A neural process-based electronic power component few-shot probabilistic prognosis method

Abstract

The embodiment of the application discloses a kind of electronic power components few sample probability prognosis methods based on neural process, it is related to electronic power components fault prediction technical field, including: obtaining historical degradation sequence, dividing context set and query set, extracting degradation feature vector;Time coding is applied to normalized cyclic index, after fusing degradation feature vector, time coding vector and real value, global degradation representation is obtained after being handled by feature projection layer and context encoder and aggregation;Global degradation representation is input into variational inference network, and reparameterization sampling obtains latent variable vector;Based on latent variable vector and global degradation representation, construct decoder input, generate life and health index prediction value by double-head decoder;Initial prediction model is obtained based on prediction loss and KL divergence joint optimization, update unfreezing parameter after being frozen and unfreezing by stratification, generate integrated prediction mean and standard deviation;According to life prediction value, calibration sample is divided into degradation stage, and confidence interval is determined in combination with standard deviation.

Description

Technical Field

[0001] This application relates to the field of electronic power component failure prediction technology, and relates to, but is not limited to, a few-sample probability prognostic method for electronic power components based on neural processes. Background Technology

[0002] Electronic power components, as core components of modern energy conversion systems, play a crucial role in many fields such as avionics, aerospace platforms, renewable energy, and industrial automation. Their performance directly determines the reliability of the entire system. However, long-term engineering practice has shown that electronic power components are continuously subjected to the coupling effects of multiple physical fields such as thermal cycling, electrical stress, mechanical vibration, and humidity during operation. Their internal key parameters drift continuously over time, eventually leading to component failure and even triggering cascading failures in downstream subsystems.

[0003] Therefore, fault prediction and health management for electronic power components, in order to accurately predict the remaining life (lifetime) and degradation trajectory of these components, has become a key supporting technology for ensuring the safe operation of modern engineering systems and reducing maintenance costs. However, in high-reliability scenarios such as avionics, the degradation observations available after a target component enters actual service are often very scarce. Maintenance personnel typically only have health monitoring data within about 10% of the early cycle range of the component, and then need to make reliable predictions about its complete degradation trajectory and the moment when the remaining life exceeds the limit. This early data-long-range extrapolation prediction paradigm poses three challenges to existing technologies: first, it is difficult to quickly adapt to individual differences with very few target samples; second, it is difficult to provide statistically reliable and coverage-guaranteed uncertainty quantification for the extrapolation interval; and third, it is difficult to maintain stable prediction accuracy across components and across degradation mechanisms.

[0004] Traditional methods for predicting the remaining lifetime of electronic power components mainly fall into two categories: physical model-based methods and data-driven methods. Physical model-based methods attempt to establish a mathematical description based on electrochemical reaction mechanisms, combining state estimation algorithms such as particle filtering and Kalman filtering for remaining lifetime prediction. These methods offer strong interpretability but rely on precise modeling of the component's internal parameters, have limited generalization ability to different operating conditions, and suffer from high computational complexity. In recent years, data-driven methods, represented by gated recurrent units, long short-term memory networks, and Transformers, have shown significant advantages in capturing long-range dependencies in degradation trends, achieving significant accuracy improvements over physical model benchmarks. However, these data-driven methods generally rely on large-scale historical training data, which is often insufficient for real-world engineering scenarios involving new models and batches of components. To address this issue, transfer learning and rapid adaptation methods have gained increasing attention, such as introducing model-independent meta-learning into remaining lifetime estimation, enabling the model to adapt with very few target samples. However, these methods often define the prediction task as a deterministic point estimation problem, meaning the model outputs a single predicted value rather than a probability distribution, which cannot directly support risk-based maintenance decisions in high-reliability scenarios such as avionics. To compensate for the lack of uncertainty quantification, probabilistic prediction methods such as Gaussian process regression and Monte Carlo Dropout have been gradually explored in the field of electronic power component reliability. Furthermore, conformal prediction, as an uncertainty quantification framework that does not require any parameterized assumptions about the data distribution, can provide a prediction interval with strictly guaranteed finite sample coverage for any basic prediction model. However, the above methods share a common structural limitation: they are designed as static prediction tools for a single target, making it difficult to simultaneously achieve rapid personalized adaptation, long-range trajectory extrapolation, and statistically reliable uncertainty quantification when there is limited early degradation data for the target component.

[0005] Therefore, there is an urgent need for a more comprehensive method for predicting the degradation trajectory of electronic power components, in order to solve the related problems existing in the current technology and achieve a synergistic improvement in long-range prediction accuracy and uncertainty quantification reliability under conditions with few samples. Summary of the Invention

[0006] This application provides a few-sample probability prognostic method for electronic power components based on neural processes.

[0007] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a few-sample probabilistic prognostic method for electronic power components based on neural processes. The method includes: acquiring historical degradation sequences of electronic power components; dividing the historical degradation sequences of each component into a context set and a query set according to a preset ratio; and extracting multidimensional degradation feature vectors of cyclic points in each sequence; applying time encoding to the normalized cyclic index; fusing the multidimensional degradation feature vector, the time encoding vector, and the true observations of the cyclic points to obtain an encoder input vector; processing the encoder input vector sequentially through a feature projection layer and a context encoder, followed by aggregation to obtain a global degradation representation; inputting the global degradation representation into a variational inference network to generate the mean vector and log-variance vector of the latent variable posterior distribution; and obtaining the latent variable vector through reparameterized sampling. The KL divergence between the approximate posterior distribution and the prior distribution followed by the latent variable vector is calculated. Based on the latent variable vector and the global degradation representation, a decoder input vector is constructed for each query point in the query set using the time-encoded vector and the normalized cyclic index. This is then processed by a dual-head decoder to generate remaining lifespan predictions and health indicator predictions. A total loss function is constructed based on the prediction loss and the KL divergence. The prediction model parameters are jointly optimized to obtain the initial prediction model after meta-training. The initial prediction model is subjected to hierarchical parameter freezing and unfreezing. The unfrozen parameters are updated using the context set of the target component to obtain the adapted model and generate the ensemble prediction mean and ensemble standard deviation. The calibration samples are divided into multiple degradation stages based on the remaining lifespan predictions. The confidence interval is determined based on the stage to which the query point belongs, combined with the ensemble standard deviation.

[0008] Optionally, the prediction model consists of a feature projection layer, a context encoder, a variational inference network, and a dual-head decoder; wherein, the context encoder is composed of... Multi-head self-attention network structure The preset positive integer is used; the variational inference network contains two independent linear transformation branches; the dual-head decoder contains a shared decoder backbone, a remaining lifetime prediction head, and a health indicator prediction head.

[0009] Optionally, the multidimensional degradation feature vector includes normalized health index values, first-order difference values, linear regression slope within the local window, standard deviation within the local window, interaction term between normalized health index and normalized cyclic index, and normalized cyclic index.

[0010] Optionally, the global degradation representation is input into the variational inference network to generate the mean vector and log-variance vector of the latent variable posterior distribution. The latent variable vector is obtained through reparameterization sampling. The KL divergence between the approximate posterior distribution and the prior distribution followed by the latent variable vector is calculated. This includes: inputting the global degradation representation into two independent linear transformation branches of the variational inference network, wherein the first branch is used to calculate the mean vector of the latent variable posterior distribution, and the second branch is used to calculate the log-variance vector of the latent variable posterior distribution; calculating the standard deviation vector based on the log-variance vector; determining the approximate posterior distribution of the component based on the mean vector and the standard deviation vector; sampling random noise vectors from the standard normal distribution according to the mean vector and the standard deviation vector, and calculating the latent variable vector through reparameterization transformation; using the standard normal distribution as the prior distribution followed by the latent variable vector, calculating the KL divergence between the approximate posterior distribution and the prior distribution.

[0011] Optionally, the remaining lifespan prediction value and health indicator prediction value are generated through dual-head decoder processing, including: inputting the decoder input vector into a shared decoder backbone to extract shared deep degradation features, wherein the shared decoder backbone is composed of a multilayer perceptron with three layers of hidden dimensions decreasing layer by layer; passing the shared deep degradation features to the remaining lifespan prediction head and the health indicator prediction head respectively, wherein in the remaining lifespan prediction head, a normalized remaining lifespan prediction value is output through mapping using a Sigmoid activation function, and in the health indicator prediction head, a normalized health indicator prediction value is output through a linear activation function; wherein the remaining lifespan prediction head and the health indicator prediction head share the shared decoder backbone.

