Component life prediction method based on multi-scale adaptive feature decomposition

Abstract

This application provides a component life prediction method based on multi-scale adaptive feature decomposition. The method includes: acquiring multiple signals during component use; decomposing the multiple signals into multiple components using empirical mode decomposition; decomposing each component into multiple sub-components using a multi-scale sliding window method, where the sliding window step size of each component is related to its dominant frequency; generating a target feature sequence that meets temporal consistency and correlation requirements based on each component and sub-component; weighting the target feature sequence according to weights related to aging stage labels and temporal evolution gains to generate a first weighted feature vector sequence representing aging-sensitive features; inputting the aging stage labels and the first weighted feature vector sequence into a recurrent neural network model to obtain a hidden state vector sequence; and inputting the hidden state vector sequence and the first weighted feature vector sequence into a dynamic regression network fused with Bayesian inference to obtain the remaining life probability distribution of the component.

Description

Technical Field

[0001] This application mainly relates to the field of nuclear power plant safety technology, and in particular to a component life prediction method based on multi-scale adaptive feature decomposition. Background Technology

[0002] Nuclear power plants operate in complex environments characterized by high temperature, high pressure, radiation, and mechanical vibration. During their service life, components are prone to material aging, performance degradation, and fatigue damage, leading to safety hazards and maintenance challenges. Online monitoring technology deploys various sensors to collect multi-source signals such as temperature, pressure, and vibration to analyze changes in component health. However, this technology is affected by factors such as on-site interference, signal noise, and changes in operating conditions, resulting in a large amount of invalid and abnormal data. This causes significant differences between signal amplitude and time scale, making it impossible to accurately extract component aging characteristics and predict component remaining life. Furthermore, the process of component health changes is complex, characterized by multiple stages, nonlinearity, and abrupt changes, making it difficult for existing technologies to provide highly reliable and interpretable component life predictions. Therefore, there is an urgent need in this field for a component life prediction method that can efficiently extract component aging characteristics and provide highly reliable predictions of component remaining life, thereby meeting the pressing needs for component health management and proactive maintenance. Summary of the Invention

[0003] The technical problem to be solved by this application is to provide a component life prediction method based on multi-scale adaptive feature decomposition, which can achieve efficient extraction of component aging features and high-reliability prediction of component remaining life, for use in component health management and proactive operation and maintenance.

[0004] To address the aforementioned technical problems, this application provides a component lifetime prediction method based on multi-scale adaptive feature decomposition, comprising the following steps: acquiring multiple signals during the component's use; decomposing the multiple signals into multiple components using an empirical mode decomposition method, each component being an intrinsic mode function; decomposing each component into multiple sub-components using a multi-scale sliding window method, wherein the sliding window step size of each component is related to the dominant frequency of each component; generating a target feature sequence that satisfies temporal consistency and correlation requirements based on each component and sub-component; weighting the target feature sequence according to weights related to aging stage labels and temporal evolution gains to generate a first weighted feature vector sequence representing aging-sensitive features; inputting the aging stage labels and the first weighted feature vector sequence into a recurrent neural network model to obtain a hidden state vector sequence; and inputting the hidden state vector sequence and the first weighted feature vector sequence into a dynamic regression network fused with Bayesian inference to obtain the remaining lifetime probability distribution of the component.

[0005] In one embodiment of this application, generating a target feature sequence that satisfies temporal consistency and correlation requirements based on each component and sub-component includes: extracting a plurality of first candidate feature sequences from each component and sub-component; filtering a second candidate feature sequence that satisfies a temporal consistency threshold from the first candidate feature sequences; and filtering the target feature sequence that satisfies a correlation threshold from the second candidate feature vector.

[0006] In one embodiment of this application, weighting the target feature sequence according to weights related to aging stage labels and temporal evolution gains to generate a first weighted feature vector sequence representing aging-sensitive features includes: calculating mutual information weights based on the target feature sequence and the aging stage labels, and calculating evolution gain weights based on the target feature sequence; calculating a weighted target feature sequence based on the mutual information weights and the evolution gain weights; and constructing a first weighted feature vector for each time point according to the time axis of the weighted target feature sequence to generate the first weighted feature vector sequence.

[0007] In one embodiment of this application, the method further includes: generating a relevant tensor principal component trend representing the aging evolution trend, a mutation index sequence representing aging evolution mutations, and an anomaly position indicating the aging evolution inflection point based on the hidden state vector sequence.

[0008] In one embodiment of this application, the method further includes: generating a tensor sequence representing the aging evolution trend, a mutation index sequence representing aging evolution mutations, and an anomaly position indicating an aging evolution inflection point based on the hidden state vector sequence, comprising: extracting state subsequences from the hidden state vector sequence using a first sliding window; calculating the relevant tensor of each state subsequence to obtain a group of relevant tensors; decomposing the relevant tensors to extract the principal components of the relevant tensors to obtain a sequence of principal components of the relevant tensors; calculating the F-norm distance between the principal components of the relevant tensors of adjacent first sliding windows in the sequence of principal components of the relevant tensors to obtain an F-norm distance sequence; identifying anomalous F-norm distances in the F-norm distance sequence and locating anomaly windows corresponding to the anomalous F-norm distances; and outputting the sequence of principal components of the relevant tensors as the tensor sequence, the F-norm distance sequence as the mutation index sequence, and the anomaly window as the anomaly position.

