Method for predicting residual life of equipment based on multi-level diffusion degradation feature fusion

Abstract

The application relates to the field of equipment life prediction, and discloses an equipment residual life prediction method based on multi-level diffusion degradation feature fusion, which acquires aero-engine multi-element monitoring data, constructs a residual service life label sequence, and establishes a prediction model comprising a multi-level diffusion degradation representation module and a deep state space model, adopts a two-stage training strategy, performs self-supervised pre-training on the multi-level diffusion degradation representation module, uses a composite loss function composed of a reconstruction error term, a Kullback-Leibler divergence term, a monotonicity constraint term and an endpoint constraint term to perform end-to-end joint optimization on the multi-level diffusion degradation representation module and the deep state space model, and outputs a residual life prediction result and an uncertainty interval, so that the application realizes the cooperation of multi-level degradation feature extraction, degradation state evolution modeling and life uncertainty prediction, and improves the accuracy and stability of residual life prediction of complex equipment.

Description

Technical Field

[0001] This invention relates to the field of equipment life prediction methods, and more specifically to a method for predicting the remaining life of equipment based on the fusion of multi-level diffusion degradation features. Background Technology

[0002] As a complex and high-value piece of equipment, aero engines are affected by various factors such as load changes, environmental factors, and component performance degradation during operation, resulting in obvious nonlinear, time-varying, and uncertain characteristics in the internal state of the engine. Accurately predicting the remaining useful life (RUL) of aero engines is of great significance for achieving predictive maintenance, reducing operating costs, and improving the safety and reliability of equipment.

[0003] Traditional methods for predicting remaining service life mainly rely on physical model-based approaches, which describe the performance degradation process of components by establishing engine degradation mechanism models. However, since the degradation process of aero-engines is affected by the coupling of multiple factors, its internal degradation mechanism is complex and difficult to fully obtain. As a result, physical model-based methods have problems such as difficulty in model establishment, strong parameter dependence, and insufficient generalization ability under actual complex working conditions.

[0004] In recent years, with the development of sensor technology and data acquisition capabilities, data-driven remaining life prediction methods based on machine learning have gradually become an important research direction in the field of aero-engine health management. These methods learn the mapping relationship between historical operation monitoring data and degradation status, and achieve life prediction without explicitly establishing a degradation mechanism model. Traditional machine learning methods rely on manually designed features, such as statistical features, frequency domain features, and trend features, but their feature expression capabilities are limited and it is difficult to fully explore the complex degradation patterns hidden in multivariate monitoring data.

[0005] With the development of deep learning technology, residual lifetime prediction methods based on neural networks have achieved good prediction results in complex time series modeling tasks by automatically extracting degradation features using deep network structures. For example, methods such as recurrent neural networks, long short-term memory networks, and deep state space models can use time series information to describe the degradation process of equipment and predict future lifetime trends through hidden state evolution. However, existing deep learning methods usually learn degradation representations of a single scale or fixed scale directly from the original monitoring data, making it difficult to simultaneously capture local change information, stage degradation patterns, and long-term evolution trends in the degradation process of aero-engines.

[0006] On the other hand, most existing machine learning-based remaining lifetime prediction methods adopt a fixed feature fusion strategy, that is, after feature extraction, information is fused by splicing, weighted averaging or attention mechanism. The fusion weight is usually learned by fixed network parameters and cannot be dynamically adjusted as the equipment degradation state changes. In the actual operation of aero engines, the importance of features at different scales varies at different degradation stages, and the fixed fusion method is difficult to adapt to the complex dynamic degradation process.

[0007] In recent years, diffusion models, as a deep learning method based on probabilistic generative modeling, have demonstrated strong capabilities in learning complex data distributions and representing high-dimensional features. Utilizing the deep network structure of diffusion models, hierarchical feature representations can be extracted from different scales, providing a new technical approach for modeling equipment degradation processes. However, existing diffusion models are mainly applied to data generation or image processing. For the task of predicting the remaining life of aero-engines, effective methods are still lacking for extracting multi-level degradation representations using pre-trained diffusion models and achieving adaptive feature fusion by combining them with the dynamic degradation state of the equipment.

[0008] Therefore, given the characteristics of multi-scale degradation information coexistence, dynamic changes in the degradation process, and uncertainties in the prediction results of aero-engine multi-dimensional monitoring data, a method is needed that can extract multi-level degradation features using a pre-trained diffusion model, dynamically fuse them step-by-step according to the equipment degradation state, and combine them with a deep state space model to realize degradation state evolution and remaining life probability prediction, so as to improve the accuracy and reliability of aero-engine remaining life prediction. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a method for predicting the remaining life of equipment based on multi-level diffusion degradation feature fusion, comprising the following steps:

[0010] Step S1: Obtain multivariate monitoring data of the aero-engine, preprocess the monitoring data, construct a piecewise linear monotonically non-increasing remaining service life label sequence, and divide it into training set and test set;

[0011] Step S2: Construct a multi-level diffusion degradation characterization module, perform self-supervised pre-training on the diffusion model to enable it to learn the degradation distribution structure and evolution law of multi-dimensional monitoring data of aero-engines, and obtain the pre-trained diffusion model; use the pre-trained diffusion model to extract multi-level diffusion degradation features from the samples to obtain original layer degradation features, shallow layer degradation features, middle layer degradation features and deep layer degradation features.

[0012] Step S3: Based on the device degradation state, perform state-conditional modeling of the original layer degradation features, shallow layer degradation features, medium layer degradation features and deep layer degradation features, and generate a unified degradation driving representation based on the time-step dynamic fusion weights.

[0013] Step S4: Construct a deep state space model, take the unified degradation driving representation as the external degradation driving input, perform time-recursive evolution on the latent degradation state, and output the remaining lifetime probability prediction result based on the latent degradation state.

[0014] Step S5: Jointly train the multi-level diffusion degradation characterization module and the deep state space model to obtain the trained remaining lifetime prediction model.

[0015] Step S6: Input the test data into the trained remaining lifetime prediction model, output the remaining lifetime probability distribution, and obtain the remaining lifetime point estimation results and the corresponding uncertainty prediction interval through Monte Carlo sampling.

