A method for constructing a digital twin model of an entire life cycle of an automobile part
By performing temporal feature fusion and dynamic causal graph analysis in a digital twin model, combined with depth-first search and the Adam optimizer, the prediction bias problem of traditional models under complex working conditions is solved, and the digital twin model is realized for accurate modeling and real-time response throughout the entire life cycle of automotive parts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 金智数字科技(苏州)有限公司
- Filing Date
- 2025-06-06
- Publication Date
- 2026-07-24
AI Technical Summary
Traditional digital twin models struggle to accurately capture the nonlinear causal relationships of component degradation under complex operating conditions in the full lifecycle management of automotive parts, resulting in significant prediction biases. Furthermore, they lack a dynamic adjustment mechanism for prioritizing key degradation characteristics, making it difficult to adapt to rapid changes in actual operating conditions.
By inputting the life cycle dataset into the initial life cycle digital twin model, temporal local feature fusion and time series analysis are performed to generate a dynamic causal graph. Key degradation path nodes are obtained using depth-first search and dynamic causal reasoning. Multi-dimensional decoupling and weight optimization are performed, and parameter incremental updates are carried out by combining the analytic hierarchy process and the Adam optimizer to construct an optimized life cycle digital twin model.
It achieves accurate modeling of dynamic causal relationships in life digital twin models, reduces prediction bias, enhances the robustness and reliability of models under complex working conditions, and can respond to changes in working conditions in real time.
Smart Images
Figure CN120632418B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital twin technology, and in particular to a method for constructing a digital twin model of the entire lifecycle of automotive parts. Background Technology
[0002] With the rapid development of intelligent manufacturing, digital twin technology, as a core means of interaction between physical entities and virtual models, has demonstrated significant value in the full lifecycle management of automotive parts. As a cyber-physical system integration method, digital twin models create virtual copies of physical products and continuously update and optimize them using real-time data. This enables precise monitoring, predictive maintenance, and performance optimization of automotive parts throughout their entire lifecycle, greatly improving product reliability and lifespan.
[0003] However, existing technologies have two key drawbacks: First, traditional methods for constructing digital twin models typically rely on static feature extraction and single-dimensional time-series analysis, making it difficult to accurately capture the nonlinear causal relationships of component degradation under complex operating conditions, resulting in significant prediction bias in the constructed digital twin models. Second, regarding the optimization and updating of digital twin model parameters, existing solutions mostly employ offline training or global optimization strategies, lacking a mechanism for dynamically adjusting the priority of key degradation features, making it difficult for the constructed digital twin models to adapt to rapid changes in actual operating conditions. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a method for constructing a digital twin model of the entire life cycle of automotive parts to solve the problems of large prediction bias and insufficient adaptability of traditional digital twin models.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0007] Firstly, this invention provides a method for constructing a digital twin model of the entire lifecycle of automotive components, comprising,
[0008] The life cycle dataset is input into the initial life digital twin model. The feature extraction layer performs temporal local feature fusion, and the time series analysis layer captures the degradation trend to form the life prediction value of automotive parts.
[0009] Extract the actual detection values from the lifecycle dataset, calculate the deviation between the predicted lifecycle values and the actual detection values of automotive parts, map the deviations to a time-series relationship, and generate a dynamic causal graph.
[0010] The dynamic causal graph is traversed to obtain key degradation path nodes. Dynamic causal reasoning is used to perform correlation analysis on the key degradation path nodes to generate degradation feature parameters. Multi-dimensional decoupling and weight optimization are performed on the degradation feature parameters to form an optimized lifetime digital twin model construction scheme.
[0011] Based on the optimized lifetime digital twin model construction scheme, the analytic hierarchy process (AHP) is used to rank the importance of features and obtain the priority weight tensor. According to the priority weight tensor, the Adam optimizer is used to incrementally update the parameters of the initial lifetime digital twin model.
[0012] As a preferred embodiment of the method for constructing a digital twin model of the entire life cycle of automotive parts as described in this invention, the life cycle dataset includes part design parameters, part production quality inspection data, part usage data, and part actual lifespan data.
[0013] As a preferred embodiment of the method for constructing a digital twin model of the entire lifecycle of automotive components according to the present invention, the step of forming the lifecycle prediction value of the automotive components specifically includes the following steps.
[0014] A global lifetime digital twin model is constructed by stacking features at multiple scales in the feature extraction layer and the temporal analysis layer through a bidirectional gated recurrent unit.
[0015] The life cycle dataset is input into the global life cycle digital twin model. The feature extraction layer performs temporal local feature fusion through multi-resolution convolution kernels and ELU activation function to generate health indicator curves.
