Optical module life prediction method of DSP integrated degradation and LSTM model

Abstract

The application relates to the technical field of modeling prediction, and discloses a light module life prediction method of a DSP integrated degradation and an LSTM model. Through a cooperative architecture of a DSP embedded mechanism model filtering an LSTM network time sequence calibration, a multi-parameter coupling degradation model and a Kalman filtering algorithm are embedded into a light module DSP at a calculation level, in-situ, real-time and safe evaluation of a life state is realized, and a traditional mode depending on external calculation resources is changed; at a model level, the explainability of a physical mechanism and the adaptive ability of data driving are fused, fundamental degradation dynamics are described through a nonlinear coupling model, and specificity degradation rules are learned from individual historical data through an LSTM network; and finally, real-time, accurate and adaptive light module life state evaluation and prediction are realized.

Description

Technical Field

[0001] This invention relates to the field of modeling and prediction technology, and in particular to a method for predicting the lifetime of optical modules using DSP-integrated degradation and LSTM models. Background Technology

[0002] In optical communication networks, the reliability of optical modules directly affects the stable operation and maintenance costs of the system. To achieve predictive maintenance, the industry has explored various technologies for predicting the lifespan of optical modules, with the following main solutions:

[0003] 1. Threshold Alarm Method: This method sets fixed alarm thresholds for key operating parameters of the optical module (such as bias current, operating temperature, and received optical power). When the monitored data exceeds the threshold, the system triggers an alarm. However, this method is essentially a reactive or critical fault indication and cannot predict degradation trends in advance. Its disadvantages include a high false alarm rate and an inability to quantify the remaining lifespan of the module, making it difficult to use for accurate maintenance planning.

[0004] 2. Cloud-based Big Data Analytics: This method centrally uploads the operational data of optical modules distributed across the network to a cloud server, utilizing the powerful computing capabilities and statistical models of the cloud for offline analysis and prediction. However, it suffers from inherent response latency, making it unsuitable for scenarios with high real-time requirements. It heavily relies on a continuous and stable network connection, and its predictive function fails when the network is interrupted. Furthermore, it poses risks to data privacy and security, and the continuous transmission and centralized processing of massive amounts of data consumes enormous amounts of network bandwidth and cloud computing resources.

[0005] 3. Simple Linear Extrapolation Method: This method is based on historical monitoring data of one or more performance parameters of the optical module, assuming that its degradation process follows a linear law, and estimates the lifetime through linear regression or extrapolation. However, the degradation of an optical module is a nonlinear physical process affected by multiple factors such as temperature, current, and optical power. Linear models cannot characterize this complex coupling effect and nonlinear degradation dynamics. Therefore, when the module enters the accelerated degradation stage, the prediction accuracy drops sharply, and the model's adaptability and extrapolation reliability are poor.

[0006] Specifically, existing technologies are unable to achieve real-time, accurate, and adaptive lifetime status assessment and prediction within or near the optical module due to isolated parameter analysis, separation of models and hardware, lack of physical mechanism guidance, and neglect of individual differences. Summary of the Invention

[0007] The purpose of this invention is to solve the technical problems of isolated parameter analysis, separation of model and hardware, lack of physical mechanism guidance, and inability to achieve accurate optical module lifetime prediction by ignoring individual differences in optical module lifetime prediction. This invention provides a method for optical module lifetime prediction using DSP-integrated degradation and LSTM models.

[0008] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0009] A method for predicting the lifetime of optical modules integrating DSP degradation and LSTM models includes the following steps:

[0010] The DSP acquires the working state of the optical module, defines the degradation state vector of the optical module, and obtains the degradation rate of the laser through the degradation state vector of the optical module.

[0011] A time-dependent degradation rate model is established by measuring the rate of change of the laser's degradation rate over time. The element values ​​of the matrix structure are fitted using accelerated aging experiments, and a nonlinear term is introduced into the time-dependent degradation rate model to form a nonlinear coupled degradation rate model.

[0012] The forward Euler method is used to discretize the nonlinear coupled degradation rate model to obtain a model predicting future degradation.

[0013] The observations are obtained by predicting the future degradation model and the observation noise. The Kalman filter recursion is performed in the DSP to perform the prediction step and the update step using the observations and the predicted future degradation model to obtain the optimal degradation state estimation sequence.

[0014] The optimal degradation state estimation sequence is processed by the LSTM lifetime prediction model to obtain the lifetime level probability distribution and remaining useful lifetime prediction value of the optical module.

[0015] The joint loss function is used to train and iteratively optimize the lifetime level probability distribution and the remaining useful lifetime prediction value. After training is completed, the lifetime prediction result of the optical module is obtained by inputting the working status of the optical module.

[0016] To address the issues of poor real-time performance and lack of coupling mechanisms caused by isolated parameter analysis and separation of model calculation from hardware in existing technologies, this invention directly integrates and runs a multi-parameter coupled degradation model and Kalman filter algorithm within the DSP chip of the optical module. By defining a degradation state vector that fuses multiple parameters and establishing a nonlinear coupled degradation rate model based on physical mechanisms, and utilizing Kalman filtering for real-time prediction and recursive updates in the DSP, online accurate quantification and optimal state estimation of the coupled degradation effects between multiple parameters such as temperature, bias current, and optical power are achieved. This migrates the core lifetime prediction calculation from the cloud to the edge of the optical module, fundamentally solving the response latency, network dependence, and data privacy risks of cloud-based solutions. It achieves sub-second real-time lifetime state assessment and overcomes the shortcomings of thresholding and linear extrapolation methods in characterizing the nonlinearity of multi-parameter coupling.

[0017] To address the problems of poor model interpretability, weak extrapolation ability, and insufficient generalization ability in existing technologies due to the lack of physical mechanism guidance and the neglect of individual differences, this invention constructs a collaborative prediction framework that uses physical model filtering to drive data-driven network calibration. The state estimation sequence, which represents the individual degradation trajectory of a specific module and is based on a physical model and optimized by Kalman filtering, is input into an LSTM network for deep time series modeling. The lifetime level probability distribution and remaining useful lifetime value are simultaneously optimized through a joint loss function, which enhances the interpretability and extrapolation reliability of the prediction process. By using the LSTM network and the online updated state sequence, the unique degradation characteristics of each module are adaptively learned, thereby significantly improving the lifetime prediction accuracy and model generalization ability of individual modules under different manufacturing tolerances and working environments.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] This invention employs a collaborative architecture of DSP embedded mechanism model filtering LSTM network timing calibration. At the computational level, it embeds a multi-parameter coupled degradation model and Kalman filtering algorithm into the optical module DSP, achieving in-situ, real-time, and secure assessment of lifetime status, thus changing the traditional model that relies on external computing resources. At the model level, it integrates the interpretability of physical mechanisms with data-driven adaptive capabilities, characterizing the fundamental degradation dynamics through a nonlinear coupled model and learning specific degradation patterns from individual historical data using an LSTM network. Ultimately, it achieves real-time, accurate, and adaptive assessment and prediction of optical module lifetime status.

[0020] Furthermore, the optical module lifetime prediction method integrating DSP degradation and LSTM models, wherein obtaining the laser degradation rate through the optical module degradation state vector includes the following steps:

[0021] The DSP collects the laser chip temperature, laser bias current, and output optical power of the optical module in real time to characterize the working status of the optical module.