[0012] Optionally, the predicted loss includes time-weighted remaining lifetime task loss and health indicator task loss. The time-weighted remaining lifetime task loss is calculated using the following formula: ; In the formula, This represents the time-weighted remaining lifetime task loss. This represents the total number of query points in the query set. Indicates the first in the query set Index for each query point; Represents normalized circular index The default monotonically increasing function; Indicates the first The first component A normalized cyclic index for each query point; Indicates the first The first component Loss function for predicting remaining lifetime at each query point; Indicates the first The first component Normalized remaining lifetime prediction for each query point; Indicates the first The first component The normalized true remaining lifespan value for each query point; the calculation formula for time-weighted health indicator task loss is expressed by the following formula: ; In the formula, Indicates time-weighted health indicator task loss Indicates the first The first component The health indicator prediction loss function for each query point; Indicates the first The first component Normalized health indicator predictions for each query point; Indicates the first The first component The normalized true values ​​of health indicators for each query point; the KL divergence is calculated using the following formula: ; In the formula, Represents the approximate posterior distribution With prior distribution The KL divergence between the variables is used to constrain the distribution shape of the latent variable space to prevent overfitting. Indicates the first The latent variable vector of each component; The approximate posterior distribution conditioned on contextual information is defined as a Gaussian distribution where each dimension is independent. ; The prior distribution of the latent variable vector is represented by the standard normal distribution. ; Indicates distribution The following expectations; The logarithmic density function represents the approximate posterior distribution; Let logarithm be the logarithmic density function representing the prior distribution; the total loss function is calculated using the following formula: ; In the formula, Represents the total loss function; This represents the time-weighted remaining lifetime task loss. Indicates time-weighted health indicators and task losses; The preset weighting coefficients represent the KL divergence; This represents the KL divergence between the approximate posterior distribution and the prior distribution.

[0013] Optionally, the initial prediction model undergoes hierarchical parameter freezing and unfreezing. The unfrozen parameters are updated using the context set of the target component to obtain the adapted model and generate the ensemble prediction mean and ensemble standard deviation. This includes: freezing all parameters of the feature projection layer, freezing all parameters of all layers in the context encoder except the last layer; unfreezing all parameters of the last layer of the context encoder, unfreezing all parameters of the variational inference network, and unfreezing all parameters of the dual-head decoder; and setting the number of ensemble iterations. and Each of the three distinct random seeds is used to initialize the random number generator for the adaptation process; the process is then executed independently using the context set of the target component. In each parameter optimization, only the unfrozen parameters are updated, while the frozen parameters remain unchanged, resulting in... An adapted model, Each adapted model is distributed in a different local optimum region in the parameter space due to different random seeds; among them... ;according to Each adapted model independently predicts the query set for the target component, resulting in... Group prediction results; calculated by arithmetic mean The final prediction result for the target component is obtained by ensemble prediction of the group prediction results; calculation The integrated standard deviation of the prediction results is used to obtain an estimate of the cognitive uncertainty of the target component prediction.

[0014] Optionally, the calibration samples are divided into multiple degradation stages based on the predicted remaining lifetime. The confidence interval is determined based on the stage to which the query point belongs, using the ensemble standard deviation and the combined ensemble standard deviation. This includes: after the meta-training and adaptation phases, the adapted model is used to predict the last 50% of all components in the training set, and the resulting ensemble prediction mean and the true values ​​are used to construct the calibration set; the corresponding degradation stage is divided based on the predicted remaining lifetime of each calibration sample in the calibration set, where a predicted remaining lifetime greater than 0.6 is considered an early stage, a predicted remaining lifetime greater than 0.25 and less than or equal to 0.6 is considered an intermediate stage, and a predicted remaining lifetime less than or equal to 0.25 is considered a late stage; and the variance-normalized inconsistency score is calculated for all calibration samples within each degradation stage, expressed by the following formula: ; In the formula, Indicates the first The non-consistency score of each calibration sample; Indicates the first The true value of each calibration sample; Indicates the first The ensemble predicted mean of 1 calibration sample; Indicates the first The integrated standard deviation of each calibration sample; Represents a preset positive decimal constant; set a significance level, and calculate the shape-preserving quantiles of the inconsistency score set within each degradation stage. The formula for calculating the shape-preserving quantiles is expressed as follows: ; In the formula, This represents the pre-defined significance level, with a corresponding confidence level of . ; Indicates the first Each stage of degradation at the confidence level Conformal quantiles below; Indicates the first The set of inconsistent scores for all calibration samples within a degradation stage; Indicates the first The number of calibration samples within each degradation stage; Represents the quantile function; Indicates that for the first The conformal quantile level is adjusted using a finite number of samples for each degradation stage; based on the predicted remaining lifetime of the query point, the degradation stage to which the query point belongs is determined, the corresponding conformal quantile is extracted, and a normalized prediction interval is constructed, which is expressed by the following formula: ; In the formula, Indicates query point The normalized prediction interval; Indicates query point The ensemble prediction mean; Indicates the degradation stage to which the query point belongs. The corresponding conformal quantiles; Indicates query point The integrated standard deviation is calculated. Based on the normalized prediction interval, the remaining life prediction interval and the health indicator prediction interval are denormalized to obtain the remaining life confidence interval and the health indicator confidence interval, respectively. The remaining life prediction interval denormalization is achieved by multiplying the upper and lower bounds of the normalized prediction interval by the total number of cycles of the target component to obtain the remaining life confidence interval at the actual cycle number scale. The health indicator prediction interval denormalization is achieved by multiplying the upper and lower bounds of the normalized prediction interval by the standard deviation of the health indicators in the training set and then adding the mean of the health indicators in the training set to obtain the health indicator confidence interval at the original physical dimensions.

[0015] The beneficial effects of the technical solutions provided in this application include at least the following: This application provides a few-shot probabilistic prognostic method for electronic power components based on neural processes. By dividing the historical degradation sequences of each component into a context set and a query set according to a preset ratio, the model can fully learn the shared degradation patterns across components during the meta-training phase. Multi-dimensional degradation feature vectors are extracted from each cycle point to construct a multi-level degradation representation system from absolute state, first-order dynamics, second-order trend, local uncertainty, stage interaction to global position. By applying time encoding to the normalized cycle index, components with different total cycle numbers have the same time encoding value at the same degradation progress, thus achieving a unified representation across components with different lifespans. By fusing the multi-dimensional degradation feature vector, time encoding vector, and the true observation values ​​of cycle points into the encoder input vector, the encoder can infer potential degradation patterns based on early true degradation values. After processing the encoder input vector sequentially through a feature projection layer and a context encoder, an aggregation operation is performed to capture the non-local dependencies between context observation points. This aggregation operation enables the model to handle different numbers of context observation points and maintains permutation invariance, significantly improving generalization ability under few-shot conditions. By inputting the global degradation representation into the variational inference network, the mean vector and log-variance vector of the latent variable posterior distribution are generated. Reparameterized sampling is used to obtain the latent variable vector, and the KL divergence is calculated. This allows the gradient to propagate back to the variational inference network from subsequent steps, and the KL divergence constrains the spatial distribution of the latent variables to prevent overfitting. Based on the latent variable vector and the global degradation representation, a decoder input vector is constructed using the temporal encoding vector and normalized cyclic index of each query point in the query set. This enables the decoder to perceive the degradation mode and current degradation progress of the component at each prediction time. A dual-head decoder simultaneously generates remaining lifespan predictions and health indicator predictions. The complementarity of these two tasks in the degradation structure is used to achieve implicit multi-task regularization, significantly improving prediction accuracy in low-sample scenarios. Furthermore, a time-weighted prediction loss is introduced to guide the model's attention to key regions in the later stages of degradation, thereby improving prediction accuracy near abrupt change points. By constructing a total loss function based on prediction loss and KL divergence to jointly optimize the prediction model parameters, an initial prediction model after meta-training is obtained, enabling the model to have good initial performance with very few target samples. By performing hierarchical parameter freezing and unfreezing on the initial prediction model, the feature projection layer and all layers of the context encoder except the last layer are frozen to retain the common degradation features shared across modules. The last layer of the context encoder, the variational inference network, and the dual-head decoder are unfrozen to adapt to the individual degradation patterns of the target components. Then, the unfrozen parameters are updated through the context set of the target components to obtain the adapted model and generate the ensemble prediction mean and ensemble standard deviation. This enables the model to quickly adapt to individual differences with very few target samples and has cross-scene transfer capabilities.By dividing the calibration samples into multiple degradation stages based on the remaining lifetime prediction, and combining the ensemble standard deviation to determine the confidence interval according to the stage to which the query point belongs, the confidence interval can adapt to the heterogeneous characteristics of the error distribution under different degradation stages. Without relying on any distribution assumptions, a confidence interval with finite sample coverage is obtained in a marginal sense, and the width of the confidence interval adapts to the ensemble standard deviation, improving the compactness of the interval while maintaining coverage. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating a few-sample probability prognostic method for electronic power components based on neural processes, provided in an embodiment of this application; Figure 2 This is a schematic diagram of the overall architecture of the prediction model provided in the embodiments of this application; Figure 3 A flowchart of few-sample adaptation based on freeze-thaw strategy and multiple sub-integrations provided for embodiments of this application; Figure 4 A schematic diagram of lithium-ion battery capacity decay curves based on NASA datasets provided for embodiments of this application; Figure 5 A schematic diagram of lithium-ion battery capacity decay curves based on the CALCE dataset provided for embodiments of this application; Figure 6 The adapted model provided in this application provides a fitted image of the prediction results on the NASA I dataset; Figure 7 The adapted model provided in this application fits the prediction results of the CALCE I dataset to the image. Figure 8 The adapted model provided in this application provides a fitted image of the prediction results on the CALCE II dataset; Figure 9 This is a schematic diagram of the accelerated degradation test platform provided in the embodiments of this application; Figure 10 This is a schematic diagram of the accelerated degradation test circuit provided in the embodiments of this application. Figure 11 This is a schematic diagram of the prediction results of Datasets I provided in an embodiment of this application; Figure 12This is a schematic diagram of the prediction results of Datasets II provided in the embodiments of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0019] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0020] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0021] The neural process combines the nonparametric probabilistic modeling capabilities of the Gaussian process (GP) with the representation learning capabilities of deep neural networks, and achieves the learning of function distribution through an encoder-decoder architecture: the encoder aggregates the context observation set into a global representation, and the decoder infers the predicted distribution of the query point based on this global representation.

