Degradation data enhancement method based on combination of conditional encoder and diffusion model

Abstract

The invention provides a degradation data enhancement method based on combination of a conditional encoder and a diffusion model, which is characterized in that a combined network of the conditional encoder and the diffusion model is designed on the basis of a physical priori guidance generation principle, and a physical rule of monotonically decreasing residual service life is embedded into a potential space through a two-stage conditional encoder network. The high-fidelity degradation data conforming to the physical law is efficiently generated by using the space-time condition diffusion model, the problems of data scarcity and unbalanced distribution faced by the existing data-driven prediction method are solved, the generalization ability and robustness of the prediction model in a complex scene are remarkably improved, and the method can be applied to prediction of the residual service life of the equipment and has a good application prospect. The problem that the prediction precision is reduced due to insufficient samples and unbalanced distribution when a traditional residual service life prediction method faces data scarcity and complex working conditions is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of encoder network and diffusion model, in particular to a conditional encoder and diffusion model combined network model for enhancing degraded data. BACKGROUND

[0002] Residual useful life (RUL) prediction, as a core technology of prognostics and health management, is of great significance for formulating maintenance strategies, reducing unexpected downtime and reducing operation and maintenance costs. In recent years, data-driven methods represented by convolutional neural networks (CNN), recurrent neural networks (RNN), Transformers and the latest Mamba architecture have become the mainstream method of RUL prediction due to their strong feature extraction capability.

[0003] However, data-driven deep learning models usually rely on a large amount of full-life degradation data. In actual engineering applications, due to the complex and variable operating conditions of equipment, high data acquisition costs and the scarcity of fault samples, the available high-quality degradation data is often severely insufficient and unevenly distributed. This "data scarcity" problem easily leads to overfitting of the model, severely limiting the generalization ability and robustness of the prediction model in complex operating conditions or small sample scenarios.

[0004] In order to alleviate the problem of data scarcity, data enhancement techniques are widely used to synthesize virtual samples to enrich the training set. Generative adversarial networks (GAN) were once the mainstream enhancement method. However, GAN-based methods generally have problems such as unstable training, mode collapse, and difficulty in maintaining long-term degradation trends or complex spatiotemporal correlations between multi-sensor data.

[0005] In contrast, diffusion probability models (DPMs) provide a more stable training process and stronger distribution modeling capability through a gradual noise-adding and noise-removing mechanism, and have shown advantages in image and time series generation. Although a small number of studies have attempted to apply diffusion models to data enhancement for RUL prediction, existing diffusion model-based degradation data enhancement methods still have the following significant limitations:

[0006] First, there is a lack of physical consistency constraints. Existing generation models mostly focus on fitting the data distribution, while ignoring the inherent physical laws of the degradation process. Specifically, the residual useful life of a device should show a monotonically decreasing trend with operating time, but the latent degradation trajectories generated by existing models often contain local cycles or non-monotonic fluctuations, making the generated samples physically uninterpretable and affecting the accuracy of the prediction model.

[0007] Second, the spatiotemporal dependency modeling capabilities are insufficient. Multi-sensor degradation data exhibits complex cross-sensor spatial correlations and long-range temporal dependencies. Existing diffusion model architectures typically focus on single-dimensional modeling or suffer from excessive computational complexity when processing long sequences, making it difficult to simultaneously and efficiently capture both local interactions between sensors and long-range degradation dynamics throughout the entire lifecycle. For example, relying solely on CNNs struggles to capture long-range dependencies, while traditional Transformers incur significant computational overhead when processing ultra-long sequences.

[0008] Therefore, there is an urgent need to develop a combined network model that can integrate physical prior knowledge and efficiently jointly model the spatiotemporal dependencies of multiple sensors to generate high-fidelity, physically consistent degradation data training samples, thereby improving the performance of remaining lifetime prediction. Summary of the Invention

[0009] To address the shortcomings of existing network models for data augmentation, such as the lack of physical consistency constraints and insufficient spatiotemporal dependency modeling capabilities, this invention employs a combined network of conditional encoders and diffusion models. This combined network is used to augment degraded data, generating high-fidelity degraded data that conforms to physical laws. This solves the problems of data scarcity and uneven distribution faced by existing data-driven prediction methods, significantly improving the accuracy and robustness of equipment remaining service life prediction.

[0010] The technical solution of this invention is as follows:

[0011] A degraded data augmentation method based on a combination of conditional encoders and diffusion models includes the following steps:

[0012] Step 1: Obtain multi-sensor degradation monitoring data of the device to be predicted, and preprocess the degradation monitoring data to obtain a basic training set containing degradation features and remaining service life labels;

[0013] Step 2: Construct a two-stage conditional encoder network, and extract latent variables representing the degree of degradation and working condition information from the basic training set through multi-task joint learning and monotonicity fine-tuning;

[0014] Step 3: Construct a spatiotemporal conditional diffusion model, which uses the latent variables as conditions to learn the inverse denoising process from Gaussian noise to real degraded data;

[0015] Step 4: Generate multi-sensor degradation time window samples based on the trained spatiotemporal conditional diffusion model, and fuse the generated multi-sensor degradation time window samples with the basic training set to construct an enhanced training set.