[0022] The degradation state vector of the optical module is defined based on the working state of the optical module, which characterizes the degradation rate of the laser.

[0023] In the above-described solution, this invention addresses the problems of isolated parameter monitoring, single and indirect degradation state representation, and inability to reflect the comprehensive health status of lasers in existing optical module lifetime prediction methods by acquiring multi-parameter operating status and constructing degradation state vectors. Existing technologies, such as threshold alarms or single-parameter trend analysis, can only monitor temperature, current, power, etc., independently or with shallow correlation. The obtained data is discrete and not integrated into a unified health indicator, failing to directly and quantitatively reflect the evolution of the laser's core performance (i.e., degradation rate), resulting in insufficient input information and unclear physical meaning for subsequent models. This invention uses a DSP to synchronously acquire three key and measurable physical quantities in real time: laser chip temperature, bias current, and output optical power. Based on these, a multi-parameter degradation state vector is defined. This vector is a high-dimensional feature constructed through preliminary physical correlation, designed to directly characterize the laser degradation rate. This is equivalent to introducing physical priors at the data acquisition layer, elevating isolated observations into comprehensive state variables with clear health indications. This invention provides a direct and reliable input foundation for subsequently establishing accurate coupled degradation models, realizing a key shift from monitoring multiple discrete signals to integrated health state characterization. It enhances the intuitiveness, comprehensiveness, and physical interpretability of state assessment, laying a solid data foundation for accurate lifetime prediction.

[0024] Furthermore, in the optical module lifetime prediction method integrating DSP degradation and LSTM models, the degradation rate of the laser is characterized by the following formula:

[0025] ;

[0026] in, Let be the degradation rate of the laser at time step t. Let t be the temperature degradation rate at time step t. Let be the current degradation rate at time step t. Let be the optical power degradation rate at time step t. For reference temperature, This is the rated bias current. For optical power, This is the temperature scaling factor. Let t be the laser chip temperature at time step t. Let be the laser bias current at time step t. The output optical power is at time step t.

[0027] In the above scheme, the formula solves the problem that existing optical module lifetime prediction methods suffer from isolated parameter degradation rate calculations, inconsistent dimensions, and ambiguous physical meanings, making unified and quantitative comprehensive evaluation impossible. Existing technologies typically calculate the absolute changes or simple ratios of each parameter independently. These indicators are greatly affected by the initial values ​​and dimensions of the parameters and are not based on the core concept of rate of change, which directly reflects aging dynamics. This results in the inability to compare and integrate the contributions of different parameters to the overall degradation, leading to high noise in the model input signal and poor comparability. The degradation rate formula defined in this invention uses the instantaneous change rates of temperature, bias current, and output optical power as core features, and respectively... , and Performing dimensional normalization and amplitude standardization of physical quantities makes... All three components are transformed into relative change rates that are independent of the initial operating point, dimensionless, or dimensionlessly proportional, quantifying the intensity of the evolution of each physical field over time. This invention unifies the characterization scale of multi-parameter degradation at the mechanistic level, generating... It not only integrates information from multiple sources, but also directly and purely characterizes the degradation dynamics of the laser's overall performance, providing accurate, standardized, and computable input for the subsequent construction of a high-fidelity nonlinear coupled degradation model. It is a key theoretical foundation for achieving accurate coupling analysis and lifetime prediction.

[0028] Furthermore, the optical module lifetime prediction method integrating DSP degradation and LSTM models, wherein forming the nonlinear coupled degradation rate model includes the following steps:

[0029] A continuous-time degradation rate model for lasers is established by considering the influence of laser degradation rate on its own changes and the direct driving effect of operating conditions on degradation.

[0030] Using a continuous-time degradation rate model, the state transition matrix and driving input matrix are determined to be 3×3 matrix structures, and the elements of the matrix structure are theoretically derived.

[0031] The specific values ​​of the elements of the state transition matrix and the driving input matrix are fitted by accelerated aging experiments;

[0032] By introducing nonlinear terms into the continuous-time degradation rate model, highly nonlinear physical effects are captured, forming a nonlinear coupled degradation rate model.

[0033] In the above-mentioned scheme, this invention solves the problems of low prediction accuracy and weak extrapolation ability of existing degradation models due to oversimplification (such as linear or single-parameter models), lack of physical coupling foundation, and empirical parameter setting. Existing technologies typically use empirical formulas or simple statistical relationships to describe degradation, and their model structures and parameters often deviate from physical mechanisms, failing to accurately characterize the dynamic process of the interaction and joint driving of degradation among temperature, current, and power, leading to prediction failure under complex or changing operating conditions. This invention first establishes a continuous-time state-space equation with the degradation rate vector as the state variable, starting from physical principles. It theoretically derives the state transition matrix reflecting the evolution of the degradation rate itself and the driving input matrix reflecting the direct driving force of operating conditions. Its 3x3 structure clarifies all possible coupling paths between parameters (such as how temperature changes affect the current degradation rate), laying a solid physical foundation for the model. By precisely fitting each element in the matrix using accelerated aging experimental data, the theoretical model acquires realistic and quantitative parameters, realizing the transformation from a theoretical framework to a concrete, computable model. Nonlinear terms are introduced into the model to capture highly nonlinear degradation phenomena present in actual physical processes (such as thermal accumulation effects and material aging acceleration points). This invention's modeling process, combining theoretical structure, experimental calibration, and nonlinear enhancement, constructs a coupled degradation rate model that possesses clear physical interpretability, closely matches real device degradation data, and can describe complex nonlinear behavior. This significantly enhances the extrapolation prediction accuracy and reliability of the model under different stress conditions.

[0034] Furthermore, the optical module lifetime prediction method integrating DSP degradation and LSTM models, wherein the formula for predicting future degradation is:

[0035] ;

[0036] in, It is a nonlinear coupled degradation rate model. To predict future degradation models, Where k is the sampling period and k is the discrete time step. This represents the true degenerate state at a discrete time step k. Let k be the input vector at the discrete time step k. Let k be the nonlinear coupling term at the discrete time step k. Let k be the noise vector for the discrete-time step k. Here is the state transition matrix. To drive the input matrix, The discrete state transition matrix, For discrete input driving matrices, For discrete nonlinear coupling terms, For discrete process noise, , , , .

[0037] In the above-described solution, this invention addresses the problem that complex continuous degradation models cannot be solved and recursively solved in real time on resource-constrained embedded hardware (such as DSPs) through model discretization based on the forward Euler method and embedded iterative computation. While existing cloud-based analysis or offline models can handle complex equations, they suffer from response latency, and simple edge computing solutions lack accuracy due to overly simplified models. This invention, through the forward Euler method, transforms the nonlinear coupled degradation rate model describing continuous-time dynamics into a discrete difference equation with a sampling period Ts, defines the discretized form of process noise, and ultimately obtains a prediction equation that can be directly iteratively computed in a DSP. The prediction model encapsulates all coupling and nonlinear effects and allows the Kalman filter algorithm to run efficiently recursively over discrete time steps. This invention transforms a theoretically accurate continuous model losslessly into an operational form suitable for real-time execution on edge computing hardware. It is the core hub for enabling complex mechanistic models to run online within the DSP chip of an optical module, completing millisecond-level state prediction and updates, fundamentally bridging the critical path between high-precision physical models and low-latency embedded applications.

