An inp optical module health self-adaptive control system based on digital twinning

Abstract

The application relates to the technical field of optical communication management and control, and discloses an InP optical module health self-adaptive control system based on digital twinning, which is used to solve the problem that the service life is limited due to multi-dimensional parameter coupling degradation and management and control lag under long-term service. The system comprises data acquisition and preprocessing, hybrid digital twinning modeling, health state evaluation and prediction, self-adaptive control strategy and closed-loop feedback modules. A high-fidelity twin body is constructed by fusing physical mechanism and residual learning to perform quantitative evaluation, and a differentiated strategy is generated according to the evaluation result, and the strategy is executed after being verified to be safe. The application realizes precise modeling and closed-loop evolution, can effectively reduce working stress, prolong the safe life of the module and improve the link reliability, and is mainly used for optical communication network device operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of optical communication control technology, and in particular to an adaptive control system for the health status of InP optical modules based on digital twins. Background Technology

[0002] This invention relates to the field of optical communication network and optical module operation and maintenance management technology, and in particular to an adaptive control system for the health status of InP optical modules based on digital twins. As the core physical layer signal conversion device of optical communication networks, InP (indium phosphide) based optical modules face complex operating conditions during long-term service due to the combined effects of aging mechanisms caused by internal carriers and heat dissipation and external environmental stress disturbances.

[0003] As optical modules age, multi-dimensional parameter coupling degradation often occurs within them. Traditional control methods struggle to address the following two core technical issues: First, existing monitoring methods cannot accurately separate the characteristics between transient environmental fluctuations and intrinsic gradual degradation of devices. This makes it difficult to perform high-fidelity virtual mapping and multi-dimensional precise quantitative assessment of the internal carrier consumption, thermal resistance evolution, and photoelectric degradation mechanisms of optical modules throughout their entire lifecycle. Second, conventional optical module control typically relies on static threshold alarms and abrupt passive service switching mechanisms. It lacks an adaptive closed-loop control architecture driven by deep-level health state prediction, making it impossible to proactively reduce device stress through differentiated dynamic optimization and pre-verification of operating points such as bias current and temperature while ensuring system communication performance. Consequently, the actual safe service life of optical modules cannot be maximized. Summary of the Invention

[0004] To address the technical problems of lacking high-fidelity virtual mapping and active closed-loop adaptive control systems in existing technologies, the purpose of this invention is to provide an adaptive control system for the health status of InP optical modules based on digital twins.

[0005] An adaptive health status control system for InP optical modules based on digital twins, comprising: The full lifecycle data acquisition and preprocessing module is used to collect various raw data generated by the InP optical module from the time it leaves the factory to the time it is decommissioned and to perform multi-scale feature extraction, providing standardized data input for digital twin modeling. The physical-data hybrid digital twin modeling module is connected to the full lifecycle data acquisition and preprocessing module. It is used to construct and maintain a high-fidelity virtual digital twin of each managed InP optical module based on the standardized data input, and output a prediction vector that integrates physical and data and its uncertainty confidence interval. The health status assessment and lifespan prediction module is connected to the physical-data hybrid digital twin modeling module. It is used to complete the multi-dimensional quantitative assessment of the health status of the InP optical module using the output of the virtual digital twin, and to make a probability estimate of the remaining lifespan based on the extrapolation of the degradation trajectory trend. The output includes the assessment results containing the comprehensive health index and the multi-dimensional health status. The gradient-based adaptive control strategy module is connected to the health status assessment and life prediction module and the physical-data hybrid digital twin modeling module. It is used to generate differentiated adaptive control strategies through a multi-objective optimization algorithm with the health assessment results as the driving signal, and output the optimized operating point and control command sequence after the control strategy is virtually pre-verified on the virtual digital twin. The closed-loop execution and feedback module, connected to the gradient adaptive control strategy module and the physical-data hybrid digital twin modeling module, is used to send the pre-verified control command sequence to the physical module for execution, collect steady-state execution effect data after control adjustment, and feed the execution effect deviation as an incentive to the physical-data hybrid digital twin modeling module for control strategy library update and digital twin incremental correction.

[0006] As a preferred technical solution of the present invention, the full life cycle data acquisition and preprocessing module includes a factory calibration data interface unit, a real-time telemetry data acquisition unit, an environmental stress sensing unit, a multi-source data time series alignment and cleaning unit, and a multi-scale feature extraction unit. The multi-scale feature extraction unit is used to perform feature extraction operations on standardized time series data at three time scales: short-time, medium-time, and long-time. It extracts instantaneous coupling relationships, daily-level degradation trend evolution, and monthly-level degradation rate to form a multi-scale feature vector, thereby enhancing the ability to capture multi-frequency band fluctuations.

[0007] As a preferred embodiment of the present invention, the physical-data hybrid digital twin modeling module includes an InP device physical mechanism model unit, a data-driven residual learning unit, a hybrid model fusion and output unit, and an adaptive fidelity synchronization and recalibration unit; the hybrid model fusion and output unit is used to adaptively adjust the fusion compensation weight coefficient between the physical mechanism model prediction value and the data-driven residual prediction value according to the prediction root mean square error in the recent time step.

[0008] As a preferred technical solution of the present invention, the InP device physical mechanism model unit is composed of a carrier-photon rate equation sub-model, a thermal conduction sub-model and a degradation evolution sub-model coupled together. The actual observation value vector required for model calculation is indirectly constructed from telemetry data using an independent physical measurement basis. The actual observation value of junction temperature is obtained by using the forward voltage method to eliminate the cyclic dependence of junction temperature estimation on thermal resistance model parameters, thereby ensuring the objectivity of residual calculation and fidelity monitoring.

[0009] As a preferred technical solution of the present invention, the adaptive fidelity synchronization and recalibration unit is used to execute a three-level progressive calibration strategy based on the deviation vector between the digital twin output and the actual telemetry data. The strategy includes a continuous fine-tuning level, a parameter re-identification level, and a model structure adaptation level that forcibly increases the residual compensation weight coefficient and expands the data-driven network capacity when the physical model equation structure fails, thereby realizing the self-healing evolution of the degraded model structure from the device's newness to its scrapping.

[0010] As a preferred technical solution of the present invention, the health status assessment and lifespan prediction module includes a multidimensional health index normalization and fusion unit, a comprehensive health index calculation unit, a degradation trajectory trend extrapolation unit, a remaining lifespan probability estimation unit, and a sudden deterioration pattern recognition unit; the remaining lifespan probability estimation unit is used to generate future evolution trajectories based on the multinomial trend posterior distribution of Bayesian linear inference, and outputs a remaining lifespan probability distribution containing different confidence intervals through Monte Carlo sampling.

[0011] As a preferred technical solution of the present invention, the comprehensive health index calculation unit adopts an adaptive weighting mechanism when integrating normalized health indicators of multiple dimensions into a comprehensive health index. This mechanism consists of a basic weight of degradation rate and a correction amplification coefficient for failure proximity. It also truncates the negative value of the time derivative of each indicator to zero, so that the comprehensive health index automatically focuses on the dimension with the most severe degradation and closest to the failure boundary, avoiding the drawback of fixed equal weight integration that may mask key degradation dimensions.

[0012] As a preferred embodiment of the present invention, the gradient-adaptive control strategy module includes a health level dynamic division unit, a level-policy mapping and multi-objective optimization unit, a digital twin pre-verification unit, a smooth transition control instruction orchestration unit, and a service switching trigger decision unit; the level-policy mapping and multi-objective optimization unit constructs a constrained bi-objective optimization problem that seeks to maximize communication performance indicators and minimize normalized device stress indicators, and the relative trade-off coefficients of the two objective functions are dynamically decayed or adjusted according to the current health level gradient.

[0013] As a preferred technical solution of the present invention, the digital twin pre-verification unit intercepts dangerous instructions by sequentially executing a virtual execution process that includes steady-state performance prediction, short-term stress evolution trend simulation and constraint safety verification. Furthermore, during the migration of services at the near-terminal level, the verification of optical power constraints is proactively reduced to the minimum sustainability threshold to avoid logical conflicts between control intentions and performance constraints.

[0014] As a preferred embodiment of the present invention, the closed-loop execution and feedback module includes a control command security verification unit, a command issuance and execution interface unit, an execution effect evaluation unit, and a closed-loop feedback and model incremental update unit. The control command security verification unit performs hardware-level interception based on the static constraints of the device datasheet, and together with the digital twin pre-verification unit, constitutes a two-layer security mechanism. The closed-loop feedback and model incremental update unit is used to extract the steady-state change quantity containing active control excitation as excitation-response data and push it to the digital twin model for nonlinear parameter re-identification, thereby supporting the long-term adaptive evolution of the system control strategy.