[0016] Preferably, in step S2:

[0017] The multi-level diffusion degradation characterization module includes a multi-level diffusion degradation feature extraction sub-module and a degradation state conditional dynamic fusion sub-module; wherein, the multi-level diffusion degradation feature extraction sub-module adopts a diffusion network based on a one-dimensional U-Net structure to perform self-supervised pre-training on multi-dimensional monitoring data of aero-engines, so that the diffusion model learns the potential distribution structure and degradation evolution law of equipment degradation data.

[0018] After pre-training is completed, the pre-training parameters of the multi-level diffusion degradation feature extraction submodule are used as initialization parameters to perform deterministic forward propagation on the input samples in order to extract degradation features corresponding to different network depth levels.

[0019] Preferably, in step S2:

[0020] In the feature extraction stage, the diffusion time step is set to zero, clean samples are forward propagated, and original layer degradation features, shallow layer degradation features, mid-layer degradation features, and deep layer degradation features are extracted from different depth levels of the diffusion network:

[0021] ;

[0022] In the formula, , , , These represent the outputs of the initial convolutional layer defined in the first stage, the first downsampling stage, the second downsampling stage, and the intermediate CBAM attention mechanism, respectively. This indicates that the diffusion time step is zero, meaning that deterministic forward propagation is performed on clean samples; , , , The first The original layer features, shallow layer features, middle layer features and deep layer features of each sample.

[0023] Preferably, in step S2:

[0024] The original layer degradation features are used to preserve the original local details and fluctuation information in the input monitoring data, the shallow layer degradation features are used to characterize local degradation details, the intermediate layer degradation features are used to characterize the stage degradation pattern, and the deep layer degradation features are used to characterize the global degradation trend.

[0025] Preferably, in step S3:

[0026] The degradation state conditional dynamic fusion submodule is used to adaptively fuse the original layer degradation features, shallow layer degradation features, mid-layer degradation features, and deep layer degradation features according to the device degradation state. First, the degradation features at different levels are aligned by time scale and mapped by channel dimension to obtain a unified dimension of layer degradation features. The hierarchical degradation features are arranged along the time dimension as follows:

[0027] ;

[0028] In the formula, For the first The first sample Hierarchy in time step The degenerate feature vector, The length of the time window;

[0029] At time step The previous moment's state was hidden. Together with the degenerate feature vectors of the four levels at the current time step, a state condition vector is constructed:

[0030] ;

[0031] Dynamic fusion weights are generated based on the state condition vector to represent the degradation features at each level at the current time step. , , , The dynamic fusion weights satisfy:

[0032] ;

[0033] Based on the dynamic fusion weights, the degradation features at each level of the current time step are weighted and fused to obtain the unified degradation driving features for the current time step:

[0034] ;

[0035] The unified degradation-driven feature sequence corresponding to the entire time window is obtained sequentially. The dynamic fusion weights, which serve as input to the deep state space model, change dynamically in the time dimension with the hidden degradation state, so that the contribution of each level of degradation features is adaptively adjusted at different time steps as the device degradation stage progresses.

[0036] Preferably, in step S4:

[0037] The deep state space model includes a degenerate state evolution module based on mLSTM and a remaining lifetime observation module based on VAE.

[0038] The mLSTM-based degenerate state evolution module uses a unified degenerate driving feature sequence as the external degenerate driving input, and combines the previous time step matrix memory and standardized state to recursively update the degenerate state of the current time step.

[0039] Preferably, in step S4:

[0040] The recursive update process of the current time step's degenerate state is represented as:

[0041] ;

[0042] In the formula, For the first Each sample at time step The unified degradation driving representation corresponds to the feature vector; For the first Each sample at time step The state of regression. and The first The matrix memory of each sample at the current time step and the previous time step and They represent the first The standardized state of a sample at the current time step and the previous time step.

[0043] Preferably, in step S4: the degenerate state Feedback is sent to the degradation state conditional dynamic fusion submodule to participate in the generation of time steps. The dynamic fusion weights enable the fusion of multi-level degradation features and the recursive evolution of latent degradation states to form a conditional closed loop that progresses along time. The VAE-based remaining lifetime observation module constructs the probability distribution of latent variables with latent degradation states as conditions and outputs the probability prediction results of remaining lifetime.

[0044] Preferably, in step S5: the multi-level diffusion degradation characterization module and the deep state space model are jointly trained. During the joint training phase, the multi-level diffusion degradation characterization module initialized with pre-trained parameters is fine-tuned, and a composite loss function is used to jointly optimize the multi-level diffusion degradation characterization module and the deep state space model.

[0045] Preferably, in step S5, the composite loss function includes:

[0046] ;

[0047] In the formula, The remaining lifetime reconstruction error is used to apply dense supervision to the remaining lifetime prediction distribution of each sample at each time step through Gaussian likelihood. The Kullback-Leibler divergence between the posterior and prior distributions of the latent variables; This is a monotonicity constraint term for remaining useful life, used to constrain the predicted remaining useful life of the same sample at adjacent time steps to not increase over time. This is the endpoint constraint term, used to ensure that the predicted remaining useful life at the end of each sample window is consistent with the actual remaining useful life label. , , These are the weighting coefficients for the Kullback–Leibler divergence, the remaining useful life monotonicity constraint, and the endpoint constraint, respectively.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. This invention utilizes a pre-trained diffusion model for degradation feature extraction, making full use of the multi-scale information contained in different depth levels of the diffusion network, and simultaneously acquiring local details, stage degradation patterns, and long-term degradation trends, thereby improving the degradation characterization capability of aero-engines.

[0050] 2. This invention proposes a state-conditional dynamic fusion mechanism, which introduces the device degradation state into the multi-level feature fusion process, so that the fusion weight can be dynamically adjusted with the time step and the device degradation state, avoiding the problem that the traditional fixed fusion method is not adaptable to complex degradation processes.

[0051] 3. This invention achieves continuous recursive evolution of the degradation state by inputting a unified degradation-driven representation into the deep state-space model, enabling the model to fully utilize long-term dependencies in the time series and improve the accuracy of remaining useful life prediction.