[0038] Furthermore, the optical module lifetime prediction method integrating DSP degradation and LSTM models, wherein obtaining the optimal degradation state estimation sequence includes the following sub-steps:

[0039] The observed values ​​are obtained based on the actual conditions and the observation noise;

[0040] In the DSP, Kalman filtering is used for the prediction step. The prediction step uses the optimal degradation state estimate at discrete time step k and the prediction model for future degradation to predict the degradation state estimate at discrete time step k+1, and updates the covariance of the estimation error.

[0041] In the DSP, Kalman filtering is used for the update step. The update step uses the observations of discrete time step k+1 to correct the degenerate state estimate of discrete time step k+1, and obtains the optimal degenerate state estimate of discrete time step k, and updates the covariance of the estimation error.

[0042] After performing Kalman filtering recursion in the prediction and update steps in the DSP, the DSP outputs the optimal degenerate state estimation sequence. .

[0043] In the above-described scheme, this invention solves the problems of inaccurate and unstable state sequences, which ultimately affect the accuracy of lifetime prediction, caused by lagging degradation state tracking, sensitivity to measurement noise, and inability to adaptively correct online, through DSP-embedded Kalman filter recursive estimation and online correction. Existing technologies typically use raw observations or simply filtered data as model input. Such data contains significant noise and does not fully utilize the system dynamics model, failing to provide optimal and smooth estimates of the true degradation state, especially with large errors under high noise or abrupt changes in operating conditions. This invention implements a complete Kalman filter recursive process within the DSP. The prediction step utilizes a discretized nonlinear coupled degradation model, starting from the optimal estimate of the previous time step to predict the state at the current time step, ensuring the physical consistency of state evolution. The update step immediately introduces the actual observations at the current time step, using a Kalman gain dynamic tradeoff model to predict and optimally correct the predicted values ​​based on the measured information. This prediction-correction closed loop is executed in real-time and cyclically in the DSP, effectively filtering out observation noise and adaptively correcting the effects of model errors and process noise online. The final output optimal degradation state estimation sequence is a high-quality time-series signal that has been dynamically optimized, smoothed, and more closely approximates the actual degradation trajectory. This invention provides a clean, reliable, and physically consistent input sequence for subsequent LSTM networks, and is a key step in extracting high-quality degradation features from noisy observation data and ensuring the accuracy of the entire prediction chain.

[0044] Furthermore, a method for predicting the lifetime of an optical module integrating DSP degradation and LSTM models is provided. The LSTM lifetime prediction model includes an input layer, two hidden layers, an attention layer, a fully connected layer, and a two-branch output layer. The two-branch output layer includes a classification output layer and a regression output layer. The two hidden layers include a first hidden layer and a second hidden layer. Both the first hidden layer and the second hidden layer include a forget gate, an input gate, a cell state module, an output gate, and a hidden state output layer. The cell state module includes a candidate cell state module and a cell state update module.

[0045] Furthermore, in the DSP-integrated degradation and LSTM model optical module lifetime prediction method, the input layer receives the optimal degradation state estimation sequence and outputs it to a double hidden layer. The double hidden layer extracts the short-term local temporal dependence and the long-term global degradation trend respectively and outputs them to the Attention layer. The Attention layer strengthens the feature weights of key degradation time steps and outputs them to the fully connected layer. The fully connected layer fuses high-order features and outputs them to a double-branch output layer. The double-branch output layer outputs the lifetime level probability distribution and the remaining useful lifetime prediction value of the optical module.

[0046] In the above-mentioned scheme, this invention solves the problems of existing data-driven lifetime prediction models, such as their simple structure, difficulty in simultaneously capturing local details and long-term trends of the degradation process, and inability to provide both probabilistic and deterministic prediction information, through hierarchical temporal feature extraction and a dual-task collaborative output network. Existing technologies often struggle to effectively handle complex temporal dependencies in long sequences, and their single output format cannot simultaneously meet the decision-making requirements of reliability assessment (lifespan level) and accurate maintenance (remaining lifetime). The LSTM lifetime prediction model constructed in this invention adopts a deep hierarchical and multi-functional collaborative design. The dual-hidden-layer LSTM structure, through a gating mechanism, focuses on extracting short-term state change patterns and long-term gradual degradation trends, respectively, achieving multi-scale deep feature mining of the degradation sequence. The introduction of the Attention layer dynamically evaluates and strengthens the contribution of features at different time steps to the final prediction, enabling the model to focus on key degradation stages and effectively improving the efficiency and discriminative power of feature utilization. The classification output layer provides the probability distribution of lifetime levels, enhancing the reliability and interpretability of the prediction results. The regression output layer directly outputs accurate remaining useful lifetime values. Both branches share deep spatiotemporal features extracted by the preceding network. Through joint optimization, the model possesses powerful capabilities in both pattern recognition and numerical regression. This invention organically combines temporal modeling, feature focusing, and multi-task learning, enabling the model to learn richer and more discriminative degradation representations from the optimal state estimation sequence. This significantly improves the accuracy, robustness, and decision support value of the final lifetime prediction.

[0047] In the aforementioned schemes, existing contrastive learning loss functions suffer from insufficient supervision signals and unreasonable negative sample weight allocation, making it difficult for the model to effectively distinguish negative samples of varying difficulty, thus limiting the discriminative performance of feature learning. The traditional InfoNCE loss function assigns the same weight to all negative samples, failing to reflect the semantic differences between different negative samples and anchor samples, and lacks a targeted learning mechanism for difficult negative samples. The improved contrastive learning loss function of this invention introduces a distance weighting factor, dynamically adjusting the contribution of negative samples based on their similarity to anchor samples, making the model focus more on challenging, difficult negative samples; by scaling the distance, it finely adjusts the distribution characteristics of the feature space, enhancing the model's ability to perceive subtle differences. This invention, through improved loss function calculation, achieves differentiated supervision of positive and negative samples, significantly improving the discriminative power of feature representation and the model's generalization ability. It retains the advantages of traditional contrastive learning while overcoming its rigid weight allocation defects, enabling the model to better adapt to the complex distribution characteristics of industrial water treatment data.

[0048] Furthermore, in the optical module lifetime prediction method integrating DSP degradation and LSTM models, the processing procedure of the first hidden layer is as follows:

[0049] The forget gate concatenates the hidden state vector at time step t-1 and the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the cell state module.

[0050] The input gate concatenates the hidden state vector at time step t-1 with the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the cell state module.

[0051] The candidate cell state module concatenates the hidden state vector at time step t-1 with the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then normalizes the data through the hyperbolic tangent activation function before outputting it to the cell state update module.

[0052] The cell state update module performs element-wise product of the output data of the forget gate at time step t and the updated cell state at time step t-1, and performs element-wise product of the output data of the input gate at time step t and the candidate cell state at time step t. The element-wise products are then added together and output to the hidden state output layer.

[0053] The output gate concatenates the hidden state vector at time step t-1 with the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the hidden state output layer.

[0054] The hidden state output layer normalizes the cell state updated at time step t using the hyperbolic tangent activation function. After normalization, it performs element-wise multiplication with the output data of the output gate at time step t and outputs it to the second hidden layer.