[0015] The beneficial effects of this invention are as follows: 1. This system effectively solves the technical problem of accurately mapping the multi-dimensional parameter coupling degradation of InP optical modules during long-term service by constructing a hybrid digital twin that combines physical mechanisms and data-driven approaches. Based on real multi-scale telemetry data, the system uses a physical sub-equation model to anchor the inherent laws of optoelectronic device decay and a temporal attention residual network to capture nonlinear local degradation characteristics. Combined with a three-level progressive fidelity synchronization mechanism, it achieves high-fidelity, multi-dimensional, and quantitative accurate assessment of the health status of optical modules throughout their entire lifecycle.

[0016] 2. Based on the quantitative assessment results of health status, this system innovatively implements gradient-based adaptive closed-loop control, effectively solving the technical problem of limited safe service life of optical modules under traditional passive control modes. The system dynamically links control objectives with health levels, and adaptively and dynamically balances communication performance assurance and device stress relief through multi-objective optimization algorithms. It also pioneers a control pre-verification and smooth transition execution mechanism based on twins, achieving safe deployment of control strategies and closed-loop evolution of the model while ensuring that the communication link does not experience transient interruptions or irreversible damage, thus maximizing the remaining service life of the optical modules. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall structural framework modules of the present invention; Figure 2 This is a schematic diagram of the specific modules of the overall structure of the present invention.

[0018] 100. Full lifecycle data acquisition and preprocessing module; 101. Factory calibration data interface unit; 102. Real-time telemetry data acquisition unit; 103. Environmental stress sensing unit; 104. Multi-source data time-series alignment and cleaning unit; 105. Multi-scale feature extraction unit; 200. Physical-data hybrid digital twin modeling module; 201. InP device physical mechanism model unit; 202. Data-driven residual learning unit; 203. Hybrid model fusion and output unit; 204. Adaptive fidelity synchronization and recalibration unit; 300. Health status assessment and lifespan prediction module; 301. Multi-dimensional health indicator normalization and fusion unit; 302. Comprehensive health... 303. Index Calculation Unit; 304. Degradation Trajectory Trend Extrapolation Unit; 305. Remaining Useful Life Probability Estimation Unit; 306. Sudden Deterioration Pattern Recognition Unit; 407. Gradient Adaptive Control Strategy Module; 408. Health Level Dynamic Classification Unit; 409. Level-Strategy Mapping and Multi-Objective Optimization Unit; 400. Digital Twin Pre-Verification Unit; 401. Smooth Transition Control Instruction Arrangement Unit; 402. Service Switching Trigger Decision Unit; 500. Closed-Loop Execution and Feedback Module; 501. Control Instruction Security Verification Unit; 502. Instruction Issuance and Execution Interface Unit; 503. Execution Effect Evaluation Unit; 504. Closed-Loop Feedback and Model Incremental Update Unit. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are only for illustrative purposes and do not constitute a limitation on the scope of protection of the present invention. All equivalent transformations and improvements based on the technical concept of the present invention in this field should be covered within the scope of protection of the present invention.

[0020] like Figure 1As shown, the InP optical module health status adaptive control system based on digital twins described in this invention addresses the multi-dimensional parameter coupling degradation problem caused by chip-level gradual degradation and external stress disturbances during long-term service of InP-based optical modules. It constructs a complete closed-loop architecture from data acquisition, digital twin modeling, health assessment and prediction to adaptive control execution. The system comprises five core functional modules: a full lifecycle data acquisition and preprocessing module 100, a physical-data hybrid digital twin modeling module 200, a health status assessment and lifespan prediction module 300, a gradient-based adaptive control strategy module 400, and a closed-loop execution and feedback module 500. The modules collaborate hierarchically through data and control flows. The full lifecycle data acquisition and preprocessing module 100 provides a standardized data foundation for the entire system. The physical-data hybrid digital twin modeling module 200 builds and maintains a high-fidelity digital twin for each managed InP optical module based on this foundation. The health status assessment and lifetime prediction module 300 utilizes the twin output to perform multi-dimensional quantitative assessments of health status and probabilistic estimations of remaining lifetime. The gradient-adaptive control strategy module 400 generates differentiated control strategies using health assessment results as driving signals and executes them after pre-verification by the twin. The closed-loop execution and feedback module 500 is responsible for collecting execution effect data and feeding it back to module 200 to form a closed-loop correction, enabling the system to continuously improve modeling accuracy and control efficiency during operation. The composition, functional implementation, and technical details of each module are described in detail below.

[0021] like Figure 2 As shown, the full lifecycle data acquisition and preprocessing module 100 is responsible for collecting various types of data generated throughout the entire process of the InP optical module from factory delivery to retirement, and performing systematic preprocessing operations on the raw data to provide standardized, high-quality data input for subsequent digital twin modeling. The full lifecycle data acquisition and preprocessing module 100 includes a factory calibration data interface unit 101, a real-time telemetry data acquisition unit 102, an environmental stress sensing unit 103, a multi-source data time-series alignment and cleaning unit 104, and a multi-scale feature extraction unit 105.

[0022] The factory calibration data interface unit 101 acquires the factory calibration dataset for each InP optical module through a standardized data interface. Optional interface types include, but are not limited to, reading the module's embedded EEPROM and querying the manufacturer's database. This dataset contains at least the following: a unique module identifier (serial number), InP chip design parameters such as active region length, cavity surface reflectivity, and confinement factor, and photoelectric characteristic parameters measured at multiple temperature points during factory calibration. The temperature range covers at least -5℃ to 70℃, with a step size not exceeding 5℃. The parameters recorded at each temperature point vary depending on the device type: for laser-type modules, the threshold current, slope efficiency, center wavelength, and side-mode rejection ratio are recorded; for modules containing modulators, the half-wave voltage, extinction ratio, and insertion loss are also recorded. Recommended bias current and operating voltage values ​​are also recorded for each temperature point. This factory calibration dataset serves as both an initial parameter source for the digital twin physical model and a benchmark for subsequent degradation assessment. The factory calibration data interface unit 101 reads the data, formats and stores it according to a predefined structured data template, stores it in the local database, and marks the corresponding module with a unique identifier so that subsequent modules can call it as needed.

[0023] The real-time telemetry data acquisition unit 102 continuously acquires key electro-optical-thermal parameters of the module during operation through the telemetry ports of the built-in monitoring photodiode, thermistor, and module drive circuit of the InP optical module, at a configurable sampling period. The default sampling period is set to 1 second. When the system detects an abnormal fluctuation in a parameter, the sampling mode automatically switches to a high-frequency sampling of 100 milliseconds to capture abnormal dynamic processes more precisely. The continuously acquired real-time operating parameters include: laser bias current. Laser forward voltage TEC drive current Internal temperature of the module (Read by internal thermistor) Monitor photodiode current (Reflecting changes in output optical power), and modulator bias voltage. (If the module contains a modulator function segment). The real-time telemetry data acquisition unit 102 adds a precise timestamp to each sampling point with an accuracy of not less than 1 millisecond, and sends the data to the multi-source data timing alignment and cleaning unit 104 in a time-series format.

[0024] The environmental stress sensing unit 103 is responsible for collecting external stress parameters of the operating environment of the InP optical module. Specifically, this includes the ambient temperature obtained through a temperature sensor deployed outside the module. The ambient relative humidity is obtained through a humidity sensor. and the power supply voltage of the module. The default sampling period of the environmental stress sensing unit 103 is 10 seconds. The environmental stress parameter plays a dual role in this system: firstly, it serves as a direct input for calculating the degradation acceleration factor in the physical model; secondly, it acts as a key criterion in the health assessment process to distinguish between transient performance fluctuations caused by environmental stress and intrinsic device degradation. The environmental stress sensing unit 103 timestamps the collected data and sends it to the multi-source data timing alignment and cleaning unit 104.

[0025] The multi-source data time-series alignment and cleaning unit 104 receives multi-source asynchronous time-series data streams from the real-time telemetry data acquisition unit 102 and the environmental stress sensing unit 103, and performs preprocessing operations on them in sequence, including time-series alignment, outlier cleaning, and missing value imputation. In the time-series alignment step, the multi-source data time-series alignment and cleaning unit 104 uses the timestamp sequence of the real-time telemetry data acquisition unit 102 as the reference time grid, and resamples the lower-frequency data of the environmental stress sensing unit 103 to a unified time reference through linear interpolation. In the outlier cleaning step, median absolute deviation detection based on a sliding window is performed on the data of all channels. The window length is set to 60 sampling points. Data points deviating from the sliding median by more than 5 times the median absolute deviation are marked as outliers, and the median within that window is used for imputation. In the missing value imputation step, when more than 10 consecutive data points are missing, the data for that period is marked as unreliable and a notification is sent to the system management module; when the number of missing points is less than 10, cubic spline interpolation is used for imputation. After the above processing, the multi-source data time-series alignment and cleaning unit 104 outputs a standardized multi-channel time-series data matrix after time alignment, outlier cleaning, and missing value filling, which is then used by subsequent modules.

