Optical fiber connector service life prediction system based on digital twinning and multi-parameter fusion

Abstract

The invention relates to the technical field of health management of optical fiber communication equipment, and discloses an optical fiber connector life prediction system based on digital twinning and multi-parameter fusion, which comprises a data acquisition module, a data preprocessing module, a feature extraction module, a model construction module, a fusion analysis module and a life prediction module. According to the optical fiber connector service life prediction system based on digital twinborn and multi-parameter fusion, multi-dimensional data is formed through the data acquisition module, complete and associated data support is provided for subsequent feature extraction and service life analysis, and a digital twinborn model is constructed based on three-dimensional structure parameters and material attributes of the optical fiber connector, so that the service life of the optical fiber connector is predicted. The virtual model can accurately reflect the structural state change of a physical entity in real time, in addition, the fusion feature vector and the structural state parameter are subjected to joint analysis, a comprehensive evaluation index capable of reflecting the association of the fusion feature vector, the structural state parameter and the comprehensive evaluation index is output, an accurate multi-dimensional evaluation basis is provided for life prediction, and then the life condition of the optical fiber connector is accurately predicted.

Description

Technical Field

[0001] This invention relates to the field of health management technology for optical fiber communication equipment, specifically a life prediction system for optical fiber connectors based on digital twins and multi-parameter fusion. Background Technology

[0002] As a core connection component in high-speed communication networks, data centers, and industrial control, the operational stability and remaining lifespan of fiber optic connectors directly determine the reliability and maintenance costs of the entire optical transmission system. With the development of optical communication technology towards higher bandwidth and longer distances, fiber optic connectors need to operate under complex conditions for extended periods. Their performance and structure are susceptible to the combined effects of environmental factors (such as temperature fluctuations, humidity changes, and mechanical vibration) and usage losses (such as insertion / removal operations and ferrule displacement), gradually leading to performance degradation issues such as increased insertion loss and abnormal return loss, ultimately resulting in connection failure.

[0003] Existing solutions often only collect environmental parameters (such as temperature and humidity) or single performance parameters (such as insertion loss), failing to simultaneously acquire multi-dimensional data related to the environment, performance, and structure (such as vibration, insertion / removal cycles, and ferrule displacement). This results in an incomplete data foundation, making it impossible to reflect the impact of multiple factors coupled on lifespan. Secondly, feature extraction from the collected data is often limited to a single dimension (such as only time-domain features), failing to achieve deep integration of time-domain, frequency-domain, and spatial features, and ignoring the temporal correlation between features, making it difficult to accurately capture the potential patterns of lifespan degradation.

[0004] Therefore, existing life prediction technologies cannot meet the high-precision and high-reliability life assessment requirements of fiber optic connectors. Inaccurate predictions can easily lead to unplanned downtime or excessive maintenance, increasing system operating costs and safety risks. Summary of the Invention

[0005] The purpose of this invention is to provide a fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion to solve the technical problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention discloses the following technical solution: a fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion, comprising:

[0007] The data acquisition module is configured to acquire multi-dimensional data of the fiber optic connector, including: temperature, humidity and vibration parameters of the environment in which the fiber optic connector is located, which are acquired in real time based on environmental sensors, and insertion loss, return loss, number of insertions and removals and ferrule displacement data of the fiber optic connector, which are acquired in real time based on performance sensors.

[0008] The data preprocessing module is configured to clean and reduce noise in the collected multi-dimensional data.

[0009] The feature extraction module is configured to: fuse the temporal, frequency, and spatial features of the preprocessed multi-dimensional data through a self-attention mechanism to extract and generate a fused feature vector with temporal correlation;

[0010] The model building module is configured to: construct a digital twin model based on the three-dimensional structural parameters and material properties of the fiber optic connector, and dynamically update the model using the physical entity dynamic parameters acquired in real time by the data acquisition module; the three-dimensional structural parameters include the geometric dimensions of the ceramic ferrule and the radius of curvature of the lens, the material properties include the elastic modulus and the coefficient of thermal expansion, and the dynamic parameters include the ferrule stress distribution and the lens offset;

[0011] The fusion analysis module is configured to: jointly analyze the fused feature vector with the structural state parameters generated by the digital twin model based on the dynamic parameters through a deep residual network, and output a comprehensive evaluation index reflecting the coupling relationship between environment, structure and performance;

[0012] The lifetime prediction module is configured to: based on the comprehensive evaluation index, train a degradation model using a long short-term memory network based on historical failure data, use the degradation model to predict the remaining lifetime of the fiber optic connector, and generate a lifetime degradation curve.

[0013] Preferably, the data preprocessing module includes:

[0014] The outlier detection unit is configured to identify and remove outliers in multi-dimensional data using either interquartile range (IQR) or Z-score (Z-score) methods.

[0015] Missing value imputation units are configured to impute missing data using any of the following algorithms: linear interpolation, polynomial interpolation, or K-nearest neighbor interpolation.

[0016] The filtering unit is configured to: reduce noise for temperature and humidity by moving average filtering, reduce noise for vibration parameters by moving average filtering and wavelet threshold filtering, and remove high-frequency noise for insertion loss and return loss by wavelet threshold filtering.

[0017] The normalization unit is configured to use Z-score normalization for temperature, humidity and vibration parameters, and Min-Max normalization for insertion loss, return loss, insertion / removal count and ferrule displacement data.