[0052] 4. This invention adopts a probabilistic lifetime prediction method, which obtains the remaining lifetime prediction distribution and uncertainty range through latent variable modeling and Monte Carlo sampling. It can not only output lifetime estimation results, but also reflect the degree of uncertainty of the prediction results, thereby improving the reliability of the prediction results.

[0053] 5. This invention achieves intelligent remaining lifetime prediction for complex degradation processes of aero-engines through the collaborative design of diffusion model pre-training, multi-level degradation representation, state conditional dynamic fusion, and deep state space prediction. Attached Figure Description

[0054] Figure 1 This is a flowchart of the present invention;

[0055] Figure 2 This is an overall structural diagram of the remaining lifetime prediction model of the present invention;

[0056] Figure 3 This is a schematic diagram of the diffusion model network of the present invention;

[0057] Figure 4 This is a schematic diagram of the depth state space model of the present invention;

[0058] Figure 5 This is a structural diagram of the mLSTM cell of the present invention;

[0059] Figure 6 This is a schematic diagram of the VAE network of the present invention;

[0060] Figure 7 This is a fitting graph of the remaining useful life prediction results of this invention on the FD001 dataset. Detailed Implementation

[0061] The present invention will be further described below with reference to the accompanying drawings and embodiments, which illustrate the above and other technical features and advantages of the present invention. However, the following embodiments are merely preferred embodiments of the present invention and are not exhaustive.

[0062] Example:

[0063] like Figures 1 to 7 As shown, the equipment remaining life prediction method based on multi-level diffusion degradation feature fusion includes the following steps:

[0064] Step S1: Obtain multivariate monitoring data of the aero-engine, preprocess the monitoring data, construct a piecewise linear monotonically non-increasing remaining service life label sequence, and divide it into training set and test set;

[0065] Step S11: Obtain multi-dimensional monitoring data of the aero-engine and perform feature filtering. In this embodiment, the multi-dimensional monitoring data comes from the FD001 dataset in the C-MAPSS dataset. This dataset contains complete time-series data of 100 turbofan engines from the start of operation to failure under the same operating conditions and a single fault mode. Each engine is equipped with 21 sensors and 3 operating condition monitoring parameters. After cleaning the acquired data, constant features and features that are not sensitive to degradation are removed. Sensor data numbered 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20 and 21 are selected from the subset of each engine and combined with the 3 operating condition monitoring parameters as input channels. In this embodiment, the total number of input channels F is 17.

[0066] Step S12: The Min-Max normalization method is used to map the data of each input channel to the range [0, 1]. The calculation formula is as follows:

[0067] ;

[0068] In the formula, represents the first... The sensor channel in the first The raw data at each time step; Indicates the first The sensor channel in the first Normalized data at each time step; Indicates the first The minimum value of each sensor, Representing the The maximum value of each sensor;

[0069] Step S13: A fixed-length sliding window is used to cut and encapsulate the continuous original features into two-dimensional matrix samples. In this embodiment, the window length is... Take 32, step size Take 1, and for data with a signal length less than 32, use linear interpolation to extend it to 32. Each sliding window sample is represented as:

[0070] ;

[0071] in, Indicates the first Each sample is located in the window position. Normalized multi-input channel data vector at the location, ;

[0072] For the remaining service life label, a piecewise linear RUL objective function is used to set it step by step over time, for the first time step of any engine's original time series. Each time step is labeled as follows:

[0073] ;

[0074] in, This represents the maximum number of cycles in the complete time series of the engine. This is the cutoff threshold for the early degradation stage. This represents the time step number in the original time series of the engine. In this embodiment, the truncation threshold is set to 125.

[0075] In generating the first A sliding window sample When, take the continuous coverage. The remaining useful life tags corresponding to each time step are arranged in chronological order to form the tag sequence of the sample. Thus forming the input matrix A monotonically non-increasing remaining useful life tag sequence with consistent time dimension;

[0076] Window samples from engines that have run to failure are used as the training set for parameter learning, while window samples from engines running in real time are used as the test set for performance verification and prediction.

[0077] Step S2: Construct a multi-level diffusion degradation characterization module, perform self-supervised pre-training on the diffusion model to enable it to learn the degradation distribution structure and evolution law of multi-dimensional monitoring data of aero-engines, and obtain the pre-trained diffusion model; use the pre-trained diffusion model to extract multi-level diffusion degradation features from the samples to obtain original layer degradation features, shallow layer degradation features, middle layer degradation features and deep layer degradation features.

[0078] Step S21, the first stage, involves self-supervised pre-training of the multi-level diffusion degradation feature extraction sub-module, enabling it to learn the potential distribution structure and degradation evolution law of multivariate monitoring data in a generative machine learning manner without relying on the remaining lifespan label.

[0079] The multi-level diffusion degradation feature extraction submodule adopts a diffusion network structure based on a one-dimensional U-shaped network (U-Net), including an encoder, intermediate layers, and a decoder, such as... Figure 2 and Figure 3As shown, the encoder uses one-dimensional convolution and gradually reduces the temporal dimension and increases the channel dimension through a downsampling path. Each downsampling stage includes residual blocks, an efficient channel attention (ECA) mechanism, and average pooling operations. The intermediate layers use residual blocks and CBAM attention mechanisms. The decoder gradually recovers the feature dimension through an upsampling path combined with skip connections.

[0080] During the forward diffusion process, the diffusion time step is randomly selected. Input Sample The data is gradually corrupted by Gaussian noise through forward diffusion, resulting in noisy data. :

[0081] ;

[0082] In the formula, For the first One original input sample, The time window length, This represents the total number of input channels. The noise is standard Gaussian and the elements are independent and identically distributed. , This is the cumulative retention factor. For the first Noisy data from each diffusion step, This represents the number of diffusion steps. The cumulative retention coefficient is obtained by multiplying the noise variance scheduling coefficients, i.e. , ;

[0083] Selecting a cosine noise scheduling method to generate noise variance scheduling coefficients. and truncate it within the interval To ensure numerical stability, the offset parameter is taken as... diffusion steps Pick Compared to linear noise scheduling, cosine noise scheduling can more gently disrupt the data structure during forward propagation, which is beneficial for the stable learning of degenerate features and the propagation time step. Sine position encoding (embedding dimension) Taking 16), we obtain the temporal embedding vector. Then, it is transformed into a high-dimensional time conditional vector by a multilayer perceptron. It is then injected into each residual block of the diffusion network to achieve adaptive feature modulation under time conditions.