[0016] The temporal analysis layer uses bidirectional gating and multi-head attention mechanisms to capture degradation trends and generate long-term dependency features;
[0017] By cascading and splicing health indicator curves and long-term dependency characteristics, the lifespan prediction values of automotive parts are formed.
[0018] As a preferred embodiment of the method for constructing a digital twin model of the entire lifecycle of automotive components according to the present invention, the calculation of the deviation between the predicted lifespan value and the actual measured value of the automotive components specifically includes the following operations:
[0019] Spline interpolation is used to dynamically segment the lifecycle dataset at multiple scales to extract actual detection values.
[0020] The mean absolute error method is used to accumulate and average the absolute deviations of the predicted and actual life values of automotive parts to obtain the deviation amount.
[0021] As a preferred embodiment of the method for constructing a digital twin model of the entire lifecycle of automotive components according to the present invention, the generation of a dynamic causal graph specifically includes the following steps.
[0022] The deviation is aligned in time series and then decomposed in the frequency domain using fast Fourier transform to form a causal intensity vector.
[0023] The temporal sequence graph version control method is used to perform spatiotemporal correlation analysis of causal chains on the causal intensity vector to obtain contribution quantification indicators.
[0024] Based on contribution metrics, time-series relationship mapping is performed using a time-series graph database to generate a dynamic causal graph.
[0025] As a preferred embodiment of the method for constructing a digital twin model of the entire lifecycle of automotive components according to the present invention, the method for forming an optimized lifecycle digital twin construction scheme specifically includes the following steps.
[0026] Use depth-first search to traverse the dynamic cause-effect graph and obtain the critical degradation path nodes;
[0027] Dynamic causal reasoning is used to perform correlation quantification on key degradation path nodes to generate a correlation causal matrix; principal component decomposition is performed on the correlation causal matrix to obtain degradation feature parameters;
[0028] The degradation feature parameters are decoupled in multiple dimensions by tensor decomposition to obtain decoupled features; the decoupled features are then weighted and fused to form an optimized lifetime digital twin model construction scheme.
[0029] As a preferred embodiment of the method for constructing a digital twin model of the entire lifecycle of automotive components according to the present invention, the step of obtaining the priority weight tensor specifically includes the following steps.
[0030] Obtaining the priority weight tensor specifically includes the following steps.
[0031] The optimized life digital twin model construction scheme is mapped to the feature space to obtain multi-dimensional feature vectors. The hierarchical analysis method is used to perform hierarchical partitioning and feature importance ranking on the multi-dimensional feature vectors to generate a hierarchical weight priority matrix.
[0032] The entropy method is applied to perform multidimensional feature weight allocation on the hierarchical weight priority matrix to obtain the priority weight tensor.
[0033] As a preferred embodiment of the method for constructing a digital twin model of the entire lifecycle of automotive components according to the present invention, the step of incrementally updating the parameters of the initial lifecycle digital twin model using the Adam optimizer specifically includes the following steps.
[0034] The loss function is weighted according to the priority weight tensor, and backpropagation is performed using the weighted loss function to obtain the global gradient tensor.
[0035] The gradient of the global gradient tensor is calculated using an exponentially weighted moving average to obtain the gradient mean and the squared mean of the gradient.
[0036] The Adam optimizer is used to correct the deviation of the gradient mean and the gradient squared mean, generating an incremental parameter matrix. Based on the incremental parameter matrix, gradient clipping is used to update the parameters incrementally, and the optimized lifetime digital twin model is output synchronously.
[0037] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the method for constructing a digital twin model of the entire life cycle of automotive parts as described in the first aspect of the present invention.
[0038] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for constructing a digital twin model of the entire life cycle of automotive parts as described in the first aspect of the present invention.
[0039] The beneficial effects of this invention are as follows: It employs depth-first search to traverse the dynamic causal graph and uses dynamic causal inference to perform correlation quantification and principal component decomposition on key degradation path nodes, achieving accurate dynamic causal correlation modeling of the lifespan digital twin model and reducing prediction bias. Furthermore, it uses the analytic hierarchy process (AHP) to incrementally update the parameters of the initial lifespan digital twin model, enabling the optimized model to respond in real-time to changes in operating conditions and significantly enhancing its robustness and reliability under complex real-world operating conditions. Attached Figure Description
[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0041] Figure 1 A flowchart illustrating the method for constructing a digital twin model of the entire lifecycle of automotive parts.
[0042] Figure 2 A flowchart of the process for generating lifetime prediction values.
[0043] Figure 3 A flowchart for constructing a dynamic causal graph.
[0044] Figure 4 A flowchart for updating parameters of the initial lifetime digital twin model. Detailed Implementation
[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0048] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for constructing a digital twin model of the entire lifecycle of automotive parts, including the following steps:
[0049] S1. Input the life cycle dataset into the initial life digital twin model, perform temporal local feature fusion in the feature extraction layer, and capture the degradation trend in the time series analysis layer to form the life prediction value of automotive parts.