[0018] Preferably, the method of fusing the temporal, frequency, and spatial features of the preprocessed multi-dimensional data using a self-attention mechanism includes the following steps:

[0019] The preprocessed multi-dimensional data is decomposed into a time-domain feature matrix, a frequency-domain feature matrix, and a spatial feature matrix. The time-domain feature matrix consists of a time-series of environmental and performance parameters, including temperature, humidity, and vibration, and insertion loss and return loss. This time-domain feature matrix has both a time dimension and a time-domain feature dimension. The frequency-domain feature matrix is ​​obtained by extracting frequency components from vibration parameters and insert displacement data using a Fast Fourier Transform. This frequency-domain feature matrix also has both a frequency dimension and a frequency-domain feature dimension. The spatial feature matrix consists of real-time insert displacement data and spatial correlation features derived from three-dimensional structural parameters. This spatial feature matrix also has both a spatial sampling dimension and a spatial feature dimension.

[0020] Multiple attention heads are set up. Each attention head performs matrix multiplication operations on the time-domain feature matrix, frequency-domain feature matrix, and spatial feature matrix with the query weight matrix, key weight matrix, and value weight matrix, respectively, and obtains the corresponding query matrix, key matrix, and value matrix.

[0021] The query matrix and the transpose of the key matrix are multiplied together, then divided by the square root of the key matrix dimension, and finally normalized using the Softmax function to obtain the attention weights. The attention weights are then multiplied with the value matrix to output the feature mapping results of a single attention head. The outputs of all attention heads are concatenated and processed through a linear transformation layer to obtain self-attention features of a single feature type, including time-domain self-attention features, frequency-domain self-attention features, and spatial self-attention features.

[0022] Preferably, the method of fusing the temporal, frequency, and spatial features of the preprocessed multi-dimensional data using a self-attention mechanism further includes the following steps:

[0023] The query matrix corresponding to the time-domain feature matrix is ​​interactively calculated with the key matrix and value matrix corresponding to the frequency-domain feature matrix to obtain the time-domain and frequency-domain cross-attention weights. Then, matrix multiplication is performed between these time-domain and frequency-domain cross-attention weights and the value matrix corresponding to the frequency-domain feature matrix to output the time-domain and frequency-domain cross-features. Similarly, the query matrix of the time-domain feature matrix is ​​interactively calculated with the key matrix and value matrix of the spatial feature matrix to obtain the time-domain and spatial cross-attention weights. Then, matrix multiplication is performed between these time-domain and spatial cross-attention weights and the value matrix corresponding to the spatial feature matrix to output the time-domain and spatial cross-features. Finally, the query matrix of the frequency-domain feature matrix is ​​interactively calculated with the key matrix and value matrix of the spatial feature matrix to obtain the frequency-domain and spatial cross-attention weights. Then, matrix multiplication is performed between these frequency-domain and spatial cross-attention weights and the value matrix corresponding to the spatial feature matrix to output the frequency-domain and spatial cross-features.

[0024] The self-attention features of the single feature type are weighted and fused with the time-domain and frequency-domain cross features, the time-domain and spatial cross features, and the frequency-domain and spatial cross features to obtain a fused feature matrix.

[0025] Preferably, the process of obtaining the fused feature vector includes the following steps:

[0026] The fused feature matrix is ​​divided into time windows. Within each time window, local temporal features are extracted through average pooling. Using a position encoding method based on sine and cosine functions, features at different time positions are represented by sine and cosine waves of different frequencies.

[0027] By using a bidirectional self-attention layer, attention weights are calculated for the feature vector of each window simultaneously with the feature vectors of the preceding and subsequent windows, establishing a bidirectional relationship between windows and realizing cross-window temporal relationship modeling.

[0028] The temporal correlation features of all windows are concatenated, and the concatenated features are compressed to a preset feature dimension through a 1×1 convolutional layer, finally generating a fused feature vector with temporal correlation and preset dimension feature information.

[0029] Preferably, the model building module includes:

[0030] The parameter acquisition unit is configured to acquire the three-dimensional structural parameters and material properties of the fiber optic connector, and the acquisition process combines the physical entity design data and measured data of the fiber optic connector to ensure that the acquired parameters are consistent with the structure and material properties of the physical entity.

[0031] The virtual model building unit is configured to: build core component sub-models of ceramic ferrules, lenses, and sleeves using parametric modeling software based on the three-dimensional structural parameters and material properties, and virtually assemble them according to the actual assembly relationship of the physical entities to form a three-dimensional geometric model of the fiber optic connector; at the same time, establish a stress-strain relationship model of the model based on the material mechanical properties through finite element simulation.

[0032] The dynamic update unit is configured to receive the dynamic parameters of the physical entity transmitted by the data acquisition module in real time, and when the dynamic parameters change beyond the preset range, synchronously adjust the relative positions of the components of the three-dimensional geometric model and the calculation parameters of the stress and strain relationship model.

[0033] Preferably, the joint analysis includes the following steps:

[0034] The fused feature vector and the structural state parameters are aligned in feature dimensions and their data formats are unified.

[0035] The adapted fused feature vector and structural state parameters are input into a deep residual network, and the fused feature vector and structural state parameters are interacted dimension by dimension through the feature fusion layer of the deep residual network.

[0036] The correlation features between the adapted fused feature vector and the structural state parameters are extracted by the residual blocks of the deep residual network. The correlation features include the influence of environmental parameters on the structural state and the relationship between the structural state and the performance parameters.