[0009] In one embodiment of this application, in the step of extracting a state subsequence from the hidden state vector sequence using a first sliding window, the length of the first sliding window is greater than the step size of the first sliding window.

[0010] In one embodiment of this application, the component is a non-metallic flexible hose used in nuclear power plants, and the various signals include temperature, pressure, vibration, and infrared spectrum.

[0011] This application also provides an electronic device, comprising: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement the method described above.

[0012] This application also provides a computer storage medium storing computer program code that, when executed by a processor, implements the method described above.

[0013] This application also provides a computer program product, including computer program code, which, when executed by one or more processors, implements the steps of the method described above.

[0014] Compared with the prior art, this application has the following advantages:

[0015] (1) This application uses a joint algorithm of empirical mode decomposition and variable step size multiscale decomposition to adaptively adjust the decomposition scale according to the signal frequency, thereby realizing customized scale analysis of high frequency noise and low frequency evolution trend, and providing a data basis for characterizing aging-sensitive features.

[0016] (2) This application uses the calculation of relevant tensor sets to characterize the dynamic evolution of the health status of components in time and multi-dimensional feature space, thereby enabling the tracking of the stage-by-stage abrupt changes and gradual changes in the aging characteristics of components, as well as the capture of key time nodes in the aging of components.

[0017] (3) Compared with the single value predicted by the prior art, this application can obtain the probability distribution of the remaining life of the component by inputting the hidden state vector sequence coupled with time and state dimensions into a dynamic regression network that is fused with Bayesian inference. In addition, the remaining life confidence interval can be used to dynamically analyze and warn of the sudden trend of early abnormal aging characteristics. Attached Figure Description

[0018] The accompanying drawings are included to provide a further understanding of this application; they are incorporated into and constitute a part of this application. The drawings illustrate embodiments of this application and, together with this specification, serve to explain the principles of this application. In the drawings:

[0019] Figure 1 This is a flowchart illustrating a component life prediction method based on multi-scale adaptive feature decomposition in one embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the process of acquiring multiple signals in one embodiment of this application;

[0021] Figure 3 This is a schematic diagram of the process of obtaining the target feature sequence in one embodiment of this application;

[0022] Figure 4 This is a schematic diagram of the process for generating the first weighted feature vector sequence in one embodiment of this application;

[0023] Figure 5 This is a flowchart illustrating a component life prediction method in another embodiment of this application;

[0024] Figure 6 This is a flowchart illustrating the modeling tensor sequence, mutation index sequence, and anomaly location in one embodiment of this application.

[0025] Figure 7 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0026] To more clearly illustrate the technical solutions of 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 merely some examples or embodiments of this application. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0027] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0028] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0029] Furthermore, it should be noted that the use of terms such as "first" and "second" to define components is merely for the purpose of distinguishing the corresponding components. Unless otherwise stated, these terms have no special meaning and therefore should not be construed as limiting the scope of protection of this application. In addition, although the terminology used in this application is selected from commonly known and used terms, some terms mentioned in this application's specification may have been chosen by the applicant according to his or her judgment, and their detailed meanings are explained in the relevant sections of this description. Moreover, this application should be understood not only through the actual terms used, but also through the meaning implied by each term.

[0030] This application uses flowcharts to illustrate the steps performed by the method according to embodiments of this application. It should be understood that the preceding or following steps are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other steps may be added to these processes, or one or more steps may be removed from these processes.

[0031] refer to Figure 1 The flowchart shown in one embodiment illustrates a component life prediction method based on multi-scale adaptive feature decomposition, which includes the following steps.

[0032] Step S101: Acquire various signals during the use of the components.

[0033] This step allows for the acquisition of multi-source data that reflects various aspects of components during use, including changes in material properties, structural stress, potential thermal aging, and mechanical damage. For example, the acquired signals may include, but are not limited to, temperature, pressure, vibration, and infrared spectroscopy.

[0034] In this step, refer to Figure 2 The flowchart of acquiring multiple signals in one embodiment is shown. Step S101 includes the following steps.

[0035] Step S201: Collect multi-source data. Multiple sensors are deployed to continuously collect multi-source data at high frequency, recording the operational status of components in real-time throughout the entire process. In some embodiments, the deployed sensors include: temperature sensors, pressure sensors, vibration sensors, and infrared spectroscopy sensors, etc.

[0036] Step S202 involves cleaning and denoising the collected data to obtain purified data, thereby reducing the impact of interference factors (e.g., abnormal data transmission, sensor malfunction, etc.) on the accuracy of the collected data. In some embodiments, an adaptive wavelet thresholding algorithm is used for data cleaning and denoising to remove high-frequency noise from the collected data. Furthermore, invalid or abnormal samples in the collected data are cleaned using manual noise identification methods, such as expert experience or anomaly detection rules. The uncleaned samples are used as purified data, thus ensuring the accuracy and usability of the data. It is understood that, in addition to the aforementioned embodiments, denoising algorithms such as autoencoder denoising, principal component analysis denoising, or deep learning denoising networks can also be used to clean and denoise the collected data.