[0050] Specifically, the operations include the following:
[0051] S1.1 Collect lifecycle datasets, which include part design parameters, part production quality inspection data, part usage data, and part actual life data; among which, part design parameters include data on geometric dimensions, material properties, and mechanical performance indicators; geometric dimension data are obtained using CAD software, material properties are measured using material testing instruments, and mechanical performance indicator data are collected using finite element analysis tools.
[0052] The quality inspection data for parts production includes data on surface roughness, dimensional accuracy, and hardness. Surface roughness is measured using a coordinate measuring machine, dimensional accuracy is collected using a surface profilometer, and hardness is collected using a hardness tester.
[0053] The component usage data includes operating temperature, load conditions, and vibration frequency data; the operating temperature is collected using a temperature sensor, the load conditions are monitored in real time by an on-board diagnostic unit (OBD), and the vibration frequency is collected using an accelerometer.
[0054] The actual life data of a component includes data on failure mode, service life duration, and failure time. Failure mode is collected by a flaw detector, service life duration is collected by a timer, and failure time is obtained by an on-board diagnostic unit.
[0055] Lifecycle datasets not only enhance understanding and control of product lifecycle management, but also optimize design to extend lifespan, ultimately improving product market competitiveness.
[0056] S1.2. Preprocessing of the lifecycle dataset: Specifically, an outlier removal algorithm is used to identify and eliminate abnormal sampling points, and linear interpolation is applied to fill in missing values. A time-series alignment algorithm is used to unify lifecycle datasets with different sampling frequencies to the same timestamp to ensure strict time synchronization. Next, the lifecycle dataset is smoothed in the temporal domain using a Savitzky-Golay filter to suppress noise. Z-score normalization and One-Hot encoding are used for standardization and discretization to eliminate dimensional differences. Finally, principal component analysis is used to decompose the lifecycle dataset into features, compressing high-dimensional features into a low-dimensional space to obtain the preprocessed lifecycle dataset.
[0057] S1.3. Construct and train the initial lifespan digital twin model. Specifically, the PyTorch framework is used to call neural networks, deploying multi-scale parallel convolutional structures of 1D-CNN (one-dimensional convolutional neural network) and 3D-CNN (three-dimensional convolutional neural network). Simultaneously, the ELU activation function is embedded to perform nonlinear transformation and feature fusion on local temporal features, completing the construction of the feature extraction layer. The torch.nn parameter is used to call a gated recurrent unit and a multi-head attention structure, setting the hidden layer dimension to 128, the number of attention heads to 8, and the key-value pair to a dimension of 64. The softmax function is used for dynamic weight allocation, completing the construction of the temporal analysis layer. Cross-layer concatenation is performed on the feature extraction layer and the temporal analysis layer using a bidirectional gated recurrent unit to obtain multi-scale spatiotemporal feature weights. Based on the spatiotemporal feature weights, the torch.cat function is used to stack multi-scale features in the feature extraction layer and the temporal analysis layer, and normalization is performed using LayerNorm, synchronously outputting the constructed initial lifespan digital twin model.
[0058] Next, the constructed initial lifespan digital twin model is trained. Specifically, the lifespan dataset is divided into a sample set, a training set, and a validation set. On the sample set, StandardScaler is used for feature scaling and data augmentation to obtain augmented samples. On the training set, the Adam optimizer is used to backpropagate the augmented samples, and a weighted loss function is used to calculate the error. Gradient clipping is used simultaneously for numerical stabilization to generate updated initial lifespan digital twin model parameters. On the validation set, the root mean square error is used to quantize the updated initial lifespan digital twin model parameters, forming the validation error. When the validation error reaches the convergence threshold for five consecutive rounds, training is stopped, and the trained initial lifespan digital twin model is saved using the torch.save function.
[0059] It should be noted that the convergence threshold is based on the definition of the exponential moving average of the verification error, and its value ranges from [0.001, 0.01].
[0060] S1.4. Lifespan predictions for automotive parts are generated using an initial lifespan digital twin model. Specifically, the lifespan dataset is input into the feature extraction layer via a data loading interface. Multi-resolution convolutional kernels are used for temporal local feature extraction. Small-sized kernels capture high-frequency local features, while large-sized kernels extract macro-trend features, resulting in multi-scale temporal features. After each multi-resolution convolutional kernel branch, an ELU activation function is applied to perform nonlinear transformation and feature enhancement on the multi-scale temporal features, outputting activated temporal local features. These activated temporal local features are then subjected to max pooling for dimensionality reduction and weighted fusion using channel attention units to generate multi-scale fused features. Subsequently, dilated convolutions are used to expand the receptive field of the multi-scale fused features, and residual connections are combined to alleviate gradient vanishing, forming enhanced features. Finally, the enhanced features are time-aligned using a time warping algorithm, and curve fitting is performed using the least squares method. After smoothing with a Savitzky-Golay filter, a health indicator curve is output.