[0022] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0023] In view of the current problems in the research on few-sample probability prognostic methods for electronic power components in the field of electronic power component failure prediction technology, this application provides a few-sample probability prognostic method for electronic power components based on neural processes.

[0024] The technical solution of this application is described below, starting with the method embodiments.

[0025] Please refer to Figure 1 It illustrates a flowchart of a few-sample probability prognostic method for electronic power components based on neural processes, as provided in an embodiment of this application. Figure 1 As shown, the method includes at least the following steps S110 to S160.

[0026] Step S110: Obtain the historical degradation sequence of the electronic power components, divide the historical degradation sequence of each component into a context set and a query set according to a preset ratio, and extract the multidimensional degradation feature vector of the cycle point in each sequence.

[0027] In this embodiment of the application, the historical degradation sequence of the electronic power component during its historical operation is obtained, the first... The historical degradation sequence of each component is as follows: ,in For the first The first component The index of a cycle, These correspond to the original health indicators collected in the corresponding cycle. This represents the total number of cycles from the initial operation of the component until it reaches the preset failure threshold. The original health indicators are normalized to eliminate differences in units and initial values ​​between different components, resulting in normalized health indicator values. Normalization is then applied to the cycle index, setting... To eliminate the differences in the lifespan of different components, a normalized remaining lifetime label is constructed based on this, so that the normalized remaining lifetime label of each component monotonically decreases from 1.0 at the initial time to 0 at the time of failure.

[0028] In this embodiment, the preset ratio is set to 1:9, meaning that the proportion of early observation data is 1:9. The ordered sequence of each component Each point is divided into context sets. , and the remaining Each point is divided into a query set. ,in Input features for the query point. For the corresponding supervision label, The context set is used to simulate the limited early observation data available in actual engineering projects, while the query set is used to supervise the extrapolation training and prediction evaluation of subsequent long-term degradation trajectories.

[0029] To effectively capture the degradation characteristics of electronic power components and achieve cross-module generalization, a multidimensional degradation feature vector is extracted for each cycle point in the historical degradation sequence. This vector includes normalized health index values, first-order difference values, linear regression slope within the local window, standard deviation within the local window, interaction terms between normalized health index and normalized cycle index, and normalized cycle index. Among them, the normalized health index value is used to directly reflect the absolute health level at the current moment; the first difference value is the difference between the health index at the current moment and the previous moment, used to reflect the instantaneous degradation rate of the component; the linear regression slope within the local window is the slope of the linear trend of the health index changing with the number of cycles within a preset local time window, used to reflect the direction of the short-term degradation trend of the component; the standard deviation within the local window is the degree of dispersion of the health index around the mean within the local window, used to reflect the local fluctuation intensity and noise level of the component degradation process; the interaction term between the normalized health index and the normalized cycle index is the product of the normalized health index value and the normalized cycle index, used to display the encoded degradation stage information, enabling the model to distinguish different degradation modes of early, middle and late stages from the input level; the normalized cycle index is used to retain the absolute progress position of the component, ensuring that the model perceives which relative position it is in the life cycle.

[0030] One point to note is that when the electronic power component is a DC-DC power module, after constructing the historical degradation sequence and before extracting the multidimensional degradation feature vector, data cleaning is also required. Specifically, voltage anomalies with an output voltage less than 0V or greater than 10V and current anomalies with an output current less than 0A are removed. The adjacent average filtering method is then used to smooth the original health index sequence, and the smoothed sequence after filtering is used as the effective value of the original health index.

[0031] Step S120: Apply time encoding to the normalized cyclic index, fuse the multidimensional degenerate feature vector, the time encoding vector, and the real observations of the cyclic points to obtain the encoder input vector; after processing the encoder input vector sequentially through the feature projection layer and the context encoder, perform an aggregation operation to obtain the global degenerate representation.

[0032] In this embodiment of the application, sine-cosine time coding is applied to each loop point in the context set according to the normalized loop index, resulting in the following time coding vector: ; In the formula, Indicates the first The first component Time-encoded vectors for each loop point; The preset frequency is used to control the oscillation period of the sine and cosine functions. The high-frequency component is used to capture local abrupt changes and instantaneous fluctuations during the degradation process, while the low-frequency component is used to capture long-term degradation trends and global patterns. In this embodiment, the preset frequency is set at logarithmic linear intervals. , ; Indicates a normalized circular index. ,in Indicates the first The first component The index of a cycle, Indicates the first The total number of cycles for each component; This represents the total dimension of the time encoding; Represents the frequency index, with a range of values. to Because the time-encoding vector uses the normalized cycle index as its independent variable, components with different total cycle numbers have the same time-encoding value at the same degradation stage, thus possessing lifetime-scale normalization consistency and scaling invariance. The multidimensional degradation feature vector, time-encoding vector, normalized remaining lifetime true value, and normalized health index true value corresponding to each cycle point in the context set are sequentially concatenated to generate the encoder input vector, enabling the context encoder to infer potential degradation patterns based on early observed true degradation values.

[0033] In this embodiment, the encoder input vector is sequentially passed through two feature projection layers and mapped to a preset model hidden dimension to obtain a projected context hidden vector sequence. The projected context hidden vector sequence is then input into a... In a context encoder composed of multi-head self-attention networks, the temporal dependencies between any two context points in each layer are globally modeled using a multi-head self-attention mechanism. This captures the nonlocal impact of early local mutations on subsequent degradation trends. After layer encoding, the context representation vector corresponding to each context point is obtained. ,in The value is a preset positive integer, for example, L is an integer from 3 to 6, preferably 4. This is the context representation vector corresponding to each context point. ,in For the number of loop points in the context set of the i-th component, perform mean aggregation to obtain the global degradation representation. This mean aggregation operation ensures that the global degenerate representation is permutation-invariant to the order of observations in the context set, enabling the model to handle different numbers of context observations and avoiding prediction results that depend on the input order of context points.

[0034] Step S130: Input the global degenerate representation into the variational inference network to generate the mean vector and log-variance vector of the latent variable posterior distribution. Obtain the latent variable vector through reparameterized sampling and calculate the KL divergence between the approximate posterior distribution and the prior distribution followed by the latent variable vector.

[0035] In this embodiment, the global degradation representation is input into two independent linear transformation branches of the variational inference network. The first branch calculates the mean vector of the posterior distribution of the latent variables through linear transformation. The second branch calculates the log-variance vector of the posterior distribution of the latent variables through linear transformation. The weight parameters of the two linear transformation branches , and bias parameters , No sharing. Furthermore, the standard deviation vector is calculated based on the log-variance vector. Based on the mean vector and the standard deviation vector, the approximate posterior distribution of the component is defined as a Gaussian distribution independent of each dimension. This approximate posterior distribution is used to represent the probabilistic estimate of the potential degradation mode of the component after the contextual information of the component has been observed.

[0036] Furthermore, random noise vectors are sampled from the standard normal distribution based on the mean vector and standard deviation vector. , The identity matrix is ​​used to calculate the latent variable vector through reparameterization transformation. ,in Representing element-wise multiplication, the reparameterization transformation converts the random sampling process into a deterministic differentiable computation of the mean and standard deviation vectors, enabling the gradient to propagate back through the subsequent decoding steps to the parameters of the variational inference network, resulting in a latent variable vector that can be trained end-to-end. Using the standard normal distribution as the prior distribution followed by the latent variable vectors, the KL divergence between the approximate posterior distribution and the prior distribution is calculated. This KL divergence serves as a regularization constraint term, used to restrict the distribution shape of the latent variable space to prevent overfitting.

[0037] Step S140: Based on the latent variable vector and global degradation representation, construct the decoder input vector for the time encoding vector and normalized cyclic index of each query point in the query set, and generate the remaining life expectancy prediction value and health indicator prediction value through dual-head decoder processing.

[0038] In this embodiment, based on latent variable vectors and global degradation representations, a decoder input vector is constructed for each query point in the query set using its temporal encoding vector and normalized cyclic index. The introduction of latent variable vectors enables the decoder to perceive the personalized degradation pattern of the component at each prediction time. The decoder input vector is fed into a dual-head decoder, which includes a shared decoder backbone, a remaining lifespan prediction head, and a health indicator prediction head. The shared decoder backbone consists of a multi-layer perceptron with three progressively decreasing hidden dimensions, and a Dropout layer with a preset Dropout rate is introduced for regularization. After forward propagation, shared deep degradation features are extracted. The preset Dropout rate is set to 0.05 to 0.10, preferably 0.08.