[0026] The multi-scale feature extraction unit 105 performs multi-time-scale feature extraction operations on the standardized data output by the multi-source data time-series alignment and cleaning unit 104, providing statistical features at different time granularities for digital twin modeling and health assessment. The multi-scale feature extraction unit 105 extracts features at three time scales. At the short time scale, the window length is set to 10 minutes, calculating the mean, standard deviation, maximum, minimum, and linear regression slope of each parameter within the window, as well as the Pearson correlation coefficient between any two parameters, to capture the rapid fluctuation characteristics and instantaneous coupling relationships of the parameters. At the medium time scale, the window length is set to 24 hours, calculating the daily average, intraday fluctuation amplitude (i.e., the difference between the maximum and minimum values), daily trend slope, and the offset between the current daily average and the factory calibration value of each parameter, to reflect the daily-level degradation trend evolution. At the long time scale, the window length is set to 30 days, calculating the monthly average, intra-month standard deviation, and rate of change between the monthly average and the previous month's average of each parameter, to reflect the monthly degradation rate. All features extracted from the above three scales constitute a multi-scale feature vector, which is output to the physical-data hybrid digital twin modeling module 200 and the health status assessment and lifespan prediction module 300, respectively.

[0027] The physical-data hybrid digital twin modeling module 200 is the implementation carrier of the first core technology point of this invention, responsible for building and continuously maintaining a virtual digital twin for each managed InP optical module. The physical-data hybrid digital twin modeling module 200 includes an InP device physical mechanism model unit 201, a data-driven residual learning unit 202, a hybrid model fusion and output unit 203, and an adaptive fidelity synchronization and recalibration unit 204.

[0028] InP device physical mechanism model unit 201 establishes a parameterized physical model based on the physical equations of InP semiconductor devices. This model consists of three mutually coupled sub-models.

[0029] The first sub-model is the carrier-photon rate equation sub-model, used to describe the carrier density in the active region of an InP laser. and photon density Steady-state relationship under a given injection current. Under steady-state conditions, i.e. and The carrier rate equation simplifies to: ; in For injection efficiency, The amount of electron charge. For the volume of the active region, For carrier lifetime, This is the gain coefficient. Given the transparent carrier density, the photon rate equation simplifies to: ; in As the light limiting factor, The spontaneous emission coupling factor, Let be the photon lifetime. The given bias current can be solved by simultaneously solving the two equations above. steady-state photon density This leads to the expression for the output optical power: ; in For optical coupling efficiency, is Planck's constant. The optical frequency is the key identifiable parameter in this sub-model. , , , and The initial values ​​of these parameters are determined by curve fitting using calibration data provided by the factory calibration data interface unit 101.

[0030] The second sub-model is the heat conduction sub-model, which is used to describe the junction temperature of the active region of the InP chip. The thermal balance relationship between the module's thermal management status and the module's thermal balance. For modules equipped with TEC, the TEC set temperature... The module case temperature is stabilized near the set value through the TEC temperature control loop, i.e., the module case temperature is maintained when the TEC is operating normally. Junction temperature The relationship with shell temperature is as follows: ; in The equivalent thermal resistance from the active region to the module housing, The active region dissipation power is given by... The calculation yielded the following result: Therefore, when the TEC is operating normally, the junction temperature... It can be expressed as bias current. and TEC set temperature The function, i.e. This relationship constitutes the basis for subsequent optimization units. This serves as the basis for the physical model of controlling the junction temperature using variables. For modules without TEC, Follow the ambient temperature Fluctuations, at which point the above equation degenerates into Junction temperature This has a direct feedback effect on the parameters in the carrier-photon rate equation sub-model. The threshold current changes exponentially with temperature: ; in Characteristic temperature, Reference temperature The threshold current. The slope efficiency as a function of temperature can be approximated linearly within the engineering accuracy range: ; in The temperature coefficient is the key identifiable parameter in this sub-model. and It is worth noting that the thermal resistance during module operation... The increase in is itself an important indicator of chip degradation.

[0031] The third sub-model is the degradation evolution sub-model, used to describe the degradation evolution trend of key parameters of InP devices over cumulative operating time. For the laser threshold current, the degradation model adopts a power-law form: ; in The initial threshold current, The degradation rate coefficient, The degradation acceleration index, The cumulative runtime is used. For slope efficiency, the degradation model adopts an exponential decay form: ; in The efficiency degradation rate coefficient. The extinction ratio is the degradation index. For modules containing modulators, the extinction ratio degradation model is: ; Degradation rate coefficient , , All are junction temperatures The function follows the Arrhenius acceleration relation: ; in To activate energy, This is the Boltzmann constant. Key identifiable parameters of this sub-model include those related to threshold degradation. , , and related to efficiency degradation , These parameters are initialized based on accelerated aging test data of the same batch of devices during the initial stage of module deployment, and are subsequently updated online via adaptive fidelity synchronization and recalibration unit 204 during operation.

[0032] For InP optical modules containing a modulator functional segment, the InP device physical mechanism model unit 201 also includes a fourth sub-model, namely the modulator electro-optic response sub-model. This sub-model describes the modulator bias voltage. The relationship between the optical transmission function and modulation performance parameters such as extinction ratio and insertion loss. For Mach-Zehnder modulators, the optical transmission function follows a cosine square relationship. ,in The half-wave voltage; the extinction ratio is determined by the ratio of the modulator's transmission rate in the "on" state to that in the "off" state, i.e. For electroabsorption modulators, their transfer function is empirically fitted. ,in Zero bias transmittance and These are the fitting parameters. Key parameters in the transfer function described above. (or , The initial value of the modulator bias voltage is determined by the factory calibration data and gradually drifts with device degradation. This sub-model establishes the modulator bias voltage. to extinction ratio The mapping relationship between the modulator insertion loss and the modulator insertion loss allows subsequent optimization units to... This sub-model is included as an effective control variable in the optimization solution. For pure laser modules without modulators, this sub-model does not participate in the calculation, and the corresponding... It is also not used as a control variable.

[0033] InP device physical mechanism model unit 201 couples the above three sub-models into a complete parameterized physical model. Given the current bias current... Ambient temperature (or TEC set temperature) ) and cumulative running time As input, the model outputs a vector of physical prediction values. ,in This is the predicted output optical power value. This is the predicted threshold current value. This is a predicted value for slope efficiency. This is the predicted junction temperature value. This is the predicted value of the center wavelength. This is the predicted equivalent thermal resistance. This is the predicted extinction ratio (applicable only to modules containing modulators). The predicted wavelength is obtained based on a temperature-wavelength linear drift relationship. ,coefficient The wavelength measurements at multiple temperature points from the factory calibration data are obtained through linear regression; typically, the value for an InP-based DFB laser is approximately 0.1 nm / ℃. For modules without a modulator, This item is not included in subsequent calculations.

[0034] It should be noted that not all components of the aforementioned physical prediction value vector can be directly measured by the real-time telemetry data acquisition unit 102; their corresponding actual observation value vectors... The actual observed value of the output optical power needs to be constructed indirectly from telemetry data in the following way. By monitoring the photodiode current This was calculated based on the photodiode responsivity coefficient determined during factory calibration. The actual observed value of the threshold current was also obtained. The bias current is obtained through periodic online LI curve scanning. Specifically, during system maintenance windows or off-peak communication periods, the bias current is scanned in small steps and the corresponding values ​​are recorded. In response, the threshold current value is extracted using an inflection point detection algorithm for the LI curve. This scan is performed every 24 hours by default. Actual observed values ​​of slope efficiency are also provided. Similarly, during the LI curve scanning process described above, the slope is determined by performing linear regression on the linear segment above the threshold. The actual observed junction temperature... The junction temperature is obtained independently using the forward voltage method, rather than indirectly calculated through the thermal resistance equation, to avoid the cyclic dependence of the junction temperature estimation on the thermal resistance model parameters. The forward voltage method utilizes the forward voltage of the InP semiconductor device. An approximately linear relationship exists between the junction temperature and the junction temperature. ,in This is the voltage-temperature sensitivity coefficient (unit: °C / V). and These represent the reference junction temperature and corresponding forward voltage under calibration conditions. The coefficients of this linear relationship... During the module factory calibration phase, multiple known shell temperature points are scanned and recorded by injecting a low duty cycle short pulse current under precise TEC temperature control (to eliminate self-heating effects) and recording the corresponding values. The value is used for calibration and determination. During operation, the real-time telemetry data acquisition unit 102 continuously collects data. The value can be directly converted. This measurement path is completely independent of the thermal resistance parameters in the physical model. The actual observed value of the equivalent thermal resistance... During each LI curve scan, the independently measured junction temperature was used. and current shell temperature (For modules equipped with TEC, Calculate using the following formula: Among them, the power dissipation .because Independently obtained through the forward voltage method Direct measurement by thermistor Calculated directly from telemetry data, none of the three parameters depend on the thermal resistance parameters in the physical model. ,therefore The calculations are based on completely independent observations, accurately reflecting the gradual increase in thermal resistance as the device degrades. There is no circular dependency problem of deriving observations from model parameters and then updating model parameters with those observations. The actual observed value of the center wavelength... Wavelength is obtained through the module's built-in wavelength monitoring function; for module types without a wavelength lockout, it is obtained through the temperature-wavelength relationship. Indirect estimation, of which The extinction ratio was obtained using the aforementioned forward voltage method. The actual observed value of the extinction ratio. (Only applicable to modules containing modulators) By sampling during the "1" and "0" levels of the modulated signal respectively. And calculate their ratio to obtain it. The above indirect construction process makes and It is strictly consistent in terms of dimension and physical meaning, and the acquisition path of each component has an independent physical measurement basis, providing a reliable data closed loop for subsequent residual calculation and fidelity monitoring.