[0061] The temporal analysis layer captures trends in the lifecycle dataset through bidirectional gated recurrent units. The forward GRU captures progressively accumulating degradation trends through forward temporal processing, while the backward GRU captures abrupt and abnormal degradation trends through backward temporal processing, thus obtaining bidirectional hidden states. Subsequently, a multi-head attention mechanism is used to take the bidirectional hidden states as inputs for queries, keys, and values. Attention is calculated by scaling the dot product, and a causal mask matrix is embedded during the calculation process to ensure temporal dependencies, generating an attention weight matrix. The specific mathematical formula is as follows.
[0062] ;
[0063] in, Represents the attention weight matrix; Represents the query matrix; Represents the key matrix; Represents the feature dimension of the key matrix; Represents the causal mask matrix;
[0064] LayerNorm is used to normalize the attention weight matrix to form normalized attention features. The normalized attention features are then fused with the bidirectional hidden state through gated residual connections. After the short-term fluctuations are eliminated by average pooling in the time dimension, long-term dependency features with long-term degradation trend awareness are output.
[0065] The health indicator curve and long-term dependency features are concatenated along the channel dimension. The health indicator curve has 3 channels, and the long-term dependency features have 256 channels. After concatenation, a 259-dimensional fused feature is generated. The fused feature is then temporally downsampled using strided convolution to obtain temporally compressed features. Gated linear units are used to perform feature selection and numerical transformation on the temporally compressed features to generate preliminary lifespan prediction values. Subsequently, physical degradation parameters of the lifespan dataset are extracted, and the physical degradation parameters are weighted and fused with the preliminary lifespan prediction values. Dimension mapping is performed through a fully connected layer, and normalization is performed synchronously using the Sigmoid activation function. The final lifespan prediction value is compressed to the [0,1] interval, and the normalized lifespan prediction value of the automotive parts is output.
[0066] Life prediction values for automotive parts can not only accurately pinpoint performance changes during critical degradation stages, but also dynamically adapt to the needs of product lifecycle management, achieving full coverage from fault repair to predictive maintenance, thereby improving overall operational efficiency and reliability.
[0067] S2. Extract the actual detection values from the lifecycle dataset, calculate the deviation between the predicted lifecycle values and the actual detection values of automotive parts, map the deviations to a time-series relationship, and generate a dynamic causal graph.
[0068] Specifically, the operations include the following:
[0069] S2.1. Dynamic multi-scale segmentation of the lifecycle dataset is performed using spline interpolation to extract actual detection values. Specifically, a non-uniform sampling grid is constructed in a three-dimensional coordinate system with the timestamp of the lifecycle dataset as the x-axis and the multi-source data in the lifecycle dataset as the y-axis. Cubic B-spline interpolation is used to perform density sampling on the non-uniform sampling grid to obtain discrete degradation feature points. The least squares method is used to continuously fit the discrete degradation feature points to obtain a smooth degradation trajectory curve. The smooth degradation trajectory curve is dynamically segmented into three stages: healthy period, degradation period, and failure period, through a sliding window. Within each stage, data time alignment is performed through dynamic time warping, and high-frequency noise is eliminated simultaneously using a Savitzky-Golay filter to generate standardized stage features. Principal component analysis is used to reduce the dimensionality of the standardized stage features, and Z-score-based numerical transformation is performed to finally generate actual detection values.
[0070] S2.2. The mean absolute error (MAE) method is used to accumulate and average the absolute deviations of the predicted and actual lifespan values of automotive parts to obtain the deviation amount. Specifically, the MAE method is used to calculate the absolute difference point-by-point between the predicted and actual lifespan values of automotive parts, obtaining an absolute deviation sequence. The time decay factor is used to accumulate the deviations in the absolute deviation sequence, generating a time-weighted deviation sum. The time-weighted deviation sum is then averaged to obtain the mean absolute error. An exponentially weighted moving average is applied to smooth the mean absolute error, and the final deviation amount is output. The specific mathematical formula is as follows.
[0071] ;
[0072] in, Indicates the deviation amount. Indicates the smoothing coefficient; Indicates the total duration; Indicates the cumulative time steps; Indicates the time step for error calculation; Indicates the time decay factor; Indicates the time step in error calculation Lifespan prediction; Indicates the time step in error calculation The actual measured value;
[0073] It should be noted that the smoothing coefficient is based on the exponential smoothing definition of the mean absolute error, and its value range is (0,1); the time decay factor is based on the exponential decay definition of the historical deviation, and its value range is (0,1).