[0055] In the above-described solution, this invention addresses the problems of gradient vanishing or exploding, difficulty in selective memorization and forgetting, and weak ability to capture long-range degradation dependencies in existing simple recurrent neural networks when modeling long-term sequences, through gated recurrent unit collaboration and refined cell state updates. Existing technologies using ordinary RNNs or simple time-series models have a single internal state update mechanism and lack refined control over information flow. They cannot dynamically determine how much historical information to retain and how much new information to update during the learning process. Therefore, when processing long and complex degradation sequences of optical modules, they struggle to maintain long-term memory of early key degradation features, affecting prediction accuracy. The first hidden layer LSTM unit of this invention achieves intelligent management of temporal information through precise coordination of the forget gate, input gate, candidate cell state, cell state update module, and output gate. The forget gate uses the sigmoid function to determine which secondary information to discard from historical cell states; the input gate works with the candidate cell state to filter and transform the new information currently input; the cell state update module completes a smooth and controllable update of long-term memory (cell state) through element-wise multiplication and summation; and the output gate controls how many of the updated cell states are output as the current hidden state based on the current input and hidden state. This invention enables the model to adaptively learn complex temporal patterns in degradation sequences, laying a solid foundation for the second hidden layer to further extract deeper and more global temporal features, thereby significantly enhancing the entire network's ability to represent the dynamic evolution of degradation.

[0056] Furthermore, the joint loss function, through training and iterative optimization of the lifetime level probability distribution and the remaining useful lifetime prediction, includes the following steps:

[0057] The classification loss of the classification output layer is calculated using the cross-entropy loss function.

[0058] The regression loss of the regression output layer is calculated using the mean squared error loss function.

[0059] The prediction accuracy of the classification output layer and the regression output layer is optimized by combining classification loss and regression loss.

[0060] In the above-described scheme, this invention addresses the problem in existing lifespan prediction models that suffer from a single training objective and an inability to simultaneously optimize classification and regression tasks, leading to a trade-off between reliability and accuracy in prediction results. Existing technologies typically employ a single loss function (e.g., using only mean squared error for regression or only cross-entropy for classification), forcing the model to prioritize the optimization of a single task while neglecting another important objective. For example, optimizing only the regression loss may result in a lack of confidence in predictions near lifespan level boundaries; while optimizing only the classification loss may fail to provide accurate remaining lifespan values, limiting the practical value of the prediction results. This invention jointly optimizes the cross-entropy loss for classification and the mean squared error loss for regression, ensuring that the model, during training, simultaneously considers two closely related but distinct objectives: correctly classifying lifespan levels and accurately predicting remaining lifespan values. By balancing these two losses, the feature representation learned by the model possesses both clear class discriminative power and maintains accurate fitting capabilities to continuous degradation. This invention achieves a synergistic improvement in classification accuracy and regression accuracy, ensuring that the final prediction results can both assess reliability risk in probabilistic form and provide accurate maintenance time references, thereby significantly enhancing the overall performance and decision support capabilities of the prediction model. Attached Figure Description

[0061] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 A flowchart illustrating a method for predicting the lifetime of optical modules by integrating DSP degradation and LSTM models.

[0063] Figure 2 A flowchart for obtaining the optimal degenerate state estimation sequence.

[0064] Figure 3 This is a schematic diagram of the LSTM lifetime prediction model. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0066] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between elements or an indirect connection via other elements.

[0067] like Figure 1 As shown, a method for predicting the lifetime of an optical module integrating DSP degradation and LSTM models includes the following steps:

[0068] S1: The DSP acquires the working status of the optical module, defines the degradation state vector of the optical module, and obtains the degradation rate of the laser through the degradation state vector of the optical module.

[0069] S1 includes:

[0070] S11: The DSP collects the laser chip temperature, laser bias current, and output optical power of the optical module in real time to characterize the operating status of the optical module. The formula is:

[0071] ;

[0072] in, The working state of the optical module at time step t. Let t be the laser chip temperature at time step t (in °C). Let t be the laser bias current at time step t (in mA). Let t be the output optical power at time step t (in mW or dBm, with a unified dimension in the calculation), where t is the continuous time.

[0073] Specifically, The temperature sensor from the TEC is fed into the DSP chip. The current monitoring points originating from the drive circuit are sent to the DSP chip. The photodetector and ADC from the optical module are then fed into the DSP chip. For the working state of the optical module, the basic modeling core is the laser chip temperature, laser bias current, and output optical power. Their degradation is not reflected in absolute values, but in the trend and rate of change, which is the key to constructing the degradation state vector.

[0074] S12: Define the optical module degradation state vector based on the optical module's operating state to characterize the laser's degradation rate:

[0075] ;

[0076] in, Let be the degradation rate of the laser at time step t. Let t be the temperature degradation rate at time step t. Let be the current degradation rate at time step t. Let be the optical power degradation rate at time step t. For reference temperature, This is the rated bias current. For optical power, This is the temperature scaling factor.

[0077] In the embodiments, , and Used for normalization to eliminate the influence of dimensions. , making and , With similar orders of magnitude, the derivative can be approximated by a difference, for example... , Let be the bias current for the discrete time step k. The bias current is the discrete time step k-1. The sampling period.

[0078] It is important to note that time step t is a continuous time, describing the continuous changes of the laser on the actual timeline, while k is the discrete time step, describing the discrete data point sequence actually processed by the DSP.

[0079] Specifically, regarding the laser's operating state: The laser generates heat during operation. Poor heat dissipation or excessively high drive temperature will lead to an increase in junction temperature. This temperature increase accelerates all aging processes. The continuous upward trend in temperature is itself a sign of thermal management failure (such as dried thermal paste or loose heatsink). Therefore, the rate of temperature degradation is more about the rate of change. Laser aging leads to a decrease in its electro-optical conversion efficiency. To maintain a constant output optical power, the control circuit slowly increases the bias current (for FP / DFB lasers). The monotonic rate of increase of the bias current... The degradation rate of the laser was quantified; even with automatic power control circuitry, laser aging can cause it to fail to reach the target optical power, at which point the optical power begins to decrease monotonically, and the rate of decrease in optical power is... The degradation rate of the laser was also quantified.

[0080] S2: A time-dependent degradation rate model is established by using the rate of change of the laser's degradation rate over time. The element values ​​of the matrix structure are fitted by accelerated aging experiments, and a nonlinear term is introduced into the time-dependent degradation rate model to form a nonlinear coupled degradation rate model.

[0081] S2 includes:

[0082] S21: By considering the influence of the laser's degradation rate on its own variation and the direct driving force of operating conditions on degradation, a continuous-time degradation rate model for the laser is established, with the following formula:

[0083] ;

[0084] in, Let be the degradation rate of the laser at time step t. This is a continuous-time degradation rate model for lasers. Let t be the degradation term at time step t, i.e., how the current degradation rate of the laser affects the future degradation rate of the laser. It is an externally driven type at time step t, that is, the absolute value of the working condition at time step t. How does it affect the degradation rate of a direct laser? Here is the state transition matrix. This is to drive the input matrix.

[0085] Specifically, because It is not a constant, its rate of change Driven by both its own changes and external working conditions, a basic linear model is established to assume a time degradation rate model for the laser.