[0016] A further preferred embodiment, the preprocessing of the degradation monitoring data in step 1 includes:

[0017] The multi-condition data is clustered into working conditions, the minimum-maximum normalization is performed on the data of each sensor channel in each working condition cluster, the degradation monitoring data is divided into time window samples with a length L and a step of 1 by using a sliding time window, and the true residual life at the end of the time window is taken as a label.

[0018] In a further preferred embodiment, the two-stage conditional encoder network in step 2 comprises a first encoder, a first decoder and a regressor.

[0019] The first encoder adopts a time series backbone network embedded with a Mamba module to map the input time window sample into the mean and variance of latent variables, and obtain the latent variables by reparameterization sampling.

[0020] The first decoder receives the latent variables and reconstructs the time window sample through a Mamba module and a transpose convolution layer.

[0021] The regressor receives the latent variables and outputs the predicted value of the residual life.

[0022] In a further preferred embodiment, the training process of the two-stage conditional encoder network in step 2 is divided into two stages.

[0023] The first stage is a joint optimization stage, and the network is trained by using a joint loss function, wherein the joint loss function comprises a KL divergence loss for constraining the distribution of latent variables, a reconstruction loss for minimizing the difference between the input and the output, a regression loss for accurately predicting the residual life, and a triplet contrast loss for optimizing the clustering characteristics of the latent space.

[0024] The second stage is a monotonicity fine-tuning stage, and the two-stage conditional encoder network is fine-tuned by introducing a monotonicity loss on the basis of the training convergence of the first stage.

[0025] In a further preferred embodiment, the KL divergence loss for constraining the distribution of latent variables in the joint optimization stage is .

[0026]

[0027] The reconstruction loss for minimizing the difference between the input and the output is .

[0028]

[0029] The regression loss for accurately predicting the residual life is .

[0030]

[0031] Triplet contrastive loss for optimizing latent space clustering properties is:

[0032]

[0033] Final joint loss function is:

[0034]

[0035] wherein denotes the number of samples after time window processing, and denote the mean and variance of the latent variable corresponding to the -th time window sample; denotes the reconstructed time window sample; denotes the true remaining useful life label, denotes the remaining useful life prediction output by the regressor based on the latent variable ; is the latent variable obtained by inputting the positive sample into the first encoder, is the latent variable obtained by inputting the negative sample into the first encoder, denotes a preset marginal distance; denotes a preset marginal distance; denote corresponding weights, respectively.

[0036] In a further preferred embodiment, in the monotonicity fine-tuning stage, the monotonicity loss is:

[0037]

[0038] wherein denotes the remaining useful life prediction output by the regressor based on the latent variable corresponding to the -th time window sample, denotes a monotonicity tolerance boundary;

[0039] Total loss function of the second stage is:

[0040]

[0041] wherein denotes a monotonicity loss weight, denotes the first stage joint loss corresponding to the -th time window sample, denotes the first stage joint loss corresponding to the -th time window sample, denotes the first stage joint loss corresponding to the Corresponding first-stage joint loss.

[0042] In a further preferred embodiment, in step 3, the spatio-temporal conditional diffusion model adopts a U-Net as a denoising backbone network, and embeds a spatio-temporal block in each layer of the encoder and the decoder; the spatio-temporal block is composed of a cascaded dilated depth separable convolution module and a Mamba state space module; the dilated depth separable convolution module is used to extract cross-channel spatial local correlation and multi-scale features in the sensor dimension; and the Mamba state space module is used to model long-term degradation trends in the time dimension.

[0043] In a further preferred embodiment, the spatio-temporal conditional diffusion model adopts a conditional injection mechanism based on feature linear modulation to inject the latent variable into the spatio-temporal conditional diffusion model; the conditional injection mechanism includes: mapping the diffusion time step into a time embedding vector, and mapping the latent variable into a condition embedding vector; fusing the time embedding vector and the condition embedding vector to generate a channel scaling vector and a channel offset vector; and performing affine transformation on the feature map of the intermediate layer of the spatio-temporal conditional diffusion model by using the channel scaling vector and the channel offset vector, so as to control the generation process of the degraded sample.

[0044] In a further preferred embodiment, the training target of the spatio-temporal conditional diffusion model is to minimize the noise prediction error: a real time window sample is sampled from the basic training set, and Gaussian noise is added to the real time window sample according to a preset noise schedule to obtain a noisy sample; the noisy sample, the current diffusion time step, and the corresponding latent variable are input into the spatio-temporal conditional diffusion model to predict the added noise; and the mean square error between the predicted noise and the real added noise is calculated as a loss function.

[0045] In a further preferred embodiment, in step 4, the multi-sensor degraded time window sample is generated based on the trained spatio-temporal conditional diffusion model, and specifically includes: using a deterministic sampling strategy of the denoising diffusion implicit model to sample an initial vector from a standard Gaussian distribution, under the condition guidance of the latent variable, using the trained spatio-temporal conditional diffusion model to predict the noise, and iteratively updating according to a deterministic non-Markov chain formula to finally obtain the denoised multi-sensor degraded time window sample.

