Method for predicting residual service life of equipment, equipment and medium

Abstract

The invention discloses an equipment remaining service life prediction method, equipment and a medium, and the method comprises the steps: firstly obtaining operation monitoring data of source domain and target domain equipment, then constructing according to the source domain data to obtain a first input time sequence, and inputting the first input time sequence into a source domain prediction model framework; a feature extraction model and a probability prediction model are arranged in the model framework. Performing feature processing on the sequence through a feature extraction model to obtain comprehensive degradation features, training a probability prediction model by using the degradation features, and obtaining a trained model framework and corresponding posterior parameters based on the trained prediction model; and after a target domain prediction model is constructed according to the target domain monitoring data and the parameters, predicting the to-be-detected equipment through the model to obtain a prediction result of the equipment and corresponding uncertainty information. According to the method, robust prediction in a cross-domain scene can be realized, and meanwhile, corresponding uncertainty information is provided, so that a more comprehensive reference basis is provided for an equipment maintenance decision.

Description

Technical Field

[0001] This application relates to, but is not limited to, the field of prediction and health management technology, and particularly to a method, device, and medium for predicting the remaining useful life of equipment. Background Technology

[0002] Remaining Useful Life (RUL) prediction is a crucial task in industrial equipment health management, significantly impacting equipment maintenance strategy optimization, cost reduction, and safety assurance. Existing methods primarily include physical model-based and data-driven approaches. Physical model-based methods require precise descriptions of physical laws, but are limited by model assumptions, making them difficult to adapt to complex, nonlinear, and cross-equipment operating conditions. Data-driven methods, often employing deep learning models for point estimation predictions, while achieving high accuracy in single-equipment or single-condition scenarios, generally suffer from issues such as neglecting prediction uncertainty modeling, failing to provide risk assessment information, insufficient generalization ability across cross-equipment or cross-condition scenarios, and a lack of systematic quantification methods for model parameter uncertainties and input data noise.

[0003] Therefore, how to achieve robust RUL prediction in cross-domain scenarios and quantify the uncertainty of the prediction has become a technical problem that urgently needs to be solved in the field of industrial health management. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0005] This application provides a method, device, and medium for predicting the remaining useful life of equipment, which can effectively improve the robustness and practicality of predicting the remaining useful life of industrial equipment.

[0006] This application provides a method for predicting the remaining useful life of equipment, comprising: acquiring first operational monitoring data of a source domain device and second operational monitoring data of a target domain device; constructing a first input time series based on the first operational monitoring data, and inputting the first input time series into a source domain prediction model framework, wherein the source domain prediction model framework includes a feature extraction model and a probabilistic prediction model; performing feature extraction and feature fusion processing on the first input time series through the feature extraction model to obtain comprehensive degradation features; training the probabilistic prediction model using the comprehensive degradation features, and obtaining a trained source domain prediction model framework and corresponding posterior parameters based on the trained probabilistic prediction model; constructing a target domain prediction model based on the second operational monitoring data and the posterior parameters; and predicting the remaining useful life of the equipment through the target domain prediction model to obtain the prediction result of the remaining useful life of the equipment and the corresponding uncertainty information.

[0007] In one embodiment of this application, the feature extraction model includes an input encoding unit, a multi-layer encoding unit, and a feature fusion unit. The step of performing feature extraction and feature fusion processing on the first input time series through the feature extraction model to obtain comprehensive degradation features includes: performing linear mapping and positional encoding processing on the first input time series through the input encoding unit to obtain an initial feature representation; performing attention feature extraction on the initial feature representation through the multi-layer encoding unit to obtain multi-layer degradation feature information; performing average pooling on each layer of degradation feature information to obtain a corresponding feature vector; and performing weighted fusion of the feature vectors according to preset layer weights through the feature fusion unit to obtain comprehensive degradation features.

[0008] In one embodiment of this application, the probabilistic prediction model includes a first fully connected layer and a second fully connected layer. The step of training the probabilistic prediction model using the comprehensive degradation features, and obtaining a trained source domain prediction model framework and corresponding posterior parameters based on the trained probabilistic prediction model, includes: predicting the comprehensive degradation features using the first and second fully connected layers to obtain a predicted value of the remaining service life of the equipment; performing variational probability modeling on the weights and biases of the first and second fully connected layers, and obtaining a variational approximate posterior distribution through weight reparameterization; optimizing the probabilistic prediction model based on the joint loss of the prediction error and KL divergence, and obtaining a trained source domain prediction model framework based on the trained probabilistic prediction model; and extracting the corresponding posterior parameters from the variational approximate posterior distribution.

[0009] In one embodiment of this application, after obtaining the variational approximate posterior distribution, the method further includes: performing multiple weight reparameterizations based on the variational approximate posterior distribution to obtain the cognitive uncertainty corresponding to the predicted value; independently modeling the uncertainty caused by observation noise to obtain the corresponding stochastic uncertainty; and fusing the cognitive uncertainty and the stochastic uncertainty to obtain the overall uncertainty quantification result of the predicted value.

[0010] In one embodiment of this application, the step of constructing a target domain prediction model based on the second operational monitoring data and the posterior parameters includes: initializing the posterior parameters as prior parameters of the target domain model to obtain an initial target domain prediction model; constructing a second input time series based on the second operational monitoring data; extracting features from the first input time series and the second input time series respectively using the initial target domain prediction model to obtain source domain feature representations and target domain feature representations; determining the adversarial domain loss and mean difference loss between the source domain feature representations and the target domain feature representations; and performing distribution alignment optimization on the initial target domain prediction model based on the adversarial domain loss and mean difference loss to obtain a target domain prediction model whose feature distribution difference meets preset requirements.

[0011] In one embodiment of this application, the source domain prediction model framework further includes a domain adaptive model, which includes an adversarial domain discriminator and a distribution difference discriminator. The adversarial domain discriminator is used to calculate the domain category difference between the source domain feature representation and the target domain feature representation to obtain the adversarial domain loss. The distribution difference discriminator is used to calculate the overall distribution deviation between the source domain feature representation and the target domain feature representation to obtain the mean difference loss.