[0086] S22: Using a continuous-time degradation rate model, the state transition matrix and driving input matrix are determined to have a 3×3 matrix structure. The elements of the matrix structure are theoretically derived as follows:

[0087] ;

[0088] in, , , , This is the temperature degradation self-damping coefficient. Current degradation rate Effect on temperature degradation acceleration Temperature degradation rate Effect on current degradation acceleration This is the current degradation self-damping coefficient. Optical power degradation rate Effect on current degradation acceleration Current degradation rate The effect on the acceleration of optical power degradation The self-damping coefficient for optical power degradation;

[0089] in, , The contribution of laser chip temperature to the acceleration of temperature degradation. The contribution of laser bias current to current degradation acceleration. The contribution of output optical power to the acceleration of optical power degradation.

[0090] It is important to note that in the process of system modeling, the system structure (in matrix form) is often determined first, and then the meaning of each element is explained.

[0091] Specifically, regarding the state transition matrix In the analysis, the diagonal elements are usually negative to indicate the cross-coupling between parameter degradations (such as current degradation acceleration). (It may gradually slow down with aging), other elements Current degradation rate The effect of temperature acceleration (thermal effect), i.e. for An increase in bias current leads to increased heat generation, which may accelerate the temperature rise. ; Temperature degradation rate The effect on current degradation acceleration, i.e. for Increased temperature accelerates the degradation of current, therefore ; Optical power degradation rate The effect on current degradation acceleration, i.e. for When optical power decreases, the APC circuit increases the current to compensate, thus accelerating current degradation. ; Current degradation rate The acceleration of optical power degradation, i.e. for Current degradation typically means a decrease in laser efficiency, leading to accelerated optical power degradation. ; , It may be small; it can be initially set to 0.

[0092] In the embodiment, the state transition matrix The example table of elements is shown in Table 1;

[0093] Table 1: State Transition Matrix Example table of elements;

[0094] ;

[0095] Specifically, for the driving input matrix The analysis is performed, and the matrix is ​​designed as a diagonal or approximately diagonal array to illustrate the effect of input (absolute operating conditions) on degradation acceleration. Specifically, this represents the laser chip temperature at time step t. acceleration of temperature degradation The contribution, namely for Operating at high temperatures accelerates the thermal aging process, therefore ; Specifically, this represents the laser bias current at time step t. Current degradation acceleration The contribution, namely for The higher the operating current, the faster the current degradation. ; Specifically, it represents the output optical power at time step t. acceleration of optical power degradation The contribution, namely for The higher the output optical power, the heavier the burden on the laser, and the faster the optical power decays. .

[0096] In the embodiment, the input matrix The example table of elements is shown in Table 2;

[0097] Table 2: Input Matrix Example table of elements;

[0098] ;

[0099] S23: Fit the specific values ​​of the elements of the state transition matrix and the driving input matrix through accelerated aging experiments.

[0100] S231: Group several optical modules to different acceleration stress conditions, with sampling period Record the laser chip temperature sequence for each group Laser bias current sequence Output optical power sequence ;

[0101] Where n is the total number of sampling points in the accelerated aging experiment;

[0102] S232: Perform differential processing on the collected raw data to obtain the observed degradation rate, using the following formula:

[0103]

[0104] in, The temperature degradation rate is normalized to the discrete time step k. The current degradation rate is normalized to the discrete time step k. The optical power degradation rate is normalized to the discrete time step k. For the discrete time step k, the degradation rate observation is... Let k be the temperature of the laser chip with a discrete time step. Let k be the laser bias current with discrete time step k. The bias current is the discrete time step k-1. Output optical power for discrete time step k;

[0105] S233: Calculate the degradation acceleration observations from the degradation rate observations, using the following formula:

[0106] ;

[0107] in, For discrete time step k, the degenerate acceleration observations are... The observed degradation rate is given at a discrete time step of k-1.

[0108] In the embodiments, different accelerated stress conditions include a high-temperature group. , High current group 1.5, 2.0, high temperature and high current combined group and 1.5, And 2.0.

[0109] S234: Substitute the observed degradation acceleration values ​​into the continuous-time degradation rate model, and solve for the specific values ​​of the elements of the matrix structure in the continuous-time degradation rate model using the coordinate descent method.

[0110] S24: Introducing nonlinear terms into the continuous-time degradation rate model captures highly nonlinear physical effects, forming a nonlinear coupled degradation rate model:

[0111] ;

[0112] in, It is a nonlinear coupled degradation rate model. It is a nonlinear coupling term. For the process noise at time step t, This represents the temperature-accelerated gain coefficient. For activation energy, Boltzmann's constant, The thermal-electric cross-coupling coefficient, This is the temperature-optical power degradation coupling coefficient.

[0113] Specifically, This is used to capture the accelerating effect of temperature on current degradation and the feedback effect of current and heat. Specifically, the accelerating effect of temperature on current degradation is that increased temperature significantly accelerates the aging process of semiconductor devices; that is, the accelerating effect of temperature on current degradation and... It is proportional to the current degradation rate at time step t; therefore, the degradation rate of the current at time step t is proportional to the current degradation rate at time step t. It should be regulated by an exponential term related to temperature, namely Temperature-accelerated gain coefficient Further explanation of the current degradation rate at time step t The current itself is amplified exponentially with increasing temperature; the feedback effect of current and heat is as follows: increased temperature leads to an increase in the laser threshold current, and the APC circuit further increases the bias current to compensate, resulting in increased heat generation, forming a positive feedback loop. Therefore, it reflects... and The product term pairs or The contribution, namely , The direct effect of temperature on the rate of optical power degradation is described, capturing the thermally induced optical decay effect caused by temperature rise, i.e., as temperature increases, the rate of decrease in output optical power accelerates, thus obtaining... .

[0114] In the embodiments, , , Initially, the parameters can be determined through accelerated aging experiments, including temperature, current, optical power, and duration. These parameters are obtained through repeated experiments. Later, fine-tuning is achieved through online learning using a DSP. The reaction model does not consider the disturbance of factors and is set as Gaussian white noise.

[0115] S3: The forward Euler method is used to discretize the nonlinear coupled degradation rate model to obtain a model for predicting future degradation.

[0116] ;

[0117] in, To predict future degradation models, Where k is the sampling period and k is the discrete time step. This represents the true degenerate state at a discrete time step k. Let k be the input vector at the discrete time step k. Let k be the nonlinear coupling term at the discrete time step k. Let k be the noise vector for the discrete-time step k. The discrete state transition matrix, For discrete input driving matrices, For discrete nonlinear coupling terms, For discrete process noise, , , , .

[0118] Specifically, It describes how a degenerate state evolves to the next time step, quantifying the impact of the current degenerate state on the degenerate state at the next time step. This describes the direct driving effect of current operating conditions on degradation evolution, primarily quantifying how absolute operating parameters (temperature, current, optical power) accelerate or slow down the degradation process. It primarily describes complex physical interactions that go beyond linear models, including temperature-accelerated effects, thermoelectric cross-coupling effects, and thermo-induced optical decay effects. That is, random factors and uncertainties that are not considered in the model for predicting future degradation.

[0119] It is important to note that since the nonlinear coupled degradation rate model cannot be directly run in a DSP, it is discretized. The forward Euler method is simple to calculate and suitable for DSPs. The model for predicting future degradation is the core degradation state model that is recursively updated in each sampling period in the DSP. , These are all constant matrices calculated offline, requiring only initialization. The DSP can then recursively calculate the predicted future degradation model based on the discrete time step k-optical module measurements in each sampling period. .

