Optical module performance control method based on multi-source state fusion and model predictive feedforward

Abstract

The application discloses a kind of based on multi-source state fusion and model prediction feedforward optical module performance control method, comprising: synchronous acquisition driving chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal and backlight monitoring current signal, generate multi-source electric heat signal sequence;With multiple optical modules in the same batch as node, the batch association graph structure is constructed;Multi-source electric heat signal sequence and batch association graph structure are input into graph enhanced long short-term memory network to aggregate cross-module information, obtain laser equivalent junction temperature estimate and threshold current aging rate estimate;Predict optical power deviation by optical performance prediction model;Solve the cooperative compensation trajectory of bias current and thermoelectric cooler temperature, generate bias current compensation sequence and thermoelectric cooler temperature set value correction sequence;Step down to driving circuit and temperature control unit, execute timing cooperative feedforward compensation, effectively improve the optical power consistency in optical module production batch.

Description

Technical Field

[0001] This application relates to the field of optical communication device manufacturing technology, and in particular to a method for controlling the performance of optical modules based on multi-source state fusion and model prediction feedforward. Background Technology

[0002] With the rapid development of data centers and 5G communication networks, the demand for high-speed optical modules has exploded. Optical modules must undergo rigorous burn-in testing before leaving the factory to screen for early failures and verify long-term operational stability. However, during burn-in testing, the threshold current of the laser drifts over time, and the cooling efficiency of the thermoelectric cooler (TEC) degrades due to thermal stress, leading to inconsistencies in optical power output. Traditional control methods rely solely on feedback control based on local sensor data from individual optical modules, neglecting the process dependence and thermal coupling relationships between modules within the same batch. This results in poor control performance and makes it difficult to guarantee the performance consistency of the entire batch of optical modules. Summary of the Invention

[0003] The main purpose of this application is to provide a performance control method for optical modules based on multi-source state fusion and model prediction feedforward, which aims to solve the technical problem of low consistency control accuracy of optical power during the aging process of optical modules.

[0004] To achieve the above objectives, this application proposes a performance control method for optical modules based on multi-source state fusion and model prediction feedforward. The performance control method for optical modules based on multi-source state fusion and model prediction feedforward includes: In the optical module aging test station, the bias voltage signal of the driver chip, the operating current signal of the thermoelectric cooler, the temperature signal of the thermoelectric cooler, and the backlight monitoring current signal are collected in real time to generate a multi-source electrothermal signal sequence. Construct a batch association graph structure with multiple optical modules within the same batch as nodes; The multi-source electrothermal signal sequence and the batch association graph structure are input into a graph-enhanced long short-term memory network for cross-module information aggregation to obtain the laser equivalent junction temperature estimate and threshold current aging rate estimate. Based on the estimated equivalent junction temperature of the laser, the estimated threshold current aging rate, and the current cumulative aging time, the optical power deviation at the future target time is predicted using an optical performance prediction model. Based on the optical power deviation, the collaborative compensation trajectory of bias current and thermoelectric cooler temperature is solved, and a bias current compensation sequence and a thermoelectric cooler temperature setpoint correction sequence are generated. The bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence are sent step by step to the drive circuit and temperature control unit of the optical module to perform timing-coordinated feedforward compensation in order to achieve consistent control of optical module performance.

[0005] Optionally, the step of synchronously acquiring the driver chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal, and backlight monitoring current signal in real time at the optical module aging test station to generate a multi-source electrothermal signal sequence includes: The sampling frequency and sampling duration are determined based on the optical module model and aging process requirements, and a multi-channel synchronous data acquisition unit is deployed at the aging test station. The multi-channel synchronous data acquisition unit synchronously acquires the driver chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal, and backlight monitoring current signal at the sampling frequency, and records the hardware timestamp in each sampling period to obtain the original signal matrix. The data packet reception time of each channel recorded by the software layer is obtained as the software reception timestamp. The transmission delay of each channel is determined based on the hardware timestamp and the software reception timestamp, and a delay compensation matrix is ​​constructed. Based on the delay compensation matrix, phase compensation is performed on the driver chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal and backlight monitoring current signal to generate a time-aligned signal set. The time-aligned signal set is subjected to sliding window mid-value filtering to generate a filtered signal set; Each channel signal in the filtered signal set is normalized to generate a standardized electrothermal signal sequence. The standardized electrothermal signal sequence is non-overlappingly segmented according to a preset duration to generate a multi-source electrothermal signal sequence.