[0012] In one embodiment of this application, each coding unit has a normalization layer, a multi-head self-attention layer and a feedforward network layer connected in sequence. A first residual connection layer is connected between the output of the multi-head self-attention layer and the input of the feedforward network layer, and a second residual connection layer is connected to the output of the feedforward network layer.

[0013] In one embodiment of this application, the step of constructing a first input time series based on the first operation monitoring data includes: normalizing the first operation monitoring data to obtain standardized source domain monitoring data; dividing the standardized source domain monitoring data into time series using a preset sliding window to obtain time series subsequences of equal length; and obtaining the first input time series based on the time series subsequences of equal length.

[0014] On the other hand, embodiments of this application provide an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the method described above.

[0015] On the other hand, embodiments of this application provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the methods described above.

[0016] This application provides a method for predicting the remaining useful life of equipment, an electronic device, and a computer-readable storage medium. The method first acquires first operational monitoring data of a source domain device and second operational monitoring data of a target domain device. Then, a first input time series is constructed based on the first operational monitoring data and input into a source domain prediction model framework, which includes a feature extraction model and a probabilistic prediction model. Subsequently, the feature extraction model performs feature extraction and feature fusion processing on the first input time series to obtain comprehensive degradation features. These comprehensive degradation features are then used to train the probabilistic prediction model. Based on the trained probabilistic prediction model, a trained source domain prediction model framework and corresponding posterior parameters are obtained. After constructing a target domain prediction model based on the second operational monitoring data and posterior parameters, the target domain prediction model is used to predict the remaining useful life of the equipment, yielding a prediction result and corresponding uncertainty information. This application utilizes source domain degradation features to train the probabilistic prediction model, quantifying the uncertainty of model parameters. Combined with the subsequent construction of the target domain prediction model, robust prediction across domain scenarios can be achieved, effectively improving the insufficient generalization ability of existing methods in cross-device and cross-operating-condition scenarios. At the same time, while outputting the remaining service life prediction results through the target domain prediction model, it can also provide corresponding uncertainty information, providing a more comprehensive reference for equipment maintenance decisions. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for predicting the remaining useful life of a device provided in one embodiment of this application; Figure 2 This is provided in one embodiment of the present application. Figure 1 The detailed flowchart of step 130; Figure 3 This is provided in one embodiment of the present application. Figure 1 The detailed flowchart of step 140; Figure 4 This is provided in one embodiment of the present application. Figure 1 The detailed flowchart of step 150; Figure 5 This is a flowchart illustrating the remaining useful life prediction and uncertainty quantification provided in one embodiment of this application; Figure 6 This is a diagram of a prediction model architecture provided in one embodiment of this application; Figure 7 This is a prediction graph of the remaining service life and its uncertainty of the engine under Task 1 provided in one embodiment of this application; Figure 8 This is a prediction graph of the engine's remaining service life and its uncertainty under Task 2 provided in one embodiment of this application; Figure 9 This is a prediction graph of the engine's remaining service life and its uncertainty under Task 3 provided in one embodiment of this application; Figure 10 This is a structural diagram of a device for predicting the remaining useful life of an equipment, provided in an embodiment of this application. Figure 11 This is a hardware structure diagram of an electronic device provided in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be noted that although the flowchart shows a logical order, in some cases, the steps shown or described may be performed in a different order than that shown in the flowchart. The terms "first," "second," etc., used in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the structures, proportions, sizes, etc., depicted in the drawings are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effects and purposes achieved by this application, should still fall within the scope of the technical content disclosed in this application. Similarly, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are only for clarity of description and are not used to limit the scope of implementation of this application. Changes or adjustments in their relative relationships, without substantially altering the technical content, should also be considered within the scope of implementation of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] Remaining useful life (RUL) prediction is a crucial task in industrial equipment health management, significantly impacting equipment maintenance strategy optimization, cost reduction, and safety assurance. Existing methods primarily include physical model-based and data-driven approaches. Physical model-based methods require precise descriptions of physical laws and are limited by model assumptions, making them difficult to adapt to complex, nonlinear, and cross-equipment operating conditions. Data-driven methods, on the other hand, often employ deep learning models for point estimation predictions. While achieving high accuracy in single-equipment or single-condition scenarios, they generally suffer from several drawbacks, including neglecting to model prediction uncertainties and thus failing to provide risk assessment information, insufficient generalization ability across cross-equipment or cross-condition scenarios, and a lack of systematic quantification methods for model parameter uncertainties and input data noise.

[0022] Therefore, how to achieve robust RUL prediction in cross-domain scenarios and quantify the uncertainty of the prediction has become a technical problem that urgently needs to be solved in the field of industrial health management.

[0023] In view of this, embodiments of this application provide a method for predicting the remaining useful life of equipment, an electronic device, and a computer-readable storage medium. In this method, firstly, first operational monitoring data of a source domain device and second operational monitoring data of a target domain device are acquired. Then, a first input time series is constructed based on the first operational monitoring data and input into a source domain prediction model framework, which includes a feature extraction model and a probabilistic prediction model. Subsequently, the first input time series is subjected to feature extraction and feature fusion processing through the feature extraction model to obtain comprehensive degradation features. These comprehensive degradation features are then used to train the probabilistic prediction model. Based on the trained probabilistic prediction model, a trained source domain prediction model framework and corresponding posterior parameters are obtained. After constructing a target domain prediction model based on the second operational monitoring data and posterior parameters, the target domain prediction model is used to predict the remaining useful life of the equipment, yielding a prediction result and corresponding uncertainty information. Embodiments of this application utilize source domain degradation features to train the probabilistic prediction model, which quantifies the uncertainty of model parameters. Combined with the subsequent construction of the target domain prediction model, robust prediction in cross-domain scenarios can be achieved, effectively improving the insufficient generalization ability of existing methods in cross-device and cross-operating condition scenarios. At the same time, while outputting the remaining service life prediction results through the target domain prediction model, it can also provide corresponding uncertainty information, providing a more comprehensive reference for equipment maintenance decisions.