[0074] S2.3. Time series alignment of the deviation is performed, and frequency domain energy decomposition is performed using Fast Fourier Transform to form a causal intensity vector. Specifically, the deviation is aligned to an equal-interval timestamp through linear interpolation to eliminate differences in device sampling frequencies and obtain the aligned deviation. Then, Fast Fourier Transform is applied to the aligned deviation to perform frequency domain energy decomposition, generating a power spectral density distribution. The power spectral density distribution is integrated with the frequency band energy to extract the energy distribution of the characteristic frequency band. A Chebyshev II bandpass filter is used to separate the energy distribution of the characteristic frequency band, and the energy components are quantized using Short-Time Fourier Transform to form a time-frequency energy matrix. The Granger causality test is used to extract the lag covariance of the time-frequency energy matrix, and causal propagation fitting is performed on the lag covariance to obtain the causal influence strength. The causal influence strength is weighted by Min-Max normalization and generated into a causal intensity vector by tensor concatenation.
[0075] S2.4. Using a temporal graph version control method, spatiotemporal correlation analysis of causal chains is performed on the causal intensity vector to obtain contribution metrics. Specifically, the causal intensity vector is sliced and version-marked according to the time dimension to construct a versioned graph structure. Each node stores the 8-dimensional causal intensity vector at the corresponding time point, and the causal chain time axis of the causal intensity vector is aligned through dynamic time warping. The temporal graph version control method is used to extract spatiotemporal features from the versioned graph structure, and graph convolution is used to aggregate node features to generate a spatiotemporal feature matrix. Causal chain spatiotemporal correlation detection is performed on the spatiotemporal feature matrix to obtain potential causal correlations. Threshold masks are applied to potential causal correlations to perform noise filtering, forming significant causal features. Multi-head attention mechanism is used to perform cross-time step weight allocation on significant causal features, and Softmax is used for weight normalization to form contribution weights. Based on the contribution weights, key contributions are accumulated through path integrals, and finally, contribution metrics are output.
[0076] S2.5. Based on the contribution metrics, perform time-series relation mapping using a time-series graph database to generate a dynamic causal graph. Specifically, the contribution metrics are imported into the Neo4j time-series graph database in batches via CSV. The APOC time extension function of the Neo4j time-series graph database is used to extract the causal efficacy features of the contribution metrics and map them as timestamped attribute nodes. Window correlation calculation is performed on the contribution metrics through time window sliding to obtain the cross-time-step association strength matrix. Time decay weighting is applied to the cross-time-step association strength matrix to obtain the decay association matrix. Nonlinear fitting is performed on the decay association matrix using multinomial regression to generate a time-series relation chain. The time-series extension plugin of the Neo4j time-series graph database is used to map the time-series relation chain as weighted spatiotemporal edges. The Louvain algorithm is used to perform real-time clustering of the timestamped attribute nodes and weighted spatiotemporal edges to obtain causal evolution patterns. The graph snapshot difference method is used to perform topological structure mapping on the causal evolution patterns, ultimately generating a dynamic causal graph.
[0077] S3. Perform graph traversal on the dynamic causal graph to obtain key degradation path nodes. Use dynamic causal reasoning to perform correlation analysis on the key degradation path nodes and generate degradation feature parameters. Perform multi-dimensional decoupling and weight optimization on the degradation feature parameters to form an optimized lifetime digital twin model construction scheme.
[0078] Specifically, the operations include the following:
[0079] S3.1. A depth-first search (DFS) is used to traverse the dynamic cause-effect graph to obtain critical degradation path nodes. Specifically, initial degradation state nodes are defined based on the component initial failure mode library. Starting from these initial degradation state nodes, a recursive DFS traversal is performed on the dynamic cause-effect graph. During the search, the path node sequence and edge weights of the current branch are recorded in real-time. The search of the current branch is terminated when the following conditions are encountered: reaching a node with no outgoing degree, the total path length exceeding 10 nodes, or the cumulative edge weight product falling below the decay threshold. After terminating the search, the reached nodes are integrated to obtain the critical degradation path nodes.
[0080] It should be noted that a node with no outgoing degree refers to a terminating node in the dynamic causal graph that has no outgoing edges; the decay threshold is defined based on the historical failure statistics of critical degradation path nodes, and its value range is [0.15, 0.35].