[0035] The core function of the data-driven residual learning unit 202 is to learn the predicted output of the InP device physical mechanism model unit 201. Compared with actual telemetry data The residuals between The physical model is based on idealized assumptions and inevitably ignores or simplifies some complex effects, such as the local degradation differences caused by the non-uniform distribution of defect density in the active region, the nonlinear interaction effects between multiple degradation mechanisms, and the gradual evolution of encapsulation stress with temperature cycling. These effects missed by the physical model will be reflected in the form of residuals and need to be captured and compensated by the data-driven model.

[0036] The data-driven residual learning unit 202 employs an encoder-decoder neural network architecture based on a temporal attention mechanism. The encoder part receives past... Multi-channel input at each time step The default value is set to 168, corresponding to hourly data for one week. The input vector for each time step contains three components: the physical model prediction residual for that time step. The parameters are: the difference between the actual observed value and the predicted value of the physical model; the short-term and medium-term feature vectors output by the multi-scale feature extraction unit 105; and the environmental stress parameters. The encoder consists of three one-dimensional causal convolutional layers with kernel sizes of 3, 5, and 7, and the number of channels is set to 64 for each layer, used to extract local features within different time ranges. The output of the causal convolutional layers is processed by a multi-head self-attention layer with 4 heads and a hidden dimension of 128. Its function is to learn the long-range dependencies between different time steps and the weighted importance of each time step to the current prediction. The output of the attention layer is compressed into a fixed-dimensional hidden state vector through a fully connected layer. The decoder part uses the hidden state vector As input, the residual prediction for the next time step is output through two fully connected layers. The two fully connected layers have hidden dimensions of 64 and 32, respectively, and the activation function is ReLU.

[0037] The neural network model is trained using an online incremental learning strategy. In the offline phase, the model's initial weights are pre-trained using historical running datasets of similar InP modules to acquire general prior knowledge about residual patterns. During the actual execution of the target module, the weights are then trained using the most recent... The data at each time step is the training set. The default value is set to 720, corresponding to 30 days of hourly data. Incremental updates are performed using the Adam optimizer with learning rate decay. The initial learning rate is 0.001, and the learning rate decay coefficient for each training round is 0.95. Training continues for 5 epochs every 24 hours based on the existing model weights, allowing the model to gradually adapt to the individualized degradation characteristics of the target module. The loss function used for training is the sum of the mean squared error loss and the L2 regularization term. ; in Represents the set of trainable weight parameters of a neural network, and the regularization coefficient. The output of the data-driven residual learning unit 202 is the residual prediction value. .

[0038] Hybrid model fusion and output unit 203 is responsible for integrating the output of InP device physical mechanism model unit 201. With the output of the data-driven residual learning unit 202 The data is then fused to obtain the final predicted output of the digital twin: ; in This is the residual compensation weighting coefficient, with a value range of [value range missing]. . The value of is determined through the following adaptive mechanism: the physical model and the hybrid model are calculated respectively in the most recent The root mean square error of the prediction within each time step The default setting is 48, corresponding to hourly data for 2 days, denoted as follows: and .like That is, the prediction accuracy is improved after adding residual compensation, then let: ; The residual compensation weight is adaptively increased according to its contribution. If If residual compensation fails to improve accuracy or even leads to a decrease in accuracy, this may be due to overfitting of the residual model. Therefore, let: ; This reduces or even disables residual compensation. This adaptive weight adjustment mechanism ensures that the prediction accuracy of the mixture model tends to be no worse than that of the pure physics model within the current evaluation window; it should be noted that, due to... The adjustment is based on statistical error backtracking calculation within the recent window. Its guarantee of future prediction accuracy is a heuristic approximation rather than a rigorous mathematical guarantee, but in engineering practice, it can effectively avoid prediction degradation caused by overfitting of the residual compensation model. The final output of the hybrid model fusion and output unit 203 is... That is, the optimal estimate of the current state of the physical module by the digital twin, where This is an estimate of the current equivalent thermal resistance. This is an extinction ratio estimate (valid only for modules containing modulators). Meanwhile, the hybrid model fusion and output unit 203 is also based on recent... The predicted residual distribution at each time step is calculated, and the 95% confidence interval is output as a quantitative indicator of the prediction uncertainty.

[0039] The adaptive fidelity synchronization and recalibration unit 204 is a key mechanism to ensure that the digital twin maintains high fidelity throughout the module's entire lifecycle. The adaptive fidelity synchronization and recalibration unit 204 continuously monitors the output of the hybrid model fusion and output unit 203. Compared with actual telemetry data Deviation vector between A three-level progressive synchronous calibration strategy is implemented based on the magnitude and persistence of the deviation.

[0040] In the first level, i.e., the continuous fine-tuning level, the deviation vector of the most recent hour is calculated every hour. The mean and standard deviation of each component. If the absolute value of the mean of all components is less than 1% of its corresponding factory calibration value, the current fidelity is considered good, and no calibration operation is performed.

[0041] In the second level, i.e., the parameter re-identification level, if the bias vector If the absolute value of the 24-hour moving average of any component exceeds 2% of its corresponding factory calibration value, and this deviation exhibits a monotonic trend (with the absolute value of the Kendall rank correlation coefficient within 24 hours greater than 0.3 as the criterion), then the physical model parameters are determined to have drifted significantly, triggering online re-identification of the physical model parameters. The re-identification process uses telemetry data from the most recent 30 days as the identification dataset and employs the Levenberg-Marquardt nonlinear least squares algorithm to re-estimate all identifiable parameters in the InP device physical mechanism model unit 201, including... , , , , , , , , This process allows the physical model's predicted output to approximate the actual observed data again. After parameter re-identification is complete, the incremental update of the data-driven residual learning unit 202 is triggered synchronously, because changes in the physical model parameters will cause changes in the residual distribution.

[0042] In the third level, the model structure adaptation level, if the deviation still fails to converge to within the tolerance range after the parameter re-identification in the second level—specifically, if the absolute value of the 24-hour moving average of the deviation after three consecutive re-identifications still exceeds 5% of the factory calibration value—then the equation structure of the current physical model is deemed insufficient to fully describe the actual degradation behavior of the device. This situation typically occurs in the later stages of device degradation, when new degradation mechanisms begin to superimpose, such as a combination of sudden degradation and gradual degradation. The adaptive fidelity synchronization and recalibration unit 204 then performs a model structure adaptation operation: adjusting the residual compensation weight coefficients in the hybrid model... The value is forcibly set to a higher range of 0.8 to 1.0. This forced setting covers the adaptive adjustment rule based on RMSE comparison in the hybrid model fusion and output unit 203 during the third-level effective period, i.e., during the model structure adaptation phase. Instead of fluctuating with the RMSE window statistics, the value is locked within the aforementioned high range until subsequent retraining is completed and the bias converges to within the first-level tolerance, at which point it reverts to the adaptive adjustment mode of 203. This aims to fully unleash the compensation capability of the data-driven model when the physical model structure fails. Simultaneously, the network capacity of the data-driven residual learning unit 202 is increased, expanding the hidden dimensions of the two fully connected layers in the decoder from 64 and 32 to 128 and 64 respectively; and a longer time window of 90 days of data is used for retraining to enhance the data-driven model's ability to capture complex degradation patterns. Furthermore, the system simultaneously sends a warning message to the management module, indicating that the module may have entered an atypical degradation stage and suggesting an increase in the frequency of on-site inspections. Through this three-level progressive strategy, the adaptive fidelity synchronization and recalibration unit 204 ensures that the digital twin maintains a precise mapping relationship with the physical module throughout the entire lifecycle of the InP optical module, from new to scrapped.

