A method and system for predicting the lifetime of an electronic component

Abstract

The application relates to the field of electronic component reliability and intelligent operation and maintenance technology, and discloses a life prediction method and system for electronic components, the method comprising: acquiring historical data and real-time data; performing a preprocessing operation on the historical data to obtain preliminary features; performing an abnormality elimination and transformation operation on the real-time data according to the preliminary features to obtain a matching vector, performing a projection and aggregation operation to obtain unified features; performing a density calculation and weighting operation on the unified features to obtain a weighted sequence, and reconstructing the weighted sequence according to a distribution metric to obtain corrected data; performing a reorganization and smoothing operation on the corrected data to obtain a stable sequence; performing a mapping and regression operation on the stable sequence by using a preset memory network to obtain a life estimate; and performing an updating operation on the memory network when the life estimate is lower than a preset life threshold to obtain an optimized system. The method can realize robust alignment of features under complex working conditions and high-precision adaptive life prediction.

Description

Technical Field

[0001] This invention relates to the field of electronic component reliability and intelligent operation and maintenance technology, and in particular to a method and system for predicting the lifespan of electronic components. Background Technology

[0002] Currently, with the widespread application of high-reliability equipment in aerospace, new energy, and industrial automation, the safe operation of electronic components is directly related to the long-term stability of the entire system. As a core component of equipment failure prediction and health management systems, it is crucial to accurately predict the remaining lifespan of components to avoid significant losses caused by sudden failures.

[0003] In existing technologies, the lifespan prediction of electronic components mainly relies on a large amount of historical operating records collected under specific operating conditions. This typically involves constructing a static statistical degradation model or an end-to-end data-driven network to establish a direct mapping relationship from underlying physical monitoring signals (such as voltage, current, and temperature) to remaining lifespan values, thereby assessing the aging and degradation trends of the components. However, this prediction method based on static models and historical experience reveals significant limitations when facing complex and ever-changing real-world industrial scenarios. On the one hand, there are severe systematic differences in the distribution of data characteristics across different equipment or operating environments (such as operating noise and environmental physical drift). When new equipment is put into use and the accumulated new operating data is extremely limited, existing models often struggle to achieve effective alignment of feature spaces between heterogeneous data, leading to the failure of cross-scenario prediction logic. On the other hand, throughout the long lifespan of components (lasting several years), the characteristic data representing early micro-aging or edge failures are extremely sparse, while health data under steady-state operation is highly dense. Existing technologies lack adaptive weight allocation and calibration mechanisms for this extremely uneven data distribution, causing the model's ability to capture early failure signs to be overwhelmed by the massive amount of normal operating data. Furthermore, when the equipment undergoes physical maintenance, the existing model lacks the closed-loop evolutionary ability to dynamically feed back the new state to fine-tune the prediction system, which further exacerbates the disconnect between the prediction results and the actual physical state.

[0004] Therefore, existing technologies suffer from poor cross-scenario generalization ability and low prediction accuracy of component life prediction models under complex working conditions and with limited new data scenarios. Summary of the Invention

[0005] This invention provides a method and system for predicting the lifespan of electronic components, in order to solve the problems of poor cross-scenario generalization ability and low prediction accuracy of component lifespan prediction models in complex working conditions and scenarios with a small amount of new data in the prior art.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for predicting the lifetime of electronic components, comprising:

[0007] Historical data is acquired, and the historical data is preprocessed to obtain preliminary features;

[0008] Acquire real-time data, perform anomaly removal operation on the real-time data according to the preliminary features to obtain clean data, perform transformation operation on the clean data to obtain a matching vector, perform projection operation on the matching vector according to the preset affine matrix to obtain an alignment vector, and perform aggregation operation on the alignment vector and the preliminary features to obtain a unified feature;

[0009] The uniform features are density-calculated to obtain a deviation metric. The uniform features are then weighted according to the deviation metric to obtain a weighted sequence. The weighted sequence is then distributed to obtain a distribution metric. If the distribution metric is lower than a preset equilibrium threshold, a calibration coefficient is calculated. The weighted sequence is then reconstructed according to the calibration coefficient to obtain corrected data. If the distribution metric is not lower than the equilibrium threshold, the weighted sequence is used as the corrected data.

[0010] The corrected data is recombined to obtain a recombined sequence, and the recombined sequence is smoothed to obtain a stable sequence.

[0011] The stable sequence is mapped using a pre-defined memory network to obtain a hidden vector, and a regression operation is performed on the hidden vector to obtain a lifetime estimate.

[0012] If the estimated lifetime is lower than a preset lifetime threshold, then operational data is acquired, and correlation calculations are performed on the operational data and preset reference data to obtain relevant values. Adaptation weights are generated based on the relevant values, and the memory network is updated using the backpropagation algorithm based on the adaptation weights to obtain an optimized system.

[0013] Secondly, the present invention provides a lifespan prediction system for electronic components, comprising:

[0014] The data preprocessing module is used to acquire historical data, perform preprocessing operations on the historical data, and obtain preliminary features;

[0015] The feature alignment module is used to acquire real-time data, perform anomaly removal on the real-time data according to the preliminary features to obtain clean data, perform transformation operation on the clean data to obtain a matching vector, perform projection operation on the matching vector according to a preset affine matrix to obtain an alignment vector, and perform aggregation operation on the alignment vector and the preliminary features to obtain a unified feature.

[0016] The distribution correction module is used to calculate the density of the uniform feature to obtain a deviation metric, perform a weighting operation on the uniform feature based on the deviation metric to obtain a weighted sequence, perform distribution calculation on the weighted sequence to obtain a distribution metric, calculate a calibration coefficient if the distribution metric is lower than a preset equilibrium threshold, reconstruct the weighted sequence based on the calibration coefficient to obtain corrected data, and use the weighted sequence as corrected data if the distribution metric is not lower than the equilibrium threshold.

[0017] A time-series smoothing module is used to reassemble the corrected data to obtain a reassembled sequence, and then smooth the reassembled sequence to obtain a stable sequence.