[0024] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0025] Reference Figure 1 , Figure 1This is a flowchart of a method for predicting the remaining useful life of equipment provided in an embodiment of this application. The process may specifically include, but is not limited to, steps 110 to 160.

[0026] Step 110: Obtain the first operational monitoring data of the source domain devices and the second operational monitoring data of the target domain devices; Step 120: Construct a first input time series based on the first operation monitoring data, and input the first input time series into the source domain prediction model framework, which includes a feature extraction model and a probability prediction model; Step 130: Perform feature extraction and feature fusion processing on the first input time series using a feature extraction model to obtain comprehensive degradation features; Step 140: Train the probabilistic prediction model using the comprehensive degradation features. Based on the trained probabilistic prediction model, obtain the trained source domain prediction model framework and the corresponding posterior parameters. Step 150: Construct a target domain prediction model based on the second operational monitoring data and posterior parameters; Step 160: Use the target domain prediction model to predict the equipment to be tested, and obtain the prediction results of the remaining service life of the equipment and the corresponding uncertainty information.

[0027] Steps 110 to 160 will be described in detail below.

[0028] In one feasible embodiment, the source domain device refers to a device that has accumulated sufficient monitoring data, has clear fault labels, or has a clear degradation pattern. The first operational monitoring data refers to the multi-sensor time-series data of the source domain device. This data can be obtained by various monitoring sensors and has complete supporting label information, covering the remaining service life label, fault occurrence time, and other relevant information throughout the entire life cycle. For example, in the task of predicting the remaining service life of a turbofan engine, the engine's operational monitoring data specifically includes multivariate time-series signals collected by various engine sensors. These multivariate time-series signals can include one or more of the following: temperature signal, pressure signal, speed signal, fuel flow signal, and vibration signal, or they can include multiple of the above signals simultaneously.

[0029] In one feasible embodiment, the target domain device refers to the device that actually needs to be predicted for its remaining useful life, but currently lacks sufficient labeled data, or whose data distribution differs from that of the source domain device. The second operational monitoring data refers to the multi-sensor time series data of the target domain device. This data can also be obtained by various monitoring sensors, but the overall data resources are relatively scarce. For example, only a small amount of daily operation data of the device can be obtained, and the complete degradation cycle data of the device cannot be obtained. At the same time, the actual operating conditions of the target domain device can be more complex and variable, and there may be certain individual differences between it and the source domain device, such as different production batches or different years of use.

[0030] In a feasible embodiment, when constructing the first input time series based on the first operational monitoring data, the first operational monitoring data can first be normalized to obtain standardized source domain monitoring data. Then, the standardized source domain monitoring data is divided into time series using a preset sliding window to obtain time series subsequences of equal length. Finally, the first input time series is obtained based on the time series subsequences of equal length. Specifically, the original first operational monitoring data is first normalized. This step is to eliminate the dimensional differences between monitoring data from different sensors. For example, the numerical ranges and units of indicators such as temperature, pressure, and rotational speed are different. After normalization, all data can be mapped to a unified numerical range, avoiding the impact of dimensional issues on the feature extraction and training effect of subsequent models. After processing, standardized source domain monitoring data can be obtained. Next, the standardized source domain monitoring data is divided into time series using a preset sliding window. The sliding window can slide and truncate segments on the continuous time series monitoring data according to a set length and step size. Each truncation will obtain a time series subsequence of consistent length. Finally, a series of time series subsequences of equal length can be obtained from the complete time series data, ensuring that the time dimension of all subsequences is consistent. Finally, by integrating these time series subsequences of equal length in their original chronological order, we can obtain the first input time series, whose format and dimensions match the input requirements of the source domain prediction model.

[0031] In one feasible embodiment, by analyzing data from each sensor Normalization can bring them to the same dimension, and the process is shown in equation (1):

[0032] in, Represents raw sensor data, This is the minimum value of the sensor. This is the maximum value of the sensor. This is the normalized data.

[0033] The normalized time series is divided into fixed lengths. Dividing the input into equal-length subsequences with stride, the final first input time series can be denoted as: ,in The length of the time series. This is the feature dimension (also the number of sensors).

[0034] In one feasible embodiment, when it is necessary to distinguish batches of the input time series, the input time series can be denoted as... ,in This refers to the batch size.

[0035] In one feasible embodiment, the feature extraction model is built on a Transformer network. Its overall model structure is similar to the coding structure of the standard Transformer, integrating an input coding unit, a multi-level coding unit, and a feature fusion unit. The input coding unit is mainly used to perform linear mapping and positional encoding on the input time-series sequence to ensure the orderliness of the time-series features. The multi-level coding unit is mainly used to extract deep, hierarchical degradation features from the encoded sequence data, mining the device degradation patterns hidden in the data layer by layer. The feature fusion unit can adaptively fuse the different levels of degradation features extracted by the multi-level coding unit through a trainable weighted fusion mechanism, fully integrating the effective information of each level of features to ultimately obtain a more comprehensive degradation feature representation with stronger representational capabilities.

[0036] In a feasible embodiment, each coding unit in the multi-layer coding unit is provided with a normalization layer, a multi-head self-attention layer and a feedforward network layer connected in sequence, wherein a first residual connection layer is connected between the output of the multi-head self-attention layer and the input of the feedforward network layer, and a second residual connection layer is connected between the output of the feedforward network layer and the input of the feedforward network layer.

[0037] In one feasible embodiment, such as Figure 2 As shown, step 130 involves performing feature extraction and feature fusion processing on the first input time series using a feature extraction model to obtain comprehensive degradation features. This process includes, but is not limited to, steps 210 to 240.

[0038] Step 210: Perform linear mapping and positional encoding on the first input time series through the input encoding unit to obtain the initial feature representation; Step 220: Attention feature extraction is performed on the initial feature representation through multi-layer coding units to obtain multi-layer degenerate feature information; Step 230: Perform average pooling on the degenerate feature information of each layer to obtain the corresponding feature vector; Step 240: The feature vectors are weighted and fused by the feature fusion unit according to the preset layer weights to obtain the comprehensive degradation features.