[0037] Preferably, the output process of the comprehensive evaluation index includes:

[0038] The output layer of the deep residual network, based on a multi-dimensional feature mapping mechanism, maps the extracted correlation features to environmental impact, structural health, and performance degradation, respectively. The environmental impact dimension integrates the correlation features between environmental parameters and structural state, characterizing the direct and indirect effects of environmental factors on the structural state. The structural health dimension integrates the correlation features between structural state and environmental feedback, and between structural state and performance, characterizing the health level of the structure under environmental influence and performance constraints. The performance degradation dimension integrates the correlation features between performance parameters and structural state, characterizing the trend of the impact of structural state changes on performance.

[0039] The final output is a comprehensive evaluation index.

[0040] Preferably, the training process of the degradation model includes:

[0041] Historical failure data is matched with the comprehensive evaluation indicators for the corresponding period to form sample pairs;

[0042] The sample pairs are divided into a training set, a validation set, and a test set. The training set is used to input the Long Short-Term Memory (LSTM) network. The temporal feature learning layer of the LSM network captures the changing patterns of the comprehensive evaluation index over time. Based on the correspondence between index changes and lifetime decay in historical failure data, a mapping model between index changes and lifetime consumption is constructed. The validation set is used to adjust the hyperparameters of the LSM network, and the test set is used to evaluate the generalization ability of the mapping model.

[0043] By iteratively adjusting the parameters of the long short-term memory network, the deviation between the predicted lifetime output by the mapping model and the actual lifetime in the historical failure data is gradually reduced until the deviation stabilizes within a preset range, ultimately obtaining a degradation model that reflects the correlation between changes in comprehensive evaluation indicators and lifetime degradation.

[0044] Preferably, the process of generating the lifetime degradation curve includes the following steps:

[0045] The real-time acquired comprehensive evaluation index is input into the trained degradation model, which outputs the remaining life prediction value corresponding to the current comprehensive evaluation index based on the historical learning of the mapping law between index changes and lifespan degradation.

[0046] The degradation model generates a complete time series based on the predicted remaining lifespan by tracing the changes in comprehensive evaluation indicators over historical time periods. The complete time series includes historical indicator change segments, current indicator nodes, and future indicator prediction segments.

[0047] The complete time series is associated with the corresponding lifespan depletion level by a nonlinear fitting method along the time dimension, so that each time node corresponds to a unique lifespan depletion level. The lifespan decay curve is obtained by presenting the changing trend of the comprehensive evaluation index over time and the decreasing process of the remaining lifespan in the form of a curve.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. The environmental and performance parameters of the fiber optic connector are collected synchronously through the data acquisition module to form multi-dimensional data input, providing complete and relevant data support for subsequent feature extraction and life analysis, and avoiding analysis bias caused by a single data dimension;

[0050] 2. The data preprocessing module cleans and reduces noise in multi-dimensional data, effectively removing abnormal data, supplementing missing data, and reducing noise interference, ensuring that the data input to the feature extraction module is accurate and effective, and laying a high-quality data foundation for subsequent feature fusion and model analysis.

[0051] 3. The feature extraction module uses a self-attention mechanism to fuse the temporal, frequency, and spatial features of the preprocessed data to generate a fused feature vector with temporal correlation. This can accurately capture the dynamic correlation between features of different dimensions and avoid the problem that a single feature dimension cannot reflect the lifespan degradation pattern.

[0052] 4. The model building module constructs a digital twin model based on the three-dimensional structural parameters and material properties of the fiber optic connector, and dynamically updates the model through real-time dynamic parameters obtained by the data acquisition module, so that the virtual model can reflect the changes in the structural state of the physical entity in real time and accurately.

[0053] 5. The fusion analysis module uses a deep residual network to jointly analyze the fused feature vectors and the structural state parameters generated by the digital twin model. This can effectively uncover the coupling relationship between the environment, structure and performance, and output a comprehensive evaluation index that reflects the relationship between the three, providing accurate multi-dimensional evaluation basis for life prediction.

[0054] 6. The lifespan prediction module, based on comprehensive evaluation indicators, trains a degradation model using a long short-term memory network combined with historical failure data, which can accurately predict the remaining lifespan of fiber optic connectors. At the same time, it generates a lifespan degradation curve, which intuitively presents the decreasing process of the remaining lifespan and the changing trend of the comprehensive evaluation indicators, making it easy for maintenance personnel to quickly grasp the equipment's lifespan status. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0056] Figure 1 The diagram shows the structure of the fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion provided in this embodiment of the invention. Detailed Implementation

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0058] In this document, the term "comprising" is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0059] This embodiment provides a method such as Figure 1 The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion shown includes:

[0060] The data acquisition module is configured to acquire multi-dimensional data of the fiber optic connector, including: temperature, humidity and vibration parameters of the environment in which the fiber optic connector is located, which are acquired in real time based on environmental sensors, and insertion loss, return loss, number of insertions and removals and ferrule displacement data of the fiber optic connector, which are acquired in real time based on performance sensors.

[0061] The data preprocessing module is configured to clean and reduce noise in the collected multi-dimensional data.

[0062] The feature extraction module is configured to: fuse the temporal, frequency, and spatial features of the preprocessed multi-dimensional data through a self-attention mechanism to extract and generate a fused feature vector with temporal correlation;

[0063] The model building module is configured to: construct a digital twin model based on the three-dimensional structural parameters and material properties of the fiber optic connector, and dynamically update the model using the physical entity dynamic parameters acquired in real time by the data acquisition module; the three-dimensional structural parameters include the geometric dimensions of the ceramic ferrule and the radius of curvature of the lens, the material properties include the elastic modulus and the coefficient of thermal expansion, and the dynamic parameters include the stress distribution of the ferrule and the lens offset;

[0064] The fusion analysis module is configured to: jointly analyze the fused feature vector with the structural state parameters generated by the digital twin model based on dynamic parameters through a deep residual network, and output a comprehensive evaluation index that reflects the coupling relationship between environment, structure and performance.