[0018] The lifetime prediction module is used to map the stable sequence using a preset memory network to obtain a hidden vector, and to perform a regression operation on the hidden vector to obtain a lifetime estimate.

[0019] The closed-loop optimization module is used to acquire operating data if the estimated lifetime is lower than a preset lifetime threshold, perform correlation calculations on the operating data and preset reference data to obtain correlation values, generate adaptation weights based on the correlation values, and use the backpropagation algorithm to update the memory network based on the adaptation weights to obtain an optimized system.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] (1) This invention obtains preliminary features by acquiring historical data, performing denoising and preprocessing operations, and then uses a preset affine matrix to project the matching vector extracted from real-time data, thereby aggregating it with the preliminary features to obtain unified features. This scheme decouples environmental deviations from historical operation records, constructs a pure benchmark feature space, and then uses spatial projection mapping technology to force real-time heterogeneous data collected under new scenarios or new equipment to be aligned to this unified feature benchmark, eliminating the systematic coordinate system differences between different operating conditions or equipment. Thus, it effectively solves the problem of prediction model failure caused by inconsistent feature spaces under complex and variable operating environments, and significantly improves the model's cross-domain generalization ability and prediction accuracy in scenarios with a small amount of new data.

[0022] (2) This invention performs density calculations on uniform features, assigns high weighting factors to sparse regions and low weighting factors to dense regions for weighting operations, and then combines equilibrium threshold judgment to dynamically scale and reconstruct the weighted sequence to obtain corrected data, and extracts autocorrelation periodic recombination to form a stable sequence. This scheme addresses the objective phenomenon of redundant steady-state health data and extremely sparse early aging features throughout the entire life cycle of components, and constructs a density-based adaptive weight allocation and spatial distribution calibration mechanism, which can automatically amplify the edge features representing internal aging, prevent key degradation information from being submerged by massive normal data, and effectively filter out transient operating condition noise through periodic recombination. Thus, it overcomes the state assessment bias caused by extreme sample imbalance and greatly enhances the sensitivity and robustness of the prediction model in capturing early weak failure signs.

[0023] (3) This invention acquires operational data when the estimated lifespan is below a preset threshold, performs correlation calculations between the operational data and reference data to generate adaptive weights, and uses the backpropagation algorithm to update the memory network based on these adaptive weights to obtain an optimized system. This scheme breaks the limitation of traditional lifespan prediction models being static for life once trained. After physical maintenance actions are triggered, operational data of the new state is immediately collected to evaluate the performance recovery trajectory, and the network parameters are safely and directionally fine-tuned by calculating the correlation with historical experience, avoiding the catastrophic forgetting that is prone to occur when the neural network is updated online. Thus, the self-evolutionary closed loop between physical equipment maintenance and digital prediction models is completely opened up, enabling the lifespan prediction system to achieve adaptive iteration with the sudden changes in the hardware state of components, ensuring the continuity and long-term accuracy of the entire life cycle management. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of the lifespan prediction method for electronic components provided in the first embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the electronic component lifetime prediction system provided in the second embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Reference Figure 1 The first embodiment of the present invention provides a method for predicting the lifetime of electronic components, comprising the following steps:

[0028] S11, acquire historical data, perform preprocessing operations on the historical data to obtain preliminary features;

[0029] S12, acquire real-time data, perform anomaly removal operation on the real-time data according to the preliminary features to obtain clean data, perform transformation operation on the clean data to obtain a matching vector, perform projection operation on the matching vector according to the preset affine matrix to obtain an alignment vector, and perform aggregation operation on the alignment vector and the preliminary features to obtain a unified feature;

[0030] S13, perform density calculation on the uniform feature to obtain a deviation metric, perform a weighting operation on the uniform feature based on the deviation metric to obtain a weighted sequence, perform distribution calculation on the weighted sequence to obtain a distribution metric, if the distribution metric is lower than a preset equilibrium threshold, calculate a calibration coefficient, perform a reconstruction operation on the weighted sequence based on the calibration coefficient to obtain corrected data, if the distribution metric is not lower than the equilibrium threshold, then use the weighted sequence as corrected data;

[0031] S14, perform a recombination operation on the corrected data to obtain a recombined sequence, and perform a smoothing operation on the recombined sequence to obtain a stable sequence;

[0032] S15, the stable sequence is mapped using a preset memory network to obtain a hidden vector, and the hidden vector is regressed to obtain a lifetime estimate;

[0033] S16, if the estimated lifetime is lower than a preset lifetime threshold, then obtain the operating data, perform correlation calculations on the operating data and preset reference data to obtain the correlation values, generate the adaptation weights based on the correlation values, and use the backpropagation algorithm to update the memory network based on the adaptation weights to obtain the optimized system.

[0034] In step S11, historical data is acquired, and the historical data is preprocessed to obtain preliminary features.

[0035] The historical data is preprocessed to obtain preliminary features, including:

[0036] The historical data is decomposed using a wavelet transform algorithm to obtain denoised data.

[0037] The denoised data is compensated according to a preset environmental reference value to obtain a correction sequence;

[0038] The covariance of the corrected sequence is calculated to obtain the covariance matrix.

[0039] The covariance matrix is ​​subjected to eigenvalue decomposition to obtain preliminary features.

[0040] In one implementation, this embodiment acquires physical signal records of multiple identical electronic components during their full lifecycle aging test at the factory using sensors as historical data. This historical data, along with subsequently acquired real-time data, includes, but is not limited to, voltage timing signals, current timing signals, and surface temperature data of the components during operation. This embodiment employs a multi-resolution analysis method, selecting the Daubechies wavelet as the mother wavelet function, and calculating and determining the decomposition level to be three levels based on the signal information entropy minimization criterion. This embodiment uses the mother wavelet function to perform a three-level decomposition operation on the historical data, separating high-frequency detail components and low-frequency approximation components. The values ​​of the detail components are set to 0, and the low-frequency approximation components are used for inverse wavelet transform reconstruction to obtain denoised data.