[0081] S3.2. Dynamic causal inference is used to perform correlation quantification on key degradation path nodes, generating a correlation causal matrix. Principal component decomposition is performed on the correlation causal matrix to obtain degradation feature parameters. Specifically, dynamic causal inference is used to perform linear regression decomposition on key degradation path nodes and their adjacent nodes to obtain time-delay related data. A sliding time window is applied to smooth the time-delay related data, generating time-delay correlation coefficients. A weighted average method is used to weight and integrate the time-delay correlation coefficients to obtain causal influence factors. L1 regularization constraints are applied to the time-delay correlation coefficients and causal influence factors for sparsification fusion, generating a sparse causal matrix. Missing value completion and noise filtering are performed on the sparse causal matrix to form a correlation causal matrix. Principal component analysis is used to decompose the correlation causal matrix into eigenvalues, extracting the top 5 largest eigenvalues. Z-score normalization is applied to the top 5 largest eigenvalues for weight allocation and normalization, generating corresponding feature vectors. The corresponding feature vectors are subjected to Varimax orthogonal rotation to eliminate inter-component correlations, and linear interpolation is used to sample the importance of the corresponding feature vectors to obtain degradation feature parameters.
[0082] S3.3. Degradation feature parameters are decoupled in multiple dimensions using tensor decomposition to obtain decoupled features. These decoupled features are then weighted and fused to form an optimized lifetime digital twin model construction scheme. Specifically, degradation feature parameters are organized into a three-dimensional tensor structure comprising samples, time series, and feature dimensions through tensor reconstruction. In the multi-dimensional decoupling stage, principal component extraction is performed on the three-dimensional tensor structure using tensor decomposition to obtain a low-dimensional approximate tensor. Principal component extraction is then performed on the low-dimensional approximate tensor to obtain the core tensor. Modal expansion of the core tensor is performed using matrix slicing, and dimensionality compression is achieved through principal component analysis to generate decoupled features. In the weight optimization stage, a two-layer attention mechanism is constructed: the first layer performs temporal attention allocation on the decoupled features and uses tanh activation for weight normalization to obtain the importance weights at each time step; the second layer performs feature dimension aggregation on the importance weights at each time step to obtain preliminary fusion weights, and applies the softmax function to standardize the preliminary fusion weights to generate the final attention weights; in the weighted fusion stage, the decoupled features are weighted and summed according to the final attention weights to obtain the fusion feature vector; the fusion feature vector is nonlinearly mapped using a fully connected function to generate an optimized lifetime digital twin model construction scheme.
[0083] S4. Based on the optimized lifetime digital twin model construction scheme, the analytic hierarchy process is used to sort the importance of features and obtain the priority weight tensor. Based on the priority weight tensor, the Adam optimizer is used to incrementally update the parameters of the initial lifetime digital twin model.
[0084] Specifically, the operations include the following:
[0085] S4.1 The optimized lifespan digital twin model construction scheme is mapped to the feature space to obtain multi-dimensional feature vectors. The hierarchical analysis method (AHP) is used to perform hierarchical partitioning and feature importance ranking on the multi-dimensional feature vectors, generating a hierarchical weight priority matrix. Specifically, cubic spline interpolation is used to sample the lifespan digital twin model construction scheme at equal intervals to obtain discrete feature point sequences. Linear discriminant analysis is used to map the discrete feature point sequences to the feature space to generate multi-dimensional feature vectors. The AHP is used to perform eigenvalue decomposition on the multi-dimensional feature vectors to obtain initial weight vectors. Simultaneously, hierarchical partitioning is performed on the initial weight vectors to obtain multi-level elements. The least squares method is used to perform linear regression fitting on the multi-level elements to obtain the feature weight distribution. Feature importance ranking is performed on the feature weight distribution to obtain the importance scale. The eigenvalue method is used to synthesize the weights of the importance scale to form a preliminary weight vector. Z-score normalization is used to normalize the preliminary weight vector to eliminate dimensional differences. Finally, matrix transformation is performed using the eigenvector method to obtain the hierarchical weight priority matrix.
[0086] S4.2. Apply the entropy method to perform multi-dimensional feature weight allocation on the hierarchical weight priority matrix to obtain the priority weight tensor. Specifically, the entropy method is applied to measure the entropy of the hierarchical weight priority matrix, measuring the dispersion of the criterion features column by column to obtain the criterion entropy vector; multi-dimensional feature weight allocation is performed on the criterion entropy vector to generate the original hierarchical weights; then, the criterion entropy vector is linearly transformed and normalized through entropy weight transformation to construct the entropy weight correction coefficient matrix; the entropy weight correction coefficient matrix is multiplied by the original hierarchical weights to achieve dynamic weight adjustment, and the proportion is adjusted through weight rebalancing to obtain the balanced weight matrix; the balanced weight matrix is tensor expanded through Kronecker product to generate the priority weight tensor.