[0046] Beneficial effects:

[0047] The present application has the following technical effects:

[0048] The degradation data enhancement method using the combination network of the conditional encoder and the diffusion model proposed in the present application has the core of proposing the combination network of the conditional encoder and the diffusion model.

[0049] Firstly, by means of a two-stage conditional encoder network, especially by introducing a monotonicity loss to fine-tune the latent space, the physical prior knowledge that the "residual life decreases monotonically with time" is explicitly embedded into the data generation process, effectively eliminating the common local cycles and non-monotonic fluctuations in the generated samples, ensuring the physical consistency and interpretability of the generated data.

[0050] And by constructing a spatio-temporal conditional diffusion model that integrates dilated depth separable convolution and Mamba state space module, the local spatial correlation between sensors is efficiently captured by dilated convolution, and the long-range temporal dependence of the whole life cycle is accurately modeled by the linear complexity advantage of the Mamba module, solving the problem that traditional methods are difficult to balance computational efficiency and long-term degradation feature capture.

[0051] In addition, the conditional injection mechanism based on feature linear modulation (Feature-wise Linear Modulation, FiLM) realizes the accurate controllable generation of degradation state, and finally utilizes the generated high-fidelity samples to construct an enhanced training set, thereby effectively solving the problem of fault data scarcity and uneven working condition distribution in actual industrial scenarios. The enhanced data set provided by the overall method can significantly reduce the error of the residual life prediction model in complex working conditions and small sample scenarios, improve the generalization ability and robustness of the model, and provide more accurate and reliable data support for equipment maintenance decisions.

[0052] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0053] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the description of the embodiments, which follows, including the accompanying drawings.

[0054] Figure 1 The degradation data enhancement method flowchart based on the combination of conditional encoder and diffusion model provided by the embodiment of the application.

[0055] Figure 2 The overall framework diagram of the degradation data enhancement method based on the combination of conditional encoder and diffusion model provided by the embodiment of the application applied to residual useful life prediction, which shows the complete process from the construction of the basic training set, the extraction of the latent representation, the spatio-temporal conditional diffusion modeling to the final RUL prediction.

[0056] Figure 3 The network architecture diagram of the two-stage conditional encoder network in the embodiment of the application, which shows the connection relationship of the encoder, the decoder, the regressor and the multi-task joint loss and the monotonicity fine-tuning loss.

[0057] Figure 4 Figure 1 is a detailed structure schematic diagram of the spatiotemporal condition diffusion model in the embodiments of the present application, which includes: (a) the overall U-Net architecture of the STCDM; (b) the spatiotemporal block structure composed of the cascaded inflation depth separable convolution module and Mamba module; (c) the condition injection mechanism based on feature linear modulation.

[0058] Figure 5 Figure 2 is a comparison schematic diagram of the generated degradation trajectory and the real degradation trajectory in the embodiments of the present application on the C-MAPSS dataset, which shows the fitting effect of the generated data on the real degradation trend and local fluctuations.

[0059] Figure 6 Figure 3 is a standard deviation comparison schematic diagram of the prediction results RMSE and Score using the basic training set and the enhanced training set in the embodiments of the present application.

[0060] Figure 7 Figure 4 is a performance comparison schematic diagram of the remaining useful life prediction CNN model under different data enhancement methods and the enhancement of the method of the present application, which shows the advantages of the present application in the RMSE and Score indicators.

[0061] Figure 8 Figure 5 is a performance comparison schematic diagram of the remaining useful life prediction RNN model under different data enhancement methods and the enhancement of the method of the present application, which shows the advantages of the present application in the RMSE and Score indicators.

[0062] Figure 9 Figure 6 is a performance comparison schematic diagram of the remaining useful life prediction Transformer model under different data enhancement methods and the enhancement of the method of the present application, which shows the advantages of the present application in the RMSE and Score indicators.

[0063] Figure 10 Figure 7 is a performance comparison schematic diagram of the remaining useful life prediction Mamba model under different data enhancement methods and the enhancement of the method of the present application, which shows the advantages of the present application in the RMSE and Score indicators.

[0064] Figure 11 Figure 8 is a latent space trajectory schematic diagram obtained by using only a single-stage training in the embodiments of the present application, which shows that the sample distribution is relatively chaotic.

[0065] Figure 12 Figure 9 is a latent space trajectory schematic diagram obtained by using the two-stage conditional encoder network in the embodiments of the present application, which shows that the sample presents clear monotonic evolution along the degradation direction. DETAILED DESCRIPTION

[0066] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0067] This embodiment addresses the shortcomings of existing data augmentation network models, such as the lack of physical consistency constraints and insufficient spatiotemporal dependency modeling capabilities. It provides a degradation data augmentation method employing a combined network of conditional encoders and diffusion models. Based on the principle of physical prior-guided generation, this method designs a combined network of conditional encoders and diffusion models. A two-stage conditional encoder network embeds the monotonically decreasing physical law of remaining useful life into the latent space, and a spatiotemporal conditional diffusion model efficiently generates high-fidelity degradation data that conforms to the physical law. This solves the problems of data scarcity and uneven distribution faced by existing data-driven prediction methods, significantly improving the generalization ability and robustness of the prediction model in complex scenarios. It can be applied to equipment remaining useful life prediction, addressing the problem of decreased prediction accuracy caused by insufficient samples and uneven distribution when traditional remaining useful life prediction methods face data scarcity and complex operating conditions.