[0006] Optionally, the construction of a batch association graph structure with multiple optical modules within the same batch as nodes includes: Obtain the identification code of each optical module, and identify optical modules in the same batch based on the identification code; The aging process formula parameters of the optical modules in the same batch are obtained and encoded into a process fingerprint vector. The aging process formula parameters include temperature cycling curve setting value, current stress level, heating rate and heat preservation time. The process fingerprint vector is used to characterize the current thermal stress environment and electrical stress environment of the optical module. Based on the thermoelectric cooler temperature signal and thermoelectric cooler operating current signal in the standardized electrothermal signal sequence, the real-time thermal state characteristics of each optical module are determined. The real-time thermal state characteristics include the current junction temperature estimate and heat flux density. The real-time thermal state characteristics are corrected based on the temperature cycling curve set value in the aging process formula parameters to generate thermal state correction characteristics. Using each optical module in the same batch as a graph node, the thermal state correction feature is used as the initial feature of the node, and the process fingerprint vector is concatenated with the initial feature of the node to form a composite feature of the node. The importance weight of nodes is set according to the current stress level in the aging process formula parameters, and a weighted graph node set is generated based on the node composite characteristics and the node importance weight. Determine the similarity of the composite features of any target nodes in the graph node set, and obtain the physical distance and process time difference of the target nodes in the aging furnace. The physical distance is determined based on the aging furnace compartment number, and the process time difference is the difference in the time of entering the aging furnace. When the similarity of the composite features of nodes is greater than a preset similarity threshold, the physical distance is less than a preset distance threshold, and the process time difference is less than a preset time difference threshold, a graph connection edge is established between the target nodes, and a graph connection edge set is generated. An adjacency matrix is ​​constructed based on the graph edge set, and the adjacency matrix is ​​then symmetricized and self-connected to generate an enhanced adjacency matrix. The enhanced adjacency matrix is ​​row normalized to obtain the message passing weight matrix; Based on the message passing weight matrix and the weighted graph node set, a batch association graph structure is constructed with multiple optical modules in the same batch as nodes.

[0007] Optionally, the step of inputting the multi-source electrothermal signal sequence and the batch association graph structure into a graph-enhanced long short-term memory network for cross-module information aggregation to obtain the laser equivalent junction temperature estimate and the threshold current aging rate estimate includes: A graph-enhanced long short-term memory network is constructed, wherein the graph-enhanced long short-term memory network includes an encoding layer, a graph message passing layer, a long short-term memory layer, and a dual decoder head structure; The multi-source electrothermal signal sequence is input into the coding layer, and the temporal features of the multi-source electrothermal signal sequence are extracted by a one-dimensional convolutional network to generate the encoded feature sequence corresponding to each node. The message passing weight matrix in the batch association graph structure and the encoded feature sequence are input together into the graph message passing layer. The graph message passing layer aggregates the hidden states of the neighbor nodes of each node according to the message passing weight matrix to generate an aggregated node representation vector. The aggregated node representation vector is input into the long short-term memory layer. During the cell state update process of the long short-term memory layer, the laser heat conduction physical equation is embedded as a bias term to generate the updated cell state and hidden state. The laser heat conduction physical equation is used to describe the dynamic relationship between the laser junction temperature and the changes in electrical power input and ambient temperature. A graph attention mechanism is introduced at the output of the long short-term memory layer. Based on the neighbor relationships of each node in the batch association graph structure, the attention weights of different neighbor nodes on the current node state update are determined. The attention weights are then weighted and fused with the hidden state to generate an attention-enhanced node state vector. The attention-enhanced node state vector is input into the dual-decoder structure. The laser equivalent junction temperature estimate is output through the fully connected network of the first decoder, and the threshold current drift is output through the fully connected network of the second decoder. The threshold current drift is numerically differentiated to obtain an estimated threshold current aging rate.