[0084] During the forward propagation process, noisy data is added. and diffusion time step A common input diffusion network structure is used to output predicted noise through an encoder, intermediate layer, and decoder structure. ,in The learnable parameter set for the diffusion network structure is processed as follows:

[0085] First, add noise to the data. The initial features are obtained by mapping an initial one-dimensional convolutional layer with a kernel size of 7 to a high-dimensional feature space:

[0086] ;

[0087] In the formula, This represents a one-dimensional convolution operation with a kernel size of 7. The initial convolutional layer reduces the number of channels from... Expand to Preservation of timing length The timing remains unchanged; in this embodiment, it is expanded from 17 to 34, while the timing length remains at 32.

[0088] The encoder consists of two downsampling modules, each containing two residual blocks, an ECA channel attention mechanism, and an average pooling operation. The calculation process of the first downsampling module is as follows:

[0089] ;

[0090] The second downsampling module has the same structure as the first downsampling module, further compressing the temporal dimension to obtain the second-level encoded features:

[0091] ;

[0092] In the formula, This represents a residual block with temporal step embedding, which includes a one-dimensional convolutional layer, a normalization layer, a nonlinear activation layer, and a skip connection branch, and embeds the diffusion temporal step. After mapping, it is injected into the convolutional features; when the number of input channels and output channels of the residual block are inconsistent, it is done through... One-dimensional convolution performs channel matching on skip connection branches; The channel attention mechanism is represented by global average pooling along the time dimension, followed by local cross-channel interaction modeling through one-dimensional convolution with adaptive kernel, and then generating attention weights for each channel through Sigmoid activation and adaptive recalibration of the feature channels, thereby enhancing the response of key degraded feature channels without dimensionality reduction.

[0093] This represents a one-dimensional average pooling operation with a window size and stride of 2, used to halve the time series length.

[0094] The first residual block and the second residual block work together to complete the task. arrive Channel expansion, timing length from Halved to The third and fourth residual blocks together complete the task assigned by... arrive Channel expansion, timing length from Halved to .

[0095] The intermediate layers encode features for the deepest layer. Two residual blocks and the CBAM attention mechanism are applied sequentially. The CBAM attention mechanism does not change the tensor shape of the features, but only performs adaptive recalibration of the features along the channel dimension and spatial dimension. Its calculation process is as follows:

[0096] ;

[0097] In the formula, This represents the CBAM attention mechanism, which consists of a channel attention submodule and a spatial attention submodule connected in series.

[0098] The decoder employs a step-by-step upsampling structure and fuses features from corresponding layers of the encoder via skip connections. The output is from an intermediate layer. As input, combined with encoder features and The calculation process for the two upsampling stages is as follows:

[0099] ;

[0100] ;

[0101] In the formula, Indicates an upsampling operation. This indicates a jump connection splicing along the channel dimension;

[0102] Finally, The first is obtained through a one-dimensional convolutional layer. Prediction noise for each sample:

[0103] ;

[0104] The decoder upsamples features twice and combines this with skip connections, transferring them from... Gradually recover to In this embodiment, the value was restored from 136×8 to 17×32.

[0105] The first stage aims to minimize prediction noise. With added standard Gaussian noise The mean squared error between them, the formula for calculating the pre-training loss function is:

[0106] ;

[0107] In the formula, Standard Gaussian noise, The predicted noise output by the diffusion network structure, Mean square error, The self-supervised pre-training loss function for the multi-level diffusion degradation feature extraction submodule;

[0108] Through the aforementioned self-supervised pre-training objectives, the multi-level diffusion degradation feature extraction submodule learns the intrinsic distribution structure of multivariate monitoring data, providing high-quality degradation feature representations for the subsequent remaining lifetime prediction stage;

[0109] In this embodiment, the first stage of pre-training uses the Adam optimizer, with an initial learning rate of 0.001, 50 training epochs, and gradient clipping to stabilize the training process.

[0110] In the second stage, the multi-level diffusion degradation feature extraction submodule, which was pre-trained in the first stage, is jointly trained with the degradation state conditional dynamic fusion submodule, the mLSTM-based hidden degradation state evolution module, and the VAE-based remaining lifetime observation module to achieve synergistic optimization of degradation feature extraction, time-varying feature fusion, hidden state evolution, and remaining lifetime prediction.

[0111] Among them, the multi-level diffusion degradation feature extraction submodule is initialized with the parameters obtained in the first stage and fine-tuned with a differential learning rate: a small learning rate is used for this submodule to preserve the degradation distribution representation learned in the pre-training, and a large learning rate is used for the deep state space model.

[0112] In the second stage of this embodiment, the Adam optimizer is used, the number of training epochs is set to 120, and gradient clipping is used to stabilize the training.

[0113] In the second stage, the diffusion time step of the pre-trained multi-level diffusion degradation feature extraction submodule is fixed at 10 ... That is, using noise-free deterministic forward sampling to obtain clean samples Deterministic forward propagation is performed using the pre-trained diffusion network structure as input;

[0114] In the sample During the forward propagation along the diffusion network structure, multi-level feature sampling is performed based on network layers of different depths, resulting in four degenerate feature branches of different depths from four positions in the forward process:

[0115] ;

[0116] In the formula, , , , These represent the outputs of the initial convolutional layer defined in the first stage, the first downsampling stage, the second downsampling stage, and the intermediate CBAM attention mechanism, respectively. This represents the value when the diffusion time step is zero. , , , The first The original layer features, shallow layer features, middle layer features, and deep layer features of each sample, with dimensions respectively. , , and .