[0041] It should be noted that the aforementioned preset environmental reference value is an objective reference parameter determined by consulting the nominal operating environment parameters marked in the datasheets provided by the component manufacturers. For example, the preset environmental temperature reference value is set to 25 degrees Celsius. In this embodiment, the actual measured temperature value corresponding to the denoised data is extracted, and the temperature difference between the actual measured temperature value and the environmental temperature reference value is calculated. In this embodiment, the inherent temperature drift coefficient of the component is obtained by reading the linear regression slope of temperature and output characteristics in the factory performance calibration curve. The temperature difference is multiplied by the temperature drift coefficient to obtain the deviation. Then, the original physical value of the denoised data is subtracted from the deviation, and a reverse compensation operation is performed to obtain a correction sequence that eliminates the influence of environmental temperature deviation.

[0042] For example, in this embodiment, the correction sequence is constructed into a multi-dimensional feature matrix according to different physical dimensions. The expected value of the data vectors in each dimension of the multi-dimensional feature matrix is ​​calculated, and the corresponding expected value is subtracted from each dimension data vector to obtain a centered matrix. The transpose of the centered matrix is ​​multiplied by the centered matrix itself, and the result is divided by the difference between the total number of samples and 1. Covariance calculation is then performed to obtain a covariance matrix reflecting the linear correlation of each physical quantity.

[0043] It is worth noting that in this embodiment, singular value decomposition is performed on the covariance matrix to obtain all its eigenvalues ​​and corresponding orthogonal eigenvectors. All eigenvalues ​​are sorted in descending order of value, and the proportion of each eigenvalue to the sum of all eigenvalues ​​is calculated to obtain the variance contribution rate. This embodiment calculates the cumulative variance contribution rate of the first N eigenvalues. When the cumulative variance contribution rate first reaches a threshold of 85%, the first N eigenvectors corresponding to this threshold are selected and concatenated to form a projection transformation matrix. The corrected sequence is multiplied by the projection transformation matrix to project the original high-dimensional data to a low-dimensional principal component space, obtaining preliminary features. The cumulative variance contribution rate threshold is determined using the scree plot method in principal component analysis. By calculating the eigenvalues ​​of each principal component, the principal components before the inflection point (elbow point) of the eigenvalue change curve are selected, and their cumulative contribution rate is used as the threshold. Those skilled in the art can adjust the sensitivity of this inflection point according to the data characteristics to determine the specific value.

[0044] In step S12, real-time data is acquired, and anomaly removal is performed on the real-time data according to the preliminary features to obtain clean data. The clean data is then transformed to obtain a matching vector. The matching vector is projected onto the matching vector according to a preset affine matrix to obtain an alignment vector. Finally, the alignment vector and the preliminary features are aggregated to obtain a unified feature.

[0045] Specifically, based on the preliminary features, an anomaly removal operation is performed on the real-time data to obtain clean data; then, a transformation operation is performed on the clean data to obtain a matching vector, including:

[0046] Based on the preliminary characteristics, a physical threshold inversion operation is performed to obtain the physical fluctuation range;

[0047] Data that exceeds the physical fluctuation range in the real-time data is removed as outliers, and the remaining real-time data is considered clean data.

[0048] Perform time-series slicing on the clean data to obtain waveform segments;

[0049] The waveform segment is subjected to parameter extraction to obtain a matching vector.

[0050] In one implementation, this embodiment acquires real-time component operating voltage and current timing signals as real-time data. This embodiment calculates the mean and standard deviation of the preliminary features (i.e., principal component scores in low-dimensional space), and determines the confidence boundary ellipsoid extreme points within the principal component space based on the 3x standard deviation statistical principle. The transpose of the projection transformation matrix generated in the previous step is obtained, and matrix multiplication is performed between the coordinates of the confidence boundary extreme points within the principal component space and the transpose matrix to reverse-map the boundary extreme points back to the original physical measurement dimension space. This embodiment extracts the maximum and minimum values ​​on each dimension coordinate after the reverse mapping, constructs an orthogonal physical fluctuation interval encompassing the reverse-mapped space, and obtains the upper and lower limits of each physical quantity.

[0051] It should be noted that this embodiment judges the value of the real-time data point by point. Data points in the real-time data whose values ​​are greater than the upper limit value or less than the lower limit value are regarded as outliers caused by external transient surges and are removed. The remaining continuous real-time data that is not removed is regarded as clean data.

[0052] For example, in this embodiment, the pulse width modulation setting parameters of the component controller are read to obtain the current operating switching frequency. The reciprocal of the operating switching frequency is calculated to obtain the standard time window length. In this embodiment, the time axis is divided into equal intervals according to the standard time window length, and the continuous clean data is truncated using the divided time intervals to obtain multiple independent waveform segments.

[0053] It is worth noting that, for each waveform segment, this embodiment extracts the maximum value in the waveform segment as the peak parameter; sums the values ​​of all sampling points within the waveform segment, and multiplies the sum by the time interval between adjacent sampling points to obtain the area parameter; finds the maximum value point in the waveform segment and its preceding zero-crossing point, calculates the quotient of the difference between these two points divided by the time difference, and obtains the rising edge slope parameter. The extracted peak parameter, area parameter, and rising edge slope parameter are concatenated and combined according to a fixed dimensional order to obtain the matching vector.

[0054] Specifically, the matching vector is projected onto a preset affine matrix to obtain an alignment vector. The alignment vector is then aggregated with the preliminary features to obtain unified features, including:

[0055] The matching vector is projected onto a preset low-dimensional space using the affine matrix to obtain an alignment vector;

[0056] The fusion weights are generated according to the preset time decay rules;

[0057] The alignment vector and the preliminary features are weighted and aggregated using the sliding window algorithm based on the fusion weights to obtain unified features.