[0039] Furthermore, the shared deep degradation features are passed to the remaining lifespan prediction head and the health indicator prediction head, respectively. In the remaining lifespan prediction head, a normalized remaining lifespan prediction value is output after mapping through the Sigmoid activation function. , The sigmoid activation function constrains the predicted values ​​to the same dimensions and range as the normalized remaining life label; in the health indicator prediction header, the normalized health indicator predicted values ​​are directly output through a linear activation function. This approach preserves the ability of health indicators to continuously change during the degradation process. The remaining lifespan prediction head outputs normalized remaining lifespan predictions, making it suitable for capturing the global trend of degradation. The output of the health indicator prediction head has direct physical interpretability. By sharing the same latent variable vector and the same decoder backbone, the remaining lifespan prediction head and the health indicator prediction head achieve implicit multi-task regularization based on the complementarity of the degradation structure, thereby improving the model's generalization ability in scenarios with few samples.

[0040] Step S150: Construct a total loss function based on the prediction loss and KL divergence, jointly optimize the prediction model parameters, and obtain the initial prediction model after meta-training; perform hierarchical parameter freezing and unfreezing on the initial prediction model, update the unfrozen parameters through the context set of the target component, obtain the adapted model, and generate the ensemble prediction mean and ensemble standard deviation.

[0041] In this embodiment, the total loss function is constructed based on the prediction loss and KL divergence. The prediction loss includes a time-weighted remaining lifetime task loss and a health indicator task loss. For each query point, the remaining lifetime prediction loss and health indicator prediction loss are calculated based on the normalized remaining lifetime prediction value and its corresponding normalized remaining lifetime actual value, as well as the normalized health indicator prediction value and its corresponding normalized health indicator actual value. A time weight dependent on its normalized circular index is applied to each query point. This time weight is a preset monotonically increasing function with respect to the normalized circular index, ensuring that prediction errors in the later stages of degradation receive a higher contribution weight to the total loss. The formula for calculating the time-weighted remaining lifetime task loss is as follows: ; In the formula, This represents the time-weighted remaining lifetime task loss. This represents the total number of query points in the query set. Indicates the first in the query set Index for each query point; Represents normalized circular index The preset monotonically increasing function is taken in this embodiment. This is used to apply a higher contribution weight to the prediction error in the later stages of degradation; in other embodiments, it can also be used... Or in the form of piecewise linear functions, Indicates the first The first component A normalized cyclic index for each query point; Indicates the first The first component The remaining lifetime prediction loss function for each query point Indicates the first The first component Normalized remaining lifetime prediction for each query point Indicates the first The first component The normalized remaining lifetime value for each query point. The formula for calculating the time-weighted health indicator task loss is: ; In the formula, Indicates time-weighted health indicators and task losses; Indicates the first The first component The health indicator prediction loss function for each query point is the mean squared error in this embodiment. Indicates the first The first component Normalized health indicator predictions for each query point; Indicates the first The first component The normalized true values ​​of health indicators for each query point. The formula for calculating KL divergence is: ; In the formula, This represents the KL divergence between the approximate posterior distribution and the prior distribution. The KL divergence serves as a regularization constraint term, used to restrict the distribution shape of the latent variable space to prevent overfitting. Indicates the first The latent variable vector of each component, The approximate posterior distribution, conditional on contextual information, is defined as a Gaussian distribution independent of each dimension. The prior distribution of the latent variable vector is represented by the standard normal distribution. ; Indicates distribution The following expectations; The logarithmic density function represents the approximate posterior distribution; This represents the logarithmic density function of the prior distribution.

[0042] Furthermore, by weighting the divergence with preset weighting coefficients, and adding the time-weighted remaining lifetime task loss, the time-weighted health indicator task loss, and the weighted KL divergence, the total loss function can be obtained. The calculation formula is as follows: ; In the formula, Represents the total loss function; The preset weighting coefficients represent the KL divergence.

[0043] Please refer to Figure 2 It shows a schematic diagram of the overall architecture of the prediction model provided in the embodiments of this application, such as Figure 2As shown, the prediction model consists of four core modules: a feature projection layer, a context encoder, a variational inference network, and a dual-head decoder. The encoder input vector is mapped to a preset model hidden dimension through the feature projection layer, resulting in a projected sequence of context latent vectors. This sequence is then input into the context encoder, where a multi-head self-attention mechanism is used to globally model the temporal dependencies between any two context points in each layer, capturing the nonlocal impact of early local mutations on subsequent degradation trends. This yields a context representation vector for each context point. Mean aggregation is then performed on the context representation vectors for each context point to obtain a global degradation representation. This global degradation representation is then input into two independent linear transformations of the variational inference network. The first branch calculates the mean vector of the posterior distribution of the latent variables through linear transformation, and the second branch calculates the log-variance vector of the posterior distribution of the latent variables through linear transformation. The latent variable vector is further calculated through reparameterization transformation. The dimension of the latent variables is 32 to 64, preferably 48. Based on the latent variable vector and global degradation representation, the decoder input vector is constructed for the time encoding vector and normalized cyclic index of each query point in the query set. In the dual-head decoder, the shared deep degradation features are obtained by sharing the decoder backbone. The remaining lifespan prediction head and the health index prediction head generate the remaining lifespan prediction value and the health index prediction value, respectively.

[0044] In this embodiment, the total loss function is used as the optimization objective. The shared decoder backbone, remaining lifetime prediction head, health index prediction head, weight parameters and bias parameters backpropagated to the variational inference network after reparameterization transformation, and parameters backpropagated to the multi-head self-attention network after context encoder are updated synchronously through gradient backpropagation. KL divergence is used as a regularization constraint term in gradient backpropagation to complete end-to-end joint optimization training and output the initial prediction model after meta-training.

[0045] After meta-training, the model has learned shared degradation priors across components. However, in real-world industrial scenarios, target components often exhibit individual differences. Directly using the initial prediction model may not adequately reflect the specific degradation patterns of the target components. Therefore, a few-shot personalized adaptation is performed on the first 10% of early observation data (i.e., the context set) of the target components. Hierarchical parameter freezing and unfreezing operations are performed on the initial prediction model. During the adaptation phase, the low-level parameters used to extract general degradation patterns are frozen, and only the high-level parameters related to the individualization of the target module are updated. Specifically, all parameters of the feature projection layer and all parameters of all layers except the last layer in the context encoder are frozen, ensuring that the frozen layers retain the shared general degradation features learned during meta-training during subsequent adaptation. All parameters of the last layer in the context encoder, all parameters of the two independent linear transformation branches in the variational inference network, and all parameters of the dual-head decoder are unfrozen, allowing the unfrozen layers to update their parameters to adapt to the individual degradation patterns of the target components during subsequent adaptation. The reason for this hierarchical freezing strategy is that the lower-level encoder mainly learns the degradation features shared across modules, which should be preserved; the higher-level encoder and variational inference network are responsible for mapping general features to the personalized degradation distribution of the target module, and should be allowed to be updated; the dual-head decoder directly determines the remaining lifetime and health indicator prediction results of the target module, so it needs to be fully adapted.

[0046] Please refer to Figure 3 It illustrates a few-sample adaptation flowchart based on a freeze-thaw strategy and multiple sub-integrations provided in an embodiment of this application, such as... Figure 3 As shown, in the selective parameter fine-tuning stage, all parameters of the feature projection layer are frozen, all parameters of all layers in the context encoder except the last layer are frozen, all parameters of the last layer of the context encoder are unfrozen, all parameters of the two independent linear transformation branches in the variational inference network are unfrozen, and all parameters of the dual-head decoder are unfrozen.

[0047] Furthermore, in the ensemble reasoning phase, to reduce the randomness of single adaptations, an ensemble strategy is adopted, and the number of ensemble iterations is set. and Each of the following is a unique random seed. Each random number generator for the adaptation process is initialized separately, and each is executed independently using the context set of the target component. In the sub-parameter optimization process, each optimization only updates the unfrozen parameters while keeping the frozen parameters unchanged, resulting in... An adapted model, Each adapted model is distributed in a different local optimum region in the parameter space due to different random seeds. According to... Each adapted model independently predicts the query set for the target component, resulting in... Group prediction results, for The prediction results are averaged arithmetically to calculate the ensemble mean, yielding the final prediction result for the target component. The ensemble standard deviation of the prediction results is calculated to obtain an estimate of the cognitive uncertainty in the prediction of the target component.

[0048] Step S160: Divide the calibration samples into multiple degradation stages based on the remaining lifetime prediction values, and determine the confidence interval based on the stage to which the query point belongs by combining the integrated standard deviation.