[0117] Among them, the original layer degradation features mainly retain the local details and original fluctuation information in the input monitoring data; the shallow layer degradation features are used to describe the local degradation details of the equipment; the intermediate layer degradation features are used to characterize the stage patterns in the degradation process; and the deep layer degradation features, after being enhanced by CBAM attention, can further reflect the overall degradation trend of the equipment throughout its entire life cycle. The four layers of degradation features depict the equipment degradation process from different depths and scales, providing rich degradation information for subsequent state-conditional dynamic fusion.

[0118] Step S3: Based on the device degradation state, perform state-conditional modeling of the original layer degradation features, shallow layer degradation features, medium layer degradation features and deep layer degradation features, and generate a unified degradation driving representation based on the time-step dynamic fusion weights.

[0119] The degenerate state conditional dynamic fusion submodule is used to condition the original layer features Shallow features Mid-layer characteristics and deep features By performing hierarchical attention-weighted fusion, a unified degradation-driven representation is obtained. The specific steps are as follows:

[0120] Since the time length and channel dimension of the degradation features at different levels are different, the time scale is first aligned and the channels are unified for each level of degradation features.

[0121] Based on the original time window length Based on this, interpolation upsampling is performed on degradation features at different levels, and then they are projected onto a uniform channel dimension through a one-dimensional convolution with a kernel size of 1. The unified hierarchical degradation feature is obtained, and its expression is:

[0122] ;

[0123] In the formula, This means linearly interpolating the input features along the time dimension to a length. Time-scale alignment operation; Indicates the first The kernel size corresponding to each level is [value]. One-dimensional convolution channel projection operation. Indicates the first The first sample The degradation features at each level are unified by the channel. After processing, the degradation features at each level all have length in the time dimension. The corresponding feature vector can be extracted according to the time step:

[0124] ;

[0125] In the formula, Indicates the first The first sample The level is at the The degenerate feature vector at each time step.

[0126] The degenerate state conditional dynamic fusion submodule of this invention uses the latent degenerate state as the system memory variable, models the multi-level feature selection process as a state conditional dynamic fusion process, and realizes the collaborative modeling of degenerate evolution and feature fusion through the recursive update of the latent degenerate state.

[0127] The fusion submodule generates time-varying fusion weights at each time step based on the previous time step's degeneracy state, enabling the weighting method of multi-level features to adaptively change with the degradation stage. Its specific implementation is as follows:

[0128] (1) Joint conditional modeling of state and multi-level features, at time step The previous moment's state was hidden. Concatenate the degenerate feature vectors from the four levels at the current time step along the feature dimension to construct a conditional vector:

[0129] ;

[0130] In the formula, For the first Each sample at time step The condition vector has dimensions of ; Degenerate state The four-level degradation feature vectors are used to characterize the degradation stage information of the system up to the previous time step. They are used to characterize the observable local details, stage patterns, and global trend information at the current time step. Together, they constitute the condition space for generating the time-varying fusion weights. At the beginning of the window, the latent degradation state is initialized with a zero vector, i.e. ;

[0131] (2) Time-varying fusion weight generation. The condition vector... The input consists of a weight generation network formed by sequentially connecting the first linear transformation, the nonlinear activation function, and the second linear transformation. This network is mapped to a four-dimensional hierarchical attention score vector, which is then normalized along the hierarchical dimension using Softmax to obtain the time-varying fusion weights for the four hierarchical degradation features corresponding to the current time step.

[0132] ;

[0133] In the formula, , and , These are the weight matrices and bias terms for the two linear transformations of the weight generation network, respectively. It is a non-linear activation function. Indicates the first Each sample at time step For the The fusion weights for hierarchical degradation features are time-varying weights generated step-by-step and sample-by-sample, dynamically adjusted according to the degradation state, and satisfying normalization constraints.

[0134] ;

[0135] (3) State-conditional weighted fusion. The four levels of degradation features at the current time step are weighted and fused using time-varying fusion weights to obtain the unified degradation driving feature for the current time step:

[0136] ;

[0137] In the formula, Indicates the first The sample at the th The unified degradation driving feature vector at each time step, along the time step By sequentially performing the above conditional modeling and weighted fusion, a unified degradation-driven feature sequence is obtained:

[0138] ;

[0139] The unified degradation-driven feature sequence is used as input to the deep state space model to drive the recursive evolution of the latent degradation state. At the same time, the latent degradation state output by the deep state space model will be fed back to the dynamic fusion module in the next time step to generate a new state condition vector and further calculate the new fusion weights.

[0140] Step S4: Construct a deep state space model, take the unified degradation driving representation as the external degradation driving input, perform time-recursive evolution on the latent degradation state, and output the remaining lifetime probability prediction result based on the latent degradation state.

[0141] like Figure 4 , Figure 5 and Figure 6 As shown, the deep state-space model includes a degenerate state evolution module based on matrix long short-term memory (mLSTM) and a remaining lifetime observation module based on variational autoencoder (VAE). In the state-space model, the dynamic evolution of the system state is characterized by state equations. The degenerate state evolution module based on mLSTM takes the unified degenerate driving representation output by the degenerate state conditional dynamic fusion submodule as the external degenerate driving input, and performs recursive modeling of the degenerate state through exponential gating mechanism and matrix memory structure.

[0142] For the Each sample, at time step Current state of degeneration Memory from the matrix of the previous time step and standardized state As a recursive state, and combined with the fusion degradation driving features of the current time step The recursive process of state updates can be represented as follows:

[0143] ;

[0144] In the formula, For the first Each sample at time step The unified degradation driving representation corresponds to the feature vector. For the first Each sample at time step The state of regression. and The first The matrix memory of each sample at the current time step and the previous time step and They represent the first The standardized state of each sample at the current time step and the previous time step;

[0145] First, from the input of the current time step Generate the query vector, key vector, and value vector respectively, and generate the input gate pre-activation value, forget gate pre-activation value, and output gate:

[0146] ;

[0147] ;

[0148] In the formula, , , The first Each sample at time step The query vector, key vector, and value vector. , These are the pre-activation values ​​for the input gate and the forget gate, respectively. For output gate, , , These are the weight matrices for the query vector, key vector, and value vector, respectively. This is the weight matrix of the output gate. , These are the weight vectors for the input gate and the forget gate, respectively. , , , , , These are the query vector, key vector, value vector, input gate pre-activation value, forget gate pre-activation value, and trainable bias term corresponding to the output gate, respectively. Let be the dimension of the key vector. This is the Sigmoid activation function.