[0120] S4: Obtain observations by predicting the future degradation model and observation noise. Use the observations and the predicted future degradation model to perform Kalman filtering recursion in the DSP for prediction and update steps to obtain the optimal degradation state estimation sequence.

[0121] like Figure 2 As shown, S41: Based on the actual state and observation noise, obtain the observed values:

[0122] ;

[0123] in, For the observations at discrete time step k, This represents the true degenerate state at a discrete time step k. The observation noise is for the discrete time step k;

[0124] S42: Kalman filtering is used in the DSP for the prediction step. The prediction step uses the optimal degradation state estimate at discrete time step k and the model predicting future degradation to predict the degradation state estimate at discrete time step k+1, updating the covariance of the estimation error:

[0125] ;

[0126] in, For prior state estimation with discrete time step k+1. Let be the prior state covariance matrix at the k+1th discrete time step (before the observations are obtained), and T be the matrix transpose. To estimate the error covariance matrix, Let be the covariance matrix of the process noise. This is the optimal degradation state estimate for discrete time step k.

[0127] S43: Kalman filtering is used in the DSP for the update step. The update step uses the observations at discrete time step k+1 to correct the degenerate state estimate at discrete time step k+1, thus obtaining the optimal degenerate state estimate at discrete time step k. The covariance of the update estimation error is:

[0128] ;

[0129] in, Here is the Kalman gain matrix with a discrete time step of k+1. Let be the prior state covariance matrix at the k+1th discrete time step (before the observations are obtained). The covariance matrix of the observed noise, This represents the optimal degenerate state estimate for a discrete time step of k+1. It is the identity matrix. This is the updated posterior covariance matrix. The observations are for a discrete time step of k+1.

[0130] S44: After performing Kalman filtering recursion in the prediction and update steps in the DSP, the DSP outputs the optimal degenerate state estimation sequence. .

[0131] It is important to note that in Kalman filtering, the prediction step uses a model predicting future degradation to predict the state, while the process noise at the discrete time step k is... Since it is zero-mean random noise, its specific value cannot be predicted. Therefore, in the prediction step, the deterministic part of the model predicting future degradation is used for prediction, while the statistical properties of the process noise (covariance matrix) are used. Considering the update of the prediction error covariance; the prediction step uses the past to predict the future, while the update step uses the present to correct the previous prediction and obtain the optimal result.

[0132] S5: The optimal degradation state estimation sequence is processed by the LSTM lifetime prediction model to obtain the lifetime level probability distribution and the remaining useful lifetime prediction value of the optical module.

[0133] Specifically, the input layer receives the optimal degradation state estimation sequence and outputs it to the two hidden layers. The two hidden layers extract the short-term local temporal dependencies and the long-term global degradation trends respectively and output them to the Attention layer. The Attention layer strengthens the feature weights of the key degradation time steps and outputs them to the fully connected layer. The fully connected layer fuses high-order features and outputs them to the dual-branch output layer. The dual-branch output layer outputs the lifetime level probability distribution and the remaining useful lifetime prediction of the optical module.

[0134] like Figure 3 As shown, the LSTM lifetime prediction model includes an input layer, two hidden layers, an attention layer, a fully connected layer, and a two-branch output layer. The two-branch output layer includes a classification output layer and a regression output layer. The two hidden layers include a first hidden layer and a second hidden layer. Both the first and second hidden layers include a forget gate, an input gate, a cell state module, an output gate, and a hidden state output layer. The hierarchical structure of the first and second hidden layers is the same, except that the parameter dimensions are adapted to the corresponding number of units. The first hidden layer has 16 units, and the second hidden layer has 8 units. The cell state module includes a candidate cell state module and a cell state update module.

[0135] S51: Using the sliding window method, for each time step k (k≥L), extract the previous... The optimal estimate of the degradation state at each time step is used as the input feature, and the input sequence is fed into the input layer of the LSTM lifetime prediction model.

[0136] ;

[0137] in, The input sequence is the LSTM lifetime prediction model with discrete time step k. The time window length, , Given the dimensional space of the input sequence, 3 includes the temperature degradation rate. Bias current degradation rate Optical power degradation rate , For the first The estimated optimal degenerate state at time t. .

[0138] In the embodiments, The specific value depends on the sampling period. Adaptation, for example, setting hour, The total time span is The input sequence covers historical data from the past 5 minutes.

[0139] The output of the input layer is connected to the input of the double hidden layer, the output of the double hidden layer is connected to the input of the Attention layer, the output of the Attention layer is connected to the input of the fully connected layer, and the output of the fully connected layer is connected to the double-branch output layer.

[0140] Specifically, at the beginning of the input sequence processing for each LSTM lifetime prediction model, the hidden states and cell states of the LSTM lifetime prediction model need to be initialized, typically using zero initialization:

[0141] Initial state of the first hidden layer: , ;

[0142] Initial state of the second hidden layer: , .

[0143] S52: LSTM lifetime prediction model with internal processing time step At time step t, the degenerate state vector of the input layer output is obtained. After splicing the hidden state at time step t-1, it passes through the forget gate, input gate, cell state module and output gate of the first hidden layer to realize the memory and update of short-term local temporal features and long-term global degradation features. The hidden state output layer is updated to obtain the state at time step t, and the state at time step t is passed to time step t+1 and output to the second hidden layer.

[0144] in, , , , Let be the degenerate state vector at time step t.

[0145] Specifically, the processing procedure for the first hidden layer is as follows:

[0146] S521: The forget gate concatenates the hidden state vector at time step t-1 and the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the cell state module.

[0147] ;

[0148] Among them, The output data of the forget gate at time step t in the first hidden layer is used to control the proportion of historical cell states retained. , Here is the weight matrix of the forget gate in the first hidden layer. This represents the hidden state of the first hidden layer at time step t-1. Let be the bias vector of the forget gate in the first hidden layer. This is the activation function.

[0149] S522: The input gate concatenates the hidden state vector at time step t-1 and the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the cell state module.

[0150] ;

[0151] in, This is the output data of the input gate in the first hidden layer at time step t, used to control the proportion of the current feature written into the cell state. , Here is the weight matrix of the input gate of the first hidden layer. is the bias vector of the input gate of the first hidden layer.

[0152] S523: The candidate cell state module concatenates the hidden state vector at time step t-1 with the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, normalizes the data through the hyperbolic tangent activation function, and then outputs it to the cell state update module.

[0153] ;

[0154] in, This represents the candidate cell state at time step t in the first hidden layer, used to store newly added features at the current time step. , This is the weight matrix for the candidate cell state modules in the first hidden layer. This is the bias vector for the candidate cell state module in the first hidden layer. It is the hyperbolic tangent activation function.

[0155] S524: The cell state update module performs element-wise product of the output data of the forget gate at time step t and the updated cell state at time step t-1, and performs element-wise product of the output data of the input gate at time step t and the candidate cell state at time step t. The element-wise products are then added together and output to the hidden state output layer.

[0156] ;

[0157] in, This represents the cell state updated at time step t in the first hidden layer. , This represents the cell state updated at time step t-1 of the second hidden layer.

[0158] S525: The output gate concatenates the hidden state vector at time step t-1 and the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the hidden state output layer.