[0068] like Figure 1 and Figure 2 As shown, the degradation data augmentation method based on the combination of conditional encoder and diffusion model proposed in this embodiment is executed by a computer device. Specifically, it can be executed by a computer device such as a terminal or server alone, or by a terminal and server together. The specific steps include:

[0069] Step 1: Acquire multi-sensor degradation monitoring data of the device to be predicted, and preprocess the degradation monitoring data to obtain a basic training set containing degradation features and remaining service life labels. .

[0070] In this embodiment, multi-sensor monitoring data of the device is first acquired. Assume there are a total of... The device unit, for the first Each unit, whose multivariate degradation sequence from the start of operation to failure is represented as follows: ,in For the first The total lifetime of each unit, This represents the number of sensors.

[0071] To eliminate the influence of unit dimensions and handle differences across multiple operating conditions, the data is first clustered by operating condition, and then min-max normalization is performed within each cluster. The normalization formula is:

[0072]

[0073] wherein, denotes the normalized measurement value of the th sensor, denotes the observation value of the th sensor at the th time instant under the working condition, denotes the observation value of the th sensor at the th time instant under the working condition, denotes the observation maximum value and the observation minimum value of the th sensor under the working condition, respectively.

[0074] Subsequently, a sliding time window strategy is adopted to construct training samples. Assuming that the length of the time window is and the step length is 1, the time window sample set intercepted from the degradation sequence is denoted as , wherein denotes the th time window of the th unit, and the label of the time window is the RUL corresponding to the last time instant in the time window. Thus, the basic training set is composed of the training samples.

[0075] Step 2: Constructing a two-stage conditional encoder network to extract latent variables representing the degradation degree and working condition information from the basic training set through multi-task joint learning and monotonicity fine-tuning.

[0076] As shown in FIG. 2, the construction and training process of the two-stage conditional encoder network includes the following steps 21 to 23. Figure 3 Step 21: Constructing a two-stage conditional encoder network architecture including a first encoder, a first decoder and a regressor.

[0077] Specifically, the first encoder adopts a time series backbone network embedded with Mamba modules. The input of the first encoder is the preprocessed time window sample

[0078] In the first encoder, the sensor channels are mapped to a high-dimensional feature space through a linear layer, and sequence modeling is performed through stacked Mamba state space modules to capture long-range temporal dependencies with linear complexity. The output of the first encoder is the mean and the variance of the distribution of the latent variable , and the latent variable is obtained by sampling using the reparameterization trick. The calculation formula of the latent variable is as follows:

[0079] ​

[0080] in, To obtain from the standard normal distribution Random noise vector sampled in the middle, This indicates element-wise multiplication.

[0081] The first decoder structure is symmetrical to the first encoder and is used to receive latent variables. The time window samples are reconstructed using the Mamba module and transposed convolutional layers. .

[0082] The regressor consists of a multi-layer fully connected network used to receive latent variables. Output the predicted remaining lifetime value .

[0083] Step 22: The two-stage conditional encoder network is trained in the first stage using a multi-task joint learning strategy.

[0084] Specifically, the first stage is the joint optimization stage, which minimizes the joint loss function to enable the latent space to simultaneously possess reconstruction, prediction, and discrimination capabilities. The joint loss function consists of KL divergence loss, reconstruction loss, regression loss, and triplet contrastive loss.

[0085] 1. KL divergence loss : Used to constrain the distribution of latent variables learned by the encoder Approximating the prior standard Gaussian distribution This is used to regularize the latent space. Its calculation formula is:

[0086]

[0087] in, Denotes KL divergence, This indicates the number of samples after processing within the time window. and They represent the first Latent variables corresponding to each sample The mean and variance of.

[0088] 2. Reconstruction Loss This is used to ensure that latent variables retain key information from the original monitoring data. Its calculation formula is:

[0089]

[0090] in, This represents the input time window sample. This represents the time window sample for reconstruction.

[0091] 3. RUL Regression Loss : used to force the latent variables to be strongly correlated with the degradation phase. Its calculation formula is:

[0092]

[0093] wherein, represents the true remaining useful life label, represents the remaining useful life prediction value output by the regressor based on the latent variables .

[0094] 4. Triplet contrastive loss : used to enhance the discriminability of the latent space for different degradation phases. For a time window sample , construct a positive sample similar to its degradation trend and a negative sample with a larger difference in degradation trend, input the positive sample and the negative sample into the first encoder to obtain the corresponding latent variables and , where the judgment criteria for the similarity of the degradation trend and the larger difference in the degradation trend are both conventional judgment criteria for those skilled in the art. The triplet contrastive loss is calculated as follows:

[0095]

[0096] wherein, represents a preset margin distance.