[0149] Subsequently, the exponential input gate is obtained by exponential operation based on the pre-activation value of the input gate, and the forget gate is obtained by sigmoid operation based on the pre-activation value of the forget gate.

[0150] ;

[0151] In the formula, For exponential input gate, For the Gate of Oblivion This is for exponential operations.

[0152] Next, the matrix memory is updated according to the covariance update rule, and the standardized state is also updated:

[0153] ;

[0154] ;

[0155] Finally, based on the updated matrix memory Standardized status Query vector and output gate Calculate the degenerate state at the current time step:

[0156] ;

[0157] In the formula, This is an element-wise multiplication operation. This indicates taking the larger value between the independent variable and 1. For the first Each sample at time step The degenerate states, exponential input gates, and matrix memory structure enable this module to have a larger memory capacity and a stronger ability to characterize complex degenerate patterns compared to traditional recurrent neural networks with scalar memory units; in this embodiment, the dimension of the degenerate states... Set the value to 80, and the number of stacking layers for mLSTM to be 1;

[0158] In this embodiment, the implementation steps of the VAE-based remaining lifetime observation module are as follows:

[0159] In the state-space model, the observation equation describes the mapping relationship between the system's hidden states and observable variables. The VAE-based remaining lifetime observation module consists of a VAE encoder, a priori network, and a VAE decoder, as shown in the specific structure below. Figure 6 As shown, all three are implemented using fully connected neural networks combined with nonlinear activation functions;

[0160] At each time step Introducing latent variables ,in, Indicates the first The latent degradation characterization variables of each sample at the current time step are used to characterize the stochastic mapping relationship between the latent degradation state and the remaining lifetime observation. The prior network is based on the current latent degradation state. The prior distribution parameters for generating latent variables include the mean parameter of the prior network output. and standard deviation parameter The prior distribution is defined as:

[0161] ;

[0162] In the formula, For the first Each sample at time step Potential variables, For the first Each sample at time step The state of regression. The prior distribution output by the prior network. and These are the mean and standard deviation parameters output by the prior network, respectively. Indicates a Gaussian distribution. These are the trainable parameters of the prior network; the prior distribution is independent of current remaining lifetime observations, modeling the uncertainty of latent variables solely based on historical degradation states, enabling the latent spatial distribution to be determined according to the latent degradation states. It adapts to the data rather than using a fixed standard Gaussian distribution as a priori.

[0163] To further constrain the degenerate state under given observation conditions, latent variables are constructed within a variational inference framework. The approximate posterior distribution of the latent degenerate state The label of the actual remaining useful life at the current time step The VAE encoder is a common input, and the VAE encoder outputs the mean parameter of the latent variables. and standard deviation parameter The approximate posterior distribution is constructed as follows:

[0164] ;

[0165] In the formula, For the first Each sample at time step The label indicating the true remaining service life. This is the approximate posterior distribution of the encoder output. and These are the mean and standard deviation parameters of the encoder output, respectively. These are the trainable parameters of the encoder, and the latent variables are obtained from the approximate posterior distribution during training through a reparameterization method. This enables end-to-end joint optimization, and the approximate posterior incorporates both historical degradation information and current observation information, thereby improving the accuracy of latent variable inference.

[0166] Obtain latent variables Afterwards, the VAE decoder... Construct a probability mapping from the potential space to the observation space for the input, and output the mean parameter of the remaining lifetime likelihood distribution. and standard deviation parameter The Gaussian generation distribution observed for the remaining lifetime is as follows:

[0167] ;

[0168] In the formula, Generate a distribution for the remaining lifetime of the decoder output. and These are the mean and standard deviation parameters output by the decoder, respectively. Provides a conditional mean estimate of the remaining lifetime at the current time step; During the training phase, it serves as an adaptive weighting factor in the Gaussian likelihood, enabling the model to automatically adjust the reconstruction penalty strength based on the sample noise level, and taking the mean parameter of the decoder output. As the predicted remaining useful life at the current time step, i.e. In this embodiment, the dimensions of the latent variables Take 30.

[0169] Step S5: Jointly train the multi-level diffusion degradation characterization module and the deep state space model to obtain the trained remaining lifetime prediction model.

[0170] To achieve end-to-end joint optimization of the multi-level diffusion degradation representation module and the deep state-space model, a composite loss function is constructed within the variational inference framework to optimize the model parameters; the composite loss function... Including remaining useful life reconstruction error The Kullback-Leibler divergence between the posterior and prior distributions of the latent variables. Remaining useful life monotonicity constraint and endpoint constraints :

[0171] ;

[0172] In the formula, , , These are the weighting coefficients for the Kullback–Leibler divergence, monotonicity constraint, and endpoint constraint, respectively.

[0173] Among them, the remaining useful life reconstruction error Dense supervision is applied to the remaining lifetime prediction distribution of each sample at each time step using Gaussian likelihood:

[0174] ;

[0175] In the formula, The total number of samples, The time window length, For the corresponding real tags, and These are the predicted mean and predicted standard deviation of the decoder output, respectively. It acts as an adaptive weighting factor in Gaussian likelihood, enabling the model to automatically adjust the reconstruction penalty intensity based on the sample noise level.

[0176] Kullback–Leibler divergence between the posterior and prior distributions of the latent variables for:

[0177] ;

[0178] In the formula, This is the approximate posterior distribution of the encoder output. The prior distribution output by the prior network. This term represents the Kullback-Leibler divergence between two distributions. It imposes a regularization constraint on the latent spatial distribution, driving the approximate posterior to converge toward the prior and suppressing overfitting of latent variables.

[0179] Remaining useful life monotonicity constraint This is used to constrain the predicted remaining lifetime of the same sample at adjacent time steps to not increase over time.