[0087] S4.3. The loss function is weighted according to the priority weight tensor, and backpropagation is performed using the weighted loss function to obtain the global gradient tensor. Specifically, the priority weight tensor and the loss function are dimensionally expanded to ensure that the shapes of the priority weight tensor and the loss function match; the dimensionally expanded priority weight tensor is multiplied with the loss function to generate the loss tensor; the loss function is dynamically weighted according to the loss tensor to form a weighted loss function; then, PyTorch's automatic differentiation framework is used to perform backpropagation on the weighted loss function. During the backpropagation process, gradient flow is injected through the chain rule to generate local gradient tensors; GroupNorm (grouping normalization) is used to aggregate the gradients of the local gradient tensors to eliminate the dimensional shift caused by priority differences, thus obtaining the global gradient tensor.
[0088] It should be noted that the chain rule refers to the chain rule for finding the derivative of a composite function in calculus, which is called through the autograd unit in PyTorch.
[0089] S4.4. Calculate the gradient of the global gradient tensor using an exponentially weighted moving average to obtain the gradient mean and the gradient squared mean. In practice, a sliding window is used to perform sliding statistics on the global gradient tensor to obtain the current gradient statistics. The current gradient statistics are initialized using a zero tensor to generate the historical gradient mean and the historical gradient squared mean.
[0090] It should be noted that a zero tensor refers to a single-dimensional tensor initialized with all zeros, obtained by calling the torch.zeros function in the PyTorch framework;
[0091] Based on the historical gradient mean and historical gradient squared mean, two decay coefficients (β1 and β2) are set through the Adam optimizer to control the update weights of the historical gradient mean and historical gradient squared mean, respectively. The gradient is calculated using an exponentially weighted moving average: the global gradient tensor is fused with the historical gradient mean according to the decay coefficient (β1) to generate the gradient mean; at the same time, the square of the global gradient tensor is fused with the historical gradient squared mean according to the decay coefficient (β2) to generate the gradient squared mean. Among them, the gradient mean records the directional trend of the gradient, and the gradient squared mean captures the magnitude change of the gradient.
[0092] S4.5. The Adam optimizer is used to correct the bias of the gradient mean and the mean squared gradient, generating an incremental parameter matrix. Based on the incremental parameter matrix, gradient pruning is used to incrementally update the parameters, simultaneously outputting the optimized lifetime digital twin model. Specifically, the Adam optimizer is used to perform an exponentially weighted average of the gradient mean and the mean squared gradient to obtain the gradient moment estimate. Time step compensation is applied to the gradient moment estimate to generate the gradient bias. Based on the gradient bias, the bias correction factor is used to correct the bias of the gradient mean and the mean squared gradient, generating an unbiased estimate. Learning rate scaling is applied to perform incremental parameter synthesis on the unbiased estimate, forming the incremental parameter matrix.
[0093] It should be noted that the deviation correction factor is based on the time step decay rate and is defined by the reciprocal of the decay coefficient, with a value range of (1,∞).
[0094] Gradient clipping is used to apply L2 norm constraints to the incremental parameter matrix to prevent extreme weight updates, generating clipped parameter increments. The clipped parameter increments are then iterated to obtain the updated lifetime digital twin model parameters. Simultaneously, an adaptive learning rate mechanism triggers the weight update of the lifetime digital twin model. The relative rate of change of parameter updates is obtained using the moving average statistical method. When the relative rate of change of parameter updates is lower than the update threshold for three consecutive rounds, the learning rate decay is automatically triggered until the initial lifetime digital twin model update stops, and the optimized lifetime digital twin model is output.
[0095] It should be noted that the update threshold is defined based on the statistical distribution of the historical lifespan digital twin model update, and the value range is [1e-5, 1e-3].
[0096] This embodiment also provides a computer device applicable to the method of constructing a digital twin model of the entire life cycle of automotive parts, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the method of constructing a digital twin model of the entire life cycle of automotive parts as proposed in the above embodiment.
[0097] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0098] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for constructing a digital twin model of the entire lifecycle of automotive parts as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0099] In summary, this invention achieves accurate modeling of dynamic causal relationships during the construction of a digital twin model by: traversing the dynamic causal graph using depth-first search and performing correlation quantification and principal component decomposition on key degradation path nodes using dynamic causal inference; and reducing prediction bias. Furthermore, the use of the analytic hierarchy process (AHP) for incremental parameter updates to the initial lifetime digital twin model enables the optimized lifetime digital twin model to respond in real-time to changes in operating conditions, significantly enhancing its robustness and reliability under complex real-world operating conditions.