[0037] Step S203 involves standardizing the purified data to obtain processed multi-source signals, thereby improving the quality of the multi-source signals. In some embodiments, a segmented normalization algorithm is used to standardize the purified data. This algorithm segments the long-time-series purified data according to the time nodes of operating condition changes or fixed time windows, and performs standardization processing separately within each segment. Specifically, the algorithm calculates the normalization parameters such as the maximum value, minimum value, mean, or standard deviation of each segment of purified data, and performs a standardization transformation on the segment of purified data according to the normalization parameters to achieve amplitude scaling and dimension elimination. The segmented normalization algorithm can preserve local mutations and abnormal fluctuations in the purified data on the one hand, and avoid the problem of important information being masked by normalization on the other hand. It can be understood that, in addition to the segmented normalization algorithm in the aforementioned embodiments, global normalization, quantile normalization, Z-Score normalization, and other normalization algorithms can also be used to normalize the cleaned and denoised data.

[0038] Step S102: Use the empirical mode decomposition method to decompose multiple signals into multiple components, each of which is an intrinsic mode function.

[0039] In this step, the Empirical Mode Decomposition (EMD) method is used to decompose each signal. The EMD formula is as follows: In the formula, For this signal, For the signal's first eigenmode functions of order 1 This represents the remaining components of the signal. After decomposing each signal, only a finite set of eigenmode function components obtained from the decomposition are retained. Therefore, in this step, the components obtained from the decomposition include the eigenmode functions obtained from the decomposition of each signal.

[0040] Step S103: Decompose each component into multiple sub-components using a multi-scale sliding window method, where the sliding window step size of each component is related to the dominant frequency of each component.

[0041] In this step, a multi-scale sliding window is used for each component, allowing the window to traverse the entire component. Within each window, the components contained within that window are further decomposed, and the decomposition results are statistically analyzed and combined to obtain multiple sub-components for each component after decomposition. The step size of the sliding window is variable; specifically, the step size for each component is related to its dominant frequency. This allows the sliding window to adapt to different frequency noises and variation characteristics within the component, thereby achieving adaptive scaling decomposition for high-frequency noise and low-frequency evolution trends.

[0042] In some embodiments, the sliding window step size of the component currently being decomposed is inversely proportional to the frequency of that component.

[0043] In some embodiments, the sliding window step size of the component is obtained by joint optimization based on the energy and spectral entropy of the component currently being decomposed. The formula for joint optimization is as follows: In the formula, This is the sliding window step size for this component. For this component of energy, The spectral entropy of this component. For energy weighting coefficients, These are the weighting coefficients for the spectral entropy. The weighting coefficients for energy and spectral entropy are adjusted based on the physical characteristics of the signal, task requirements (such as decomposition target requirements), and the actual effectiveness of the method. For example, energy and spectral entropy are initially set to have equal weight during the decomposition process, i.e. and All are set to 0.5 to ensure a comprehensive consideration of energy distribution and frequency distribution.

[0044] In some embodiments, the decomposition method within each sliding window includes methods such as empirical mode decomposition and wavelet decomposition.

[0045] Step 104: Generate a target feature sequence that meets the requirements of temporal consistency and correlation based on each component and sub-component.

[0046] Specifically, the components and sub-components obtained from steps 102 and 103 are subjected to feature optimization, screening, and redundancy removal to generate a target feature sequence that meets the requirements of temporal consistency and correlation. (Reference) Figure 3 The flowchart of obtaining the target feature sequence in one embodiment is shown. Step 104 includes the following steps.

[0047] Step 301: Extract multiple first candidate feature sequences from each component and sub-component. In some embodiments, for each component, extract the statistical features of the components within each second sliding window according to the second sliding window. Similarly, for each sub-component, extract the statistical features of the sub-components within each second sliding window according to the second sliding window. The statistical features include mean, variance, and kurtosis, etc. The second sliding windows are sorted according to time order, and the same statistical feature values ​​within all second sliding windows constitute the first candidate feature sequence. For example, the first candidate feature sequence includes the mean sequence, variance sequence, and kurtosis sequence of the component and sub-components, etc. It can be understood that at least one first candidate feature sequence can be extracted from each component and sub-component.

[0048] Step 302: Select second candidate feature sequences that meet the temporal consistency threshold from the first candidate feature sequences, thereby eliminating features with large temporal fluctuations, strong randomness, and lack of sustained representativeness. Specifically, firstly, a temporal consistency index is defined, and the temporal consistency index of each first candidate feature sequence between adjacent second sliding windows is calculated, thus reflecting the changing trend of the feature in the time dimension. The formula for calculating the temporal consistency index is: In the formula, For the first The first candidate feature sequences at time node eigenvalues, For the first The first candidate feature sequences at time node The eigenvalues. This is understandable. The closer the value is to 1, the more stable and representative the feature is in the temporal evolution.

[0049] In some embodiments, the Pearson correlation coefficient is used as a time series consistency metric. In other embodiments, dynamic time series consistency is defined as a time series consistency metric.

[0050] Subsequently, a time-series consistency threshold is preset, and the calculated time-series consistency index is used to select the first candidate feature sequence that meets the time-series consistency threshold (i.e., the time-series consistency index is greater than the time-series consistency threshold) as the second candidate feature sequence, and discard the first candidate feature sequence that does not meet the time-series consistency threshold (i.e., the time-series consistency index is less than or equal to the time-series consistency threshold). In some embodiments, the time-series consistency threshold is preset to 0.8.