[0058] In one implementation, this embodiment retrieves a preset affine matrix. In offline mode, a benchmark matching vector set generated by the new component under standard test conditions is acquired, and a historical preliminary feature set generated under the same standard conditions is obtained from a historical reference database. The cross-covariance matrix of the two sets is calculated, and singular value decomposition is performed on the cross-covariance matrix. The resulting right singular matrix is ​​multiplied by the transpose of the left singular matrix to obtain an orthogonal transformation matrix that minimizes the Frobenius norm distance between the two datasets, which is used as the preset affine matrix. In this embodiment, the matching vector extracted online is multiplied by the preset affine matrix to obtain an alignment vector projected into a preset low-dimensional space.

[0059] It should be noted that this embodiment uses an exponential decay mathematical model as the preset time decay rule. The decay coefficient involved in the rule is determined by offline hyperparameter optimization using a grid search method combined with historical data, selecting the parameter that minimizes the overall prediction root mean square error. This embodiment calculates the absolute time difference between the current acquisition timestamp of the alignment vector and the historical acquisition timestamp of the preliminary feature. Using the natural constant as the base, the negative of the product of the absolute time difference and the decay coefficient is used as the exponent for power operation to calculate the fusion weight.

[0060] For example, this embodiment sets a sliding window. The length of the sliding window is a fixed length determined by evaluating the feature fusion classification accuracy at different time steps offline, taking the minimum number of time steps when the accuracy convergence rate is less than 1%. For the alignment vector and its corresponding preliminary feature at each time step within the window, the preliminary feature is multiplied by a first weight (value 1 minus the fusion weight), the alignment vector is multiplied by a second weight (the fusion weight), and the products of the two are added together. A weighted aggregation operation is then performed to obtain a unified feature.

[0061] In step S13, the density of the uniform feature is calculated to obtain a deviation metric. The uniform feature is weighted according to the deviation metric to obtain a weighted sequence. The distribution of the weighted sequence is calculated to obtain a distribution metric. If the distribution metric is lower than a preset equilibrium threshold, a calibration coefficient is calculated. The weighted sequence is reconstructed according to the calibration coefficient to obtain corrected data. If the distribution metric is not lower than the equilibrium threshold, the weighted sequence is used as corrected data.

[0062] Specifically, the uniform features are density-calculated to obtain a deviation metric, and the uniform features are weighted according to the deviation metric to obtain a weighted sequence, including:

[0063] Perform a neighbor count operation on the unified feature to obtain the local density value;

[0064] Based on the local density values, the unified feature is spatially divided to obtain sparse and dense regions.

[0065] Distance calculation is performed between the feature points of the sparse region and the preset pattern center to obtain the deviation metric;

[0066] Based on the deviation metric, high-weight factors and low-weight factors are generated, and the high-weight factors and low-weight factors are combined to obtain the initial weights.

[0067] The uniform features are multiplied element-wise using the initial weights to obtain a weighted sequence.

[0068] In one implementation, this embodiment retrieves a preset search radius parameter. This parameter is determined by performing k-distance graph analysis on the historical full feature dataset and extracting the value corresponding to the point where the distance curve shows the most significant exponential growth inflection point. This embodiment calculates the Euclidean distance between each feature point in the unified feature and all other feature points, counts the number of feature points whose Euclidean distance is less than the search radius parameter, and obtains the local density value.

[0069] It should be noted that in this embodiment, all local density values ​​are sorted in ascending order, and the local density value corresponding to the 5th percentile is selected as the dividing boundary. This percentile is determined by statistically analyzing the objective proportion of historical failure samples to the total lifecycle samples. Specifically, the percentage of sample points clearly identified as early failure signs in the total lifecycle data of the same type of component is used as the dividing boundary. In this embodiment, feature points with local density values ​​greater than the dividing boundary are classified as dense regions, and feature points with local density values ​​not greater than the dividing boundary are classified as sparse regions. In this embodiment, the arithmetic mean of the coordinates of all feature points in the dense region is calculated to generate the preset pattern center.

[0070] For example, for each feature point in the sparse region, this embodiment calculates the Euclidean distance between its spatial coordinates and the preset pattern center coordinates as a deviation metric. This embodiment multiplies the deviation metric by a value of 1 as a scaling factor, and adds a base value of 1 to generate a high-weight factor. For feature points in the dense region, this embodiment uniformly generates a low-weight factor equal to 1. This embodiment combines the high-weight factors and low-weight factors corresponding to all feature points according to their sampling indices to obtain an initial weight matrix. This embodiment performs element-wise multiplication of the initial weight matrix with the uniform feature to obtain a weighted sequence.

[0071] It should be noted that, in this embodiment, for feature points in the dense region, no distance calculation is required. The deviation metric is uniformly set to zero, and a low weight factor with a value equal to the preset benchmark value is generated accordingly. For feature points in the sparse region, a high weight factor greater than the preset benchmark value is generated based on the deviation metric. The high weight factor and the low weight factor are combined according to the sampling index of the feature points to obtain the initial weight covering all feature points.

[0072] The process involves calculating the distribution of the weighted sequence to obtain a distribution metric. If the distribution metric is lower than a preset equilibrium threshold, a calibration coefficient is calculated. The weighted sequence is then reconstructed based on the calibration coefficient to obtain corrected data. If the distribution metric is not lower than the equilibrium threshold, the weighted sequence is used as the corrected data. This includes:

[0073] The weighted sequence is subjected to a balance calculation using a spatial evaluation algorithm to obtain a distribution metric.

[0074] If the distribution metric is lower than the equilibrium threshold, an optimization algorithm is used to perform a dimension scaling calculation to obtain a calibration coefficient. The calibration coefficient is then used to dynamically scale the weighted sequence to obtain corrected data.

[0075] If the distribution metric is not lower than the equilibrium threshold, then the weighted sequence is used as correction data.

[0076] In one implementation, this embodiment employs a Gaussian kernel density estimation algorithm as the spatial evaluation algorithm. This embodiment uses a Gaussian kernel function to fit the discrete weighted sequence with a continuous probability density, obtaining a differentiable probability density function. Based on the fitted probability density function, this embodiment performs numerical integration on the entire feature space to calculate the spatial information entropy, and uses the calculated information entropy as a distribution metric.