[0039] The processing flow of steps 210 to 240 is described below using an example.

[0040] In this embodiment, the first input time series is first processed. The linear mapping process is shown in equation (2):

[0041] in, The linear mapping weight matrix, For bias vectors, Dimensions of the feature extraction model.

[0042] After further positional encoding, the initial feature representation can be obtained:

[0043] in, It is a learnable positional code that can be used to preserve sequence order information.

[0044] Furthermore, the processing procedure of the multi-head self-attention mechanism is shown in equations (4) and (5):

[0045]

[0046] in, For the number of attention heads; , For the first Size search, For the first The key of the size, For the first The value weight matrix of each head; For multi-head output linear transformation matrix, .

[0047] Furthermore, the process of residual connectivity and layer normalization is shown in equation (6):

[0048] Furthermore, the processing procedure of the feedforward network is shown in equations (7) and (8):

[0049]

[0050] in, For trainable feedforward network weights, Here are the trainable bias parameters, and GELU is the activation function.

[0051] Furthermore, by performing average pooling on the degenerate feature information of each layer, the corresponding fixed-dimensional feature vector can be obtained, as shown in equation (9):

[0052] Furthermore, through trainable weights The feature vectors are weighted and fused using the preset layer weights to obtain the comprehensive degenerate features. The calculation process is shown in equations (10) and (11):

[0053]

[0054] in, To train learnable weights, This represents the number of coding unit layers.

[0055] In one feasible embodiment, the source domain prediction model framework also includes a domain adaptation model, which comprises an adversarial discriminator and a distribution difference discriminator. These two discriminators can be pre-trained during the source domain training phase to prepare for subsequent cross-domain alignment in the target domain. When the model is transferred to the target domain, these two discriminators can be directly used, thereby improving the efficiency and effectiveness of cross-domain adaptation.

[0056] In one feasible embodiment, the probabilistic prediction model is built on a Bayesian Prediction Network (BPN). Its model structure is consistent with that of the standard Bayesian prediction network. The model integrates a prediction head, a first fully connected layer, and a second fully connected layer. The prediction head is used to probabilistically model the comprehensive degradation features of the input, providing an adapted feature basis for subsequent variational inference. The first and second fully connected layers can work together to achieve hierarchical processing of feature transformation and mapping, avoiding insufficient feature fitting at a single level.

[0057] In one feasible embodiment, such as Figure 3 As shown, the execution process of step 140 may include, but is not limited to, steps 310 to 340.

[0058] Step 310: Predict the comprehensive degradation characteristics through the first fully connected layer and the second fully connected layer to obtain the predicted value of the remaining service life of the equipment; Step 320: Perform variational probability modeling on the weights and biases of the first and second fully connected layers, and obtain the variational approximate posterior distribution by reparameterizing the weights; Step 330: Optimize and train the probabilistic prediction model based on the joint loss of prediction error and KL divergence. Based on the trained probabilistic prediction model, obtain the trained source domain prediction model framework. Step 340: Extract the corresponding variational approximate posterior parameters from the variational approximate posterior distribution.

[0059] The processing flow of steps 310 to 340 is described below with reference to a specific embodiment.

[0060] This embodiment is based on the aforementioned specific embodiments. Comprehensive degradation characteristics. The prediction head of the probabilistic prediction model can be fed into the fully connected layer, and then the degenerate features can be processed through the first fully connected layer. By performing linear transformations and activation processes, intermediate features can be obtained. ; intermediate features are processed through a second fully connected layer By mapping, a predicted value of the remaining service life of the equipment can be obtained. The processing procedure is shown in equations (12) and (13):

[0061]

[0062] in, It is a point estimate (i.e., a predicted value) of the remaining useful life of the equipment.

[0063] Furthermore, by modeling the weights and biases of the fully connected layer using Gaussian distribution variational probabilities, a variational approximate posterior distribution can be obtained, which is used to approximate the weights. The true posterior probability distribution is used to quantify the uncertainty of the model parameters. The execution process is shown in equations (14) and (15):

[0064]

[0065] in, The mean of the weighted distribution; These are trainable parameters; Standard deviation; As a numerically stable term, its value can be set as follows: .

[0066] At the same time, weights can also be adjusted. Reparameterization sampling is performed, which quantifies the uncertainty of the model weights and provides a basis for obtaining prediction results with uncertainty. The process is shown in Equation (16):

[0067] It should be understood that weight reparameterization does not directly generate predicted values. Instead, after determining the probability distribution of the weights in this way, the weights are sampled multiple times and the calculation process of the fully connected layer is repeated to obtain multiple remaining lifetime predicted values. Then, the probability distribution of the prediction results is fitted to achieve the quantification of prediction uncertainty.

[0068] Furthermore, the deviation between the predicted and actual values ​​(i.e., the prediction error) can be calculated using the root mean square error (RMSE) and the mean absolute error (MAE), thereby obtaining the corresponding RMSE loss. and MAE loss The processing procedure is shown in equations (17) and (18):

[0069]

[0070] Furthermore, the KL divergence regularization process is shown in equation (19):

[0071] Based on this, the total source domain loss (i.e., joint loss) can be obtained:

[0072] in, These are the weighting coefficients for the MAE loss. These are the weighting coefficients for the divergence loss.

[0073] In one implementation, the weighting factor for the RMSE loss can be set to... This is to control the proportion of RMSE loss in the total loss.

[0074] During training, the model can adjust the mean of the weight distribution. and standard deviation While ensuring accurate predictions, the probability distribution of the weights maintains a reasonable level of uncertainty. When the joint loss converges to a stable value, the trained probabilistic prediction model is obtained. Replacing the original probabilistic prediction model with the trained model yields the trained source domain prediction model framework.

[0075] Furthermore, the mean can be extracted from the optimized variational approximation posterior distribution. with standard deviation As posterior parameters, they can be transferred as pre-trained weight distributions when constructing the target domain prediction model, thereby enabling the target domain model to inherit the degradation patterns and uncertainty quantification capabilities learned from the source domain.