[0097] The total loss function of the first stage is the weighted sum of the above four terms:

[0098]

[0099] wherein, is the corresponding weight.

[0100] Step 23, introduce monotonicity loss to conduct second-stage fine-tuning on the two-stage conditional encoder network.

[0101] Specifically, after the first-stage training converges, in order to solve the problem of possible local loops or non-monotonic fluctuations in the latent trajectory, a physical prior constraint is introduced for fine-tuning. For two adjacent time window samples of the same device in the same operating cycle, their corresponding RUL prediction values should physically satisfy . A monotonicity loss is constructed to punish predictions that violate monotonicity, and its calculation formula is:

[0102]

[0103] wherein is the th time window sample corresponding output remaining life prediction value, denotes monotonicity tolerance boundary, used to tolerate a certain degree of prediction noise.

[0104] The total optimization objective of the second stage is to add a monotonicity constraint on the basis of maintaining the representation ability learned in the first stage, and obtain the total loss function of the second stage :

[0105]

[0106] wherein is the th time window sample corresponding first stage joint loss, is the th time window sample corresponding first stage joint loss, is the monotonicity loss weight, through small step gradient update based on the loss function, adjust the encoder parameters, force the degenerative trajectory in the latent space to follow the physical monotonicity, finally output the physical consistent latent variable .

[0107] Step 3: constructing a spatio-temporal conditional diffusion model, which learns the inverse denoising process from Gaussian noise to real degradation data with the latent variable as a condition.

[0108] As Figure 4 shown, the construction and training process of the spatio-temporal conditional diffusion model includes the following steps 31 to 33.

[0109] Step 31, constructing a spatio-temporal conditional diffusion model architecture based on a U-Net backbone network.

[0110] Specifically, the spatio-temporal conditional diffusion model adopts U-Net as a denoising backbone network, and embeds a spatio-temporal block at each level of its encoder and decoder to jointly model the dependence relationship of time dimension and sensor dimension. The spatio-temporal block is composed of a cascaded dilated depth separable convolution module and a Mamba state space module.

[0111] The dilated depth separable convolution module is used to extract the cross-channel spatial local correlation of the sensor dimension, and capture local feature patterns of different scales through an adjustable dilation rate; the Mamba state space module is used to receive the output of the convolution module, and model the long-term degradation trend of the time dimension by using the recursive property of its linear complexity.

[0112] The encoder of the spatiotemporal conditional diffusion model gradually aggregates semantic information through downsampling, and the decoder of the spatiotemporal conditional diffusion model gradually restores the spatiotemporal resolution through a skip connection, and finally outputs a noise prediction result with the same dimension as the input through a convolution layer.

[0113] Step 32, the latent variable is injected into the spatiotemporal conditional diffusion model by using a feature linear modulation based conditional injection mechanism.

[0114] Specifically, in order to realize accurate control of the generation process, a feature linear modulation mechanism is used to modulate the intermediate feature maps in the spatiotemporal conditional diffusion model. First, the diffusion time step is mapped into a time embedding vector, and the latent variable is mapped into a conditional embedding vector. Then, the two are fused through a linear layer to generate a channel scaling vector and a channel offset vector .

[0115] For the intermediate feature map F in the spatiotemporal conditional diffusion model, the calculation formula of the FiLM based modulation is:

[0116]

[0117] wherein, represents an element-wise multiplication along the channel dimension. By dynamically adjusting the statistical characteristics of the feature map, the model can adaptively generate samples that conform to specific degradation states and working conditions.

[0118] Step 33, the spatiotemporal conditional diffusion model is trained based on the noise prediction target.

[0119] In this embodiment, the training process includes constructing a variational lower bound of the forward diffusion process and the minimum reverse denoising process, specifically including the following steps:

[0120] Step 331: Construct the forward diffusion process: sample a real time window sample from the basic training set, and add Gaussian noise to it according to a predetermined noise schedule to obtain a noisy sample;

[0121] Specifically, a real time window sample is sampled from the basic training set, a Markov chain with a fixed parameter is defined, and Gaussian noise is gradually added to the data through steps. For each time step , the forward transition probability is defined as:

[0122]

[0123] wherein, is the noisy sample at time step , is the time step is the noisy sample, is the preset variance schedule parameter, denotes the identity matrix, denotes the Gaussian distribution.

[0124] Define and define the cumulative product . Then the marginal distribution from to satisfies the closed-form solution:

[0125]

[0126] Accordingly, the noisy sample can be directly calculated by sampling the standard Gaussian noise as the real added noise:

[0127]

[0128] Step 332: Define the conditional reverse denoising process: input the noisy sample , the current diffusion time step and the corresponding latent variable into the spatio-temporal conditional diffusion model to predict the added noise.

[0129] Specifically, the generation process of the diffusion model corresponds to the inverse process of the above forward process. Since the real inverse distribution is difficult to calculate directly, the spatio-temporal conditional diffusion model is used to approximate it. Under the condition of introducing the latent variable , the reverse transition probability is parameterized as a Gaussian distribution:

[0130]

[0131] where is the mean value predicted by the spatio-temporal conditional diffusion model, is the cumulative value predicted by the spatio-temporal conditional diffusion model.