[0077] It is worth noting that the preset equilibrium threshold is determined by constructing a completely uniformly distributed ideal dataset offline and calculating its ideal continuous information entropy, taking 80% of this ideal information entropy value. Specifically, a large number of random points are uniformly generated in the feature space through Monte Carlo simulation, and their information entropy is calculated as the ideal uniform information entropy. Considering the distribution non-uniformity allowed in actual engineering, 80% of this ideal value is taken as the threshold for determining whether calibration is needed. This ratio can be optimized by cross-validation, using model prediction accuracy as the indicator. In this embodiment, the distribution metric is compared with the equilibrium threshold. If the distribution metric is lower than the equilibrium threshold, this embodiment executes an optimization algorithm based on gradient ascent. Under the constraint that the total volume of the feature space remains unchanged (i.e., the product of the scaling factors of each dimension is equal to a constant 1), with maximizing the spatial information entropy as the objective function, the initial value of each scaling factor is uniformly set to 1. The partial derivatives of the objective function with respect to each scaling factor are calculated, and iterative updates are performed along the gradient ascent direction until the derivative approaches 0, obtaining the calibration coefficients corresponding to each feature dimension. In this embodiment, the calibration coefficient is multiplied by the corresponding dimension value of the weighted sequence, and a dynamic scaling operation is performed to obtain the corrected data. If the distribution metric is not lower than the equilibrium threshold, this embodiment does not perform the dynamic scaling operation and directly outputs the weighted sequence as the corrected data downstream.

[0078] In step S14, the corrected data is recombined to obtain a recombined sequence, and the recombined sequence is smoothed to obtain a stable sequence, including:

[0079] The variance of the corrected data is calculated using a sliding window algorithm to obtain the fluctuation range;

[0080] Autocorrelation analysis was performed on the fluctuation amplitude to obtain the period length;

[0081] The correction data is phase-aligned according to the period length to obtain the recombined sequence;

[0082] The boundary of the recombinant sequence is smoothed using a moving average algorithm to obtain a stable sequence.

[0083] In one implementation, this embodiment sets a sliding window of a preset length, which is the product of the reciprocal of the rated operating frequency of the component controller and the sampling frequency. The sliding window is shifted along the time axis in steps of a single sampling point. The arithmetic mean of all data within the window is calculated. Then, the square of the difference between each data point value and the arithmetic mean is calculated. The squared values ​​in each dimension are added together to obtain the L2 norm. All L2 norm values ​​are summed and divided by the total number of data points. A variance calculation operation is performed to output a one-dimensional continuously evolving sequence of fluctuation amplitude values.

[0084] It should be noted that in this embodiment, the inner product of the fluctuation amplitude numerical sequence and itself is calculated by shifting them asynchronously with long delays on the time axis to generate an autocorrelation coefficient curve. This embodiment finds the first maximum point in the autocorrelation coefficient curve other than the zero-delay point, converts the delay steps corresponding to this maximum point into a specific time length, and determines it as the period length. This embodiment performs a time-axis segmentation operation on the correction data according to the period length, extracts the voltage zero-crossing point in each data segment as a reference phase, aligns all data segments relative to the reference phase, and arranges the aligned one-dimensional data segments in chronological order as row vectors, concatenating them to construct a two-dimensional matrix to obtain the recombined sequence.

[0085] It is worth noting that in this embodiment, each row of the two-dimensional matrix is ​​taken as an independent time step input sequence, the number of rows is the number of time steps, and the values ​​of each column in the row are the multi-dimensional feature vectors at that time step; after boundary smoothing, the stable sequence maintains the same three-dimensional tensor structure and is directly used as the time step input of the memory network.

[0086] For example, this embodiment explicitly quantifies the boundary region to address the data discontinuity at the segment splicing boundary of the recombined sequence. A boundary region is formed by taking sampling points equal to 10% of the period length in both directions, centered on the splicing point. Within this boundary region, an exponentially weighted moving average algorithm is introduced for calculation. The current sampling point value is multiplied by a 20% smoothing factor, and then the smoothed value of the previous sampling point is multiplied by 80% to obtain a smoothed value. Subsequently, this embodiment uses a linear interpolation algorithm to perform a linear numerical transition calculation between the endpoints of the smoothed boundary region and the adjacent endpoints of the original unsmoothed sequence, performing a smooth transition operation to eliminate quadratic jumps and obtain a stable sequence.

[0087] In step S15, a preset memory network is used to map the stable sequence to obtain a hidden vector, and a regression operation is performed on the hidden vector to obtain a lifetime estimate, including:

[0088] The memory network's forgetting gate mechanism is used to filter preset cell states to obtain key historical information;

[0089] The stable sequence is superimposed onto the key historical information using the input gate mechanism of the memory network to obtain the updated state;

[0090] The updated state is filtered using the output gate mechanism of the memory network to obtain the hidden vector;

[0091] The hidden vector is input into a preset regression layer and subjected to a dot product operation to obtain the lifetime estimate.

[0092] In one implementation, the preset memory network in this embodiment adopts a Long Short-Term Memory (LSTM) network model. The network parameters of this model are determined through offline iterative optimization and pre-training. Specifically, the construction and training process involves collecting time-series data of the entire lifecycle of components from initial use to physical failure as input samples. The total running time is extracted from the full lifecycle data, and the time difference from the current moment to the final failure moment is used as the basic label. In the early health stage, the time difference is truncated to an upper limit of the nominal lifespan value as the final supervision label to prevent excessive penalties from small early fluctuations. The time step of the network input layer is dynamically set according to the period length obtained from the aforementioned autocorrelation analysis. The hidden layer dimension is preset to 64 dimensions, which is objectively determined by comparing the convergence speed of the mean square error of the validation set under different dimensions. An adaptive moment estimation optimization algorithm is used to minimize the mean square error loss function and update the network parameters. This embodiment obtains the pre-trained memory network and extracts the preset cell states initialized to zero vectors.