[0076] In a feasible embodiment, after obtaining the variational approximate posterior distribution, the uncertainty of the predicted value can be quantified based on the distribution. The process includes: first, performing multiple weight reparameterizations based on the variational approximate posterior distribution to obtain the cognitive uncertainty corresponding to the predicted value of the remaining service life of the equipment; then, independently modeling the uncertainty caused by observation noise to obtain the corresponding stochastic uncertainty; and finally, fusing the cognitive uncertainty and the stochastic uncertainty to obtain the overall uncertainty quantification result of the predicted value.

[0077] The calculation process of the overall uncertainty quantification result is explained below with a specific example.

[0078] This embodiment is based on the aforementioned specific embodiment.

[0079] In one implementation, cognitive uncertainty stems from the probability distribution of the model weights, reflecting the inherent uncertainty of the model. T Monte Carlo samplings are performed on the model weights, and the prediction process is repeated to obtain T predicted values. By calculating the variance of these predicted values, the corresponding cognitive uncertainty can be obtained. The calculation process is shown in equations (21) and (22):

[0080]

[0081] The larger this variance value, the greater the fluctuation in the model's prediction of the sample, and the higher the cognitive uncertainty.

[0082] In one implementation, random uncertainty originates from noise in the observed data, reflecting fluctuations in the data itself. When calculating random uncertainty, one can... Input to a fully connected layer specifically designed to model stochastic uncertainty, the learnable weight matrix of that layer. Can be Perform a linear transformation, plus a learned bias. This yields an intermediate result. Then, this result is input into... In the activation function, by calculating Ensure the output is non-negative; this value represents the variance of the random uncertainty. This variance reflects the magnitude of data fluctuation caused by observation noise. The specific calculation process is shown in equation (23):

[0083] In one embodiment, the calculation process for the overall uncertainty quantification result is shown in equation (24):

[0084] In a feasible embodiment, after obtaining the optimized posterior parameters, these parameters can be used as a pre-trained weight distribution and transferred to the target domain model. This allows the target domain model to inherit the device degradation patterns learned from the source domain, while retaining the uncertainty quantification capabilities of the source domain. Figure 4 As shown, the execution process of step 150 may include, but is not limited to, steps 410 to 440.

[0085] Step 410: Initialize the posterior parameters to the prior parameters of the target domain model to obtain the initial target domain prediction model; Step 420: Construct the second input time series based on the second operation monitoring data; Step 430: Extract features from the first input time series and the second input time series using the initial target domain prediction model to obtain source domain feature representation and target domain feature representation; Step 440: Determine the adversarial loss and mean difference loss between the source domain feature representation and the target domain feature representation; Step 450: Based on the adversarial domain loss and mean difference loss, optimize the distribution alignment of the initial target domain prediction model to obtain a target domain prediction model whose feature distribution difference meets the preset requirements.

[0086] It should be noted that the target domain prediction model is essentially an adapted implementation model of the source domain prediction model after cross-domain adaptation of the unified architecture. It fully retains all the core functions of the unified architecture of the source domain and also has the same feature extraction (such as encoding the input time series to obtain a high-dimensional feature representation), domain adaptation, probability prediction, and dual uncertainty quantization functions as the source domain model. In other words, the target domain prediction model is not a completely new model trained independently, but a prediction model with target domain generalization ability formed after the model architecture weights of the source domain are adapted to the target domain data through parameter transfer initialization and feature distribution alignment optimization.

[0087] In a feasible embodiment, step 410 is the initialization step for cross-domain transfer, which involves using the posterior parameters obtained from training in the source domain. , Transferred to the target domain, serving as prior parameters for the target domain model. , This completes the construction of the initial target domain prediction model. Subsequently, the target domain model uses the posterior parameters of the source domain... , Use these initial values ​​to initialize their own variational posterior parameters. , This prepares for subsequent adaptive optimization in the target domain. It is the posterior mean of the Gaussian distribution of the weights after the source domain model has been trained; It is the expected value of the posterior standard deviation of the Gaussian distribution of the source domain model weights, reflecting the magnitude of the uncertainty of the source domain model weights; These are trainable parameters in the source domain used to calculate the standard deviation; It is the prior mean of the target domain model; It is the prior standard deviation of the target domain model; It is the variational posterior mean of the target domain model itself; These are the variational posterior parameters of the target domain model itself.

[0088] In a feasible embodiment, in step 420, when constructing the second input time series, the original second operational monitoring data can first be normalized to obtain standardized source domain monitoring data. Then, a preset sliding window is used to divide the standardized source domain monitoring data into time series segments. Segments are progressively extracted from the continuous time series monitoring data, with each extraction yielding a time series subsequence of uniform length. Ultimately, a series of equal-length time series subsequences can be obtained from the complete time series data. Integrating these equal-length time series subsequences according to their original chronological order yields the second input time series.

[0089] In a feasible embodiment, in step 430, feature extraction is performed on the training sets of the first input time series and the second input time series respectively using the initial target domain prediction model to obtain the source domain feature table. and target domain training feature representation .

[0090] In one feasible embodiment, the adversarial domain discriminator can be used to calculate the domain category difference between the source domain feature representation and the target domain feature representation to obtain the adversarial domain loss; the distribution difference discriminator can be used to calculate the overall distribution deviation between the source domain feature representation and the target domain feature representation to obtain the mean difference loss.

[0091] In a feasible embodiment, in step 440, the source domain feature representation is calculated by combining a cross-domain feature distribution alignment mechanism. and target domain training feature representation The adversarial domain loss and mean difference loss are used to quantify the difference in feature distribution between the source and target domains.

[0092] Among them, adversarial domain loss The calculation expression is:

[0093] in, The number of source domain samples (which can be obtained from the first input time series). The number of samples in the target domain (which can be obtained from the partitioning of the second input time series); The domain label is used to distinguish between source domain samples and target domain samples; It is a characteristic of unified representation, including and ; Domain discriminator predicts features The probability of originating from the source domain.

[0094] Maximum mean difference loss The calculation expression is:

[0095] in, It is a feature representation of the source domain samples. It is the feature representation of the target domain sample.