[0159] ;

[0160] in, This is the output data of the output gate in the first hidden layer at time step t. , This is the weight matrix of the output gate of the first hidden layer. is the bias vector of the output gate of the first hidden layer.

[0161] S526: The hidden state output layer normalizes the cell state updated at time step t by using the hyperbolic tangent activation function. After normalization, it is multiplied element-wise with the output data of the output gate at time step t and then output to the second hidden layer.

[0162] ;

[0163] in, This represents the output data of the hidden state output layer at time step t in the first hidden layer.

[0164] It is important to note that The dimension is 16×(16+3), where 16 corresponds to the number of hidden layer units and 4 corresponds to the input dimension. The dimension is 16×1, and it is initially set as a 0 vector. Used to map output data to interval, Used to normalize the output data to Interval.

[0165] S53: The second hidden layer concatenates the hidden state at time step t-1 with the output data of the hidden state output layer at time step t. After concatenation, the data passes through the forget gate, input gate, cell state module, and output gate of the second hidden layer to extract high-order temporal features and multi-parameter coupling rules. The hidden state output layer updates the state at time step t, passes the state at time step t to time step t+1, and outputs it to the Attention layer.

[0166] Specifically, the processing procedure for the second hidden layer is as follows:

[0167] S531: The forget gate concatenates the hidden state at time step t-1 with the output data of the hidden state output layer at time step t, multiplies the vector by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the cell state module.

[0168] ;

[0169] in, This is the output data of the forget gate in the second hidden layer at time step t. , This is the weight matrix of the forget gate in the second hidden layer. This represents the hidden state of the second hidden layer at time step t-1. is the bias vector of the forget gate in the second hidden layer.

[0170] S532: The input gate concatenates the hidden state at time step t-1 with the output data of the hidden state output layer at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the cell state module.

[0171] ;

[0172] in, This is the output data of the input gate in the second hidden layer at time step t. , Here is the weight matrix of the input gate for the second hidden layer. is the bias vector of the input gate of the second hidden layer.

[0173] S533: The candidate cell state module concatenates the hidden state at time step t-1 with the output data of the hidden state output layer at time step t, multiplies it by the weight matrix, adds the bias vector, and then normalizes the data through the hyperbolic tangent activation function before outputting it to the cell state update module.

[0174] ;

[0175] in, Let be the candidate cell state at time step t in the second hidden layer. , This is the weight matrix for the candidate cell state modules in the second hidden layer. is the bias vector of the candidate cell state module in the second hidden layer;

[0176] S534: The cell state update module performs element-wise product of the output data of the forget gate at time step t and the updated cell state at time step t-1, and performs element-wise product of the output data of the input gate at time step t and the candidate cell state at time step t. The element-wise products are then added together and output to the hidden state output layer.

[0177] ;

[0178] in, This represents the cell state updated at time step t in the second hidden layer. , This represents the cell state updated at time step t-1 of the second hidden layer.

[0179] S535: The output gate concatenates the vector of the hidden state at time step t-1 and the output data of the hidden state output layer at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the hidden state output layer.

[0180] ;

[0181] in, This is the output data of the output gate in the second hidden layer at time step t. , This is the weight matrix for the output gate of the second hidden layer. This is the bias vector for the output gate of the second hidden layer.

[0182] S536: The hidden state output layer normalizes the cell state updated at time step t by using the hyperbolic tangent activation function. After normalization, it is element-wise multiplied with the output data of the output gate at time step t and then output to the Attention layer.

[0183] ;

[0184] in, This is the output data of the hidden state output layer at time step t in the second hidden layer.

[0185] It is important to note that The dimension is 8×(8+16), where 8 corresponds to the number of hidden layer units and 16 corresponds to the input dimension. The dimension is 8×1, and it is initially set as a 0 vector. Used to map output data to interval, Used to normalize the output data to Interval.

[0186] S54: The Attention layer strengthens the feature weights of key degradation time steps in the output data of the hidden state output layer at time step t. The attention-weighted global feature vector is output to the fully connected layer, as shown in the formula:

[0187] ;

[0188] in, The attention score at time step t. This is the attention weight matrix. The output data of the hidden state output layer at time step L in the second hidden layer (i.e., the final time feature of the sequence), where T is the transpose. Let be the attention weight at time step t. , It is an exponential function. For global feature vectors, .

[0189] It is important to note that Used to measure the correlation between the feature at time step t and the feature at the final time step L. The dimension is 8×8, and the sum of the weights of all time steps is 1. Used to amplify score differences at highly correlated time steps. The dimension is 8×1, which integrates key feature information from all time steps.

[0190] S55: The fully connected layer performs feature fusion and dimension transformation on the global feature vector, and outputs it to the dual-branch output layer. The formula is as follows:

[0191] ;

[0192] in, This is the output vector of the fully connected layer. This is the weight matrix of the fully connected layer. is the bias vector of the fully connected layer; ReLU is the ReLU activation function.

[0193] It is important to note that The dimension is 64×1. The dimensions are 64×8. With a dimension of 64×1, ReLU is used to enhance the nonlinear feature representation of the model and adapt to the nonlinear characteristics of the degradation process.

[0194] S56: The classification output layer predicts the lifetime level probability distribution of the optical module based on the output vector of the fully connected layer, and outputs the lifetime level probability distribution vector. The formula is:

[0195] ;

[0196] in, This represents the probability distribution vector for lifespan levels; The weight matrix for the classification output layer; is the bias vector of the classification output layer, and Softmax is the Softmax activation function.

[0197] It is important to note that The three dimensions correspond to three lifespan levels (0 = normal state, 1 = warning state, 2 = severe state). The dimension is 3×64. With a dimension of 3×1, Softmax is used to transform the output into a probability distribution form to meet the classification decision requirements.

[0198] In the embodiments, That is, the probability of the optical module being in a normal state is 85%, the probability of being in a warning state is 13%, and the probability of being in a critical state is 2%.

[0199] S57: Simultaneously, the regression output layer, based on the output vector of the fully connected layer, predicts the remaining useful lifetime of the optical module quantitatively and outputs the predicted remaining useful lifetime value at time step t, using the following formula:

[0200] ;

[0201] in, The remaining useful life prediction at time step t (in hours); It is a linear activation function; This is the weight matrix of the regression output layer; This is the bias vector for the regression output layer.

[0202] It is important to note that A quantitative scale used to retain remaining lifetime. The dimension is 1×64. The dimension is 1×1.

[0203] S6: The joint loss function is used to train and iteratively optimize the lifetime level probability distribution and the remaining useful lifetime prediction value. After training is completed, the optical module's working status is input to obtain the optical module's lifetime prediction result.

[0204] S61: Calculate the classification loss of the classification output layer using the cross-entropy loss function:

[0205] ;

[0206] in, The loss is for classification, and N is the number of training samples. The true rank label (one-hot encoded) for training sample j. Let j be the predicted probability of training sample j belonging to lifespan level c, where c = 0, 1, 2;

[0207] S62: Calculate the regression loss of the regression output layer using the mean squared error loss function:

[0208] ;

[0209] in, To regress the loss, To determine the true remaining lifetime of training sample j, Predict the remaining lifetime for training sample j;

[0210] S63: The joint classification loss and regression loss optimize the prediction accuracy of the classification output layer and the regression output layer. The formula is as follows:

[0211] ;

[0212] in, For the total loss function, This represents all trainable parameters (including weight matrices and bias vectors for each layer) of the LSTM lifetime prediction model. For classification loss function, Represents the regression loss function. To regress the loss weights, This is an L2 regularization term.