[0093] It should be noted that this embodiment utilizes the forgetting gate mechanism of the memory network, concatenating the hidden vector from the previous time step with the current stable sequence, multiplying it by the forgetting gate weight matrix and adding a bias, and then calculating the forgetting vector with each dimension between 0 and 1 using the logistic activation function. The forgetting vector is then element-wise multiplied with the preset cell state, discarding dimensional features with values ​​close to 0 to obtain key historical information.

[0094] For example, this embodiment utilizes the input gate mechanism of the memory network to calculate the input weight vector and candidate state vector by applying the logistic activation function and the hyperbolic tangent activation function to the concatenated sequence, respectively. Element-wise multiplication of the two vectors is performed to obtain incremental information, and element-wise addition of the incremental information and the key historical information is performed to obtain the updated state.

[0095] It is worth noting that this embodiment utilizes the output gate mechanism of the memory network to calculate the output mask vector from the concatenated sequence using the logistic activation function. The updated state is then processed using the hyperbolic tangent function and multiplied element-wise with the output mask vector to obtain the hidden vector. In this embodiment, the hidden vector is input into a preset regression layer with fixed weights. A dot product operation is then performed between the hidden vector and the weight matrix, and a bias term is added to obtain a specific scalar value, which serves as the lifetime estimate.

[0096] In step S16, if the estimated lifetime is lower than a preset lifetime threshold, then operational data is acquired, correlation calculations are performed between the operational data and preset reference data to obtain correlation values, adaptation weights are generated based on the correlation values, and the memory network is updated using the backpropagation algorithm based on the adaptation weights to obtain an optimized system, including:

[0097] The similarity between the running data and the preset reference data is calculated using the Pearson algorithm to obtain relevant numerical values;

[0098] If the relevant value is greater than a preset reference threshold, then an adaptive weight with a value greater than a preset benchmark is generated.

[0099] If the relevant value is not greater than the reference threshold, then an adaptive weight with a value equal to the preset benchmark is generated;

[0100] A loss function is constructed using the adaptive weights, and the gradient of the loss function is calculated to obtain the gradient value.

[0101] The connection weights of the memory network are updated using the backpropagation algorithm based on the gradient values ​​to obtain the optimized system.

[0102] In one implementation, this embodiment retrieves a preset lifetime threshold. This threshold is objectively determined by performing Weibull distribution statistical analysis on historical accelerated aging test data to extract the lower limit of remaining lifetime corresponding to a component failure probability of 5%. This embodiment compares the currently output lifetime estimate with the lifetime threshold. If the comparison result shows that the lifetime estimate is lower than the preset lifetime threshold, physical maintenance is performed. After physical maintenance is completed, this embodiment acquires continuous real-time parameters, using the cycle length determined by prior autocorrelation analysis as a unit, and extracts 50 cycles of real-time parameters as operating data.

[0103] It should be noted that this embodiment retrieves preset reference data stored in a pre-built historical reference database. This database is pre-built by collecting performance recovery timing waveforms of historical components of the same model after performing standard maintenance actions and extracting features. This embodiment uses the Pearson algorithm to calculate the covariance between the running data vector and the reference data vector as the numerator, calculate the product of the standard deviations of the two data vectors as the denominator, and perform a division operation to obtain a correlation value between -1 and 1.

[0104] For example, this embodiment sets a preset reference threshold, which is objectively determined by analyzing the receiver operating characteristic curves of the test set and selecting the correlation critical value corresponding to the maximization of the Youden index, for example, 0.8; the preset benchmark is set to a constant 1. This embodiment determines whether the correlation value is greater than the reference threshold. If the correlation value is greater than 0.8, it indicates that the maintenance recovery trajectory is highly consistent with the historical benchmark. In this embodiment, the correlation value is added to the value 1 to generate an adaptation weight with a value greater than the preset benchmark 1; if the correlation value is not greater than 0.8, this embodiment directly generates an adaptation weight with a value equal to the preset benchmark 1.

[0105] It is worth noting that, since the actual remaining lifetime after online maintenance cannot be obtained immediately, this embodiment retrieves the standard lifetime extension value of the historical maintenance record corresponding to the relevant value in the historical reference database as the pseudo-label observation value. Specifically, in the historical reference database, multiple sets of post-maintenance operating data and their corresponding, actually verified effective lifetime extension values ​​are pre-stored. After calculating the correlation values ​​between the current operating data and the historical data online, the set of historical data with the highest correlation to the current operating data is selected from the database, and its corresponding effective lifetime extension value is used as the pseudo-label observation value for this model update, ensuring that the supervision signal used for model fine-tuning originates from the most similar historical recovery pattern. The squared difference between the current network prediction value and the pseudo-label observation value is collected to obtain the basic mean square error. The basic mean square error is multiplied by the adaptation weight to construct the loss function. In this embodiment, the chain rule of calculus is applied to calculate the partial derivative of the loss function with respect to the internal connection weights of the memory network, and gradient calculation is performed to obtain the gradient value matrix. Subsequently, this embodiment utilizes the backpropagation algorithm, combined with a learning rate step size of 0.001 pre-set through the grid search method, to update the connection weights of the memory network by subtracting the product of the learning rate step size and the gradient value matrix from the current connection weights, thereby obtaining the optimized system.

[0106] It should be noted that the correspondence with the relevant value refers to retrieving the historical maintenance record with the smallest absolute value of the difference with the current relevant value from the historical reference database. If there are multiple records with the same difference, the arithmetic mean of the standard life extension values ​​of these records is taken. The standard life extension value is a robust estimate obtained by calculating the median of the remaining life of the same type of component after performing standard maintenance actions, based on the amount of extension of its remaining life relative to the predicted life before maintenance.