[0065] The lifetime prediction module is configured to: train a degradation model based on historical failure data through a long short-term memory network based on comprehensive evaluation indicators, and use the degradation model to predict the remaining lifetime of the fiber optic connector and generate a lifetime degradation curve.

[0066] Based on the above, the fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion in this embodiment constructs a monitoring system that includes full-link data on environment, structure, and performance through the collaborative acquisition of environmental and performance sensors, solving the problem that traditional single-parameter monitoring cannot reflect the complex failure mechanisms of fiber optic connectors. At the same time, a digital twin model is constructed based on three-dimensional structural parameters and material properties, and a bidirectional accurate mapping between the physical entity and the virtual model is achieved through real-time dynamic parameter updates. In addition, a deep residual network fuses self-attention feature vectors and digital twin structural state parameters to output a comprehensive evaluation index that reflects the coupling relationship between environment, structure, and performance, breaking through the limitations of traditional methods that rely on only single-dimensional data.

[0067] In this embodiment, the data preprocessing module includes:

[0068] The outlier detection unit is configured to identify and remove outliers in multi-dimensional data using either interquartile range (IQR) or Z-score (Z-score) methods.

[0069] Missing value imputation units are configured to impute missing data using any of the following algorithms: linear interpolation, polynomial interpolation, or K-nearest neighbor interpolation.

[0070] The filtering unit is configured to: reduce noise for temperature and humidity by moving average filtering, reduce noise for vibration parameters by moving average filtering and wavelet threshold filtering, and remove high-frequency noise for insertion loss and return loss by wavelet threshold filtering.

[0071] The normalization unit is configured to use Z-score normalization for temperature, humidity and vibration parameters, and Min-Max normalization for insertion loss, return loss, insertion / removal count and ferrule displacement data.

[0072] Based on the aforementioned data processing module design, moving average filtering is used for temperature and humidity, moving average and wavelet threshold filtering are combined for vibration parameters, and wavelet threshold filtering is used for insertion loss / return loss. Customized noise reduction schemes are implemented for different data characteristics to improve the signal-to-noise ratio. Simultaneously, Z-score standardization is used for environmental parameters, and Min-Max normalization is used for performance parameters to address model training bias caused by differences in the dimensions of data from different sensors. Furthermore, outliers are removed using interquartile range (INR) or Z-score detection, and missing values ​​are filled using linear interpolation / K-nearest neighbor interpolation to ensure data integrity and avoid misjudgments in lifetime prediction due to data contamination.

[0073] In this embodiment, the temporal, frequency, and spatial features of the preprocessed multi-dimensional data are fused using a self-attention mechanism, including the following steps:

[0074] The preprocessed multi-dimensional data is decomposed into a time-domain feature matrix, a frequency-domain feature matrix, and a spatial feature matrix. The time-domain feature matrix consists of time-series environmental and performance parameters, including temperature, humidity, and vibration, and insertion loss and return loss. It possesses both a time dimension and a time-domain feature dimension. The frequency-domain feature matrix is ​​obtained by extracting frequency components from vibration parameters and insert displacement data using a Fast Fourier Transform. It also possesses both a frequency dimension and a frequency-domain feature dimension. The spatial feature matrix consists of real-time insert displacement data and spatial correlation features derived from three-dimensional structural parameters. It also possesses both a spatial sampling dimension and a spatial feature dimension.

[0075] Multiple attention heads are set up. Each attention head performs matrix multiplication operations on the time-domain feature matrix, frequency-domain feature matrix, and spatial feature matrix with the query weight matrix, key weight matrix, and value weight matrix, respectively, and obtains the corresponding query matrix, key matrix, and value matrix.

[0076] The query matrix and the transpose of the key matrix are multiplied together, then divided by the square root of the key matrix dimension, and finally normalized using the Softmax function to obtain the attention weights. These attention weights are then multiplied with the value matrix to output the feature mapping results of a single attention head. The outputs of all attention heads are concatenated and processed through a linear transformation layer to obtain self-attention features of a single feature type, including time-domain self-attention features, frequency-domain self-attention features, and spatial self-attention features.

[0077] Furthermore, the fusion of temporal, frequency, and spatial features of the preprocessed multi-dimensional data using a self-attention mechanism also includes the following steps:

[0078] The query matrix corresponding to the time-domain feature matrix is ​​interactively calculated with the key matrix and value matrix corresponding to the frequency-domain feature matrix to obtain the time-domain and frequency-domain cross-attention weights. Then, matrix multiplication is performed between these time-domain and frequency-domain cross-attention weights and the value matrix corresponding to the frequency-domain feature matrix to output the time-domain and frequency-domain cross-features. Similarly, the query matrix of the time-domain feature matrix is ​​interactively calculated with the key matrix and value matrix of the spatial feature matrix to obtain the time-domain and spatial cross-attention weights. Then, matrix multiplication is performed between these time-domain and spatial cross-attention weights and the value matrix corresponding to the spatial feature matrix to output the time-domain and spatial cross-features. Finally, the query matrix of the frequency-domain feature matrix is ​​interactively calculated with the key matrix and value matrix of the spatial feature matrix to obtain the frequency-domain and spatial cross-attention weights. Then, matrix multiplication is performed between these frequency-domain and spatial cross-attention weights and the value matrix corresponding to the spatial feature matrix to output the frequency-domain and spatial cross-features.