[0213] In the embodiments, A value of 0.3 is used to balance the training priority between classification and regression tasks. This is used to prevent the LSTM lifetime prediction model from overfitting and to adapt to small sample data scenarios in accelerated aging experiments.

[0214] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0215] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0216] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the lifetime of optical modules integrating DSP degradation and LSTM models, characterized in that, Includes the following steps: The DSP acquires the working state of the optical module, defines the degradation state vector of the optical module, and obtains the degradation rate of the laser through the degradation state vector of the optical module. A time-dependent degradation rate model is established by measuring the rate of change of the laser's degradation rate over time. The element values ​​of the matrix structure are fitted using accelerated aging experiments, and a nonlinear term is introduced into the time-dependent degradation rate model to form a nonlinear coupled degradation rate model. The forward Euler method is used to discretize the nonlinear coupled degradation rate model to obtain a model predicting future degradation. The observations are obtained by predicting the future degradation model and the observation noise. The Kalman filter recursion is performed in the DSP to perform the prediction step and the update step using the observations and the predicted future degradation model to obtain the optimal degradation state estimation sequence. The optimal degradation state estimation sequence is processed by the LSTM lifetime prediction model to obtain the lifetime level probability distribution and remaining useful lifetime prediction value of the optical module. The joint loss function is used to train and iteratively optimize the lifetime level probability distribution and the remaining useful lifetime prediction value. After training is completed, the lifetime prediction result of the optical module is obtained by inputting the working status of the optical module. The process of obtaining the laser degradation rate through the optical module degradation state vector includes the following steps: The DSP collects the laser chip temperature, laser bias current, and output optical power of the optical module in real time to characterize the working status of the optical module. The degradation state vector of the optical module is defined based on the working state of the optical module to characterize the degradation rate of the laser; The degradation rate of the laser is characterized by the following formula: ； in, Let be the degradation rate of the laser at time step t. Let t be the temperature degradation rate at time step t. Let be the current degradation rate at time step t. Let be the optical power degradation rate at time step t. For reference temperature, This is the rated bias current. For optical power, This is the temperature scaling factor. Let t be the laser chip temperature at time step t. Let be the laser bias current at time step t. The output optical power is at time step t.

2. The optical module lifetime prediction method based on DSP integration degradation and LSTM model according to claim 1, characterized in that, The formation of the nonlinear coupled degradation rate model includes the following steps: A continuous-time degradation rate model for lasers is established by considering the influence of laser degradation rate on its own changes and the direct driving effect of operating conditions on degradation. Using a continuous-time degradation rate model, the state transition matrix and driving input matrix are determined to be 3×3 matrix structures, and the elements of the matrix structure are theoretically derived. The specific values ​​of the elements of the state transition matrix and the driving input matrix are fitted by accelerated aging experiments; By introducing nonlinear terms into the continuous-time degradation rate model, highly nonlinear physical effects are captured, forming a nonlinear coupled degradation rate model.

3. The optical module lifetime prediction method based on DSP integration degradation and LSTM model according to claim 1, characterized in that, The formula for the model predicting future degradation is: ； in, It is a nonlinear coupled degradation rate model. To predict future degradation models, Where k is the sampling period and k is the discrete time step. This represents the true degenerate state at a discrete time step k. Let k be the input vector at the discrete time step k. Let k be the nonlinear coupling term at the discrete time step k. Let k be the noise vector for the discrete-time step k. Here is the state transition matrix. To drive the input matrix, The discrete state transition matrix, For discrete input driving matrices, For discrete nonlinear coupling terms, For discrete process noise, , , , .

4. The optical module lifetime prediction method based on DSP integration degradation and LSTM model according to claim 1, characterized in that, Obtaining the optimal degenerate state estimation sequence includes the following sub-steps: The observed values ​​are obtained based on the actual conditions and the observation noise; In the DSP, Kalman filtering is used for the prediction step. The prediction step uses the optimal degradation state estimate at discrete time step k and the prediction model for future degradation to predict the degradation state estimate at discrete time step k+1, and updates the covariance of the estimation error. In the DSP, Kalman filtering is used for the update step. The update step uses the observations of discrete time step k+1 to correct the degenerate state estimate of discrete time step k+1, and obtains the optimal degenerate state estimate of discrete time step k, and updates the covariance of the estimation error. After performing Kalman filtering recursion in the prediction and update steps in the DSP, the DSP outputs the optimal degenerate state estimation sequence. .

5. The optical module lifetime prediction method based on DSP integration degradation and LSTM model according to claim 1, characterized in that, The LSTM lifetime prediction model includes an input layer, two hidden layers, an attention layer, a fully connected layer, and a two-branch output layer. The two-branch output layer includes a classification output layer and a regression output layer. The two hidden layers include a first hidden layer and a second hidden layer. Both the first hidden layer and the second hidden layer include a forget gate, an input gate, a cell state module, an output gate, and a hidden state output layer. The cell state module includes a candidate cell state module and a cell state update module.

6. The optical module lifetime prediction method based on DSP integration degradation and LSTM model according to claim 5, characterized in that, The input layer receives the optimal degradation state estimation sequence and outputs it to the dual hidden layers. The dual hidden layers extract the short-term local temporal dependencies and the long-term global degradation trends respectively and output them to the Attention layer. The Attention layer strengthens the feature weights of the key degradation time steps and outputs them to the fully connected layer. The fully connected layer fuses high-order features and outputs them to the dual-branch output layer. The dual-branch output layer outputs the lifetime level probability distribution and the remaining useful lifetime prediction value of the optical module.

7. The optical module lifetime prediction method based on DSP integration degradation and LSTM model according to claim 5, characterized in that, The processing procedure of the first hidden layer: The forget gate concatenates the hidden state vector at time step t-1 and the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the cell state module. The input gate concatenates the hidden state vector at time step t-1 with the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the cell state module. The candidate cell state module concatenates the hidden state vector at time step t-1 with the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then normalizes the data through the hyperbolic tangent activation function before outputting it to the cell state update module. The cell state update module performs element-wise product of the output data of the forget gate at time step t and the updated cell state at time step t-1, and performs element-wise product of the output data of the input gate at time step t and the candidate cell state at time step t. The element-wise products are then added together and output to the hidden state output layer. The output gate concatenates the hidden state vector at time step t-1 with the degenerate state vector at time step t, multiplies it by the weight matrix, adds the bias vector, and then maps the data through the Sigmoid activation function before outputting it to the hidden state output layer. The hidden state output layer normalizes the cell state updated at time step t using the hyperbolic tangent activation function. After normalization, it performs element-wise multiplication with the output data of the output gate at time step t and outputs it to the second hidden layer.

8. The optical module lifetime prediction method based on DSP integration degradation and LSTM model according to claim 1, characterized in that, The joint loss function is trained and iteratively optimized using the lifetime rank probability distribution and the remaining useful lifetime prediction, including the following steps: The classification loss of the classification output layer is calculated using the cross-entropy loss function. The regression loss of the regression output layer is calculated using the mean squared error loss function. The prediction accuracy of the classification output layer and the regression output layer is optimized by combining classification loss and regression loss.

Images