[0051] Step 303: Select target feature sequences that meet the relevance threshold from the second candidate feature sequences to solve the information redundancy problem. Specifically, calculate the relevance matrix of the second candidate feature sequences. Each element in the matrix For the first The second candidate feature sequence and the first The correlation between two second candidate feature sequences. A high correlation between two second candidate feature sequences indicates that they are highly correlated. Using a clustering algorithm, the second candidate feature sequences are grouped according to the correlation matrix. Each second candidate feature sequence belongs to a unique group, and the second candidate feature sequences within each group are highly correlated.

[0052] In some embodiments, the Pearson correlation coefficient between the second candidate feature sequences is calculated as the correlation degree to construct the correlation matrix of the second candidate feature sequences.

[0053] In some embodiments, clustering algorithms include hierarchical clustering algorithms, spectral clustering algorithms, etc. It is understood that clustering algorithms capable of grouping second candidate feature sequences according to the correlation matrix are all within the protection scope of this application.

[0054] Subsequently, within each cluster obtained from clustering, the average relevance of the second candidate feature sequence within that cluster is calculated using the following formula: In the formula, This represents the average relevance of the group. The number of second candidate feature sequences in this group. and These are the indices of the second candidate feature sequences within each group. Weights are assigned to each group based on the calculated average relevance; the weights are calculated using the following formula: In the formula, The weight of this group, This is a smoothing factor to avoid division by zero. In high-weight groups, the redundancy between second-candidate feature sequences is low.

[0055] In some embodiments, the second candidate features are sorted according to their weights, and the top-ranked second candidate feature sequences are retained as target feature sequences that meet the correlation threshold.

[0056] In some embodiments, only the most representative second candidate feature sequence is retained in the lowest-weighted group. "Most representative" means that the second candidate feature sequence has the lowest correlation with other second candidate feature sequences in the group, and the other second candidate feature sequences in the group are removed as redundancy. In high-weighted groups, all second candidate feature sequences are retained to preserve features with strong complementarity.

[0057] In some embodiments, a larger number of second candidate feature sequences are retained in groups with high weights, while a smaller number of second candidate feature sequences are removed as redundancy.

[0058] In some embodiments, a second candidate feature sequence is retained within a group with moderate weight.

[0059] Step 105: Based on the weights related to the aging stage label and the temporal evolution gain, the target feature sequence is weighted to generate the first weighted feature vector sequence representing the aging-sensitive features.

[0060] The aging stage labels accurately reflect the aging stage of the component at the current time point. In some embodiments, the aging stage labels include the initial stage, slow degradation stage, accelerated aging stage, and pre-failure stage. In this step, the aging stage labels of the component at each time point are manually labeled according to the time sequence. The correlation and contribution of the target feature sequence with the aging stage labels are calculated as weights, and the target feature sequence is weighted according to these weights to generate a first weighted feature vector sequence representing aging-sensitive features. Simultaneously, the first weighted feature vector sequence has the advantages of high expressiveness, low redundancy, and a clear physical mechanism, improving the ability to detect early and abrupt aging stages of components.

[0061] refer to Figure 4 The flowchart illustrating the generation of the first weighted feature vector sequence in one embodiment is shown. Step 105 includes the following steps.

[0062] Step 401: Calculate the mutual information weights based on the target feature sequence and the aging stage label, and calculate the evolutionary gain weights based on the target feature sequence. In this step, for each target feature sequence, the mutual information between the target feature sequence and the corresponding aging stage label is calculated as the mutual information weight used in subsequent weighting. The formula for calculating the mutual information weights can be found in relevant technologies and will not be elaborated here. Simultaneously, the evolutionary gain weights are calculated based on the target feature sequence. The formula for calculating the evolutionary gain weights is: In the formula, For the first Evolutionary gain weights of each target feature sequence For the first The target feature sequence at time node eigenvalues, For the first The target feature sequence at time node The eigenvalues. On the one hand, mutual information weights can measure the amount of effective information the target feature sequence provides for distinguishing different aging stages; on the other hand, evolutionary gain weights can reflect the sensitivity of the target feature sequence to changes in different aging stages.

[0063] Step 402: Calculate the weighted target feature sequence based on the mutual information weight and evolutionary gain weight, thereby highlighting features with high importance. In this step, the weighted weight of the target feature sequence is calculated based on the mutual information weight and evolutionary gain weight calculated in step 401. The formula for calculating the weighted weight is: In the formula, For the first The weighted weights of the target feature sequences, For the first The mutual information weights between the weights of the target feature sequences and the aging stage labels. and These are the weighting coefficients for mutual information weights and evolutionary gain weights, respectively, and the weighting coefficients are adjusted according to feature requirements. In some embodiments, The value is 0.7. The value is set to 0.3. Based on the calculated weights, each target feature sequence is weighted to obtain a weighted target feature sequence.

[0064] Step 403: Construct the first weighted feature vector for each time node according to the time axis of the weighted target feature sequence, generating a first weighted feature vector sequence. The elements of the weighted target feature sequence are arranged in chronological order; the number of elements in the weighted target feature sequence represents the number of time nodes, and all time nodes form a continuous time axis. In this step, the time axis of the weighted target feature sequence is traversed, and the feature values ​​of each weighted target feature sequence within each time node are used as elements of the first weighted feature vector to construct the first weighted feature vector corresponding to each time node. Thus, the first weighted feature vectors corresponding to all time nodes are used to generate the first weighted feature vector sequence in chronological order.