[0132] Step 333: Construct a loss function based on noise prediction and optimize the spatio-temporal conditional diffusion model.

[0133] In order to simplify the training, the predicted mean value is converted into the predicted added noise . The loss function based on noise prediction is the mean square error loss:

[0134]

[0135] minimizing the loss function by a stochastic gradient descent algorithm updating network parameters of the spatio-temporal conditional diffusion model to learn to utilize latent variables recovering structured degraded data conforming to physical degradation rules from noise.

[0136] Step 4: generating multi-sensor degradation time window samples based on the trained spatio-temporal conditional diffusion model, and fusing the generated multi-sensor degradation time window samples with the basic training set to construct an enhanced training set.

[0137] Specifically, the generation and fusion process includes the following steps 41 to 42.

[0138] Step 41: generating multi-sensor degradation time window samples using a deterministic sampling strategy of a denoising diffusion implicit model (DDIM).

[0139] Specifically, in order to solve the problem of traditional diffusion model that has many reverse sampling steps and low generation efficiency, the DDIM deterministic sampling strategy is adopted to establish an implicit mapping from the latent variable to the degraded sample. First, an initial noise vector is sampled from a standard Gaussian distribution . Then, under the condition guidance of the latent variable , the trained spatio-temporal conditional diffusion model is used to predict noise , and iterative update is performed according to the deterministic non-Markov chain formula.

[0140] The formula for estimating noiseless data using the predicted noise is:

[0141]

[0142] wherein, is the cumulative product of the preset noise variance scheduling parameters in the forward process.

[0143] Based on and the predicted noise, the sampling formula for the state at the previous time is:

[0144]

[0145] Through step-by-step iteration from to 1, the denoised generated sample is finally obtained, which constitutes a virtual degradation time window sample set . Since the process is deterministic and constrained by the physically consistent latent variable , the generated sample can maintain high fidelity and conform to the physical degradation rules.

[0146] Step 42, merge the generated samples with the base training set to construct an enhanced training set.

[0147] Specifically, the virtual degradation time window sample set generated in step 41 is regarded as an additional training sample, and the original base training set is merged to construct an enhanced training set . The enhanced training set expands the scale of the training data and enriches the diversity of the data distribution, especially for the degradation samples under rare fault modes and complex working conditions, thereby alleviating the imbalance problem of the original data.

[0148] Through the above process, the enhanced training set is finally obtained, and the application effect of the enhanced training set proposed in the embodiment in the remaining life prediction is verified through related experiments.

[0149] First, the remaining life prediction model is constructed, and the enhanced training set is used for supervised training.

[0150] Specifically, in order to verify the universality and effectiveness of the data enhancement method proposed in the present application, any one of CNN, RNN, Transformer (denoted as Tran in all subsequent tables) or Mamba (denoted as Mam in all subsequent tables) network is selected as the downstream remaining life prediction model.

[0151] The enhanced training set constructed in step 4 is used for model training. Different time window lengths are set for device data with different working condition complexities : for data sets with relatively simple working conditions (such as FD001 and FD003 in C-MAPSS), the time window length is set to ; for data sets with complex working conditions (such as FD002 and FD004), the time window length is set to . The mean square error is used as the loss function in the training process, and the model parameters are updated by the Adam optimizer.

[0152] Secondly, the target data to be predicted is input into the trained model to output the prediction result.

[0153] Specifically, after the model training is completed, the real-time multi-sensor monitoring sequence of the target device to be predicted is obtained. According to the same preprocessing standard as step 1, the monitoring sequence is subjected to working condition clustering, clustering normalization processing, and the time window sample at the current time is intercepted by using the sliding window strategy.

[0154] The sample is input into the trained residual life prediction model, and the model directly outputs the current residual life prediction value of the device. Since the high-fidelity samples generated by the present application that meet the physical monotonicity constraint are fused in the training data, the model can more accurately identify the degradation trend, especially showing stronger robustness under complex working conditions.

[0155] Finally, the quality of the generated data and the prediction error indicators are used to evaluate the improvement effect of the enhanced training set constructed in this embodiment on the residual life prediction performance.

[0156] Specifically, in order to comprehensively verify the beneficial effects of the present application, experimental verification was carried out on the NASA C-MAPSS standard data set.

[0157] First, the quality of the generated data is qualitatively evaluated. As shown in Figure 5 generated degradation trajectories are compared with real degradation trajectories. Taking the FD003 subset as an example, the generated sensor data sequence accurately reproduces the subtle fluctuation characteristics of the real signal in the early approximate stationary stage; in the later accelerated degradation stage, it successfully captures the obvious downward trend and slope change, and there is no physically unreasonable mutation or artifact. This shows that the two-stage conditional encoder network and the spatiotemporal conditional diffusion model of the present embodiment can generate high-fidelity physically consistent samples.