[0079] We perform weighted fusion of self-attention features of single feature types with time-domain and frequency-domain cross features, time-domain and spatial cross features, and frequency-domain and spatial cross features to obtain a fused feature matrix.

[0080] By employing the aforementioned self-attention mechanism feature fusion and refinement process, the data is decomposed into time-domain, frequency-domain, and spatial feature matrices to capture temporal variation patterns (such as temperature fluctuations over time), frequency components (such as the dominant frequency and harmonics of vibration signals), and spatial geometric relationships (such as the three-dimensional coordinates of ferrule displacement), thereby achieving refined modeling of multimodal information. Simultaneously, multiple attention heads process different feature matrices in parallel, capturing cross-dimensional correlations through querying, key-value interactions, and enhancing the global relevance of feature representations. Furthermore, through cross-attention calculations in the time-domain, time-domain, and frequency-domain, higher-order coupling relationships that are difficult to capture using traditional methods (such as the nonlinear impact of the spatial distribution of ferrule displacement on optical path performance) are uncovered.

[0081] A feasible approach is to obtain the fused feature vectors by including the following steps:

[0082] The fused feature matrix is ​​divided into time windows. Within each time window, local temporal features are extracted through average pooling. The position encoding method based on sine and cosine functions is used to represent the features of different time positions through sine and cosine waves of different frequencies.

[0083] By using a bidirectional self-attention layer, attention weights are calculated for the feature vector of each window simultaneously with the feature vectors of the preceding and subsequent windows, establishing a bidirectional relationship between windows and realizing cross-window temporal relationship modeling.

[0084] The temporal correlation features of all windows are concatenated, and the concatenated features are compressed to a preset feature dimension through a 1×1 convolutional layer, finally generating a fused feature vector with temporal correlation and preset dimension feature information.

[0085] By refining the fused feature vector generation process described above, time window segmentation and average pooling are used to extract local temporal features. Combined with position encoding based on sine / cosine functions, the feature vectors are given absolute temporal position information, which solves the problem of insufficient modeling ability of recurrent neural networks for long sequence dependencies. At the same time, a bidirectional self-attention layer establishes bidirectional dependency relationships between windows, realizes global temporal context awareness, and improves the ability to capture gradual failure processes. In addition, a 1×1 convolutional layer compresses the multi-window concatenation features to a preset dimension, which reduces computational complexity while retaining key information and improves the real-time performance of the model.

[0086] In this embodiment, the model building module includes:

[0087] The parameter acquisition unit is configured to acquire the three-dimensional structural parameters and material properties of the fiber optic connector, and the acquisition process combines the physical entity design data and measured data of the fiber optic connector to ensure that the acquired parameters are consistent with the structure and material properties of the physical entity.

[0088] The virtual model building unit is configured to: build core component sub-models of ceramic ferrules, lenses, and sleeves based on three-dimensional structural parameters and material properties using parametric modeling software, and virtually assemble them according to the actual assembly relationship of physical entities to form a three-dimensional geometric model of the fiber optic connector; at the same time, through finite element simulation, establish a stress-strain relationship model of the model based on the mechanical properties of materials.

[0089] The dynamic update unit is configured to receive the dynamic parameters of the physical entity transmitted by the data acquisition module in real time, and when the dynamic parameters change beyond the preset range, synchronously adjust the relative positions of the components of the three-dimensional geometric model and the calculation parameters of the stress and strain relationship model.

[0090] Through the aforementioned digital twin model construction process, the parameter acquisition unit combines design data and measured data to calibrate the three-dimensional structural parameters (such as the geometric dimension error of the ceramic ferrule ≤ ±0.01mm), ensuring the consistency of the geometric and material properties between the digital twin model and the physical entity, and avoiding simulation distortion caused by parameter deviations. At the same time, a stress-strain relationship model is established based on the mechanical properties of materials, and the mechanical response of the ferrule under dynamic loads is simulated through finite element simulation, realizing a physical-level simulation of the structural health state. In addition, when the dynamic parameters (such as the ferrule stress) change beyond a preset threshold, the three-dimensional geometric model and stress calculation parameters are automatically adjusted to ensure that the digital twin model always reflects the real-time state of the physical entity.

[0091] Furthermore, the joint analysis includes the following steps:

[0092] Align the feature dimensions and unify the data format of the fused feature vector and structural state parameters;

[0093] The adapted fused feature vector and structural state parameters are input into a deep residual network, and the fused feature vector and structural state parameters are interacted dimension by dimension through the feature fusion layer of the deep residual network.

[0094] The correlation features between the adapted fused feature vector and the structural state parameters are extracted by the residual blocks of the deep residual network. The correlation features include the influence of environmental parameters on the structural state and the relationship between the structural state and the performance parameters.

[0095] Secondly, the output process of the comprehensive evaluation indicators includes:

[0096] The output layer of the deep residual network, based on a multi-dimensional feature mapping mechanism, maps the extracted correlation features to environmental impact, structural health, and performance degradation, respectively. The environmental impact dimension integrates the correlation features between environmental parameters and structural state to characterize the direct and indirect effects of environmental factors on the structural state. The structural health dimension integrates the correlation features between structural state and environmental feedback, and between structural state and performance, to characterize the health level of the structure under environmental influence and performance constraints. The performance degradation dimension integrates the correlation features between performance parameters and structural state to characterize the trend of the impact of structural state changes on performance.