[0180] ;

[0181] In the formula, A penalty is imposed for forecasts that increase, while no additional constraints are imposed for forecasts that decrease or remain unchanged.

[0182] Endpoint constraint To ensure that the predicted remaining useful life at the end of each sample window is consistent with the actual remaining useful life label:

[0183] ;

[0184] In the formula, and The first Each sample at the end of the window time step The predicted value and the true label, This term represents the absolute value and imposes a strong constraint on the accuracy of end-point prediction, which helps improve the model's prediction accuracy at critical failure moments.

[0185] In this embodiment, the weighting coefficient is taken as , , To ensure effective adaptation to downstream remaining lifespan prediction tasks while retaining the degradation features obtained from the pre-trained diffusion model, a differentiated learning rate strategy is adopted: a smaller learning rate is used for fine-tuning of the diffusion degradation feature extraction module, and a relatively larger learning rate is used for training the deep state space model. In this embodiment, the second stage of joint training uses the Adam optimizer, with 120 training epochs, and gradient pruning is used to stabilize the training process.

[0186] Step S6: Input the test data into the trained remaining useful life prediction model, output the remaining useful life probability distribution, and obtain the remaining useful life point estimation results and the corresponding uncertainty prediction interval through Monte Carlo sampling.

[0187] During the testing phase, the model relies solely on the prior network and decoder for predictions, without requiring actual remaining lifetime labels. To quantify prediction uncertainty, time steps at the end of the window are used. The remaining useful life prediction at the location is performed using multiple Monte Carlo sampling methods, for the first... For each Monte Carlo sampling, the mLSTM first recursively derives the latent degeneracy state at the end of the window time step based on the unified degeneracy-driven feature sequence. Then, based on the prior network The generated prior distribution is sampled to obtain the latent variables. Finally Input the VAE decoder and retrieve the mean parameter of the decoder output. As the predicted remaining useful life value for this sampling :

[0188] ;

[0189] Repeat sampling The predicted sample set is obtained this time. The sample mean is used as the point estimate of the remaining useful life, and the 2.5% and 97.5% quantiles of the sample are used as the lower and upper bounds of the 95% prediction interval, respectively.

[0190] ;

[0191] In the formula, The number of Monte Carlo samplings is 1000 in this embodiment; and These represent the 2.5% and 97.5% quantiles of the empirical prediction distribution, respectively. The uncertainty in the prediction interval stems from latent variables. Random sampling from the prior distribution: the prior distribution is in a degenerate state Parameterization reflects the uncertainty in our understanding of the actual degradation state of equipment; different The decoder maps the values ​​to different predicted means. This results in the predicted sample set exhibiting a distribution pattern consistent with the uncertainty of the degradation state. The prediction interval characterizes the uncertainty of the remaining useful life prediction results, providing a risk measurement basis for condition-based maintenance decisions.

[0192] Furthermore, the effectiveness of the remaining lifetime prediction model training is evaluated based on the root mean square error (RMSE) and the score function. When the effectiveness of the remaining lifetime prediction model meets the standard, a well-trained complete remaining lifetime prediction model is obtained.

[0193] The performance of the model is measured by two metrics: Root Mean Square Error (RMSE) and Score. The calculation formulas are as follows:

[0194] ;

[0195] ;

[0196] In the formula, The number of test samples, For the first The true remaining useful life value of each sample For the first Predicted remaining useful life for each sample;

[0197] The test set data was input into the complete remaining useful life prediction model to predict the remaining useful life of the equipment. The remaining useful life of the equipment is shown in Table 1.

[0198] Table 1. Predicted Remaining Useful Life of Equipment

[0199] RMSE Score FD001 FD001 12.35 255.85

[0200] Where RMSE is the root mean square error, used to assess the deviation between the predicted and actual values; Score is the scoring function value, which applies a weighted penalty to the prediction error of RUL, with a smaller penalty coefficient for early predictions and a larger penalty coefficient for delayed predictions.

[0201] To visually represent the error between the predicted and actual values, Figure 7 The image shows the fitted plot of the FD001 prediction results. The horizontal axis of the fitted plot represents the test engine number, and the vertical axis represents the remaining service life at the corresponding time. The blue line represents the actual remaining service life, the red line represents the predicted remaining service life, and the shaded area represents the 95% prediction interval.

[0202] Table 1 and Figure 7 In combination, this invention can effectively predict the remaining service life of aircraft turbofan engines and provide a quantitative characterization of prediction uncertainty. This allows for the arrangement of appropriate maintenance strategies and operating times based on the predicted RUL and prediction interval, providing users with a suitable decision-making basis for taking preventive measures before equipment failure occurs.

[0203] The above are merely preferred embodiments of the present invention and are illustrative rather than restrictive. Those skilled in the art will understand that many changes, modifications, and even equivalents can be made within the spirit and scope defined by the claims of the present invention, all of which will fall within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of equipment based on the fusion of multi-level diffusion degradation features, characterized in that, Includes the following steps: Step S1: Obtain multivariate monitoring data of the aero-engine, preprocess the monitoring data, construct a piecewise linear monotonically non-increasing remaining service life label sequence, and divide it into training set and test set; Step S2: Construct a multi-level diffusion degradation characterization module, perform self-supervised pre-training on the diffusion model to enable it to learn the degradation distribution structure and evolution law of multi-dimensional monitoring data of aero-engines, and obtain the pre-trained diffusion model; use the pre-trained diffusion model to extract multi-level diffusion degradation features from the samples to obtain original layer degradation features, shallow layer degradation features, middle layer degradation features and deep layer degradation features. Step S3: Based on the device degradation state, perform state-conditional modeling of the original layer degradation features, shallow layer degradation features, medium layer degradation features and deep layer degradation features, and generate a unified degradation driving representation based on the time-step dynamic fusion weights. Step S4: Construct a deep state space model, take the unified degradation driving representation as the external degradation driving input, perform time-recursive evolution on the latent degradation state, and output the remaining lifetime probability prediction result based on the latent degradation state. Step S5: Jointly train the multi-level diffusion degradation characterization module and the deep state space model to obtain the trained remaining lifetime prediction model. Step S6: Input the test data into the trained remaining lifetime prediction model, output the remaining lifetime probability distribution, and obtain the remaining lifetime point estimation results and the corresponding uncertainty prediction interval through Monte Carlo sampling.