[0158] Secondly, the root mean square error (RMSE) and the asymmetric scoring function Score are used as quantitative indicators to compare the prediction performance of training using only original data (Base configuration) and using the enhanced data of the present application (DA configuration). The experimental results are shown in Table 1.

[0159] Table 1 Comparison of experimental results

[0160]

[0161] The experimental results show that:

[0162] 1. The prediction accuracy is significantly improved: on the most complex subset FD004 containing multiple working conditions and multiple failure modes, after using the enhanced data of the present embodiment, the RMSE of the CNN model is reduced from 22.75 to 20.39 (reduced by about 10.37%), and the RMSE of the Mamba model is reduced from 16.28 to 14.62 (reduced by about 10.19%).

[0163] 2. The safety of the prediction is greatly enhanced: The Score index imposes more severe punishment on the overestimation of RUL and can better reflect the safety of the prediction. Experiments show that the Score under the DA configuration generally decreases more than the RMSE. For example, on the FD003 subset, the Score of the RNN model decreases from 416.38 to 202.42 (about 51.38% reduction); on the FD002 subset, the Score of the Mamba model decreases by about 31.49%. This proves that the physically consistent data generated by the embodiment effectively suppresses the dangerous overestimation of RUL by the model.

[0164] 3. The stability of the model is enhanced: As shown in the standard deviation analysis results in Figure 6 , the fluctuation of the results of multiple repeated experiments under the DA configuration is significantly smaller than that under the Base configuration, proving that the method enhances the robustness of the model under different initialization conditions.

[0165] In addition, the necessity of the related technical features in the method of the embodiment is verified through comparative and ablation experiments of similar methods. Specifically, in order to prove the advancement of the technical solution, the following comparative and ablation experiments are performed:

[0166] 1. Comparison with existing data enhancement methods:

[0167] The method of the present application (STCDM) is compared with existing representative enhancement methods, including the generative adversarial network (CR-GAN) and the diffusion-based method (DiffAT, DiffRUL, DegraDiff-Transformer / DDT). As shown in Figure 7 , when combined with CNN and RNN models, the RMSE and Score indexes of the present application method are at the lowest or near the lowest level on FD001, FD003 and FD004, and the advantage is obviously better than CR-GAN and other diffusion-based methods; when combined with Transformer and Mamba models, the present application method still leads on the multiple working condition subsets of FD002 and FD004, and the Score index is significantly better than the comparative methods. This proves that the spatio-temporal joint modeling strategy of the present application has superiority in processing complex working condition data.

[0168] 2. Ablation verification of the two-stage conditional encoder network:

[0169] As shown in Figure 11 and Figure 12 , the potential space trajectory visualization results of using only the first stage training (single stage) and the complete two-stage training (the present application) are compared:

[0170] When only using the first stage training, the sample distribution in the latent space is loose, the degradation trajectory appears local loop and self-intersection, and the RUL evolution is not strictly monotonic. After using the two-stage training (introducing monotonicity fine-tuning) of the present application, the latent space presents a clear and smooth manifold structure, and the samples strictly evolve monotonically along the direction of RUL reduction. This confirms that introducing monotonicity loss plays a key role in ensuring the physical consistency of generated data.

[0171] 3. Ablation verification of the spatiotemporal conditional diffusion model architecture:

[0172] As shown in Table 2, three architecture variants are compared: removing dilated depthwise separable convolution (Method I), removing Mamba state space module (Method II), and the complete model of the present application (STCDM).

[0173] Table 2. Ablation verification results

[0174]

[0175] The experimental results show that the complete STCDM has the lowest RMSE on all test subsets. For example, when using the Mamba predictor under the FD004 subset, the RMSE of STCDM is 14.62, which is better than 14.98 of Method I and 15.33 of Method II. This proves the necessity of jointly modeling the cross-sensor spatial correlation and long-range temporal dependence in the diffusion model.

[0176] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0177] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. The database involved in each embodiment provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, and is not limited thereto.

[0178] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not contradict, they should be considered as the scope of the present disclosure.

[0179] The principles and implementations of the present application are described in detail with specific examples in this paper, and the above examples are only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation and application range will be changed. Therefore, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A degraded data augmentation method based on a combination of conditional encoder and diffusion model, characterized in that: Includes the following steps: Step 1: Obtain multi-sensor degradation monitoring data of the device to be predicted, and preprocess the degradation monitoring data to obtain a basic training set containing degradation features and remaining service life labels; Step 2: Construct a two-stage conditional encoder network, and extract latent variables representing the degree of degradation and working condition information from the basic training set through multi-task joint learning and monotonicity fine-tuning; Step 3: Construct a spatiotemporal conditional diffusion model, which uses the latent variables as conditions to learn the inverse denoising process from Gaussian noise to real degraded data; Step 4: Generate multi-sensor degradation time window samples based on the trained spatiotemporal conditional diffusion model, and fuse the generated multi-sensor degradation time window samples with the basic training set to construct an enhanced training set.