[0043] The health status assessment and lifespan prediction module 300 utilizes the output of the physical-data hybrid digital twin modeling module 200 to perform multi-dimensional quantitative assessment of the health status of the InP optical module, predict degradation trends, and estimate remaining lifespan. The health status assessment and lifespan prediction module 300 includes a multi-dimensional health indicator normalization and fusion unit 301, a comprehensive health index calculation unit 302, a degradation trajectory trend extrapolation unit 303, a remaining lifespan probability estimation unit 304, and a sudden degradation pattern recognition unit 305.

[0044] The multidimensional health indicator normalization and fusion unit 301 extracts five dimensions of health indicators from the output of the digital twin and its comparison with the factory calibration values, and normalizes each indicator to a unified standard. Within the interval, 0 represents a completely new state, and 1 represents a failure boundary. The five normalized health indicators are defined as follows.

[0045] Threshold current degradation index The calculation formula is: ; in The factory-calibrated threshold current, The upper limit of the defined threshold current failure is usually taken as... 2 times.

[0046] Slope efficiency degradation index The calculation formula is: ; in The slope efficiency is the factory-calibrated value. The lower limit of slope efficiency failure is usually taken as... 50%.

[0047] Wavelength drift index The calculation formula is: ; in The center wavelength is calibrated at the factory. The maximum allowable wavelength offset is determined by the channel spacing of the communication system; for example, it is 0.4 nm for a dense wavelength division multiplexing system with a 100 GHz spacing.

[0048] thermal resistance increase index The calculation formula is: ; in The current thermal resistance value identified for the digital twin. The thermal resistance is the factory-calibrated value. To determine the upper limit of thermal resistance failure, take... 1.5 times.

[0049] Output optical power attenuation index The calculation formula is: ; in The factory-calibrated rated optical power, The minimum optical power threshold required to maintain the communication link. The calculation results of all metrics are pruned and limited to... Within the interval, values ​​less than 0 are assigned the value 0, and values ​​greater than 1 are assigned the value 1.

[0050] The comprehensive health index calculation unit 302 integrates the five normalized health indicators output by the multidimensional health indicator normalization and fusion unit 301 into a comprehensive health index. The fusion method employs an adaptive weighted summation based on degenerate correlation: ; Among them, weight satisfy and Its value is dynamically determined based on two factors. The first is the degradation rate factor: the base weights are allocated according to the degradation rate of each indicator, with indicators having a faster degradation rate receiving higher weights to reflect the weakest link effect. This is because the time derivatives of each indicator in actual operation... Negative values ​​may occur due to measurement noise or transient operating condition fluctuations (corresponding to a brief performance rebound), and all derivatives may approach zero during the system's stable period. To avoid negative weights and division-by-zero anomalies, the basic weight calculation first truncates each derivative to a lower limit: Let The negative derivative is set to zero, retaining only the contribution of the degenerate direction. Based on this, the formula for calculating the basic weights is: ; in It is a very small positive number (default value) This is used to prevent the denominator from being zero in the extreme case where the derivative is zero after all truncations. This approach ensures that when the system is in a fully stable state (all...), the denominator is zero. When the basic weights degenerate into a uniform distribution (equal weights in each dimension), the basic weights degenerate into a uniform distribution. The system will not generate incorrect health assessment results due to abnormal weight calculations during the stable period.

[0051] Secondly, there is the failure proximity correction factor: an amplification factor is introduced for each dimension. When the indicator No amplification is applied before the failure boundary is reached. Once the value exceeds 0.7 and enters the failure proximity region, the weight is gradually increased. The definition is presented in a segmented format: ; That is when hour The weight of this dimension is not affected by the failure proximity correction; when hour From 1 to 3 linearly (corresponding to) This is to highlight the impact of dimensions that are about to become obsolete on the overall health index. The final weights, after normalization, are: ; This adaptive weighting mechanism enables the comprehensive health index to automatically focus on the dimensions that are most severely degraded and closest to the failure boundary, avoiding the drawbacks of fixed equal-weight fusion that may mask key degraded dimensions.

[0052] Degeneration trajectory trend extrapolation unit 303 on comprehensive health index Historical time-series data is used for trend modeling and extrapolation to predict trends in the future. The evolutionary trajectory. The degenerative trajectory trend extrapolation unit 303 is implemented using Bayesian linear regression: assuming... The degradation trend in the short to medium term can be approximated by a polynomial function. From first to third order polynomials, the order optimal according to the Bayesian information criterion is selected as the trend model. Using the most recent 90 days... Using daily average data as the training set, the posterior distribution of the multinomial coefficients, including the mean and variance, is estimated using Bayesian inference methods, thereby simultaneously obtaining future... The predicted point values ​​and prediction uncertainty ranges are provided. The time frame for trend extrapolation is the next 180 days. The trend extrapolation results for degradation trajectory trend extrapolation unit 303 are updated daily.

[0053] The remaining lifetime probability estimation unit 304 estimates the remaining lifetime of the InP optical module based on the output of the degradation trajectory trend extrapolation unit 303. Here, the remaining lifetime is defined as the total lifetime from the current moment based on the comprehensive health index. First time reaching the preset failure threshold The time required The default value is set to 0.85. Due to the inherent uncertainty of trend extrapolation, the remaining useful life output by the remaining useful life probability estimation unit 304 is given in the form of a probability distribution: 1000 lines are generated through Monte Carlo sampling using the posterior distribution of the trend model coefficients. The future evolutionary trajectory, statistically analyzing the first crossing of each trajectory. The time points are used to form the empirical probability distribution of remaining useful life. The final output includes three statistics: median estimate. 10th percentile estimate (i.e., the shortest remaining life at a 90% confidence level) and the 90% quantile estimate .

[0054] The Sudden Degradation Pattern Recognition Unit 305 is specifically designed to detect sudden performance degradation events that differ from gradual degradation trends. These events may be caused by factors such as electrostatic discharge damage or sudden contamination of the fiber optic endface. The Sudden Degradation Pattern Recognition Unit 305 employs a dual-mechanism detection strategy. The first mechanism is based on a cumulative sum detection algorithm, continuously monitoring various health indicators. The first-order difference sequence is used to determine a mean mutation event when the cumulative deviation exceeds a preset detection sensitivity threshold. The second mechanism is Hotelling based on multidimensional statistical process control. The detection method compares the current five-dimensional health indicator vector with its recent (past 7 days) mean vector and covariance matrix. The statistic exceeds the given significance level. When the F-distribution reaches the critical value, a multidimensional joint anomaly is determined to have occurred. Once a sudden deterioration event is detected, the sudden deterioration pattern recognition unit 305 immediately sends an emergency alarm signal to the gradient adaptive control strategy module 400. The alarm information includes a sudden deterioration type label (single-dimensional mutation or multidimensional joint anomaly) and a severity assessment (divided into three levels: mild, moderate, and severe).

[0055] The gradient-based adaptive control strategy module 400 is the implementation carrier of the second core technical point of this invention. It is responsible for generating differentiated adaptive control strategies based on health assessment results and performing pre-verification on the digital twin before issuing them to the physical module for execution. The gradient-based adaptive control strategy module 400 includes a health level dynamic classification unit 401, a level-policy mapping and multi-objective optimization unit 402, a digital twin pre-verification unit 403, a smooth transition control instruction orchestration unit 404, and a service switching trigger decision unit 405.

[0056] The dynamic health level classification unit 401 outputs the comprehensive health index calculation unit 302. The value dynamically divides the current health status of the InP optical module into five gradient levels. Level A is the healthy state, corresponding to... At this point, the module is in normal working condition, with all parameters close to factory settings, requiring no special control intervention. Level B represents a slight degradation state, corresponding to... The module has shown detectable signs of degradation but has not yet affected communication performance indicators; the system has entered preventative optimization mode. Level C is a moderate degradation state, corresponding to... The module degradation has had a measurable impact on some performance indicators, such as a decrease in output optical power, but still within the system margin. The system has entered active compensation mode. Level D is a severely degraded state, corresponding to... The module has degraded to near its performance limit. Without intervention, it may impact business quality in the short term, and the system enters a performance-lifespan trade-off mode. Level E is the final stage, corresponding to... The module is nearing the end of its service life, and the system is entering a graceful degradation and switchover preparation mode.