[0065] Step 106: Input the aging stage label and the first weighted feature vector sequence into the recurrent neural network model to obtain the hidden state vector sequence.

[0066] Each first-weighted feature vector in the first-weighted feature vector sequence has a unique corresponding aging stage label. In this step, the aging stage label and the first-weighted feature vector sequence are input into a recursive neural network (RNN) model. The first-weighted feature vector sequence serves as the input data for the RNN model, and the aging stage label serves as the supervision signal for the output layer of the RNN model. The hidden layers of the RNN model can construct and output hidden state vectors that are coupled in both time and state dimensions. The hidden state vectors output at each time point form a hidden state vector sequence in chronological order. In this way, the hidden state vector sequence can track the gradual changes in the aging characteristics of components over time and the abrupt changes at turning points, thereby capturing the stage-wise changes in the state of components. The training process of the RNN model is described in detail here, referring to relevant technologies.

[0067] In some embodiments, the recurrent neural network is a bidirectional long short-term memory network (Bi-LSTM), which can simultaneously capture historical evolution information and future trends of aging characteristics.

[0068] In some embodiments, the recurrent neural network is a self-attention mechanism, which can capture long-range dependencies of aging features and highlight key change signals in the hidden state vector.

[0069] Step 107: Input the hidden state vector sequence into the dynamic regression network that integrates Bayesian inference to obtain the remaining lifetime probability distribution of the components.

[0070] In this step, the input to the dynamic regression network fused with Bayesian inference is a sequence of hidden state vectors, and the output is the probability distribution of remaining lifetime. This network combines physical interpretability and adaptive prediction capabilities, and can model the nonlinear and multi-stage abrupt changes in the aging process of components. Specifically, at each time point, the dynamic regression network fused with Bayesian inference is driven by the hidden state output from the previous time point and the hidden state vector input at the current time point. It initially predicts the probability distribution of remaining lifetime at the current time point, and then performs a Bayesian conditional posterior update on the initially predicted probability distribution of remaining lifetime using Bayesian inference, outputting the Bayesian conditionally updated probability distribution of remaining lifetime at the current time point. The architecture of the dynamic regression network fused with Bayesian inference is not the focus of this application; please refer to the relevant technical specifications for details.

[0071] In some embodiments, the remaining lifetime probability distribution is a Gaussian distribution, which can be represented as: In the formula, Time node The predicted remaining lifespan probability distribution Time node The expected value of the predicted remaining lifespan. Time node The variance of the predicted remaining lifespan.

[0072] In some embodiments, the confidence interval for the remaining life of a component is obtained based on the remaining life probability distribution. This is used for operation and maintenance early warning.

[0073] In some embodiments, the remaining lifetime probability distribution is a multimodal distribution, such as a Gaussian mixture distribution, which can predict multiple lifetime peaks.

[0074] In some embodiments, to improve the robustness of prediction, multiple structurally differentiated lifetime prediction models are used to predict the lifetime probability distribution, resulting in a fused lifetime confidence interval and a mixed probability distribution. Each model predicts its own lifetime prediction interval and probability distribution at its respective confidence level based on the hidden state vector sequence. An adaptive confidence weight allocation algorithm is used to adjust the confidence weights of each model, and the probability distributions and confidence intervals output by each model are weighted and fused according to these weights. Specifically, the upper bound of the confidence intervals output by each model is calculated based on the confidence weights to obtain the upper bound of the fused lifetime confidence interval; the lower bound of the confidence intervals output by each model is calculated based on the confidence weights to obtain the lower bound of the fused lifetime confidence interval; and the mixed probability distribution is calculated based on the confidence weights. In other words, the fused lifetime confidence interval is... In the formula, For the first Each model at the time node Confidence weights For the first The upper bound of the confidence interval output by the model at this time point. For the first The lower bound of the confidence interval output by each model at this time point. The fused mixture probability distribution is as follows: In the formula, For the first Each model at the time node The probability distribution of the output. In this way, the failure and overfitting problems of a single model in special stages such as local anomalies and missing information can be alleviated.

[0075] In some embodiments, the remaining lifespan prediction model includes a Bayesian neural network model, a dynamic factor regression model, and a probabilistic graphical model.

[0076] In some embodiments, the formula for calculating the confidence weight in the adaptive confidence weight allocation algorithm is as follows: In the formula, For the first The time points predicted by the model The relative change between the center of the confidence interval and the center of the confidence interval of the previous time point. To adjust the parameters.

[0077] In some embodiments, in the adaptive confidence weight allocation algorithm, the confidence weights are adjusted using the feature change rate of the feature vector generated in the previous step of obtaining the probability distribution from the remaining life expectancy prediction model. The larger the feature change rate of the remaining life expectancy prediction model, the higher its corresponding confidence weight adjustment. Here, the feature change rate is a quantitative indicator that measures the rate of change of the feature vector between two adjacent time points. Considering the different feature abstraction mechanisms of various remaining life expectancy prediction models, for Bayesian neural network models, their feature vectors are vectors composed of the mean outputs of the model's feature extraction layers; for dynamic factor regression models, their feature vectors are the state vectors of the dominant dynamic factors in each window; and for probabilistic graphical models, their feature vectors are vectors composed of the state probabilities of key nodes or the mean of marginal distributions. In this way, not only can each model specifically perceive drastic changes in health status, but the confidence weights of each model during the fusion process are also highly correlated with their dynamic performance, significantly improving the robustness and reliability of remaining life expectancy prediction.