[0097] The final output is a comprehensive evaluation index.

[0098] By leveraging the aforementioned process of joint analysis and outputting comprehensive evaluation metrics using deep residual networks, semantic association modeling of environmental, structural, and performance parameters is achieved through feature dimension alignment (e.g., unifying the fused feature vectors and structural state parameters into a 128-dimensional tensor) and dimensional interaction. Simultaneously, residual blocks capture nonlinear correlation patterns (e.g., the exponential relationship between ferrule stress and insertion loss) through skip connections, solving the gradient vanishing problem of traditional neural networks and improving the modeling accuracy of complex coupling relationships. Furthermore, the multi-dimensional output of environmental impact, structural health, and performance degradation provides highly interpretable evaluation metrics, supporting maintenance personnel in locating specific failure points.

[0099] In this embodiment, the training process of the degradation model includes:

[0100] Historical failure data is matched with the comprehensive evaluation indicators for the corresponding period to form sample pairs;

[0101] The sample pairs are divided into training, validation, and test sets. The training set is used to input the Long Short-Term Memory (LSTM) network. The temporal feature learning layer of the LSM network captures the changing patterns of the comprehensive evaluation index over time. Based on the correspondence between index changes and lifetime decay in historical failure data, a mapping model between index changes and lifetime consumption is constructed. The validation set is used to adjust the hyperparameters of the LSM network, and the test set is used to evaluate the generalization ability of the mapping model.

[0102] By iteratively adjusting the parameters of the long short-term memory network, the deviation between the predicted lifetime output by the mapping model and the actual lifetime in historical failure data is gradually reduced until the deviation stabilizes within a preset range, ultimately yielding a degradation model that reflects the correlation between changes in comprehensive evaluation indicators and lifetime degradation.

[0103] Furthermore, the process of generating the lifespan degradation curve includes the following steps:

[0104] The real-time acquired comprehensive evaluation indicators are input into the trained degradation model. Based on the historical learning of the mapping pattern between indicator changes and lifespan degradation, the degradation model outputs the remaining lifespan prediction value corresponding to the current comprehensive evaluation indicator.

[0105] The degradation model generates a complete time series by tracing the changes in comprehensive evaluation indicators over historical time periods and based on the predicted remaining lifespan. The complete time series includes historical indicator change segments, current indicator nodes, and future indicator prediction segments.

[0106] By using a nonlinear fitting method, the complete time series is correlated and mapped with the corresponding degree of lifespan consumption along the time dimension, so that each time node corresponds to a unique degree of lifespan consumption. The trend of the comprehensive evaluation index over time and the process of the decrease in remaining lifespan are presented in the form of a curve to obtain the lifespan decline curve.

[0107] By employing the aforementioned lifetime prediction process, historical failure data is matched with the time series of comprehensive evaluation indicators, dividing the dataset into training, validation, and test sets. A long short-term memory network is used to capture the changing patterns of the indicators over time, constructing an indicator evolution-lifetime depletion mapping model to avoid overfitting risks. Simultaneously, based on real-time comprehensive evaluation indicators, combined with historical evolution trajectories and future prediction segments, a complete time series containing historical, current, and future data is generated, enabling dynamic updates and visualization of remaining lifetime. Furthermore, through nonlinear fitting methods such as multinomial fitting or neural network mapping, the comprehensive evaluation indicators are correlated with the degree of lifetime depletion, generating a continuously traceable degradation curve, thus addressing the problem that traditional threshold alarms cannot reflect the progressive failure process.

[0108] In summary, the fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion in this embodiment achieves accurate lifetime prediction through a closed-loop process encompassing data acquisition, preprocessing, feature fusion, twin modeling, joint analysis, and lifetime prediction. The core logic is as follows:

[0109] First, the data acquisition module, in conjunction with environmental and performance sensors, acquires environmental parameters such as temperature, humidity, and vibration of the fiber optic connector in real time, as well as performance parameters such as insertion loss, return loss, insertion / removal cycles, and ferrule displacement, thus constructing a multi-dimensional data source encompassing environment, structure, and performance. Second, the data preprocessing module purifies the data and eliminates dimensional differences through outlier detection, missing value imputation, targeted filtering, and differential normalization, providing a high-quality data foundation for subsequent analysis.

[0110] Subsequently, the feature extraction module, based on the self-attention mechanism, first decomposes the preprocessed data into temporal, frequency, and spatial feature matrices. It then calculates single-feature-type self-attention features using a multi-head attention head. Next, through cross-attention interactions between the time and frequency domains, time and space domains, and frequency and space domains, combined with time window segmentation, bidirectional self-attention modeling, and 1×1 convolutional dimension compression, it generates a fusion feature vector with strong temporal correlation. Simultaneously, the model building module combines physical entity design data and measured data to obtain three-dimensional structural parameters and material properties. It then builds a three-dimensional geometric model using parametric modeling software and couples it with a finite element stress and strain model. Finally, based on real-time dynamic parameters, it dynamically updates the model, constructing a digital twin model that accurately maps to the physical entity.