2. The equipment remaining life prediction method based on multi-level diffusion degradation feature fusion according to claim 1, characterized in that: In step S2: The multi-level diffusion degradation characterization module includes a multi-level diffusion degradation feature extraction sub-module and a degradation state conditional dynamic fusion sub-module; wherein, the multi-level diffusion degradation feature extraction sub-module adopts a diffusion network based on a one-dimensional U-Net structure to perform self-supervised pre-training on multi-dimensional monitoring data of aero-engines, so that the diffusion model learns the potential distribution structure and degradation evolution law of equipment degradation data. After pre-training is completed, the pre-training parameters of the multi-level diffusion degradation feature extraction submodule are used as initialization parameters to perform deterministic forward propagation on the input samples in order to extract degradation features corresponding to different network depth levels.

3. The method for predicting remaining equipment lifetime based on multi-level diffusion degradation feature fusion according to claim 2, characterized in that: In step S2: In the feature extraction stage, the diffusion time step is set to zero, clean samples are forward propagated, and original layer degradation features, shallow layer degradation features, mid-layer degradation features, and deep layer degradation features are extracted from different depth levels of the diffusion network: ； In the formula, , , , These represent the outputs of the initial convolutional layer defined in the first stage, the first downsampling stage, the second downsampling stage, and the intermediate CBAM attention mechanism, respectively. This indicates that the diffusion time step is zero, meaning that deterministic forward propagation is performed on clean samples; , , , The first The original layer features, shallow layer features, middle layer features and deep layer features of each sample.

4. The method for predicting remaining equipment lifetime based on multi-level diffusion degradation feature fusion according to claim 3, characterized in that: In step S2: The original layer degradation features are used to preserve the original local details and fluctuation information in the input monitoring data, the shallow layer degradation features are used to characterize local degradation details, the intermediate layer degradation features are used to characterize the stage degradation pattern, and the deep layer degradation features are used to characterize the global degradation trend.

5. The method for predicting remaining equipment lifetime based on multi-level diffusion degradation feature fusion according to claim 4, characterized in that: In step S3: The degradation state conditional dynamic fusion submodule is used to adaptively fuse the original layer degradation features, shallow layer degradation features, mid-layer degradation features, and deep layer degradation features according to the device degradation state. First, the degradation features at different levels are aligned by time scale and mapped by channel dimension to obtain a unified dimension of layer degradation features. The hierarchical degradation features are arranged along the time dimension as follows: ； In the formula, For the first The first sample Hierarchy in time step The degenerate feature vector, The length of the time window; At time step The previous moment's state was hidden. Together with the degenerate feature vectors of the four levels at the current time step, a state condition vector is constructed: ； Dynamic fusion weights are generated based on the state condition vector to represent the degradation features at each level at the current time step. , , , The dynamic fusion weights satisfy: ； Based on the dynamic fusion weights, the degradation features at each level of the current time step are weighted and fused to obtain the unified degradation driving features for the current time step: ； The unified degradation-driven feature sequence corresponding to the entire time window is obtained sequentially. The dynamic fusion weights, which serve as input to the deep state space model, change dynamically in the time dimension with the hidden degradation state, so that the contribution of each level of degradation features is adaptively adjusted at different time steps as the device degradation stage progresses.

6. The method for predicting remaining equipment lifetime based on multi-level diffusion degradation feature fusion according to claim 1, characterized in that: In step S4: The deep state space model includes a degenerate state evolution module based on mLSTM and a remaining lifetime observation module based on VAE. The mLSTM-based degenerate state evolution module uses a unified degenerate driving feature sequence as the external degenerate driving input, and combines the previous time step matrix memory and standardized state to recursively update the degenerate state of the current time step.

7. The method for predicting remaining equipment lifetime based on multi-level diffusion degradation feature fusion according to claim 6, characterized in that: In step S4: The recursive update process of the current time step's degenerate state is represented as: ； In the formula, For the first Each sample at time step The unified degradation driving representation corresponds to the feature vector; For the first Each sample at time step The state of regression. and The first The matrix memory of each sample at the current time step and the previous time step and They represent the first The standardized state of a sample at the current time step and the previous time step.

8. The method for predicting remaining equipment lifetime based on multi-level diffusion degradation feature fusion according to claim 7, characterized in that: In step S4: Degenerate state Feedback is sent to the degradation state conditional dynamic fusion submodule to participate in the generation of time steps. The dynamic fusion weights enable the fusion of multi-level degradation features and the recursive evolution of latent degradation states to form a conditional closed loop that progresses along time. The VAE-based remaining lifetime observation module constructs the probability distribution of latent variables with latent degradation states as conditions and outputs the probability prediction results of remaining lifetime.

9. The method for predicting remaining equipment lifetime based on multi-level diffusion degradation feature fusion according to claim 1, characterized in that: In step S5: the multi-level diffusion degradation representation module and the deep state space model are jointly trained. During the joint training phase, the multi-level diffusion degradation representation module initialized with pre-trained parameters is fine-tuned, and a composite loss function is used to jointly optimize the multi-level diffusion degradation representation module and the deep state space model.

10. The method for predicting remaining equipment lifetime based on multi-level diffusion degradation feature fusion according to claim 9, characterized in that: In step S5, the composite loss function includes: ； In the formula, The remaining lifetime reconstruction error is used to apply dense supervision to the remaining lifetime prediction distribution of each sample at each time step through Gaussian likelihood. The Kullback-Leibler divergence between the posterior and prior distributions of the latent variables; This is a monotonicity constraint term for remaining useful life, used to constrain the predicted remaining useful life of the same sample at adjacent time steps to not increase over time. This is the endpoint constraint term, used to ensure that the predicted remaining useful life at the end of each sample window is consistent with the actual remaining useful life label. , , These are the weighting coefficients for the Kullback–Leibler divergence, the remaining useful life monotonicity constraint, and the endpoint constraint, respectively.

Images