[0078] For example, the formula for calculating the characteristic rate of change is: In the formula, For the first Each model at the time node The characteristic rate of change, For the first The time point of model generation eigenvectors, For the first The time points predicted by the model eigenvectors, This is a minimal constant to prevent the denominator from being zero. The characteristic change rates of each remaining lifetime prediction model are normalized to obtain the confidence weights of each model. The normalization formula is as follows: In the formula, For the first Each model at the time node Confidence weights For adjustable hyperparameters, This represents the number of remaining lifetime prediction models used.

[0079] In some embodiments, the abrupt trend of early abnormal aging of components is analyzed based on the fused remaining lifetime confidence interval, and an early warning is issued. Specifically, the optimal value is selected from the outputs of each remaining lifetime prediction model, and the fused remaining lifetime confidence interval is judged to be abnormal (e.g., the interval narrows sharply and the lower bound decreases rapidly) and to have a multimode peak structure. If anomalies and multimode peak structures are found, an anomaly warning signal is automatically triggered to predict component failure in advance and provide proactive operation and maintenance decisions.

[0080] In some embodiments, the criteria for anomaly detection include: In the formula, This represents the optimal value among the outputs of each remaining lifetime prediction model. This is the abnormal threshold parameter.

[0081] In some embodiments, the criteria for anomaly detection include the fact that the interval width at the current time node is more than half smaller than the interval width at the previous time node, wherein the interval width is the difference between the upper bound and the lower bound of the interval.

[0082] In some embodiments, the criteria for anomaly detection include the lower bound of the interval at the current time point shrinking by more than 30 days compared to the lower bound of the interval at the previous time point.

[0083] In some embodiments, the criterion for the multimode peak structure is the number of main peaks in the mixed probability distribution after detection and fusion. If the number of main peaks is greater than one, an abnormal warning signal is automatically triggered.

[0084] In some embodiments, the criterion for determining the multimode peak structure is that the main peak probability (also known as the main peak quality) is greater than or equal to a probability threshold, for example, a probability threshold of 0.15.

[0085] For reference Figure 5 A flowchart illustrating the component life prediction method in another embodiment is shown below. Figure 1 Unlike the embodiments shown, the component life prediction method also includes:

[0086] Step 507 involves generating a tensor sequence characterizing the aging evolution trend, a mutation index sequence characterizing aging evolution abrupt changes, and anomaly locations indicating inflection points in aging evolution based on the hidden state vector sequence. This step enables the detection and early warning of potential nonlinear mutations during component aging, thereby providing a health baseline for remaining life prediction in step 508.

[0087] For reference Figure 6 The flowchart of the modeling tensor sequence, mutation index sequence and anomaly location in one embodiment is shown. Step 507 includes the following steps.

[0088] Step 601: Extract state subsequences from the hidden state vector sequence using a first sliding window. On the timeline of the hidden state vector sequence obtained in step S506, use a window of length... The first sliding window, with a step size The hidden state vector sequence is traversed by sliding. For each first sliding window, a state subsequence is extracted. The extracted state subsequences are sorted according to the time order of the first sliding window. The length of the state subsequence is equal to the length of the first sliding window.

[0089] In some embodiments, the length of the first sliding window is greater than the step size of the first sliding window, so that the extracted state subsequences have overlap, in order to ensure the continuity of the features of key aging evolution stages.

[0090] Step 602: Calculate the correlation tensor for each state subsequence to obtain a set of correlation tensors. In this step, for each state subsequence of the first sliding window, the feature correlation of each dimension between the state vectors at each time point in the subsequence is calculated sequentially, and a correlation tensor is constructed based on the calculated feature correlation. Specifically, the first... The relevant tensor of the state subsequence of the first sliding window is: ,in, For the first The length of the state subsequence of the first sliding window For the first The dimension of the state vector in the state subsequence of the first sliding window. In the state subsequence of the first sliding window, the state within the first sliding window is... The first time point, the Dimensional features and the first The correlation between dimensional features is encoded as a correlation tensor in The element at position, i.e. ,in, For the first sliding window The state vector at time node i is the first dimensional features, For the first sliding window The state vector at time node i is the first Dimensional features. The relevant tensors are sorted according to the time sequence of the first sliding window to obtain a set of relevant tensors.

[0091] In some embodiments, the correlation calculation method is a higher-order quantization method such as correlation coefficient quantization or mutual information quantization.

[0092] Step 603: Decompose the relevant tensor to extract its principal components, obtaining a sequence of principal components. Within each first sliding window, the principal components of the relevant tensor from step 602 are extracted using a tensor decomposition algorithm. The order of the principal components is third, and the dimension of each order of the principal components does not exceed the dimension of each order of the relevant tensor, thus transforming the high-dimensional, complex nonlinear feature correlations into interpretable evolutionary factors and trends.

[0093] In some embodiments, tensor decomposition is performed using CP decomposition (CANDECAMP / PARAFAC). The third-order correlation tensor decomposition is a weighted sum of several third-order rank-1 tensors, and the result is obtained by selecting the tensor with the largest weight. We obtain the principal components of the relevant tensors by weighted combination of three rank-1 tensors.