[0111] Next, the fusion analysis module uses a deep residual network to first align the dimensions and unify the format of the fused feature vector and the structural state parameters output by the digital twin model. Then, through the feature fusion layer, it interacts dimension by dimension and extracts related features from the residual blocks. Finally, it outputs a comprehensive evaluation index that includes environmental impact, structural health, and performance degradation through multi-dimensional feature mapping. Finally, the lifetime prediction module forms sample pairs by matching historical failure data with the comprehensive evaluation index in time series. After dividing the data into training, validation, and test sets, it trains the degradation model through a long short-term memory network. Based on the real-time comprehensive evaluation index, it outputs the remaining lifetime prediction value and generates a lifetime degradation curve by combining the historical index trajectory with the future prediction segment, thus completing the full life cycle health management of the fiber optic connector.

[0112] In the embodiments provided by this invention, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor may be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments may be performed by a computer program instructing the associated hardware. During implementation, the program may be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media may be any available medium accessible to a computer. Computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0113] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion, characterized in that, include: The data acquisition module is configured to acquire multi-dimensional data of the fiber optic connector, including: temperature, humidity and vibration parameters of the environment in which the fiber optic connector is located, which are acquired in real time based on environmental sensors, and insertion loss, return loss, number of insertions and removals and ferrule displacement data of the fiber optic connector, which are acquired in real time based on performance sensors. The data preprocessing module is configured to clean and reduce noise in the collected multi-dimensional data. The feature extraction module is configured to: fuse the temporal, frequency, and spatial features of the preprocessed multi-dimensional data through a self-attention mechanism to extract and generate a fused feature vector with temporal correlation; The model building module is configured to: construct a digital twin model based on the three-dimensional structural parameters and material properties of the fiber optic connector, and dynamically update the model using the physical entity dynamic parameters acquired in real time by the data acquisition module; the three-dimensional structural parameters include the geometric dimensions of the ceramic ferrule and the radius of curvature of the lens, the material properties include the elastic modulus and the coefficient of thermal expansion, and the dynamic parameters include the ferrule stress distribution and the lens offset; The fusion analysis module is configured to: jointly analyze the fused feature vector with the structural state parameters generated by the digital twin model based on the dynamic parameters through a deep residual network, and output a comprehensive evaluation index reflecting the coupling relationship between environment, structure and performance; The lifetime prediction module is configured to: based on the comprehensive evaluation index, train a degradation model using a long short-term memory network based on historical failure data, use the degradation model to predict the remaining lifetime of the fiber optic connector, and generate a lifetime degradation curve.

2. The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion according to claim 1, characterized in that, The data preprocessing module includes: The outlier detection unit is configured to identify and remove outliers in multi-dimensional data using either interquartile range (IQR) or Z-score (Z-score) methods. Missing value imputation units are configured to impute missing data using any of the following algorithms: linear interpolation, polynomial interpolation, or K-nearest neighbor interpolation. The filtering unit is configured to: reduce noise for temperature and humidity by moving average filtering, reduce noise for vibration parameters by moving average filtering and wavelet threshold filtering, and remove high-frequency noise for insertion loss and return loss by wavelet threshold filtering. The normalization unit is configured to use Z-score normalization for temperature, humidity and vibration parameters, and Min-Max normalization for insertion loss, return loss, insertion / removal count and ferrule displacement data.

3. The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion according to claim 1, characterized in that, The method of fusing the temporal, frequency, and spatial features of preprocessed multi-dimensional data using a self-attention mechanism includes the following steps: The preprocessed multi-dimensional data is decomposed into a time-domain feature matrix, a frequency-domain feature matrix, and a spatial feature matrix. The time-domain feature matrix consists of a time-series of environmental and performance parameters, including temperature, humidity, and vibration, and insertion loss and return loss. This time-domain feature matrix has both a time dimension and a time-domain feature dimension. The frequency-domain feature matrix is ​​obtained by extracting frequency components from vibration parameters and insert displacement data using a Fast Fourier Transform. This frequency-domain feature matrix also has both a frequency dimension and a frequency-domain feature dimension. The spatial feature matrix consists of real-time insert displacement data and spatial correlation features derived from three-dimensional structural parameters. This spatial feature matrix also has both a spatial sampling dimension and a spatial feature dimension. Multiple attention heads are set up. Each attention head performs matrix multiplication operations on the time-domain feature matrix, frequency-domain feature matrix, and spatial feature matrix with the query weight matrix, key weight matrix, and value weight matrix, respectively, and obtains the corresponding query matrix, key matrix, and value matrix. The query matrix and the transpose of the key matrix are multiplied together, then divided by the square root of the key matrix dimension, and finally normalized using the Softmax function to obtain the attention weights. The attention weights are then multiplied with the value matrix to output the feature mapping results of a single attention head. The outputs of all attention heads are concatenated and processed through a linear transformation layer to obtain self-attention features of a single feature type, including time-domain self-attention features, frequency-domain self-attention features, and spatial self-attention features.

4. The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion according to claim 3, characterized in that, The method of fusing the temporal, frequency, and spatial features of preprocessed multi-dimensional data using a self-attention mechanism further includes the following steps: The query matrix corresponding to the time-domain feature matrix is ​​interactively calculated with the key matrix and value matrix corresponding to the frequency-domain feature matrix to obtain the time-domain and frequency-domain cross-attention weights. Then, matrix multiplication is performed between these time-domain and frequency-domain cross-attention weights and the value matrix corresponding to the frequency-domain feature matrix to output the time-domain and frequency-domain cross-features. Similarly, the query matrix of the time-domain feature matrix is ​​interactively calculated with the key matrix and value matrix of the spatial feature matrix to obtain the time-domain and spatial cross-attention weights. Then, matrix multiplication is performed between these time-domain and spatial cross-attention weights and the value matrix corresponding to the spatial feature matrix to output the time-domain and spatial cross-features. Finally, the query matrix of the frequency-domain feature matrix is ​​interactively calculated with the key matrix and value matrix of the spatial feature matrix to obtain the frequency-domain and spatial cross-attention weights. Then, matrix multiplication is performed between these frequency-domain and spatial cross-attention weights and the value matrix corresponding to the spatial feature matrix to output the frequency-domain and spatial cross-features. The self-attention features of the single feature type are weighted and fused with the time-domain and frequency-domain cross features, the time-domain and spatial cross features, and the frequency-domain and spatial cross features to obtain a fused feature matrix.