[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for constructing a digital twin model of the entire lifecycle of automotive parts, characterized in that: include, The life cycle dataset is input into the initial life digital twin model. The feature extraction layer performs temporal local feature fusion, and the time series analysis layer captures the degradation trend to form the life prediction value of automotive parts. The process involves extracting actual detection values from the lifecycle dataset, calculating the deviation between the predicted lifecycle values and actual detection values of automotive parts, mapping the deviations to temporal relationships, and generating a dynamic causal graph. The specific steps include the following: The deviation is aligned in time series and then decomposed in the frequency domain using fast Fourier transform to form a causal intensity vector. The temporal sequence graph version control method is used to perform spatiotemporal correlation analysis of causal chains on the causal intensity vector to obtain contribution quantification indicators. Based on contribution metrics, time-series relationship mapping is performed using a time-series graph database to generate a dynamic causal graph. A depth-first search algorithm is used to recursively search the dynamic causal graph starting from the initial degenerate state node. After the search is terminated, the reached nodes are integrated to obtain the key degenerate path nodes. Dynamic causal reasoning is used to perform correlation analysis on key degradation path nodes to generate degradation characteristic parameters; Multi-dimensional decoupling and weight optimization of degradation feature parameters are performed to form an optimized lifetime digital twin model construction scheme; Based on the optimized lifetime digital twin model construction scheme, the analytic hierarchy process (AHP) is used to rank feature importance and obtain a priority weight tensor. Then, based on this priority weight tensor, the Adam optimizer is used to incrementally update the parameters of the initial lifetime digital twin model. The specific steps include the following: The loss function is weighted according to the priority weight tensor, and backpropagation is performed using the weighted loss function to obtain the global gradient tensor. The gradient of the global gradient tensor is calculated using an exponentially weighted moving average to obtain the gradient mean and the squared mean of the gradient. The Adam optimizer is used to correct the deviation of the gradient mean and the gradient squared mean, generating an incremental parameter matrix. Based on the incremental parameter matrix, gradient clipping is used to update the parameters incrementally, and the optimized lifetime digital twin model is output synchronously.
2. The method for constructing a digital twin model of the entire lifecycle of automotive parts as described in claim 1, characterized in that: The lifecycle dataset includes part design parameters, part production quality inspection data, part usage data, and part actual lifespan data.
3. The method for constructing a digital twin model of the entire lifecycle of automotive parts as described in claim 2, characterized in that: The process of generating the life prediction value for automotive components specifically includes the following steps: A global lifetime digital twin model is constructed by stacking features at multiple scales in the feature extraction layer and the temporal analysis layer through a bidirectional gated recurrent unit. The life cycle dataset is input into the global life cycle digital twin model. The feature extraction layer performs temporal local feature fusion through multi-resolution convolution kernels and ELU activation function to generate health indicator curves. The temporal analysis layer uses bidirectional gating and multi-head attention mechanisms to capture degradation trends and generate long-term dependency features; By cascading and splicing health indicator curves and long-term dependency characteristics, the lifespan prediction values of automotive parts are formed.
4. The method for constructing a digital twin model of the entire lifecycle of automotive parts as described in claim 3, characterized in that: The calculation of the deviation between the predicted lifespan and the actual measured lifespan of automotive parts specifically includes the following operations: Spline interpolation is used to dynamically segment the lifecycle dataset at multiple scales to extract actual detection values. The mean absolute error method is used to accumulate and average the absolute deviations of the predicted and actual life values of automotive parts to obtain the deviation amount.
5. The method for constructing a digital twin model of the entire lifecycle of automotive parts as described in claim 4, characterized in that: The optimized lifetime digital twin construction scheme specifically includes the following steps: Use depth-first search to traverse the dynamic cause-effect graph and obtain the critical degradation path nodes. Dynamic causal reasoning is used to perform correlation quantification on key degradation path nodes to generate a correlation causal matrix; principal component decomposition is performed on the correlation causal matrix to obtain degradation feature parameters; Decoupled features are obtained by multi-dimensional decoupling of degenerate feature parameters using tensor decomposition. We optimize the weights and perform weighted fusion on the decoupled features to form an optimized scheme for constructing a lifespan digital twin model.
6. The method for constructing a digital twin model of the entire lifecycle of automotive parts as described in claim 5, characterized in that: Obtaining the priority weight tensor specifically includes the following steps. The optimized life digital twin model construction scheme is mapped to the feature space to obtain multi-dimensional feature vectors. The hierarchical analysis method is used to perform hierarchical partitioning and feature importance ranking on the multi-dimensional feature vectors to generate a hierarchical weight priority matrix. The entropy method is applied to perform multidimensional feature weight allocation on the hierarchical weight priority matrix to obtain the priority weight tensor.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method for constructing a digital twin model of the entire life cycle of automotive components as described in any one of claims 1 to 6.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for constructing a digital twin model of the entire life cycle of automotive parts as described in any one of claims 1 to 6.