[0049] In this embodiment, after the meta-training and adaptation phases, the adapted model is used to predict the last 50% of loops for all components in the training set. The resulting ensemble prediction mean and the true values ​​are used to construct a calibration set. The degradation stage is determined based on the remaining lifetime prediction value of each calibration sample in the calibration set. Specifically, an early stage is defined as a remaining lifetime prediction value greater than 0.6, a mid-stage as a remaining lifetime prediction value greater than 0.25 and less than or equal to 0.6, and a late stage as a remaining lifetime prediction value less than or equal to 0.25. The thresholds of 0.25 and 0.6 are empirical values ​​determined based on the statistical characteristics of the error distribution of the calibration samples in the training set. At these values, the error distribution within each stage is relatively homogeneous, thereby improving the accuracy of conformal calibration. In an alternative embodiment, the above thresholds can also be adaptively determined through cluster analysis of the error distribution of the calibration samples in the training set (such as K-means or Gaussian mixture model), without the need for manual presetting. Based on all calibration samples within each degradation stage, a variance-normalized inconsistency score is calculated using the following formula: ; In the formula, Indicates the first The inconsistency score of each calibration sample; Indicates the first The true value of each calibration sample; Indicates the first The ensemble predicted mean of 1 calibration sample; Indicates the first The integrated standard deviation of each calibration sample; This represents a preset positive decimal constant used to avoid a denominator of zero; when the actual value For the remaining lifetime, the inconsistency score corresponds to the remaining lifetime prediction task, when the true value The non-consistent scores for health indicators correspond to health indicator prediction tasks.

[0050] Furthermore, a significance level is set. The corresponding confidence level is The conformal quantiles of the non-conformal fraction set within each degradation stage are calculated using the following formula: ; In the formula, This represents the preset significance level, with a value range of [value range missing]. ; Indicates the first The set of inconsistencies in the scores of all calibration samples within a degradation stage, and each element in the set ; Indicates the first The number of calibration samples within each degradation stage is a positive integer greater than or equal to 1; Represents the quantile function; Indicates that for the first The conformal quantile level is adjusted by a finite number of samples for each degradation stage, when... A value greater than 1 indicates that the quantile level exceeds 100%, at which point the conformal quantile is used. Take the maximum value of the set of inconsistent scores, i.e. This ensures that the prediction interval maintains coverage even with a very small number of calibration samples. Furthermore, while the formula is theoretically applicable to any number of calibration samples (at least one), when the calibration sample size is extremely small, the obtained conformal quantiles tend to be conservative, increasing the randomness of the prediction interval. This limitation needs to be considered when analyzing the prediction results, for example, when the number of calibration samples is small. Significance level At this time , Take the maximum value from the set of inconsistent scores; Indicates the first Each stage of degradation at the confidence level The form-preserving quantiles are calculated independently in the early, middle, and late stages, allowing the prediction interval to adapt to the heterogeneity of error distributions at different degradation stages. The formula for this finite-sample corrected form-preserving quantile level is derived from the theoretical framework of form-preserving prediction, as detailed in Vovk V, Gammerman A, Shafer G. Algorithmic Learning in a Random World. Springer, 2005. This correction ensures finite-sample coverage of the prediction interval under finite calibration sample conditions.

[0051] Furthermore, based on the predicted remaining lifetime of the query point, the degradation stage to which the query point belongs is determined, the conformal quantile corresponding to that stage is extracted, and a normalized prediction interval is constructed. The prediction interval is expressed by the following formula: ; In the formula, Indicates query point The normalized prediction interval; Indicates query point The ensemble prediction mean; Indicates the degradation stage to which the query point belongs. The corresponding conformal quantiles; Indicates query point The integrated standard deviation; This represents a preset positive decimal constant, and the normalized prediction interval width varies with the ensemble standard deviation. The positive correlation changes mean that when the model's predictions for a certain query point have large discrepancies, the interval automatically widens, while when the prediction consistency is high, the interval remains tight. The phased calibration mechanism provided in this application embodiment can make the prediction interval more consistent with the phased error characteristics of the degradation process, thereby improving the tightness of the interval while maintaining coverage.

[0052] Finally, based on the normalized prediction interval, in-normalization of the remaining lifetime prediction interval and the health indicator prediction interval are performed respectively. Specifically, for the remaining lifetime prediction interval, the upper and lower bounds of the normalized remaining lifetime prediction interval are multiplied by the total number of cycles of the target component to obtain the remaining lifetime confidence interval at the actual cycle number scale; for the health indicator prediction interval, the upper and lower bounds of the normalized health indicator prediction interval are multiplied by the standard deviation of the health indicators in the training set and then added to the mean of the health indicators in the training set to obtain the health indicator confidence interval at the original physical dimensions. The final output is the predicted value of the target component at each query point on the query set, as well as the remaining lifetime confidence interval and health indicator confidence interval with finite sample coverage, thus providing a reliable basis for risk perception and maintenance decisions for electronic power components.

[0053] Please refer to Figure 4 It shows a schematic diagram of the capacity decay curve of a lithium-ion battery based on NASA dataset provided in an embodiment of this application, such as... Figure 4 As shown, this embodiment uses a publicly available NASA dataset to systematically validate the proposed neural process-based few-sample probability prognostic method for electronic power components. The horizontal axis represents the cyclic index, and the vertical axis represents the battery capacity. It displays the capacity decay curves of four battery groups (B0005, B0006, B0007, and B0018) from the NASA dataset. Please refer to... Figure 5 It shows a schematic diagram of the capacity decay curve of a lithium-ion battery based on the CALCE dataset provided in the embodiments of this application, such as Figure 5 As shown, this embodiment selects the University of Maryland CALCE public dataset to systematically verify the proposed neural process-based few-sample probability prognostic method for electronic power components. The horizontal axis in the figure represents the cyclic index, and the vertical axis represents the battery capacity. It shows the capacity decay curves of four battery groups (CS2_35, CS2_36, CS2_37, CS2_38) in the CALCE dataset.

[0054] This embodiment extracts four samples (B0005, B0006, B0007, and B0018) from the NASA dataset, with initial capacities of 1.86 Ah, 2.04 Ah, 1.89 Ah, and 1.86 Ah, respectively, and a rated capacity of 2.4 Ah for each. These batteries are charged to 4.2V using a constant current of 2A, then switched to a constant voltage mode until the current drops to 20mA, with a constant discharge current of 2A. The battery is considered to have reached its failure threshold when its usable capacity decays to 70% of its rated capacity. For the CALCE dataset, four batteries (CS2_35, CS2_36, CS2_37, and CS2_38) with a rated capacity of 1.1 Ah are selected. Their charging strategy is a constant current of 0.5A to 4.2V, followed by a constant voltage mode with a cutoff current of 0.05A; the discharge current is a constant 1A with a cutoff voltage of 2.7V, and the lifespan endpoint is defined as a capacity drop to 0.77 Ah. The NASA dataset allows direct extraction of discharge capacity values ​​from the data structure, while the CALCE dataset requires filtering out discharge stage data with Step_Index=7 and calculating the discharge capacity using the Coulomb integral method. To enhance the robustness of the model, Z-score normalization is applied to the data. The battery health state is defined as the normalized capacity; when the capacity decays to 70% of the rated capacity, the battery is considered to have reached the end of its lifespan. Based on this, the health state is linearly mapped to the [0,1] interval to obtain the normalized remaining lifespan label. To comprehensively evaluate the model's generalization performance and ensure the reliability of the statistical results, this embodiment uses leave-one-out cross-validation, with the specific partitioning scheme as follows: The training data is B0005, B0006, and B00018, and the test data is B0007; CALCE Ⅰ uses CS2_36, CS2_37, and CS2_38 as training data and CS2_35 as test data; CALCE Ⅱ uses CS2_35, CS2_36, and CS2_38 as training data and CS2_37 as test data.

[0055] In this embodiment, the core hyperparameters of the model are set as follows: encoder hidden layer dimension is 192, latent variable dimension is 48, attention layer number L is 4, temporal embedding dimension is 48; decoder dropout rate is 0.08; KL divergence weight coefficient is 0.003; pre-training learning rate and adaptive learning rate are both 3e -4The number of integrations during the inference and calibration phases was 50. The parameter ranges were determined through cross-dataset sensitivity analysis, demonstrating stable predictive performance and uncertainty quantification capabilities on the NASA, CALCE, and DC-DC power module datasets. All experimental models ran on the PyTorch 3.13 framework. This embodiment uses three industry-standard metrics to measure the model's point prediction accuracy: mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). To further evaluate the model's ability to quantify uncertainty, interval prediction metrics Prediction Interval Coverage (PICP) and Prediction Interval Normalized Mean Width (PINAW) were introduced. Furthermore, the Continuous Ranked Probability Score (CRPS) was used to comprehensively evaluate the overall quality of probabilistic predictions.