[0096] Furthermore, the total domain adaptation loss can be obtained as:

[0097] Where L is the number of iterations. These are the weighting coefficients of the MMD loss.

[0098] In a feasible embodiment, in step 450, after obtaining the domain adaptation total loss, this loss can be fused with the previously calculated prediction loss (RMSE, MAE) and KL divergence loss to obtain the final total loss. for:

[0099] in, , .

[0100] Subsequently, backpropagation is performed on the model based on the final total loss, adaptively updating the weights of the feature extraction network to gradually align the feature distributions of the source and target domains. This process is iterated until the feature distribution difference meets a preset threshold or the maximum number of training epochs is reached. After iteration, a target domain prediction model with target domain generalization ability is obtained. Then, a target domain test set partitioned from the second input time series can be used to independently evaluate the model's prediction performance and verify the effectiveness of cross-domain transfer.

[0101] In a feasible embodiment, after constructing the target domain prediction model, the operational monitoring data of the device under test (such as time-series data of vibration, temperature, and pressure) can be input into the target domain prediction model that has already undergone cross-domain adaptation. The model first encodes the raw monitoring data into high-dimensional features using a feature extraction network that has already achieved distribution alignment. Then, the prediction head calculates the point prediction value of the device's remaining useful life (RUL) based on these features. Simultaneously, the model calculates cognitive uncertainty and stochastic uncertainty. Specifically, by sampling the model weights multiple times, multiple prediction values ​​are generated and their variances are calculated, reflecting the uncertainty of the model's perception of the current device state; a dedicated uncertainty modeling branch learns from the input features and outputs a variance value representing the magnitude of data noise, reflecting the uncertainty caused by noise in the monitoring data. Finally, the model outputs the RUL prediction value and the corresponding uncertainty information. This uncertainty information is the overall uncertainty obtained by combining cognitive uncertainty and stochastic uncertainty, and is usually presented in the form of confidence intervals (such as 95% confidence intervals) to assess the reliability of the prediction results.

[0102] The processing procedure of this method will be described below with a specific embodiment.

[0103] like Figure 5As shown, the source domain dataset (i.e., the operational monitoring data of source domain devices) can first be preprocessed to obtain standardized time-series data suitable for model input. This data is then fed into the model framework. Within the framework, a feature extraction model can extract multi-level time-series features with contextual dependencies (i.e., multi-layer degradation feature information). These deep features are then integrated through feature fusion to obtain comprehensive degradation features. Simultaneously, the adversarial discriminator and distribution difference discriminator in the model can construct virtual domain differences within the source domain. Through adversarial training and distribution measurement, the feature extraction module outputs more generalized source domain features, while pre-training the domain discrimination and distribution measurement capabilities of these two discriminators. The probabilistic prediction model can probabilistically model the comprehensive degradation features, outputting the predicted value of the device's remaining lifespan through feature transformation and mapping of the first and second fully connected layers. Variational probability modeling is performed on the weights and biases of these two layers, and variational approximate posterior distributions are obtained through weight reparameterization. Furthermore, RMSE loss, MAE loss, and KL divergence loss can be calculated to drive backpropagation training of the model. After iterative optimization, a trained source domain prediction model framework and corresponding variational approximate posterior parameters can be obtained. Next, we can proceed to the target domain adaptation phase. The target domain dataset can also undergo data preprocessing, and then, combined with posterior parameters transferred from the source domain, prior parameters for the target domain can be learned. During this process, the domain adaptation module can calculate the adversarial domain loss and MMD loss of the features in the source and target domains, driving feature distribution alignment. The probabilistic prediction model can also fine-tune its weights based on the target domain data, allowing the initial target domain model to gradually adapt to the degradation patterns of the target domain, ultimately forming a trained target domain prediction model. In the inference phase, the trained target domain prediction model can process the monitoring data of the device to be detected, outputting the point prediction results of the device's remaining service life (RUL prediction results). Simultaneously, it can quantify the cognitive and stochastic uncertainties in the prediction process, conduct reliability assessments of the prediction results, and provide risk references for operational decisions.

[0104] The processing procedure of this method will be described below using another specific embodiment.

[0105] like Figure 6As shown, the feature extraction model contains multiple coding units, from Block 1 to Block N. Each coding unit contains a normalization layer, a multi-head self-attention layer, a residual connection layer, and a feedforward network layer. The input time series is processed by these coding units, and through operations such as attention feature extraction and average pooling, multi-layer degradation feature information can be obtained. These multi-layer degradation feature information are then fused to obtain a comprehensive degradation feature. This comprehensive degradation feature can be used to train a probabilistic prediction model, which is built on a Bayesian network. During training, the probabilistic prediction model performs probabilistic modeling on the comprehensive degradation feature. Through feature transformation and mapping of the first and second fully connected layers, it outputs the predicted value of the device's remaining lifespan. Simultaneously, it performs variational probabilistic modeling on the weights and biases of these two layers, obtaining a variational approximate posterior distribution through weight reparameterization. The model can construct a joint loss function based on the prediction error and KL divergence, driving backpropagation iterative optimization, ultimately yielding a trained source domain prediction model framework and the corresponding variational approximate posterior parameters. Next, in the target domain adaptation phase, the variational approximate posterior parameters obtained from the source domain training can be transferred and initialized as prior parameters for the target domain, and target domain prior learning can be carried out based on this. During this process, the total domain adaptation loss can be obtained by calculating the adversarial domain loss and the maximum mean difference loss, and then fused with the RMSE loss, MAE loss, and KL divergence loss to obtain the total loss. The total loss can drive the model to perform backpropagation, adaptively updating the weights of the feature extraction network and the probabilistic prediction model, so that the feature distributions of the source domain and the target domain are gradually aligned, and finally, a target domain prediction model adapted to the target domain data can be obtained.

[0106] The effectiveness of this method will be illustrated below with a specific embodiment.