[0057] The hierarchical classification introduces a hysteresis mechanism to avoid frequent transitions of modules near level boundaries: the aforementioned standard threshold is used when transitioning from a lower level to a higher level, while recovery from a higher level to a lower level requires... Below the transition threshold minus 0.03 hysteresis. For example, when a module recovers from level B to level A, it needs... Instead of directly using 0.20 as the boundary. Furthermore, if the sudden degradation pattern recognition unit 305 issues an emergency alarm signal, then regardless of the current... Regardless of the numerical value, the health level is directly upgraded to at least level C; when the alarm severity assessment is "severe", it is directly upgraded to level D to trigger the rapid response control strategy.

[0058] The level-policy mapping and multi-objective optimization unit 402 generates corresponding control policy optimization objectives and constraints based on the current health level determined by the health level dynamic partitioning unit 401, and solves for the optimal operating point through a multi-objective optimization algorithm. The level-policy mapping and multi-objective optimization unit 402 models the control problem of the InP optical module as a constrained bi-objective optimization problem.

[0059] Objective function 1 is a performance objective, aiming to maximize the normalized communication performance index. For modules containing only lasers, , Defined as the ratio of the output optical power predicted by the digital twin model to the minimum optical power required by the system, the optimization variable is... For modules containing modulators, , Defined as the ratio of the predicted extinction ratio to the minimum extinction ratio requirement, the optimization variable is... ,in Set the temperature for the TEC. Objective function 2 is a stress objective, aiming to minimize the normalized device stress index: ; in and These are the weights for temperature stress and current stress, respectively.

[0060] The constraints include four items: bias current constraint. Ensure the bias current does not exceed the device's maximum rating; TEC temperature constraint. Ensure the TEC set temperature is within the TEC module's operating range; output optical power constraint. Among them, optical power threshold The minimum requirements for normal business operations are applied under levels A to D. Under level E, take the minimum maintenance threshold. (default is) 50%); junction temperature constraint Ensure that the junction temperature does not exceed the device's maximum rated junction temperature.

[0061] At different health levels, the relative weights of the two objective functions are determined by the trade-off parameters. Adjustments are made. At level A. Highly biased towards performance targets, as the device health is good and stress factors need not be considered excessively; at level B. Level C Performance and stress are equally important; at grade D Significantly biased towards stress targets, extending remaining service life by actively reducing workload; at level E. The goal is almost entirely to minimize stress, maintaining only the most basic communication functions. The overall optimization objective is expressed as: ; This optimization problem is solved using a sequential quadratic programming algorithm. For the module containing the modulator, the optimization variables are: For modules containing only lasers, If a variable is not used as an optimization variable in the solution, the optimization variable degenerates into an optimization variable. Since the calculation of the objective function relies on the predicted output of the digital twin model, the sequential quadratic programming algorithm calls the hybrid model fusion and output unit 203 to calculate the objective function value in each iteration. The gradient is approximated using the finite difference method, and the step size is 0.1% of the current value of the optimization variable. The convergence accuracy of the algorithm is set to be less than the change in the objective function. The maximum number of iterations is 100. The level-policy mapping and multi-objective optimization unit 402 outputs the optimized target operating point vector, which is used for modules containing modulators. For pure laser modules .

[0062] The digital twin pre-verification unit 403 is a key security mechanism in this invention. After the level-policy mapping and multi-objective optimization unit 402 outputs the optimized working point, the control scheme is not directly sent to the physical module, but is first performed on the digital twin for virtual pre-verification. The pre-verification process consists of four steps performed sequentially.

[0063] In the first step, the working point will be optimized. As input, the hybrid model in the physical-data hybrid digital twin modeling module 200 is invoked to predict the steady-state performance parameters of the module at this operating point, including the predicted output optical power. Predicting the junction temperature and predicting extinction ratio (If the module contains a modulator).

[0064] In the second step, the short-term stress impact of control adjustments is assessed. Using a degradation evolution sub-model, assuming the operating point shifts from the current value to the optimal value, the evolution trend of each health indicator over the next 7 days is simulated, and the predicted comprehensive health index after 7 days is calculated. .

[0065] In the third step, security checks are performed one by one, requiring all of the following conditions to be met simultaneously. For optical power constraints, the optical power threshold applicable to pre-verification is set differently based on the current health level: for levels A to D, the optical power threshold is the minimum system requirement for normal operations. At level E, since the service migration process has been initiated by the service switching trigger decision unit 405, the module no longer assumes full responsibility for communication quality assurance. At this time, the optical power threshold is reduced to the minimum sustaining threshold. Its value is taken as At 50%, it is sufficient to ensure that the optical link can still be detected by the peer device. Let the currently applicable optical power threshold be uniformly denoted as... The optical power verification condition is: This differentiated threshold design resolves the logical conflict between significantly derating the optimizer to extend its lifespan in Class E and the fixed optical power constraint. If a normal service-level optical power threshold is still enforced in Class E, the derating scheme will inevitably be blocked by pre-verification, forcing the module to maintain a high-stress operating state and accelerating failure, creating a paradox with the intention of extending lifespan through derating. Other verification conditions include: the predicted junction temperature has a sufficient safety margin, i.e. The single bias current adjustment range shall not exceed 20% of the current value, that is... Preliminary verification confirmed that the adjustment would not cause the health index to deteriorate by more than 0.02 within 7 days, i.e. For solutions involving TEC temperature adjustment, the single temperature adjustment range shall not exceed 3℃, i.e. .

[0066] In the fourth step, a judgment is made based on the results of the security verification. If all conditions are met, the optimization working point is marked as having passed pre-verification and is allowed to be deployed for execution. If any condition is not met, it is marked as having failed pre-verification, and the specific condition and the amount of violation are fed back to the level-policy mapping and multi-objective optimization unit 402. The level-policy mapping and multi-objective optimization unit 402 then resolves the optimization problem after tightening the corresponding constraints. For example, if the junction temperature verification fails, the constraint will be... The upper limit is further reduced, and the solution is recalculated. A maximum of 3 recalculation iterations are allowed. If a feasible solution that has passed the pre-validation cannot be found after 3 iterations, the current working point is maintained, and "No feasible optimization solution is found in the current state" is recorded in the system log.

[0067] The smooth transition control instruction orchestration unit 404 is responsible for converting the pre-verified optimized operating point into an executable sequence of control instructions, and ensuring that the transition from the current operating point to the target operating point is smooth and gradual to avoid transient optical power fluctuations that may be caused by abrupt changes in the operating point. The smooth transition control instruction orchestration unit 404 employs a linear ramp transition strategy, dividing the adjustment process into... Equally spaced steps, The adjustment will be determined dynamically based on the adjustment range. ; in This represents the maximum current change per step, with a default value of 0.1mA. The time interval between each step. The default value is 1 second. In the... step( The control command value for the bias current is: ; The transition method for other control parameters is similar. During the transition, the smooth transition control command arrangement unit 404 continuously monitors the output optical power signal fed back by the real-time telemetry data acquisition unit 102. If the output optical power is detected at any intermediate step... The transient drop exceeds the currently applicable optical power threshold. If the optical power drops below 5%, immediately pause subsequent steps and hold the current step for 3 seconds to wait for the system to stabilize. If the optical power recovers to the normal range within 3 seconds, continue with subsequent steps; if the optical power does not recover, revert to the working point of the previous step and terminate this adjustment, while recording the anomaly log.

[0068] Service switchover trigger decision unit 405 triggers the service switchover preparation process when any of the following conditions are met: the health level reaches level E, i.e. or the output of the remaining useful life probability estimation unit 304 Less than the preset minimum safety margin time (Default value is 30 days), or the sudden degradation pattern identification unit 305 issues an emergency alarm with a severity level of "Severe". The service switchover trigger decision unit 405 sends a structured switchover suggestion message to the upper-level network management system. The message content includes the module identifier, current health level, and current... Values, remaining lifetime probability distribution, and the urgency level of the recommended switchover. Urgency is divided into two categories: planned switchover (recommended in...). (Completed within the time window) and emergency switchover (immediate execution recommended). During the transition period while waiting for the service switchover to complete, the service switchover triggering decision unit 405 instructs the level-policy mapping and multi-objective optimization unit 402 to adopt the most conservative control strategy, that is, to... This is to maximize the available time of the module before the switchover is complete.

[0069] The closed-loop execution and feedback module 500 is responsible for securely issuing control commands to the physical module for execution, monitoring the execution effect, and feeding back the effect data to the physical-data hybrid digital twin modeling module 200 to form a complete closed loop. The closed-loop execution and feedback module 500 includes a control command security verification unit 501, a command issuance and execution interface unit 502, an execution effect evaluation unit 503, and a closed-loop feedback and model incremental update unit 504.

