maximum FFT amplitude, frequency-domain energy), and time-frequency-domain features (energy proportion of 6-layer wavelets, energy proportion of 6-layer wavelet detail coefficients, relative variance of 6-layer wavelets). For each feature dimension, we calculated the KS test statistic, Wasserstein distance, and Jensen-Shannon divergence, and obtained the single-feature similarity score by equal weighted summation. Then, we took the arithmetic mean of all features to obtain the global feature average similarity. =0.931.

[0108] The above evaluation results show that the generated enhanced fault-state vibration data and the real fault-state vibration data maintain a high degree of consistency in multiple dimensions, including statistical distribution, time-domain features, frequency-domain features, and time-frequency-domain features. Among them, the MMD value is close to 0, the FID value is low, and the average similarity of global features is high. The result of 0.931 fully verifies that the unknown working condition fault data generated by the method of the present invention has good authenticity and credibility, and can effectively replace the vibration data of real fault state for the training and verification of fault diagnosis model. It solves the problem of scarce fault samples under unknown working conditions and provides reliable data support for cross-working condition fault diagnosis.

[0109] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion, characterized in that, include: S1. Obtain vibration data of ball screw pairs in different states operating under M working conditions to construct source domain dataset and target domain dataset; where M is a positive integer and M>2; S2. Constructing an unsupervised conditional diffusion model: One-dimensional domain translator T AB One-dimensional domain translator T BA Coarse translations are performed on the source domain dataset and the target domain dataset respectively; a conditional encoder is used to generate conditional vectors; a source domain conditional diffusion denoising network and a target domain conditional diffusion denoising network are used to achieve cross-domain conditional diffusion and denoising generation under the guidance of the conditional vectors. S3. Extract training sets based on the source domain dataset and the target domain dataset, use the training sets as input to the unsupervised conditional diffusion model, and use a weighted joint loss function to train the unsupervised conditional diffusion model in an unsupervised manner, establish a bidirectional mapping between the data in the source domain dataset and the data in the target domain dataset, and obtain the mapping model. S4. Obtain the health status data to be enhanced; based on the mapping model, generate enhanced fault state vibration data corresponding to the working conditions of the health status data to be enhanced; The unsupervised conditional diffusion model specifically includes: Two one-dimensional domain translators: One-dimensional domain translator T AB One-dimensional domain translator T BA The one-dimensional domain translator T AB One-dimensional domain translator T BA They have the same structure, including a domain translator encoder, a domain translator decoder, and an output layer. The one-dimensional domain translator T... AB Using the source domain dataset as input, a preliminary domain transformation from the source domain to the target domain is achieved; a one-dimensional domain translator T BA Using the target domain dataset as input data, a preliminary domain transformation from the target domain to the source domain is achieved. The domain translator encoder structure includes a sequentially connected one-dimensional convolutional block, two stacked "one-dimensional convolutional blocks + first residual blocks", a first residual block, and a one-dimensional convolutional block. Downsampling is introduced into the one-dimensional convolutional block of the two stacked "one-dimensional convolutional blocks + first residual blocks". The domain translator decoder structure includes a sequentially connected one-dimensional convolutional block, two stacked "first residual blocks + one-dimensional deconvolutional blocks", and upsampling is introduced into the one-dimensional deconvolutional block of the two stacked "first residual blocks + one-dimensional deconvolutional blocks". The output of the domain translator decoder is used as the input of the output layer, and the output of the output layer is the output of the one-dimensional domain translator. The conditional encoder comprises three sequentially connected one-dimensional convolutional blocks, an adaptive average pooling layer, and a fully connected layer, ultimately mapping to a fixed-dimensional conditional vector. The source domain conditional diffusion denoising network and the target domain conditional diffusion denoising network have the same structure, both based on a one-dimensional U-Net framework, including a second residual block, a downsampling layer, an upsampling layer, a bottleneck block, and an output layer.