[0107] This embodiment conducted system experiments across multiple devices and operating conditions. The main evaluation metrics included root mean square error (RMSE) and mean absolute error (MAE), defined as follows:

[0108]

[0109] in, This represents the actual remaining useful life value. Indicates the predicted value. and These represent the number of predicted samples, respectively. The smaller the RMSE and MAE values, the smaller the error.

[0110] In this embodiment, the encoding unit (Block) has a hidden dimension of 128, with 4 layers. The multi-head attention mechanism contains 4 heads, the feedforward network has a hidden dimension of 256, and the dropout rate is 0.2. In the probabilistic prediction model, the posterior scaling factor is 0.1, and the KL weights are 0.01. In the domain adaptation model, the discriminator has 256 hidden units, and the dropout rate is 0.3. The optimizer uses Adam with a learning rate of 1e-3 and a weight decay of 1e-4.

[0111] This embodiment, based on a turbofan engine dataset, aims to predict and analyze the remaining service life (RUL) of an engine under different operating conditions, as shown in Table 1. The experimental task comprises three specific prediction tasks, corresponding to different failure modes and operating condition transitions. Task 1 involves the transition from failure mode 1 to 2, while the operating condition changes from 1 to 6, simulating the engine degradation process under changing environments. Task 2 keeps the failure mode unchanged but the operating condition fixed at 6, examining the degradation patterns under high-load stable conditions. Task 3 keeps the failure mode at 2 while changing the operating condition from 1 to 6, verifying the model's adaptability and generalization ability under different combinations of modes and conditions. Overall, these tasks aim to comprehensively evaluate the performance of the prediction model under multiple operating conditions and failure modes, providing reliable data support for turbofan engine RUL prediction.

[0112] Table 1

[0113] In the experimental phase, the target domain prediction model of this application was compared with several control models in the task of predicting the remaining service life of a turbofan engine. The experiment covered three prediction tasks, corresponding to different failure modes and operating condition variation scenarios. The relevant results are summarized in Table 2. The values ​​in the table represent prediction errors; the lower the value, the higher the prediction accuracy of the model. Furthermore, Figure 7 , Figure 8 , Figure 9 The visualization results of the model's predictions of the engine's remaining service life and its uncertainties are presented under Task 1, Task 2, and Task 3, respectively.

[0114] Table 2

[0115] Experimental results show that models lacking domain adaptation experience increased errors across all tasks, indicating that domain adaptation effectively mitigates the distributional differences between the source and target domains, improving prediction accuracy. Simultaneously, models lacking target domain prior learning exhibit higher errors in Tasks 1 and 2, demonstrating that target domain prior learning enhances the model's adaptability to the specific degradation modes of the target device. When both domain adaptation and target domain prior learning are lacking, prediction performance further declines, especially for models that use no aids at all, which exhibit the highest error across all tasks. This verifies that relying solely on the basic prediction network cannot achieve stable and reliable prediction results under multi-condition and multi-failure-mode environments.

[0116] This application significantly improves the prediction accuracy of the remaining service life of turbofan engines under different operating conditions and failure modes by continuously learning the dynamic alignment and degradation laws of the feature distributions of the source and target domains. This method fully leverages the advantages of a multi-module collaborative mechanism, achieving efficient modeling and reliable prediction of engine degradation processes under complex operating conditions. It not only verifies the model's adaptability and robustness in multi-task and multi-condition environments but also provides an effective technical means for turbofan engine life management, maintenance decisions, and reliability assurance.

[0117] See Figure 10 , Figure 10 This is a structural diagram of a device for predicting the remaining useful life of equipment provided in an embodiment of this application. The device 1000 can implement the aforementioned method for predicting the remaining useful life of equipment, and its structure includes: Data acquisition module 1001 is used to acquire the first operational monitoring data of the source domain device and the second operational monitoring data of the target domain device; Data processing module 1002 is used to construct a first input time series based on the first operation monitoring data, and input the first input time series into the source domain prediction model framework. The source domain prediction model framework is equipped with a feature extraction model and a probability prediction model. Feature processing module 1003 is used to perform feature extraction and feature fusion processing on the first input time series through a feature extraction model to obtain comprehensive degradation features; Training module 1004 is used to train the probabilistic prediction model using comprehensive degradation features. Based on the trained probabilistic prediction model, the trained source domain prediction model framework and the corresponding posterior parameters are obtained. Modeling module 1005 is used to construct a target domain prediction model based on the second operation monitoring data and posterior parameters; The prediction module 1006 is used to predict the equipment to be tested through the target domain prediction model, and obtain the prediction result of the remaining service life of the equipment and the corresponding uncertainty information.

[0118] This device employs a modular design to form a complete logical closed loop from data acquisition to reliable prediction. Each module works collaboratively to accurately predict and quantify the remaining lifespan of the equipment. Specific functions are as follows: Data acquisition module 1001 acquires the first operational monitoring data from the source domain equipment and the second operational monitoring data from the target domain equipment, providing unified and standardized raw data input for subsequent full-process analysis; Data processing module 1002 serializes the received raw monitoring data to generate corresponding input time series, and simultaneously performs normalization preprocessing of the series, inputting the processed standardized series into the source domain prediction model framework; Feature extraction model 1003 performs deep feature mining on the preprocessed time series, extracting multi-level time-series features with contextual dependencies, and then integrates various deep features through a feature fusion strategy to form a comprehensive feature that accurately characterizes the equipment's degradation state. The training module 1004 is responsible for training and optimizing the probabilistic prediction model based on the comprehensive degradation characteristics, obtaining a trained probabilistic prediction model, and simultaneously generating corresponding variational approximate posterior parameters during model training; the modeling module 1005 relies on the second operational monitoring data of the target domain equipment, combined with the posterior parameters obtained from the above training, to complete the construction and adaptation optimization of the target domain prediction model; the prediction module 1006 inputs the monitoring data of the equipment to be tested into the target domain prediction model for inference prediction, outputs the point prediction results of the remaining service life of the equipment, and provides corresponding uncertainty quantification information; this device can effectively decompose and accurately quantify the cognitive uncertainty and random uncertainty in the prediction process, and based on this, conduct a comprehensive and systematic reliability assessment of the point prediction results output by the prediction module, ultimately providing a scientific and key quantitative basis for risk decision-making in equipment operation and maintenance.