[0070] The control command safety verification unit 501 acts as the final safety gate before control commands are issued, performing hardware-level safety verification on each control command output by the smooth transition control command orchestration unit 404. Verification includes: whether the bias current command value is within the absolute maximum rated value range specified in the device datasheet, whether the TEC temperature setpoint is within the operating range of the TEC module, and whether the modulator bias voltage is within the safe range. Any command exceeding the hardware safety range is intercepted and triggers a safety alarm. The control command safety verification unit 501 and the digital twin pre-verification unit 403 complement each other functionally: the former performs static safety checks based on device hardware specifications, forming the last hard protection barrier in the control link; the latter provides intelligent soft safety protection through dynamic effect pre-evaluation based on digital twin simulation. These two protections together constitute the system's dual-layer safety mechanism.

[0071] The instruction issuance and execution interface unit 502 issues security-verified control instructions to the driver circuit of the InP optical module through a standardized hardware control interface. Optional hardware control interfaces include I²C bus, SPI bus, or MDIO interface. For bias current adjustment, the bias current reference value of the laser driver chip is set via a digital-to-analog converter; for TEC temperature adjustment, the temperature reference value of the TEC controller is set; and for modulator bias voltage adjustment, the digital-to-analog converter value of the modulator bias circuit is set. After each instruction is issued, the instruction issuance and execution interface unit 502 reads the readback register of the driver circuit to confirm that the instruction has been correctly written and executed.

[0072] After the control adjustment is completed, the performance evaluation unit 503 waits for a steady-state establishment time (default 30 seconds to ensure that both the TEC temperature loop and the optical power control loop reach a new steady-state equilibrium). Then, it collects telemetry data for 5 consecutive minutes after the steady-state condition is reached and compares it with the data before adjustment and the predicted data from the digital twin pre-verification unit 403. The first comparison is an evaluation of the adjustment effectiveness: calculating the difference between the actual performance indicators after adjustment and before adjustment to verify whether the adjustment has produced the expected effect. For example, if the control objective is to reduce device stress while maintaining optical power no less than [a certain value], then [the following is a separate, unrelated sentence:] Then, check whether the adjusted optical power truly meets the requirements, and verify whether the junction temperature or bias current has actually decreased. Comparison 2 is the prediction-actual consistency assessment: calculate the deviation between the adjusted actual telemetry value and the predicted value from the digital twin pre-validation stage. If the deviation is within the prediction confidence interval, the prediction accuracy of the digital twin model is considered good; if the deviation exceeds the prediction confidence interval, it indicates that the model's prediction of the control adjustment response has errors, requiring the triggering of subsequent model calibration procedures. The performance evaluation unit 503 encodes the evaluation results into a structured report for use by the closed-loop feedback and model incremental update unit 504.

[0073] The closed-loop feedback and model incremental update unit 504 serves as a feedback channel connecting control execution and digital twin modeling, enabling continuous self-evolution of the system. When the prediction-actual consistency evaluation result of the execution effect evaluation unit 503 shows a deviation exceeding the confidence interval, the closed-loop feedback and model incremental update unit 504 packages the change in the operating point before and after control adjustment (i.e., the actively applied control excitation signal) and the corresponding change in module response (i.e., the change in actual telemetry data) into a set of excitation-response data pairs, and pushes them to the adaptive fidelity synchronization and recalibration unit 204 as training data for online re-identification of physical model parameters and data-driven incremental model updates. This type of data containing active control excitation has a higher information density than data under natural operating conditions because it provides direct observation of the module's response behavior at different operating points, helping to more accurately identify nonlinear parameters in the physical model, such as gain compression factor and the dependence of thermal resistance on dissipated power.

[0074] The closed-loop feedback and model incremental update unit 504 also maintains a historical database of control strategy effects, systematically recording complete contextual information for each control adjustment, including the health level before adjustment and The database contains the value, the current operating point status, the target operating point for this adjustment, and the actual effect after the adjustment (covering performance changes and health index changes). This database supports the strategy backtracking analysis function of the level-policy mapping and multi-objective optimization unit 402. When the same type of control adjustment has been executed multiple times in history but the effect shows a decreasing trend (e.g., the increase in optical power by increasing the bias current becomes smaller and smaller), the system recognizes that the compensation method is approaching saturation. In the subsequent optimization solution process, it automatically reduces the dependence on this control variable and instead explores other feasible compensation paths, such as adjusting the TEC temperature setting or modulator bias voltage, thereby achieving long-term adaptive evolution of the control strategy.

[0075] The complete workflow of the system of the present invention is executed cyclically according to the following steps.

[0076] During the initialization phase, when a new InP optical module is connected to the nanotube, the factory calibration data interface unit 101 reads the module's factory calibration dataset. The InP device physical mechanism model unit 201 uses this calibration data to initialize all identifiable parameters of the physical model through curve fitting, including carrier lifetime, gain coefficient, transparent carrier density, thermal resistance, and characteristic temperature, establishing an initial physical model instance for the module. The data-driven residual learning unit 202 loads the weights of a general residual model pre-trained for similar InP modules. The hybrid model fusion and output unit 203 then applies the residual compensation weights. Initialize to 0.3 to allow the physical model to dominate in the initial stages, as the accumulated runtime data is insufficient to fully support the residual model's effectiveness. Overall Health Index Initialize to 0, and initialize health level to level A. The system then begins the data acquisition process with the default sampling period.

[0077] After the system enters steady-state operation, the real-time telemetry data acquisition unit 102 continuously acquires telemetry data such as laser bias current, forward voltage, TEC current, module temperature, and monitoring photodiode current at a 1-second cycle, while the environmental stress sensing unit 103 acquires ambient temperature and humidity data at a 10-second cycle. The multi-source data time-series alignment and cleaning unit 104 performs time alignment, outlier detection and cleaning, and missing value filling on the multi-source asynchronous data streams, outputting a standardized multi-channel time-series data matrix. The multi-scale feature extraction unit 105 extracts statistical features at short-term, medium-term, and long-term scales and constructs a multi-scale feature vector. The above acquisition and preprocessing processes run continuously, providing uninterrupted data input for subsequent modules.

[0078] In the real-time mapping and synchronization step of the digital twin, the InP device physical mechanism model unit 201 takes the current bias current, ambient temperature, and cumulative operating time as inputs, and calculates the physical prediction output based on the coupled rate equation, heat conduction equation, and degradation evolution equation. The data-driven residual learning unit 202 takes the residual history and multi-scale features from the past 168 time steps as input and predicts the residual at the current time step through a temporal attention network. The hybrid model fusion and output unit 203 is adjusted according to the current adaptive weights. The final prediction of the digital twin is output after merging the two. and its confidence interval. Adaptive fidelity synchronization and recalibration unit 204 synchronous monitoring. The deviation between the actual telemetry value and the actual telemetry value is used to determine whether to perform a first-level, second-level, or third-level calibration operation based on the magnitude and trend characteristics of the deviation, so as to ensure that the digital twin always maintains high-fidelity synchronization.

[0079] In the health status assessment and lifespan prediction steps, the multidimensional health indicator normalization and fusion unit 301 calculates five normalized health indicators from the digital twin output and factory calibration values. The comprehensive health index calculation unit 302 calculates the comprehensive health index through adaptive weighted fusion. The degradation trajectory trend extrapolation unit 303 uses Bayesian linear regression to... Historical time series data are used for trend modeling to extrapolate the evolution trajectory over the next 180 days. The remaining lifetime probability estimation unit 304 is estimated using Monte Carlo sampling. The time distribution of the first crossing of the failure threshold is used to output a probability estimate of the remaining service life. The sudden degradation pattern recognition unit 305 detects and performs accumulation and hotelling... Detection and continuous monitoring of sudden deterioration events.

[0080] In the health level classification and control strategy generation step, the health level dynamic classification unit 401, based on the current... The module's health level is determined based on the health value and sudden degradation detection results, and a hysteresis mechanism is applied. The level-policy mapping and multi-objective optimization unit 402 sets trade-off parameters based on the health level. A constrained bi-objective optimization problem is constructed, and the Pareto optimal operating point is found using a sequential quadratic programming algorithm. When the health level is level A and... When the value is below 0.15, skip the optimization solution to reduce unnecessary control adjustments and maintain the current operating point.

[0081] In the digital twin pre-verification step, the digital twin pre-verification unit 403 performs virtual pre-execution verification on the optimized operating point, predicts the adjusted steady-state performance and junction temperature, simulates the health index evolution trend over 7 days, and verifies the five safety conditions one by one. If all conditions pass, the solution is approved and issued; if any condition fails, the solution is fed back to the optimization unit for resolving after tightening the constraints, with a maximum of 3 iterations. If no feasible solution is found after 3 iterations, the current operating point remains unchanged.