2. The method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion according to claim 1, characterized in that, S1 includes: S11. Under constant load on the ball screw pair fault simulation test bench, collect vibration data of healthy state and vibration data of fault state under various speed conditions respectively. S12. Using a non-overlapping sliding sampling method, select healthy state vibration data segments and fault state vibration data segments with equal data volume from various speed conditions as source domain sample sources and target domain sample sources, respectively. S13. Calculate the global mean and standard deviation of the source domain sample sources according to the working conditions, and use the global mean and standard deviation of the source domain sample sources to perform working condition-by-working-condition Z-score normalization on the source domain sample sources and the target domain sample sources, so as to construct the source domain dataset and the target domain dataset.

3. The method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion according to claim 1, characterized in that, The input data passes through two second residual blocks and then enters the first downsampling layer. After passing through two more second residual blocks, it enters the second downsampling layer, completing the downsampling of the input data. The output features of the second downsampling layer pass through two bottleneck blocks and then enter the first upsampling layer. The output of the first upsampling layer is concatenated with the input of the second downsampling layer in the channel dimension and then connected to two second residual blocks. It then enters the second upsampling layer. The output of the second upsampling layer is concatenated with the input of the first downsampling layer in the channel dimension and then passes through two more second residual blocks before finally being output by the output layer.

4. The method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion according to claim 1, characterized in that, The first residual block consists of two stacked "one-dimensional convolutional layers + group normalization layers". The first "one-dimensional convolutional layer + group normalization layer" is followed by the SiLU activation function. The output of the second "one-dimensional convolutional layer + group normalization layer" is concatenated with the original input of the first residual block and then input into the Tanh activation function to obtain the output of the first residual block.

5. The method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion according to claim 1, characterized in that, The second residual block includes two blocks of "one-dimensional convolutional layer + group normalization layer + SiLU activation function" and regularization operation. The output of the first block is concatenated with the output of the time-encoded vector after projection by the time embedding projection layer. The concatenated feature is regularized and used as the input of the second block. The output of the group normalization layer of the second block is concatenated with the original input of the second residual block and input into the SiLU activation function in the second block to obtain the final output of the second residual block. The time-encoded vector is obtained by embedding the time embedding vector generated by sine-cosine position encoding and adding it to the conditional vector output by the conditional encoder.

6. The method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion according to claim 1, characterized in that, The bottleneck block consists of a second residual block followed by a self-attention mechanism.

7. The method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion according to claim 1, characterized in that, The S3 includes: S31. From the source domain dataset and the target domain dataset, extract all samples of N working conditions as the training set, and the rest as the test set; where N is a positive integer, and 1 <N<M; S32. Training of the unsupervised conditional diffusion model: The training set is used as the input of the unsupervised conditional diffusion model. The AdamW optimizer is used, and a weighted joint loss function is adopted to train the unsupervised conditional diffusion model in an unsupervised manner. A bidirectional mapping between the data in the source domain dataset and the data in the target domain dataset is established to obtain the mapping model. The weighted joint loss function is constructed in a weighted manner based on the noise prediction mean square error loss, the cycle consistency L1 loss, and the multi-resolution STFT log spectrum L1 loss.

8. The method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion according to claim 7, characterized in that, The weighted joint loss function is expressed as follows: ； in, The mean square error loss for noise prediction is respectively L1 loss for cycle consistency and multi-resolution STFT logarithmic spectrum L1 loss The weighting coefficients.

9. The method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion according to claim 8, characterized in that, The weighting coefficient values ​​satisfy .

10. The method for cross-domain generation of ball screw pair fault data based on unsupervised conditional diffusion according to claim 1, characterized in that, The S4 includes: S41. Utilize a one-dimensional domain translator T in a mapping model to process enhanced health status data. AB A rough translation, used as input to the source domain conditional diffusion denoising network in the mapping model, is performed according to a preset number of release steps. Perform forward diffusion to generate intermediate noise state signals. ; S42. Using the health status data to be enhanced as input to the condition encoder in the mapping model, a condition vector is obtained. In the condition vector Under the guidance of the mapping model, the target domain conditional diffusion denoising network from time step Start with intermediate noise state signal Perform reverse denoising sampling until Gradually generate enhanced fault state vibration data corresponding to the working conditions of the health state data to be enhanced.

Images