5. The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion according to claim 4, characterized in that, The process of obtaining the fused feature vector includes the following steps: The fused feature matrix is ​​divided into time windows. Within each time window, local temporal features are extracted through average pooling. Using a position encoding method based on sine and cosine functions, features at different time positions are represented by sine and cosine waves of different frequencies. By using a bidirectional self-attention layer, attention weights are calculated for the feature vector of each window simultaneously with the feature vectors of the preceding and subsequent windows, establishing a bidirectional relationship between windows and realizing cross-window temporal relationship modeling. The temporal correlation features of all windows are concatenated, and the concatenated features are compressed to a preset feature dimension through a 1×1 convolutional layer, finally generating a fused feature vector with temporal correlation and preset dimension feature information.

6. The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion according to claim 1, characterized in that, The model building module includes: The parameter acquisition unit is configured to acquire the three-dimensional structural parameters and material properties of the fiber optic connector, and the acquisition process combines the physical entity design data and measured data of the fiber optic connector to ensure that the acquired parameters are consistent with the structure and material properties of the physical entity. The virtual model building unit is configured to: build core component sub-models of ceramic ferrules, lenses, and sleeves using parametric modeling software based on the three-dimensional structural parameters and material properties, and virtually assemble them according to the actual assembly relationship of the physical entities to form a three-dimensional geometric model of the fiber optic connector; at the same time, establish a stress-strain relationship model of the model based on the material mechanical properties through finite element simulation. The dynamic update unit is configured to receive the dynamic parameters of the physical entity transmitted by the data acquisition module in real time, and when the dynamic parameters change beyond the preset range, synchronously adjust the relative positions of the components of the three-dimensional geometric model and the calculation parameters of the stress and strain relationship model.

7. The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion according to claim 6, characterized in that, The joint analysis includes the following steps: The fused feature vector and the structural state parameters are aligned in feature dimensions and their data formats are unified. The adapted fused feature vector and structural state parameters are input into a deep residual network, and the fused feature vector and structural state parameters are interacted dimension by dimension through the feature fusion layer of the deep residual network. The correlation features between the adapted fused feature vector and the structural state parameters are extracted by the residual blocks of the deep residual network. The correlation features include the influence of environmental parameters on the structural state and the relationship between the structural state and the performance parameters.

8. The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion according to claim 7, characterized in that, The output process of the comprehensive evaluation indicators includes: The output layer of the deep residual network, based on a multi-dimensional feature mapping mechanism, maps the extracted correlation features to environmental impact, structural health, and performance degradation, respectively. The environmental impact dimension integrates the correlation features between environmental parameters and structural state, characterizing the direct and indirect effects of environmental factors on the structural state. The structural health dimension integrates the correlation features between structural state and environmental feedback, and between structural state and performance, characterizing the health level of the structure under environmental influence and performance constraints. The performance degradation dimension integrates the correlation features between performance parameters and structural state, characterizing the trend of the impact of structural state changes on performance. The final output is a comprehensive evaluation index.

9. The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion according to claim 1, characterized in that, The training process of the degradation model includes: Historical failure data is matched with the comprehensive evaluation indicators for the corresponding period to form sample pairs; The sample pairs are divided into a training set, a validation set, and a test set. The training set is used to input the Long Short-Term Memory (LSTM) network. The temporal feature learning layer of the LSM network captures the changing patterns of the comprehensive evaluation index over time. Based on the correspondence between index changes and lifetime decay in historical failure data, a mapping model between index changes and lifetime consumption is constructed. The validation set is used to adjust the hyperparameters of the LSM network, and the test set is used to evaluate the generalization ability of the mapping model. By iteratively adjusting the parameters of the long short-term memory network, the deviation between the predicted lifetime output by the mapping model and the actual lifetime in the historical failure data is gradually reduced until the deviation stabilizes within a preset range, ultimately obtaining a degradation model that reflects the correlation between changes in comprehensive evaluation indicators and lifetime degradation.

10. The fiber optic connector lifetime prediction system based on digital twin and multi-parameter fusion according to claim 9, characterized in that, The process of generating the lifetime degradation curve includes the following steps: The real-time acquired comprehensive evaluation index is input into the trained degradation model, which outputs the remaining life prediction value corresponding to the current comprehensive evaluation index based on the historical learning of the mapping law between index changes and lifespan degradation. The degradation model generates a complete time series based on the predicted remaining lifespan by tracing the changes in comprehensive evaluation indicators over historical time periods. The complete time series includes historical indicator change segments, current indicator nodes, and future indicator prediction segments. The complete time series is associated with the corresponding lifespan depletion level by a nonlinear fitting method along the time dimension, so that each time node corresponds to a unique lifespan depletion level. The lifespan decay curve is obtained by presenting the changing trend of the comprehensive evaluation index over time and the decreasing process of the remaining lifespan in the form of a curve.

Images