[0056] Please refer to Figures 6-8 The images shown are fitted images of the prediction results of the adapted model on the NASA I dataset, the CALCE I dataset, and the CALCE II dataset, respectively, provided in the embodiments of this application. Figures 6-8 In the graph, the horizontal axis represents the circular index, and the vertical axis represents the predicted remaining lifetime (RUL). Figure 6 The remaining lifetime prediction results for NASA I partition (test battery B0007) are presented. Figure 7 The remaining lifetime prediction results for CALCE I partitioning (test battery CS2_35) are presented. Figure 8 The remaining lifetime prediction results for CALCE II partitioning (test battery CS2_37) are presented. Based on Figures 6-8As can be seen, the model in this embodiment exhibits high fitting ability on all three sets of data. The corresponding quantitative evaluation results are as follows: In NASA I partition, the model achieved MAE of 0.0286, RMSE of 0.0363, R² of 98.35%, PICP of 100%, PINAW of 0.3910, and CRPS of 0.0217; in CALCE I partition, the model achieved MAE of 0.0227, RMSE of 0.0316, R² of 97.69%, PICP of 90.98%, PINAW of 0.0963, and CRPS of 0.0201; in CALCE II partition, the model achieved MAE of 0.0151, RMSE of 0.0237, R² of 98.69%, PICP of 92.18%, PINAW of 0.0989, and CRPS of 0.0209. The MAE and RMSE of the CALCE data sets are generally better than those of the NASA data set, which is consistent with the relatively smooth battery degradation curve and smaller abrupt changes in the CALCE data set. While the NASA data shows a slightly higher point prediction error due to more significant capacity rebound and local noise, the R² still reaches 98.35%, indicating that the model's grasp of the main degradation trend is not affected by local disturbances. From the perspective of uncertainty quantification, the model's PICP on all three data sets is close to or exceeds 90%, indicating that staged conformal calibration can provide coverage guarantees consistent with the set confidence level. The NASA data set achieves the highest coverage while also having a significantly higher PINAW than the CALCE data set. This difference stems from the dramatic fluctuations in the later stages of battery degradation in the NASA data set: to maintain sufficient coverage in this segment, the conformal quantiles adaptively expand the interval width, reflecting the adaptive characteristic of this application's method of trading width for reliability. The CRPS index, which comprehensively reflects the overall quality of probabilistic prediction, is below 0.022 on all three data sets, further verifying the high degree of agreement between the model's predicted distribution and actual observations. From the period-by-period prediction curves, the predicted mean curve and the actual value curve are generally closely matched, and the 95% confidence interval can basically encompass the actual degradation trajectory. Figures 6-8In the diagram, the dashed line marks the prediction starting point, and the area before the dashed line represents the early adaptation window. The model completes its personalized adaptation using only limited observations within this window, and all subsequent cycles are extrapolated predictions. In NASA I, the actual remaining lifetime curve exhibits significant capacity rebound and local fluctuations between cycles 30 and 80. Although the model's predicted curve shows slight fluctuations locally, it accurately captures the critical moment when the remaining lifetime exceeds the 70% failure threshold, and the confidence interval adaptively widens in this fluctuating region to maintain coverage. In CALCE I and CALCE II, the capacity degradation curve shows a slow, approximately linear decay over the first 500 cycles, followed by a significant accelerated decay phase around cycle 600, exceeding the EOL threshold. The model's predicted curve provides a high-fidelity reconstruction of this inflection point and the subsequent steep decline trend, with the confidence interval consistently closely following the predicted mean.

[0057] To further verify the effectiveness of the proposed method, this embodiment selects several representative remaining lifetime prediction methods as benchmarks for comparison, and conducts experiments under the same data partitioning and evaluation protocol. The selected benchmarks cover three mainstream paradigms: recurrent neural networks represented by LSTM and GRU, used to compare temporal modeling capabilities; attention-based models represented by Transformer and iTransformer, used to compare global dependency modeling capabilities; and probabilistic prediction methods represented by Gaussian process regression and MC-Dropout, used to compare uncertainty quantification capabilities. All benchmark models and the proposed method use completely identical feature inputs, training and test set partitions, and the same early observation ratio for prediction. Comparative experimental results show that, on the NASA I dataset partition, the proposed method achieves a MAE of 0.0286, an RMSE of 0.0363, an R² of 98.35%, a PICP of 100%, a PINAW of 0.2910, and a CRPS of 0.0217. In comparison, Transformer-GPR, which performed well among the comparative methods, had an R² of 98.48%, close to that of our method. However, its PICP was only 73.33%, far lower than our method's 100%, indicating that our method maintains high-precision point prediction while significantly outperforming existing methods in terms of confidence interval coverage. Autoformer-GL had an MAE of 0.0235, slightly better than our method, but its R² was only 94.65%, and its PICP was 82.61%, also lower than our method, verifying the comprehensive advantage of our method in balancing point prediction accuracy and coverage. On the CALCE I dataset, our method achieved an MAE of 0.0227, an RMSE of 0.0316, an R² of 97.69%, a PICP of 90.98%, a PINAW of 0.0963, and a CRPS of 0.0201. In the comparison, LSTM-MCDropout's PICP is 92.34%, slightly higher than the method in this application, but its CRPS is 0.0267, higher than the 0.0201 of the method in this application, indicating that the overall quality of the probability prediction distribution of the method in this application is better. Transformer-GPR's R² is 99.65%, higher than the method in this application, but its PICP is 88.33%, lower than the method in this application, and its CRPS is 0.0276, higher than the method in this application, indicating that the method in this application has a significant advantage in uncertainty quantification. In the CALCE II dataset partitioning, the method in this application has MAE of 0.0151, RMSE of 0.0237, R² of 98.69%, PICP of 92.18%, PINAW of 0.0989, and CRPS of 0.0209.Among the comparative methods, GRU-MCDropout achieved a PICP of 97.86%, higher than the method in this application, and its CRPS was 0.0195, slightly better than the method in this application. However, its MAE was 0.0372 and RMSE was 0.0458, both significantly higher than the method in this application. Autoformer-GL achieved an MAE of 0.0167, close to the method in this application, but its PICP was only 75.37%, far lower than the 92.18% of the method in this application, and its PINAW was 0.2320, significantly higher than the method in this application, indicating that it has insufficient coverage and a large interval width. The comparative experimental results show that the method in this application can achieve accurate point prediction with high R² and low MAE and RMSE on all three datasets, while ensuring statistical coverage of the prediction interval with a PICP close to or exceeding 90%, and maintaining the lowest or near-optimal level in CRPS. The overall point prediction accuracy and uncertainty quantification capability are significantly better than the comparative benchmarks.

[0058] To verify the necessity of the core designs of the method in this application, this embodiment designed five ablation experiments on three sets of data partitions, removing or replacing key components of the model respectively. The first set is w / o Dual-Head, removing the dual-head decoder and retaining only a single remaining lifetime prediction head; the second set is w / o Variationalz, removing variational latent variables and replacing posterior sampling with deterministic representations; the third set is w / o Context Aggregation, removing context mean aggregation and destroying the permutation-invariant global representation; the fourth set is Global Conformal, replacing staged calibration with global conformal calibration; and the fifth set is NoAdaptation, skipping the few-shot adaptation stage and directly using the meta-trained model for prediction. Ablation experiments showed that removing the dual-head decoder caused the most significant performance degradation, with MAE increasing from 0.0286 to 0.0641 and RMSE from 0.0363 to 0.0801 on NASA I. This indicates that the implicit multi-task regularization formed by the remaining lifetime prediction head and health index prediction head through shared latent variables plays a key supporting role in few-shot generalization ability. Removing variational latent variables and removing context aggregation also led to a significant decline in accuracy, verifying the necessity of variational inference and context aggregation in cross-module knowledge transfer. After replacing phased calibration with global calibration, PICP decreased from 90.98% to 80.05% on CALCE I and from 92.18% to 66.98% on CALCE II, indicating that ignoring the qualitative differences in errors during the degradation phase leads to a systematic deviation of coverage from the nominal level. Skipping the few-shot adaptation phase resulted in a loss of point prediction accuracy and coverage reliability on most partitions, indicating that selective parameter fine-tuning is indispensable for the personalized adaptation of target components. By combining both point prediction and interval prediction metrics, the fully adapted model achieved the best balance between accuracy and reliability across the three sets of data.

[0059] Please refer to Figure 9 It shows a schematic diagram of the accelerated degradation test platform provided in the embodiments of this application, such as Figure 9 As shown, the test platform consists of a computer, a data acquisition card, a programmable power supply, a temperature and humidity test chamber, and a data recording and storage terminal. Please refer to [reference needed]. Figure 10 It shows a schematic diagram of the accelerated degradation test circuit provided in the embodiment of this application, such as Figure 10As shown, the power supply terminal of the electronic power component under test (U1) is connected to a 24V DC power supply, and the output terminal is connected to a load (5V) to simulate the load state under actual working conditions. The load (5V) is connected in series in the output circuit to consume the output power and maintain the current loop. The current sensor (LTS25NP) is connected in series in the main circuit to collect the current value of the power drive device. The data acquisition card is connected to the output terminal of the current sensor through a signal line to receive the analog signal transmitted by the current sensor and convert it into a digital signal. The computer is connected to the data acquisition card through a data bus to collect and store the current data in real time at a sampling frequency of 5KHz. The internal circuit of the electronic power component includes resistors (R1, R2) to build an internal voltage sampling and feedback loop, providing internal circuit connection for the electronic power component under test (U1).

[0060] This embodiment uses a certain type of airborne electronic power component DC-DC power module, numbered DC1-DC4, as the test object, and conducts an accelerated degradation test under the combined effects of temperature, humidity, and electrical stress. According to the manual, the optimal operating environment for this device is a temperature of 20-40℃ and a humidity of 40%-60%. To accelerate the degradation process while ensuring normal operation, the stress conditions applied to the device were a temperature of 80℃, a relative humidity of 90%, and an electrical stress of 24V DC voltage. The accelerated degradation test lasted 912 hours, and a total of 16,700 data samples were collected at evenly spaced time intervals.