[0119] It should be noted that the implementation logic of each functional module in this device is basically the same as the processing flow corresponding to the aforementioned method embodiments. For details, please refer to the relevant descriptions in the aforementioned method embodiments, which will not be repeated here.

[0120] This application also discloses an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable by the processor. When the computer program is executed by the processor, it implements the device remaining useful life prediction method as described above.

[0121] See Figure 11 , Figure 11 This is a hardware structure diagram of an electronic device provided in one embodiment of this application. The electronic device includes: The processor 1101 can be implemented using a general-purpose central processing unit, microprocessor, application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0122] The memory 1102 can be implemented as a read-only memory, static storage device, dynamic storage device, or random access memory. The memory can store the operating system and other application programs. When the method of this application is implemented by software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0123] Input / output interface 1103 is used to implement information input and output.

[0124] A bus is used to transmit information between the various components of a device.

[0125] This application also discloses a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the device remaining useful life prediction method described above.

[0126] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the remaining service life of equipment, characterized in that, include: Acquire the first operational monitoring data of the source domain devices and the second operational monitoring data of the target domain devices; A first input time series is constructed based on the first operational monitoring data, and the first input time series is input into the source domain prediction model framework, which includes a feature extraction model and a probability prediction model. The first input time series is subjected to feature extraction and feature fusion processing through the feature extraction model to obtain comprehensive degradation features; The probabilistic prediction model is trained using the comprehensive degradation features. Based on the trained probabilistic prediction model, the trained source domain prediction model framework and the corresponding posterior parameters are obtained. A target domain prediction model is constructed based on the second operational monitoring data and the posterior parameters; The target domain prediction model is used to predict the remaining service life of the device under test, and the corresponding uncertainty information is obtained.

2. The method for predicting the remaining service life of equipment according to claim 1, characterized in that, The feature extraction model includes an input encoding unit, a multi-layer encoding unit, and a feature fusion unit. The step of performing feature extraction and feature fusion processing on the first input time series using the feature extraction model to obtain comprehensive degradation features includes: The first input time series is linearly mapped and positionally encoded by the input encoding unit to obtain an initial feature representation. Attention feature extraction is performed on the initial feature representation through the multi-layer coding unit to obtain multi-layer degradation feature information; Average pooling is performed on the degenerate feature information of each layer to obtain the corresponding feature vector; The feature fusion unit performs weighted fusion of the feature vectors according to preset layer weights to obtain comprehensive degradation features.

3. The method for predicting the remaining service life of equipment according to claim 1, characterized in that, The probabilistic prediction model has a first fully connected layer and a second fully connected layer; The process involves training the probabilistic prediction model using the comprehensive degradation features, and based on the trained probabilistic prediction model, obtaining the trained source domain prediction model framework and corresponding posterior parameters, including: The comprehensive degradation characteristics are predicted by the first fully connected layer and the second fully connected layer to obtain the predicted value of the remaining service life of the device; Variational probability modeling is performed on the weights and biases of the first and second fully connected layers, and the variational approximate posterior distribution is obtained by reparameterizing the weights. The probabilistic prediction model is optimized and trained based on the joint loss of prediction error and KL divergence. Based on the trained probabilistic prediction model, a trained source domain prediction model framework is obtained. Extract the corresponding posterior parameters from the variational approximation posterior distribution.

4. The method for predicting the remaining service life of equipment according to claim 3, characterized in that, After obtaining the variational approximate posterior distribution, the method further includes: Based on the variational approximation posterior distribution, multiple weight reparameterizations are performed to obtain the cognitive uncertainty corresponding to the predicted value. Independently model the uncertainty caused by observation noise to obtain the corresponding stochastic uncertainty; By fusing the cognitive uncertainty and the random uncertainty, the overall uncertainty quantification result of the predicted value is obtained.

5. The method for predicting the remaining service life of equipment according to claim 1, characterized in that, The step of constructing a target domain prediction model based on the second operational monitoring data and the posterior parameters includes: The posterior parameters are initialized as prior parameters of the target domain model to obtain the initial target domain prediction model; The second input time series is constructed based on the second operational monitoring data; The initial target domain prediction model is used to extract features from the first input time series and the second input time series respectively to obtain source domain feature representation and target domain feature representation; Determine the adversarial domain loss and mean difference loss between the source domain feature representation and the target domain feature representation; Based on the adversarial domain loss and mean difference loss, the initial target domain prediction model is optimized by distribution alignment to obtain a target domain prediction model whose feature distribution difference meets the preset requirements.

6. The method for predicting the remaining service life of equipment according to claim 1, characterized in that, The source domain prediction model framework also includes a domain adaptive model, which comprises an adversarial domain discriminator and a distribution difference discriminator. The adversarial domain discriminator is used to calculate the domain category difference between the source domain feature representation and the target domain feature representation to obtain the adversarial domain loss. The distribution difference discriminator is used to calculate the overall distribution deviation between the source domain feature representation and the target domain feature representation to obtain the mean difference loss.

7. The method for predicting the remaining service life of equipment according to claim 2, characterized in that, Each coding unit has a normalization layer, a multi-head self-attention layer and a feedforward network layer connected in sequence. The output of the multi-head self-attention layer and the input of the feedforward network layer are connected by a first residual connection layer, and the output of the feedforward network layer is connected by a second residual connection layer.

8. The method for predicting the remaining service life of equipment according to claim 1, characterized in that, The step of constructing the first input time series based on the first operational monitoring data includes: The first operational monitoring data is normalized to obtain standardized source domain monitoring data; The standardized source domain monitoring data is divided into time series using a preset sliding window to obtain time series subsequences of equal length. The first input time series is obtained based on the equal-length time series subsequences.

9. An electronic device, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein the computer program, when executed by the processor, implements the device remaining useful life prediction method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, It contains a processor-executable computer program, which, when executed by the processor, is used to implement the device remaining useful life prediction method as described in any one of claims 1 to 8.

Images