2. The degradation data augmentation method based on the combination of conditional encoder and diffusion model according to claim 1, characterized in that: Step 1, the preprocessing of the degradation monitoring data, includes: Multi-condition data are clustered; within each condition cluster, the data from each sensor channel are normalized to minimum and maximum; a sliding time window is used to divide the degradation monitoring data into time window samples of length L and step size 1, and the actual remaining life at the end of the time window is used as the label.

3. The degradation data augmentation method based on the combination of conditional encoder and diffusion model according to claim 1, characterized in that: In step 2, the two-stage conditional encoder network includes a first encoder, a first decoder, and a regressor. The first encoder uses a temporal backbone network with embedded Mamba modules to map the input time window samples to the mean and variance of latent variables, and obtains the latent variables through reparameterized sampling. The first decoder receives the latent variables and reconstructs the time window samples through Mamba modules and transposed convolutional layers. The regressor receives the latent variables and outputs the predicted value of the remaining lifetime.

4. The degradation data augmentation method based on the combination of conditional encoder and diffusion model according to claim 3, characterized in that: In step 2, the training process of the two-stage conditional encoder network is divided into two stages: The first stage is the joint optimization stage, in which the network is trained using a joint loss function. The joint loss function includes: KL divergence loss for constraining the distribution of latent variables, reconstruction loss for minimizing the difference between input and output, regression loss for accurately predicting remaining lifetime, and triple contrast loss for optimizing the clustering characteristics of the latent space. The second stage is the monotonicity fine-tuning stage. Based on the convergence of the first stage training, monotonicity loss is introduced to fine-tune the two-stage conditional encoder network.

5. The degradation data augmentation method based on the combination of conditional encoder and diffusion model according to claim 4, characterized in that: In the joint optimization phase, the KL divergence loss is used to constrain the distribution of latent variables. for: Reconstruction loss used to minimize the difference between input and output for: Regression loss used to accurately predict remaining lifetime for: Triple contrastive loss for optimizing latent spatial clustering properties for: Final joint loss function for: in This indicates the number of samples after processing within the time window. and They represent the first Sample of a time window Corresponding latent variables The mean and variance; This represents the time window samples for reconstruction; Indicates the true remaining lifespan label. The regressor is based on latent variables. Output of the predicted remaining lifetime value; To make positive samples Input the latent variables obtained from the first encoder. To make negative samples Input the latent variables obtained from the first encoder. Indicates the preset boundary distance; These are the corresponding weights.

6. The degradation data augmentation method based on the combination of conditional encoder and diffusion model according to claim 5, characterized in that: Monotonicity fine-tuning stage, monotonicity loss for: in For the first Sample of a time window The corresponding output is the predicted remaining lifetime value. Indicates the monotonicity tolerance boundary; The total loss function in the second stage for: in For monotonic loss weights, For the first Sample of a time window The corresponding first-stage joint loss, For the first Sample of a time window The corresponding first-stage joint loss.

7. The degradation data augmentation method based on the combination of conditional encoder and diffusion model according to claim 1, characterized in that: In step 3, the spatiotemporal conditional diffusion model uses U-Net as the denoising backbone network and embeds spatiotemporal blocks in each layer of its encoder and decoder. The spatiotemporal blocks are composed of cascaded dilated depthwise separable convolutional modules and Mamba state space modules. The dilated depthwise separable convolutional modules are used to extract cross-channel spatial local correlations and multi-scale features in the sensor dimension. The Mamba state space modules are used to model the long-term degradation trend in the time dimension.

8. The degradation data augmentation method based on the combination of conditional encoder and diffusion model according to claim 7, characterized in that: The spatiotemporal conditional diffusion model employs a conditional injection mechanism based on characteristic linear modulation to inject the latent variables into the spatiotemporal conditional diffusion model. The conditional injection mechanism includes: mapping the diffusion time step to a time embedding vector, and mapping the latent variable to a conditional embedding vector; The temporal embedding vector and the conditional embedding vector are fused to generate channel scaling vector and channel offset vector. The channel scaling vector and the channel offset vector are then used to perform a channel-by-channel affine transformation on the feature map of the intermediate layer of the spatiotemporal conditional diffusion model in order to control the generation process of degraded samples.

9. The degradation data augmentation method based on the combination of conditional encoder and diffusion model according to claim 7, characterized in that: The training objective of the spatiotemporal conditional diffusion model is to minimize the noise prediction error: sample real time window samples from the basic training set, add Gaussian noise to them according to the preset noise schedule to obtain noisy samples; input the noisy samples, the current diffusion time step and the corresponding latent variables into the spatiotemporal conditional diffusion model to predict the added noise; calculate the mean square error between the predicted noise and the actual added noise as the loss function.

10. The degradation data augmentation method based on the combination of conditional encoder and diffusion model according to claim 1, characterized in that: In step 4, multi-sensor degradation time window samples are generated based on the trained spatiotemporal conditional diffusion model. Specifically, this includes: adopting a deterministic sampling strategy of the denoising diffusion implicit model, sampling initial vectors from the standard Gaussian distribution, predicting noise using the trained spatiotemporal conditional diffusion model under the guidance of latent variables, and iteratively updating according to the deterministic non-Markov chain formula to finally obtain denoised multi-sensor degradation time window samples.

Images