[0094] In some embodiments, tensor decomposition is performed using Tucker decomposition (also known as higher-order singular value decomposition). The third-order correlation tensor decomposition is a multilinear product of the core tensor and the factor matrix, where the core tensor serves as the principal component of the extracted correlation tensor.

[0095] Step 604: Calculate the F-norm distance between the principal tensor components of adjacent first sliding windows in the relevant tensor principal component sequence to obtain the F-norm distance sequence. The F-norm distance sequence is a mutation index for aging evolution mutations, used to observe the mutation status of relevant tensor principal components between adjacent first sliding windows. Specifically, the formula for calculating the F-norm (Frobenius norm) distance between relevant tensor principal components between two adjacent first sliding windows is: In the formula, For the adjacent first The first sliding window and the first -1 F-norm distance between the principal components of the relevant tensors of the first sliding window For the first The principal components of the tensors related to the first sliding window. For the first The principal components of the tensors related to the first sliding window are obtained. The F-norm distances calculated for each first sliding window are sorted according to the time order of the first sliding window to obtain the F-norm distance sequence.

[0096] Step 605: Identify anomalous F-norm distances in the F-norm distance sequence and locate the anomalous window corresponding to the anomalous F-norm distance.

[0097] In some embodiments, when an anomaly occurs in the F-norm distance of a first sliding window in the F-norm distance sequence obtained in step 604 (e.g., an F-norm distance jumps, with the current F-norm distance being much higher than earlier F-norm distances), the first sliding window is the anomaly window corresponding to the anomaly F-norm distance. The timing of the first sliding window location corresponds to the inflection point of a sudden change in the component's state. Sudden changes in component state include transitioning from slow aging to accelerated degradation or from latent damage to overt damage. Thus, during the aging evolution of components, risk nodes can be continuously, dynamically, and with high sensitivity identified, thereby achieving early warning of inflection points in component performance degradation and potential failures.

[0098] In some embodiments, a planar graph is established with time nodes on the horizontal axis and F-norm distance on the vertical axis. Based on the F-norm distance sequence obtained in step 604, a smooth or abrupt change curve of the F-norm distance over time is plotted on the planar graph, and inflection points in the change curves are detected as anomaly windows corresponding to abnormal F-norm distances. This planar graph can detect potential nonlinear abrupt changes and risk inflection points of components in real time.

[0099] Step 606: Output the relevant tensor principal component sequence as the tensor sequence, the F-norm distance sequence as the mutation index sequence, and the anomaly window as the anomaly location. Specifically, output the relevant tensor principal component sequence obtained in step 603 as the tensor sequence characterizing the aging evolution trend, output the F-norm distance sequence obtained in step 604 as the mutation index sequence characterizing aging evolution mutations, and output the anomaly window located in step 605 as the anomaly location indicating the inflection point of aging evolution.

[0100] An embodiment of this application also proposes a method such as Figure 7 The electronic device 700 is shown. According to... Figure 7 The electronic device 700 may include an internal communication bus 701, a processor 702, a read-only memory (ROM) 703, a random access memory (RAM) 704, and a communication port 705. When used in a personal computer, the electronic device may also include a hard disk 706.

[0101] The internal communication bus 701 enables data communication between components of the electronic device 700. The processor 702 can perform judgments and issue prompts. In some embodiments, the processor 702 may consist of one or more processors. The communication port 705 enables data communication between the electronic device 700 and external devices. In some embodiments, the electronic device 700 can send and receive information and data from a network through the communication port 705.

[0102] The electronic device 700 may also include different forms of program storage units and data storage units, such as a hard disk 706, a read-only memory (ROM) 703, and a random access memory (RAM) 704, capable of storing various data files used for computer processing and / or communication, as well as possible program instructions executed by the processor 702. The processor executes these instructions to implement the main parts of the method as described above. The results of the processor processing are transmitted to the user equipment via a communication port and displayed on the user interface.

[0103] In another aspect, this application proposes a computer-readable medium storing computer program code that, when executed by a processor, implements the method described above.

[0104] In addition, this application also proposes a computer program product, including computer program code, which, when executed by one or more processors, can implement the steps in the method described above.

[0105] Some aspects of this application can be executed entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The aforementioned hardware or software may be referred to as a "data block," "module," "engine," "unit," "component," or "system." The processor may be one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DAPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, or combinations thereof. Furthermore, aspects of this application may manifest as computer products residing in one or more computer-readable media, including computer-readable program code. For example, computer-readable media may include, but are not limited to, magnetic storage devices (e.g., hard disks, floppy disks, magnetic tapes, etc.), optical discs (e.g., compressed CDs, digital multifunction DVDs, etc.), smart cards, and flash memory devices (e.g., cards, sticks, key drives, etc.).

[0106] A computer-readable medium may contain a propagated data signal containing computer program code, for example, on baseband or as part of a carrier wave. This propagated signal may take various forms, including electromagnetic, optical, and so on, or suitable combinations thereof. A computer-readable medium can be any computer-readable medium other than a computer-readable storage medium, which can be connected to an instruction execution system, apparatus, or device to enable communication, propagation, or transmission of a program for use. The program code located on the computer-readable medium can be propagated through any suitable medium, including radio, cable, fiber optic cable, radio frequency signals, or similar media, or any combination of the above media.