[0061] During the experiment, outliers may appear in the experimental data due to sensor malfunctions or transient impacts. Therefore, data preprocessing is necessary to eliminate interference as much as possible and improve the accuracy of subsequent modeling. First, outlier removal is performed. Ideally, the DC-DC module output voltage is 5V and the output current is positive. Therefore, data points with voltage values ​​less than 0V or greater than 10V and current values ​​less than 0A need to be removed to eliminate interference. Next-neighbor averaging filtering is then used to eliminate random fluctuations caused by sensor noise. The preprocessed data sequence is then used as the model input.

[0062] Since the output voltage of a DC-DC power module remains almost constant under the action of a voltage regulator, the average values ​​of voltage, current, or resistance alone are insufficient to effectively reflect its health degradation. Considering the failure mechanism of the power module, this embodiment selects the instantaneous fluctuation intensity of current, resistance, and power within each thermal cycle segment as degradation characterization quantities. The three index sequences are normalized and averaged point by point to construct a comprehensive health index. This comprehensive health index starts at 1.0 at the initial moment and monotonically increases with the aging process of the module, simultaneously reflecting degradation information in three dimensions: load current, internal impedance, and output power. Referring to the construction method of the lithium-ion battery remaining life tag in the previous embodiment, the moment when the health index exceeds a preset failure threshold is defined as the module failure point. The health index sequence is converted into a normalized remaining life tag using a linear mapping, monotonically decreasing from 1.0 at the initial moment to 0 at the failure moment, thus maintaining complete semantic consistency with the battery data in the previous embodiment. The preset failure threshold is greater than 1.0, preferably 1.12.

[0063] Similar to the previous embodiment, three of the four tested modules are selected for training, with the remaining one used as the test subject. Two experimental datasets are constructed using leave-one-out cross-validation: Dataset I uses DC1, DC2, and DC3 as training data and DC4 as test data; Dataset II uses DC1, DC2, and DC4 as training data and DC3 as test data. The core structure, hyperparameters, and training strategy of the DC-DC power module in this embodiment remain consistent with the previous embodiment. Key designs such as the number of multi-head self-attention layers in the context encoder, the dimension of latent variables, and the hierarchical structure of the dual-head decoder are not adjusted in any way, directly verifying the model's adaptability across different scenarios.

[0064] Please refer to Figures 11-12 It shows a schematic diagram of the prediction results of Datasets I and Datasets II provided in the embodiments of this application. Figures 11-12 The horizontal axis represents the circular index, and the vertical axis represents the predicted remaining useful life (RUL). Figure 11 The remaining lifetime prediction results of Datasets I (test module DC4) are displayed. Figure 12The remaining lifetime prediction results for Datasets II (test module DC3) are presented. The model completed a full probabilistic prediction of the remaining 90% of the degradation trajectory using only the first 10% of the degradation data. From the fitted graphs, the predicted remaining lifetime curves closely match the actual remaining lifetime curves, showing a completely consistent overall trend. Especially during the rapid degradation phase, the model accurately captured the inflection point of the degradation acceleration trend. Specifically, in Datasets I, the model achieved almost perfect tracking of the actual remaining lifetime during the rapid degradation phase, and also accurately captured the trend inflection point during the rapid decline phase. In Datasets II, the model maintained a high degree of fit for the predictions in the early and middle stages of degradation, but some deviation appeared in the later stages, with the predicted remaining lifetime slightly higher than the actual value, reflecting the differences in the final degradation path of individual DC-DC modules. It should be noted that in high-reliability scenarios, the device under test is usually replaced before the failure threshold; therefore, the prediction deviation in the later stages of degradation has a limited impact on maintenance decisions in practical engineering.

[0065] The quantitative evaluation results are summarized as follows: On Datasets I, the model achieved a prediction accuracy of R²=99.60%, RMSE=0.0234, MAE=0.0146, PICP=70.50%, PINAW=0.0287, and CRPS=0.0109; on Datasets II, it achieved a prediction accuracy of R²=98.75%, RMSE=0.0293, MAE=0.0215, PICP=78.3%, PINAW=0.0320, and CRPS=0.0147. The slightly lower coverage in Datasets I is mainly due to the local non-monotonic rebound in the DC4 degradation path of the test module during the middle stage, while the rebound characteristics of the training set module are relatively weak, causing the conformal calibration quantiles to slightly underestimate the error magnitude in this interval. Datasets II, due to systematic bias in the later stages of degradation, appropriately expanded the confidence interval in this segment to cover greater prediction uncertainty; therefore, its PICP is slightly higher than that of Datasets I. In summary, the method presented in this application can provide confidence intervals that adaptively match the error level on both datasets. The prediction accuracy of the model on Datasets I is at the same level as the best result on the CALCE dataset (RMSE=0.0237, MAE=0.0151), indicating that the method does not show significant accuracy loss when crossing the domain boundaries of battery data and industrial power module data, demonstrating its strong transferability and engineering practical value under different degradation mechanisms and acquisition environments. It is worth noting that although the coverage rate on DC-DC modules did not reach 90% or higher, the experimental results of the adapted model on public datasets, the comparative experimental results with existing models, and the ablation experiment results have effectively verified the accuracy and reliability of the model constructed in this application. Furthermore, from... Figures 11-12 The provided prediction results diagram shows that the coverage is mainly low in the later stages of degradation. However, in real industrial scenarios, equipment will be replaced directly before its lifespan is reduced to the failure threshold. Therefore, the deviation at this stage has a limited impact on maintenance decisions in actual engineering.

[0066] In summary, the present application provides a few-shot probabilistic prognostic method for electronic power components based on neural processes. This method divides the historical degradation sequences of each component into a context set and a query set according to a preset ratio, enabling the model to fully learn the shared degradation patterns across components during the meta-training phase. It extracts multi-dimensional degradation feature vectors from each cycle point to construct a multi-level degradation representation system from absolute state, first-order dynamics, second-order trend, local uncertainty, stage interaction to global position. By applying time encoding to the normalized cycle index, components with different total cycle counts have the same time encoding value at the same degradation progress, thus achieving a unified representation across components with different lifespans. By fusing the multi-dimensional degradation feature vector, time encoding vector, and the true observation values ​​of cycle points into the encoder input vector, the encoder can infer potential degradation patterns based on early true degradation values. After processing the encoder input vector sequentially through a feature projection layer and a context encoder, an aggregation operation is performed to capture the non-local dependencies between context observation points. This aggregation operation enables the model to handle different numbers of context observation points and maintains permutation invariance, significantly improving generalization ability under few-shot conditions. By inputting the global degradation representation into the variational inference network, the mean vector and log-variance vector of the latent variable posterior distribution are generated. Reparameterized sampling is used to obtain the latent variable vector, and the KL divergence is calculated. This allows the gradient to propagate back to the variational inference network from subsequent steps, and the KL divergence constrains the spatial distribution of the latent variables to prevent overfitting. Based on the latent variable vector and the global degradation representation, a decoder input vector is constructed using the temporal encoding vector and normalized cyclic index of each query point in the query set. This enables the decoder to perceive the degradation mode and current degradation progress of the component at each prediction time. A dual-head decoder simultaneously generates remaining lifespan predictions and health indicator predictions. The complementarity of these two tasks in the degradation structure is used to achieve implicit multi-task regularization, significantly improving prediction accuracy in low-sample scenarios. Furthermore, a time-weighted prediction loss is introduced to guide the model's attention to key regions in the later stages of degradation, thereby improving prediction accuracy near abrupt change points. By constructing a total loss function based on prediction loss and KL divergence to jointly optimize the prediction model parameters, an initial prediction model after meta-training is obtained, enabling the model to have good initial performance with very few target samples. By performing hierarchical parameter freezing and unfreezing on the initial prediction model, the feature projection layer and all layers of the context encoder except the last layer are frozen to retain the common degradation features shared across modules. The last layer of the context encoder, the variational inference network, and the dual-head decoder are unfrozen to adapt to the individual degradation patterns of the target components. Then, the unfrozen parameters are updated through the context set of the target components to obtain the adapted model and generate the ensemble prediction mean and ensemble standard deviation. This enables the model to quickly adapt to individual differences with very few target samples and has cross-scene transfer capabilities.By dividing the calibration samples into multiple degradation stages based on the remaining lifetime prediction, and combining the ensemble standard deviation to determine the confidence interval according to the stage to which the query point belongs, the confidence interval can adapt to the heterogeneous characteristics of the error distribution under different degradation stages. Without relying on any distribution assumptions, a confidence interval with finite sample coverage is obtained in a marginal sense, and the width of the confidence interval adapts to the ensemble standard deviation, improving the compactness of the interval while maintaining coverage.