[0082] In the smooth execution control adjustment step, the pre-validated optimization workpoint is decomposed into... A linear ramp transition sequence is used. The control command security verification unit 501 performs hardware-level security checks on each step of the command. The command issuance and execution interface unit 502 issues commands step by step to the InP optical module driver circuit through the hardware interface. During the transition, the optical power output is continuously monitored, and if an abnormal drop is detected, the process is paused or rolled back.

[0083] In the execution effect evaluation and closed-loop feedback steps, after the control adjustment is completed and steady state is established, the execution effect evaluation unit 503 collects 5 minutes of steady-state telemetry data for effectiveness evaluation and prediction consistency evaluation. The closed-loop feedback and model incremental update unit 504 pushes the stimulus-response data pair to the adaptive fidelity synchronization and recalibration unit 204 of the physical-data hybrid digital twin modeling module 200 for fine-tuning of physical model parameters and incremental learning of the residual model, while recording the adjustment effect in the control strategy effect history database.

[0084] After completing the above steps, the system returns to the data acquisition step and continues to execute in a loop. The routine update frequency for health assessment is once per hour. The execution frequency of control strategy optimization is dynamically adjusted according to the health level: once every 24 hours for levels A and B, once every 6 hours for level C, and once every hour for levels D and E. When a sudden degradation event is detected, the system immediately triggers a complete emergency response process from health assessment to control execution. With accumulated operating time, the digital twin model continuously receives correction and optimization through continuous high-fidelity synchronization and closed-loop feedback data. The control strategy continuously adapts to the evolution of individual device degradation characteristics through retrospective analysis of the historical effect database. The entire system thus achieves continuous self-evolution from initial general capabilities to individualized precise control.

[0085] When a service switchover triggers a switchover recommendation from decision unit 405, the system enters the end-of-life management phase. During this phase, the most conservative control strategy is employed to maintain the module's basic communication functions, while coordinating with the upper-layer network management system to complete the orderly migration of service traffic. After the module is finally decommissioned, its entire lifecycle operational data, the evolution history of digital twin model parameters, and the execution records of control strategies are fully archived into a knowledge base. This knowledge is used for pre-trained model optimization of similar modules and for group-level degradation pattern analysis, thereby enabling cross-module transfer and accumulation of operational knowledge.

[0086] It should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. 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 digital twin-based adaptive control system for the health status of InP optical modules, characterized in that, include: The full lifecycle data acquisition and preprocessing module is used to collect various raw data generated by the InP optical module from the time it leaves the factory to the time it is decommissioned and to perform multi-scale feature extraction, providing standardized data input for digital twin modeling. The physical-data hybrid digital twin modeling module is connected to the full lifecycle data acquisition and preprocessing module. It is used to construct and maintain a high-fidelity virtual digital twin of each managed InP optical module based on the standardized data input, and output a prediction vector that integrates physical and data and its uncertainty confidence interval. The health status assessment and lifespan prediction module is connected to the physical-data hybrid digital twin modeling module. It is used to complete the multi-dimensional quantitative assessment of the health status of the InP optical module using the output of the virtual digital twin, and to make a probability estimate of the remaining lifespan based on the extrapolation of the degradation trajectory trend. The output includes the assessment results containing the comprehensive health index and the multi-dimensional health status. The gradient-based adaptive control strategy module is connected to the health status assessment and life prediction module and the physical-data hybrid digital twin modeling module. It is used to generate differentiated adaptive control strategies through a multi-objective optimization algorithm with the health assessment results as the driving signal, and output the optimized operating point and control command sequence after the control strategy is virtually pre-verified on the virtual digital twin. The closed-loop execution and feedback module, connected to the gradient adaptive control strategy module and the physical-data hybrid digital twin modeling module, is used to send the pre-verified control command sequence to the physical module for execution, collect steady-state execution effect data after control adjustment, and feed the execution effect deviation as an incentive to the physical-data hybrid digital twin modeling module for control strategy library update and digital twin incremental correction.

2. The adaptive health status control system for InP optical modules based on digital twins according to claim 1, characterized in that: The full lifecycle data acquisition and preprocessing module includes a factory calibration data interface unit, a real-time telemetry data acquisition unit, an environmental stress sensing unit, a multi-source data time series alignment and cleaning unit, and a multi-scale feature extraction unit. The multi-scale feature extraction unit is used to perform feature extraction operations on standardized time series data at three time scales: short-time, medium-time, and long-time, to extract instantaneous coupling relationships, daily-level degradation trend evolution, and monthly-level degradation rate, and to construct multi-scale feature vectors.

3. The adaptive health status control system for InP optical modules based on digital twins according to claim 1, characterized in that: The physical-data hybrid digital twin modeling module includes an InP device physical mechanism model unit, a data-driven residual learning unit, a hybrid model fusion and output unit, and an adaptive fidelity synchronization and recalibration unit. The hybrid model fusion and output unit is used to adaptively adjust the fusion compensation weight coefficient between the physical mechanism model prediction value and the data-driven residual prediction value based on the prediction root mean square error in the recent time step.

4. The adaptive health status control system for InP optical modules based on digital twins according to claim 3, characterized in that: The InP device physical mechanism model unit is composed of a carrier-photon rate equation sub-model, a thermal conduction sub-model, and a degradation evolution sub-model coupled together. The actual observation value vector required for model calculation is indirectly constructed from telemetry data using an independent physical measurement basis. The actual observation value of junction temperature is obtained by the forward voltage method to eliminate the cyclic dependence of junction temperature estimation on thermal resistance model parameters.

5. The adaptive health status control system for InP optical modules based on digital twins according to claim 3, characterized in that: The adaptive fidelity synchronization and recalibration unit is used to execute a three-level progressive calibration strategy based on the deviation vector between the digital twin output and the actual telemetry data. This includes a continuous fine-tuning level, a parameter re-identification level, and a model structure adaptation level that forcibly increases the residual compensation weight coefficient and expands the data-driven network capacity when the physical model equation structure fails.

6. The adaptive health status control system for InP optical modules based on digital twins according to claim 1, characterized in that: The health status assessment and lifespan prediction module includes a multidimensional health index normalization and fusion unit, a comprehensive health index calculation unit, a degradation trajectory trend extrapolation unit, a remaining lifespan probability estimation unit, and a sudden deterioration pattern recognition unit. The remaining lifespan probability estimation unit is used to generate a future evolution trajectory based on a multinomial trend posterior distribution with Bayesian linear inference, and outputs a remaining lifespan probability distribution containing different confidence intervals through Monte Carlo sampling.

7. The adaptive health status control system for InP optical modules based on digital twins according to claim 6, characterized in that: The comprehensive health index calculation unit adopts an adaptive weighting mechanism when integrating normalized health indicators from multiple dimensions into a comprehensive health index. This mechanism consists of a basic weight for degradation rate and a correction amplification coefficient for failure proximity. It also truncates the negative values ​​of the time derivatives of each indicator to zero, so that the comprehensive health index automatically focuses on the dimension with the most severe degradation and closest to the failure boundary.

8. The adaptive health status control system for InP optical modules based on digital twins according to claim 1, characterized in that: The gradient-based adaptive control strategy module includes a health level dynamic division unit, a level-policy mapping and multi-objective optimization unit, a digital twin pre-verification unit, a smooth transition control instruction orchestration unit, and a service switching trigger decision unit. The level-policy mapping and multi-objective optimization unit constructs a constrained bi-objective optimization problem that seeks to maximize communication performance indicators and minimize normalized device stress indicators. The relative trade-off coefficients of the two objective functions are dynamically decayed or adjusted according to the current health level gradient.

9. The adaptive health status control system for InP optical modules based on digital twins according to claim 8, characterized in that: The digital twin pre-verification unit intercepts dangerous instructions by sequentially executing a virtual execution process that includes steady-state performance prediction, short-term stress evolution trend simulation, and constraint safety verification. When the service is in a near-terminal state, the verification of optical power constraints is proactively reduced to the minimum sustainability threshold to avoid logical conflicts between control intentions and performance constraints.

10. The adaptive health status control system for InP optical modules based on digital twins according to claim 8, characterized in that: The closed-loop execution and feedback module includes a control command security verification unit, a command issuance and execution interface unit, an execution effect evaluation unit, and a closed-loop feedback and model incremental update unit. The control command security verification unit performs hardware-level interception based on the static constraints of the device datasheet, forming a two-layer security mechanism together with the digital twin pre-verification unit. The closed-loop feedback and model incremental update unit is used to extract the steady-state changes containing active control excitations as excitation-response data and push them to the digital twin model for nonlinear parameter re-identification.

Images