[0008] Optionally, the step of predicting the optical power deviation at a future target time using an optical performance prediction model based on the estimated equivalent junction temperature of the laser, the estimated threshold current aging rate, and the current cumulative aging time includes: The estimated equivalent junction temperature of the laser is subjected to a moving average filter to generate a filtered junction temperature sequence. Based on the junction temperature values ​​at each time point in the filtered junction temperature sequence, a temperature-accelerated aging factor is constructed based on the Arrhenius equation, generating a temperature-accelerated aging factor sequence corresponding to each time point. The temperature-accelerated aging factor is used to characterize the exponential acceleration effect of junction temperature on the aging rate. Obtain the current cumulative aging time, determine the equivalent aging rate based on the estimated threshold current aging rate and the factor value at the corresponding time in the temperature accelerated aging factor sequence, and integrate the equivalent aging rate over the prediction time domain to obtain the predicted cumulative aging amount. The junction temperature value at the predicted start time in the filtered junction temperature sequence and the predicted aging accumulation are input into the optical performance prediction model to obtain the predicted optical power at the future target time. The optical performance prediction model includes the exponential decay relationship between laser optical power and junction temperature and the linear decay relationship between optical power and aging accumulation. The optical power deviation at the future target time is determined based on the predicted optical power and the target optical power.

[0009] Optionally, the step of solving the collaborative compensation trajectory of bias current and thermoelectric cooler temperature based on the optical power deviation, and generating a bias current compensation sequence and a thermoelectric cooler temperature setpoint correction sequence, includes: An electrothermal coupling dynamic model for an optical module is constructed. The electrothermal coupling dynamic model includes the laser junction temperature state equation, the thermoelectric cooler heat flow transport equation, and the bias current-optical power nonlinear mapping relationship. The bias current-optical power nonlinear mapping relationship is used to correct the gain coefficient in real time using the current threshold current aging rate estimate. Based on the optical power deviation and the electrothermal coupling dynamic model, a rolling optimization objective function in the finite time domain is constructed. The rolling optimization objective function includes a quadratic term for optical power tracking error, a control increment smoothness penalty term, and a thermoelectric cooler power consumption economy weight term. The estimated equivalent junction temperature of the laser is used as the initial value for state feedback. The bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence are used as optimization decision variables. The state is deduced in the prediction time domain in combination with the electrothermal coupling dynamic model to generate a predicted state trajectory. The predicted state trajectory includes a predicted junction temperature sequence, a predicted optical power sequence, and a predicted thermoelectric cooler power consumption sequence. Based on the maximum cooling power of the thermoelectric cooler, the maximum output current of the driver chip, and the maximum allowable junction temperature of the laser, a set of time-varying constraints is constructed. Based on the predicted state trajectory and the set of time-varying constraints, the rolling optimization objective function is solved by the sequential quadratic programming algorithm to generate the optimal control sequence. The optimal control sequence includes the optimal subsequence for bias current compensation and the optimal subsequence for thermoelectric cooler temperature setpoint correction. The optimal control sequence is subjected to feasibility pruning to generate a pruned feasible control sequence. The first control quantity is extracted from the pruned feasible control sequence as the bias current compensation value and thermoelectric cooler temperature setpoint correction value at the current moment. The remaining control sequence is used as the optimized hot start initial value for the next sampling period until the prediction time domain ends, generating the bias current compensation sequence and thermoelectric cooler temperature setpoint correction sequence.

[0010] Optionally, the step of sending the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence to the drive circuit and temperature control unit of the optical module in stages to perform timing-coordinated feedforward compensation includes: Obtain the hardware response bandwidth parameters of the current optical module driver circuit and temperature control unit, including the bias current adjustment response time and the thermoelectric cooler temperature adjustment response time. Based on the ratio of the bias current adjustment response time to the thermoelectric cooler temperature adjustment response time, a time misalignment coefficient for compensation is determined. Then, based on the time misalignment coefficient, the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence are aligned on the time axis to generate a time misalignment compensation sequence pair. In the time misalignment compensation sequence pair, the thermoelectric cooler temperature setpoint correction sequence is issued earlier than the bias current compensation sequence, and the advance amount is equal to the difference between the thermoelectric cooler temperature adjustment response time and the bias current adjustment response time. The bias current compensation sequence in the time misalignment compensation sequence pair is encapsulated in the driving circuit instruction format to generate the first instruction frame sequence. The thermoelectric cooler temperature setpoint correction sequence in the time misalignment compensation sequence pair is encapsulated in a temperature control unit instruction configuration manner to generate a second instruction frame sequence. The first instruction frame sequence and the second instruction frame sequence are sent to the driving circuit and temperature control unit of the optical module in batches to perform timing collaborative feedforward compensation.