[0107] The basic concepts have been described above. Obviously, for those skilled in the art, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore remain within the spirit and scope of the exemplary embodiments of this application.

[0108] Furthermore, this application uses specific terms to describe embodiments of the application. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of the application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different locations in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of the application can be appropriately combined.

[0109] Similarly, it should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this disclosure method does not imply that the subject matter of the present application requires more features than those mentioned in the claims. In fact, the embodiments contain fewer features than all the features of the single embodiments disclosed above.

[0110] In some embodiments, numbers describing the quantity of components and attributes are used. It should be understood that such numbers used in the description of embodiments are modified in some examples with the terms "approximately," "approximately," or "generally." Unless otherwise stated, "approximately," "approximately," or "generally" indicates that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may be changed depending on the characteristics required by individual embodiments. In some embodiments, numerical parameters should take into account specified significant digits and employ a general method of digit reservation. Although the numerical ranges and parameters used to confirm their breadth of scope in some embodiments of this application are approximate values, in specific embodiments, such values ​​are set as precisely as feasible.

[0111] Although this application has been described with reference to specific embodiments, those skilled in the art should recognize that the above embodiments are only used to illustrate this application, and various equivalent changes or substitutions can be made without departing from the spirit of this application. Therefore, any changes or modifications to the above embodiments within the essential spirit of this application will fall within the scope of the claims of this application.

Claims

1. A component life prediction method based on multi-scale adaptive feature decomposition, characterized in that, Includes the following steps: Acquire various signals during the use of the components; The empirical mode decomposition method is used to decompose the various signals into multiple components, each of which is an intrinsic mode function. Each component is decomposed into multiple sub-components using a multi-scale sliding window method, wherein the sliding window step size of each component is related to the dominant frequency of each component; Generate a target feature sequence that satisfies the requirements of temporal consistency and correlation based on each component and sub-component; The target feature sequence is weighted according to the weights related to the aging stage label and the temporal evolution gain to generate a first weighted feature vector sequence representing aging-sensitive features; The aging stage label and the first weighted feature vector sequence are input into a recurrent neural network model to obtain a hidden state vector sequence. The hidden state vector sequence is input into a dynamic regression network that integrates Bayesian inference to obtain the remaining lifetime probability distribution of the component.

2. The method as described in claim 1, characterized in that, Generating a target feature sequence that satisfies the requirements of temporal consistency and correlation based on each component and sub-component includes: Extract multiple first candidate feature sequences from each of the components and sub-components; Select a second candidate feature sequence that meets the temporal consistency threshold from the first candidate feature sequence; The target feature sequence that meets the correlation threshold is selected from the second candidate feature sequence.

3. The method as described in claim 1, characterized in that, The target feature sequence is weighted according to weights related to aging stage labels and temporal evolution gains to generate a first weighted feature vector sequence representing aging-sensitive features, including: Based on the target feature sequence and the aging stage label, calculate the mutual information weight and the evolution gain weight based on the target feature sequence; Calculate the weighted target feature sequence based on the mutual information weight and the evolutionary gain weight; Based on the time axis of the weighted target feature sequence, construct the first weighted feature vector for each time point to generate the first weighted feature vector sequence.

4. The method as described in claim 1, characterized in that, Also includes: Based on the hidden state vector sequence, a tensor sequence representing the aging evolution trend, a mutation index sequence representing aging evolution mutations, and anomaly positions indicating aging evolution inflection points are generated.

5. The method as described in claim 4, characterized in that, Also includes: The generation of tensor sequences characterizing aging evolution trends, mutation index sequences characterizing aging evolution abrupt changes, and anomaly locations indicating aging evolution inflection points based on the hidden state vector sequence includes: Use a first sliding window to extract a state subsequence from the hidden state vector sequence; Calculate the relevant tensor for each of the state subsequences to obtain a set of relevant tensors; The relevant tensors are decomposed to extract their principal components, resulting in a sequence of relevant tensor principal components. Calculate the F-norm distance between the principal components of the related tensors of adjacent first sliding windows in the related tensor principal component sequence to obtain the F-norm distance sequence; Identify anomalous F-norm distances in the F-norm distance sequence and locate the anomalous window corresponding to the anomalous F-norm distance; The relevant tensor principal component sequence is output as the tensor sequence, the F-norm distance sequence is output as the mutation index sequence, and the anomaly window is output as the anomaly location.

6. The method as described in claim 5, characterized in that, In the step of extracting state subsequences from the hidden state vector sequence using a first sliding window, the length of the first sliding window is greater than the step size of the first sliding window.

7. The method as described in claim 1, characterized in that, The component is a non-metallic flexible hose used in nuclear power plants, and the various signals include temperature, pressure, vibration, and infrared spectrum.

8. An electronic device, comprising: Memory is used to store instructions that can be executed by the processor; as well as A processor for executing the instructions to implement the method as described in any one of claims 1-7.

9. A computer storage medium storing computer program code, said computer program code implementing the method as claimed in any one of claims 1-7 when executed by a processor.

10. A computer program product comprising computer program code, wherein when the computer program code is executed by one or more processors, the one or more processors implement the steps of the method as described in any one of claims 1-7.

Images