[0107] In summary, this invention constructs a pure baseline feature space through wavelet denoising and environmental bias compensation, and uses affine projection technology to force coordinate system alignment of limited real-time heterogeneous data in new scenarios. Based on this, an adaptive weighting and spatial distribution calibration mechanism based on local density accurately amplifies extremely sparse early microscopic aging features. Furthermore, a stable sequence smoothed by multi-cycle recombination is fed into a memory network for accurate regression prediction. After physical maintenance of the equipment, the network connection weights are safely and dynamically fine-tuned based on the correlation between the actual performance recovery trajectory and historical experience. This achieves robust feature mapping of heterogeneous data across scenarios under complex and variable operating conditions and effective correction of extreme sample imbalances. It completely establishes an adaptive evolutionary closed loop between sudden changes in physical equipment state and digital prediction models, ultimately achieving high generalization, high early warning accuracy, and long-term reliable assessment of the entire lifecycle for electronic component life prediction in scenarios with limited new data.

[0108] Reference Figure 2 The second embodiment of the present invention provides a lifespan prediction system for electronic components, comprising:

[0109] The data preprocessing module is used to acquire historical data, perform preprocessing operations on the historical data, and obtain preliminary features;

[0110] The feature alignment module is used to acquire real-time data, perform anomaly removal on the real-time data according to the preliminary features to obtain clean data, perform transformation operation on the clean data to obtain a matching vector, perform projection operation on the matching vector according to a preset affine matrix to obtain an alignment vector, and perform aggregation operation on the alignment vector and the preliminary features to obtain a unified feature.

[0111] The distribution correction module is used to calculate the density of the uniform feature to obtain a deviation metric, perform a weighting operation on the uniform feature based on the deviation metric to obtain a weighted sequence, perform distribution calculation on the weighted sequence to obtain a distribution metric, calculate a calibration coefficient if the distribution metric is lower than a preset equilibrium threshold, reconstruct the weighted sequence based on the calibration coefficient to obtain corrected data, and use the weighted sequence as corrected data if the distribution metric is not lower than the equilibrium threshold.

[0112] A time-series smoothing module is used to reassemble the corrected data to obtain a reassembled sequence, and then smooth the reassembled sequence to obtain a stable sequence.

[0113] The lifetime prediction module is used to map the stable sequence using a preset memory network to obtain a hidden vector, and to perform a regression operation on the hidden vector to obtain a lifetime estimate.

[0114] The closed-loop optimization module is used to acquire operating data if the estimated lifetime is lower than a preset lifetime threshold, perform correlation calculations on the operating data and preset reference data to obtain correlation values, generate adaptation weights based on the correlation values, and use the backpropagation algorithm to update the memory network based on the adaptation weights to obtain an optimized system.

[0115] It should be noted that the electronic component lifetime prediction system provided in this embodiment of the invention is used to execute all the process steps of the electronic component lifetime prediction method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0116] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0117] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting the lifespan of electronic components, characterized in that, include: Historical data is acquired, and the historical data is preprocessed to obtain preliminary features; The process involves acquiring real-time data, performing anomaly removal on the real-time data based on the preliminary features to obtain clean data, and then transforming the clean data to obtain a matching vector. This includes: performing time-series slicing on the clean data to obtain waveform segments; extracting the maximum value from the waveform segments as a peak parameter; summing the values ​​of all sampling points within the waveform segment and multiplying the sum by the time interval between adjacent sampling points to obtain an area parameter; finding the maximum value point in the waveform segment and its preceding zero-crossing point, calculating the quotient of the difference between these two points divided by the time difference to obtain a rising edge slope parameter; concatenating the extracted peak parameter, area parameter, and rising edge slope parameter according to a fixed dimensional order to obtain a matching vector; and projecting the matching vector onto a preset affine matrix to obtain an alignment vector. The preset affine matrix is ​​calculated by taking the cross-covariance matrix of the benchmark matching vector set generated by the new component under standard test conditions and the historical preliminary feature set generated under the same standard conditions in the historical reference database, and then performing singular value analysis on the cross-covariance matrix. The decomposition operation involves multiplying the obtained right singular matrix by the transpose of the left singular matrix. The alignment vector and the preliminary features are then aggregated to obtain unified features. This includes: generating fusion weights according to a preset time decay rule, where the preset time decay rule is an exponential decay mathematical model, and the decay coefficient involved in the rule is determined by offline hyperparameter optimization using a grid search method combined with historical data, selecting the parameter that minimizes the overall prediction root mean square error; calculating the absolute time difference between the current acquisition timestamp of the alignment vector and the historical acquisition timestamp of the preliminary features; using the natural constant as the base, exponentiating the product of the absolute time difference and the decay coefficient to obtain the fusion weights; and using a sliding window algorithm to perform a weighted aggregation operation on the alignment vector and the preliminary features based on the fusion weights: multiplying the preliminary features by a first weight and the alignment vector by a second weight, then adding them together to perform the weighted aggregation operation to obtain unified features; wherein the first weight is the value 1 minus the fusion weight, and the second weight is the fusion weight itself. Density calculation is performed on the uniform features to obtain a deviation metric. The uniform features are then weighted according to the deviation metric to obtain a weighted sequence. Distribution calculation is performed on the weighted sequence to obtain a distribution metric. If the distribution metric is lower than a preset equilibrium threshold, a calibration coefficient is calculated. The weighted sequence is then reconstructed according to the calibration coefficient to obtain corrected data. The preset equilibrium threshold is determined by constructing a perfectly uniformly distributed ideal dataset offline and calculating its ideal continuous information entropy, taking 80% of the ideal continuous information entropy value. If the distribution metric is not lower than the equilibrium threshold, the weighted sequence is used as corrected data. The process involves recombining the corrected data to obtain a recombined sequence, and then smoothing the recombined sequence to obtain a stable sequence. This includes: calculating the variance of the corrected data using a sliding window algorithm to obtain the fluctuation amplitude; performing autocorrelation analysis on the fluctuation amplitude to obtain the period length; performing phase alignment on the corrected data based on the period length to obtain the recombined sequence; and smoothing the boundaries of the recombined sequence using a moving average algorithm to obtain a stable sequence. The stable sequence is mapped using a pre-defined memory network to obtain a hidden vector, and a regression operation is performed on the hidden vector to obtain a lifetime estimate. If the estimated lifetime is lower than a preset lifetime threshold, then operational data is acquired, correlation calculations are performed on the operational data and preset reference data to obtain correlation values, adaptation weights are generated based on the correlation values, and the memory network is updated using the backpropagation algorithm based on the adaptation weights to obtain an updated memory network. This includes: using the Pearson algorithm to perform similarity calculations on the operational data and the preset reference data to obtain correlation values; if the correlation value is greater than a preset reference threshold, then adaptation weights with values ​​greater than a preset baseline are generated; if the correlation value is not greater than the reference threshold, then adaptation weights with values ​​equal to the preset baseline are generated; a loss function is constructed using the adaptation weights, and gradient calculations are performed on the loss function to obtain gradient values; the connection weights of the memory network are updated using the backpropagation algorithm based on the gradient values ​​to obtain an updated memory network.