[0067] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0068] The features disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0069] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A few-sample probability prognostic method for electronic power components based on neural processes, characterized in that, The method includes: The historical degradation sequence of electronic power components is obtained, and the historical degradation sequence of each component is divided into a context set and a query set according to a preset ratio. The multidimensional degradation feature vector of the cycle point in each sequence is extracted. Temporal encoding is applied to the normalized cyclic index, and the multidimensional degenerate feature vector, temporal encoding vector, and real observations of the cyclic points are fused to obtain the encoder input vector. The encoder input vector is then processed sequentially by the feature projection layer and the context encoder, and then aggregated to obtain the global degenerate representation. The global degenerate representation is input into the variational inference network to generate the mean vector and log-variance vector of the latent variable posterior distribution. The latent variable vector is obtained through reparameterized sampling, and the KL divergence between the approximate posterior distribution and the prior distribution followed by the latent variable vector is calculated. Based on latent variable vectors and global degradation representations, a decoder input vector is constructed for the time encoding vector and normalized cyclic index of each query point in the query set. The remaining life expectancy prediction and health index prediction are generated by a dual-head decoder. The total loss function is constructed based on the prediction loss and KL divergence, and the prediction model parameters are jointly optimized to obtain the initial prediction model after meta-training. The initial prediction model is subjected to hierarchical parameter freezing and unfreezing. The unfrozen parameters are updated through the context set of the target component to obtain the adapted model and generate the ensemble prediction mean and ensemble standard deviation. The calibration samples are divided into multiple degradation stages based on the predicted remaining lifetime. Using the integrated standard deviation, confidence intervals are determined according to the stage to which the query point belongs, including: After the meta-training and adaptation phases are completed, the adapted model is used to predict the last 50% of all components in the training set. The ensemble prediction mean and the true values ​​are used to construct a calibration set. The corresponding degradation stage is divided according to the remaining lifetime prediction value of each calibration sample in the calibration set. When the remaining lifetime prediction value is greater than 0.6, it is the early stage; when the remaining lifetime prediction value is greater than 0.25 and less than or equal to 0.6, it is the middle stage; and when the remaining lifetime prediction value is less than or equal to 0.25, it is the late stage. Based on all calibration samples within each degradation stage, the variance-normalized inconsistency score is calculated, expressed by the following formula: ； In the formula, Indicates the first The non-consistency score of each calibration sample; Indicates the first The true value of each calibration sample; Indicates the first The ensemble predicted mean of 1 calibration sample; Indicates the first The integrated standard deviation of each calibration sample; This represents a pre-defined positive decimal constant. Setting a significance level, calculate the shape-preserving quantiles of the inconsistency score set within each degradation stage. The formula for calculating the shape-preserving quantiles is expressed as follows: ； In the formula, This represents the pre-defined significance level, with a corresponding confidence level of . ; Indicates the first Each stage of degradation at the confidence level Conformal quantiles below; Indicates the first The set of inconsistent scores for all calibration samples within a degradation stage; Indicates the first The number of calibration samples within each degradation stage; Represents the quantile function; Indicates that for the first The conformal quantile level is determined by a finite sample size adjustment for each degradation stage sample. When greater than 1, the conformal quantile Take the maximum value of the set of inconsistent scores, i.e. ; Based on the predicted remaining lifetime of the query point, determine the degradation stage to which the query point belongs, extract the conformal quantiles corresponding to that stage, and construct a normalized prediction interval, which is expressed by the following formula: ； In the formula, Indicates query point The normalized prediction interval; Indicates query point The ensemble prediction mean; Indicates the degradation stage to which the query point belongs. The corresponding conformal quantiles; Indicates query point The integrated standard deviation; Based on the normalized prediction interval, the remaining lifespan prediction interval and the health indicator prediction interval are denormalized separately to obtain the remaining lifespan confidence interval and the health indicator confidence interval. Specifically, the remaining lifespan prediction interval denormalization is achieved by multiplying the upper and lower bounds of the normalized prediction interval by the total number of cycles of the target component, respectively, to obtain the remaining lifespan confidence interval at the actual cycle number scale. The health indicator prediction interval denormalization is achieved by multiplying the upper and lower bounds of the normalized prediction interval by the standard deviation of the health indicators in the training set, respectively, and then adding the mean of the health indicators in the training set, to obtain the health indicator confidence interval at the original physical dimensions.

2. The method according to claim 1, characterized in that, The prediction model consists of a feature projection layer, a context encoder, a variational inference network, and a dual-head decoder; the context encoder is composed of... Multi-head self-attention network structure The preset positive integer is used; the variational inference network contains two independent linear transformation branches; the dual-head decoder contains a shared decoder backbone, a remaining lifetime prediction head, and a health indicator prediction head.

3. The method according to claim 1, characterized in that, The multidimensional degradation feature vector includes normalized health index values, first-order difference values, linear regression slope within the local window, standard deviation within the local window, interaction term between normalized health index and normalized cyclic index, and normalized cyclic index.

4. The method according to claim 1, characterized in that, The global degenerate representation is input into the variational inference network to generate the mean vector and log-variance vector of the latent variable posterior distribution. The latent variable vector is obtained through reparameterized sampling. The KL divergence between the approximate posterior distribution and the prior distribution to which the latent variable vector follows is calculated, including: The global degradation representation is input into two independent linear transformation branches of the variational inference network. The first branch is used to calculate the mean vector of the posterior distribution of the latent variables, and the second branch is used to calculate the log-variance vector of the posterior distribution of the latent variables. The standard deviation vector is calculated based on the logarithmic variance vector, and the approximate posterior distribution of the component is determined based on the mean vector and the standard deviation vector. Based on the mean vector and standard deviation vector, random noise vectors are sampled from the standard normal distribution, and latent variable vectors are calculated through reparameterization transformation; Using the standard normal distribution as the prior distribution followed by the latent variable vector, calculate the KL divergence between the approximate posterior distribution and the prior distribution.

5. The method according to claim 2, characterized in that, The dual-head decoder processes and generates predicted remaining life expectancy and health indicators, including: The decoder input vector is fed into the shared decoder backbone to extract shared deep degradation features. The shared decoder backbone is composed of a multilayer perceptron with three layers of hidden dimensions decreasing layer by layer. The shared deep degradation features are passed to the remaining lifespan prediction head and the health indicator prediction head, respectively. In the remaining lifespan prediction head, a normalized remaining lifespan prediction value is output through mapping using a Sigmoid activation function. In the health indicator prediction head, a normalized health indicator prediction value is output through a linear activation function. The remaining lifespan prediction head and the health indicator prediction head share the shared decoder backbone.

6. The method according to claim 1, characterized in that, The predicted loss includes time-weighted remaining lifespan task loss and health indicator task loss. The time-weighted remaining lifespan task loss is calculated using the following formula: ； In the formula, This represents the time-weighted remaining lifetime task loss. This represents the total number of query points in the query set. Indicates the first in the query set Index for each query point; Represents normalized cyclic index The default monotonically increasing function; Indicates the first The first component A normalized cyclic index for each query point; Indicates the first The first component Loss function for predicting remaining lifetime at each query point; Indicates the first The first component Normalized remaining lifetime prediction for each query point; Indicates the first The first component Normalized true remaining lifetime value for each query point; The formula for calculating time-weighted health indicator task loss is expressed as follows: ； In the formula, Indicates time-weighted health indicators of task loss; Indicates the first The first component The health indicator prediction loss function for each query point; Indicates the first The first component Normalized health indicator predictions for each query point; Indicates the first The first component Normalized true values ​​of health indicators for each query point; The formula for calculating the KL divergence is expressed as follows: ； In the formula, Represents the approximate posterior distribution With prior distribution The KL divergence between the variables is used to constrain the distribution shape of the latent variable space to prevent overfitting. Indicates the first The latent variable vector of each component; The approximate posterior distribution conditioned on contextual information is defined as a Gaussian distribution where each dimension is independent. ; The prior distribution of the latent variable vector is defined as the standard normal distribution. ; Indicates distribution The following expectations; The logarithmic density function represents the approximate posterior distribution; The logarithmic density function represents the prior distribution; The total loss function is calculated using the following formula: ； In the formula, Represents the total loss function; This represents the time-weighted remaining lifetime task loss. Indicates time-weighted health indicators of task loss; The preset weighting coefficients represent the KL divergence; This represents the KL divergence between the approximate posterior distribution and the prior distribution.

7. The method according to claim 1, characterized in that, The initial prediction model undergoes hierarchical parameter freezing and unfreezing. Unfrozen parameters are updated using the context set of the target component to obtain the adapted model and generate the ensemble prediction mean and ensemble standard deviation, including: Freeze all parameters of the feature projection layer, freeze all parameters of all layers in the context encoder except the last layer; unfreeze all parameters of the last layer of the context encoder, unfreeze all parameters of the variational inference network, and unfreeze all parameters of the dual-head decoder. Set the number of integrations and Each of the three distinct random seeds is used to initialize the random number generator for the adaptation process; the process is then executed independently using the context set of the target component. In each parameter optimization, only the unfrozen parameters are updated, while the frozen parameters remain unchanged, resulting in... An adapted model, Each adapted model is distributed in a different local optimum region in the parameter space due to different random seeds; among them... ; according to Each adapted model independently predicts the query set for the target component, resulting in... Group prediction results; calculated by arithmetic mean The final prediction result for the target component is obtained by ensemble prediction of the group prediction results; calculation The integrated standard deviation of the prediction results is used to obtain an estimate of the cognitive uncertainty of the target component prediction.

Images