2. The method for predicting the lifespan of electronic components according to claim 1, characterized in that, The preprocessing operation on the historical data to obtain preliminary features includes: The historical data is decomposed using a wavelet transform algorithm to obtain denoised data. The denoised data is compensated according to a preset environmental reference value to obtain a correction sequence; The covariance of the corrected sequence is calculated to obtain the covariance matrix. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain preliminary features.

3. The method for predicting the lifespan of electronic components according to claim 1, characterized in that, The step of performing anomaly removal on the real-time data based on the preliminary characteristics to obtain clean data includes: Based on the preliminary characteristics, a physical threshold inversion operation is performed to obtain the physical fluctuation range; Data points exceeding the physical fluctuation range in the real-time data are identified as outliers and removed, leaving the remaining real-time data as clean data.

4. The method for predicting the lifespan of electronic components according to claim 1, characterized in that, The step of projecting the matching vector onto a preset affine matrix to obtain an alignment vector includes: The matching vector is projected onto a preset low-dimensional space using the affine matrix to obtain an alignment vector.

5. The method for predicting the lifespan of electronic components according to claim 1, characterized in that, The process of calculating the density of the uniform features to obtain a deviation metric, and then performing a weighted operation on the uniform features based on the deviation metric to obtain a weighted sequence includes: Perform a neighbor count operation on the unified feature to obtain the local density value; Based on the local density values, the unified feature is spatially divided to obtain sparse and dense regions. Distance calculation is performed between the feature points of the sparse region and the preset pattern center to obtain the deviation metric; Based on the deviation metric, high-weight factors and low-weight factors are generated, and the high-weight factors and low-weight factors are combined to obtain the initial weights. The uniform features are multiplied element-wise using the initial weights to obtain a weighted sequence.

6. The method for predicting the lifespan of electronic components according to claim 1, characterized in that, The process involves calculating the distribution of the weighted sequence to obtain a distribution metric. If the distribution metric is lower than a preset equilibrium threshold, a calibration coefficient is calculated. The weighted sequence is then reconstructed based on the calibration coefficient to obtain corrected data. If the distribution metric is not lower than the equilibrium threshold, the weighted sequence is used as the corrected data. This includes: The weighted sequence is subjected to a balance calculation using a spatial evaluation algorithm to obtain a distribution metric. If the distribution metric is lower than the equilibrium threshold, an optimization algorithm is used to perform a dimension scaling calculation to obtain a calibration coefficient. The calibration coefficient is then used to dynamically scale the weighted sequence to obtain corrected data. If the distribution metric is not lower than the equilibrium threshold, then the weighted sequence is used as correction data.

7. The method for predicting the lifespan of electronic components according to claim 1, characterized in that, The process of mapping the stable sequence using a pre-defined memory network to obtain a hidden vector, and then performing a regression operation on the hidden vector to obtain a lifetime estimate, includes: The memory network's forgetting gate mechanism is used to filter preset cell states to obtain key historical information; The stable sequence is superimposed onto the key historical information using the input gate mechanism of the memory network to obtain the updated state; The updated state is filtered using the output gate mechanism of the memory network to obtain the hidden vector; The hidden vector is input into a preset regression layer and subjected to a dot product operation to obtain the lifetime estimate.

8. A lifespan prediction system for electronic components, characterized in that, For implementing the method as described in any one of claims 1-7, comprising: The data preprocessing module is used to acquire historical data, perform preprocessing operations on the historical data, and obtain preliminary features; The feature alignment module is used to acquire real-time data, perform anomaly removal on the real-time data according to the preliminary features to obtain clean data, perform transformation operation on the clean data to obtain a matching vector, perform projection operation on the matching vector according to a preset affine matrix to obtain an alignment vector, and perform aggregation operation on the alignment vector and the preliminary features to obtain a unified feature. The distribution correction module is used to calculate the density of the uniform feature to obtain a deviation metric, perform a weighting operation on the uniform feature based on the deviation metric to obtain a weighted sequence, perform distribution calculation on the weighted sequence to obtain a distribution metric, calculate a calibration coefficient if the distribution metric is lower than a preset equilibrium threshold, reconstruct the weighted sequence based on the calibration coefficient to obtain corrected data, and use the weighted sequence as corrected data if the distribution metric is not lower than the equilibrium threshold. A time-series smoothing module is used to reassemble the corrected data to obtain a reassembled sequence, and then smooth the reassembled sequence to obtain a stable sequence. The lifetime prediction module is used to map the stable sequence using a preset memory network to obtain a hidden vector, and to perform a regression operation on the hidden vector to obtain a lifetime estimate. The closed-loop optimization module is used to acquire operating data if the estimated lifetime is lower than a preset lifetime threshold, perform correlation calculations on the operating data and preset reference data to obtain correlation values, generate adaptation weights based on the correlation values, and use the backpropagation algorithm to update the memory network based on the adaptation weights to obtain an optimized system.

Images