[0011] Optionally, the step of sending the first instruction frame sequence and the second instruction frame sequence to the driving circuit and temperature control unit of the optical module in batches to perform timing collaborative feedforward compensation includes: Based on the current communication link status, a batching strategy is determined. According to the batching strategy, the first instruction frame sequence is sent to the driving circuit of the optical module in batches at a preset first time interval, and the second instruction frame sequence is sent to the temperature control unit of the optical module in batches at a preset second time interval. After each batch is sent, the corresponding hardware handshake confirmation signal is waited for. After receiving the handshake confirmation signal from the drive circuit, the deviation between the actual execution timestamp of the bias current and the preset execution timestamp in the bias current compensation sequence is read, and a bias current timing deviation compensation factor is generated. After receiving the handshake confirmation signal from the temperature control unit, the deviation between the actual execution timestamp of the thermoelectric cooler temperature and the preset execution timestamp in the thermoelectric cooler temperature setpoint correction sequence is read, and a thermoelectric cooler timing deviation compensation factor is generated. The bias current timing deviation compensation factor and the thermoelectric cooler timing deviation compensation factor are input to the timing synchronization regulator so that the timing synchronization regulator dynamically adjusts the first time interval and the second time interval according to the difference between the bias current timing deviation compensation factor and the thermoelectric cooler timing deviation compensation factor until the actual adjustment time of the bias current and the thermoelectric cooler temperature reaches a cooperative convergence state within a preset synchronization error threshold.

[0012] Furthermore, to achieve the above objectives, this application also proposes an optical module performance control device based on multi-source state fusion and model prediction feedforward. The optical module performance control device based on multi-source state fusion and model prediction feedforward includes: The acquisition module is used to synchronously acquire the bias voltage signal of the driver chip, the operating current signal of the thermoelectric cooler, the temperature signal of the thermoelectric cooler, and the backlight monitoring current signal in real time at the optical module aging test station, and generate a multi-source electrothermal signal sequence. The building module is used to construct a batch association graph structure with multiple optical modules in the same batch as nodes; The aggregation module is used to input the multi-source electrothermal signal sequence and the batch association graph structure into the graph-enhanced long short-term memory network for cross-module information aggregation, so as to obtain the laser equivalent junction temperature estimate and the threshold current aging rate estimate.

[0013] The prediction module is used to predict the optical power deviation at a future target time based on the estimated equivalent junction temperature of the laser, the estimated threshold current aging rate, and the current cumulative aging time using an optical performance prediction model.

[0014] The generation module is used to solve the collaborative compensation trajectory of bias current and thermoelectric cooler temperature based on the optical power deviation, and generate bias current compensation sequence and thermoelectric cooler temperature setpoint correction sequence.

[0015] The compensation module is used to send the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence to the drive circuit and temperature control unit of the optical module in stages to perform timing-coordinated feedforward compensation in order to achieve consistent control of the optical module performance.

[0016] In addition, to achieve the above objectives, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the optical module performance control method based on multi-source state fusion and model prediction feedforward as described above.

[0017] The one or more technical solutions proposed in this application realize cross-module information aggregation by constructing a batch association graph structure and a graph-enhanced long short-term memory network, thereby improving the accuracy of junction temperature estimation and aging rate estimation. Based on the collaborative compensation trajectory solution of model predictive control, it realizes the early prediction and compensation of future optical power deviation. By solving the electro-thermal response difference through time misalignment alignment, it significantly improves the consistency of optical power and control accuracy within the batch of mass-produced optical modules. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an embodiment of the optical module performance control method based on multi-source state fusion and model prediction feedforward provided in this application. Figure 2 This is a schematic diagram of the module structure of the optical module performance control device based on multi-source state fusion and model prediction feedforward according to an embodiment of this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone; or an electronic device capable of performing the above functions, such as an optical module performance control device based on multi-source state fusion and model prediction feedforward. The following description uses an optical module performance control device based on multi-source state fusion and model prediction feedforward as an example to illustrate this embodiment and the subsequent embodiments.

[0025] Based on this, embodiments of this application provide a method for controlling the performance of optical modules based on multi-source state fusion and model prediction feedforward, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the optical module performance control method based on multi-source state fusion and model prediction feedforward in this application.

[0026] In this embodiment, the optical module performance control method based on multi-source state fusion and model prediction feedforward includes steps S10~S60: Step S10: Real-time synchronous acquisition of driver chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal and backlight monitoring current signal at the optical module aging test station to generate multi-source electrothermal signal sequence.

[0027] It should be noted that the optical module aging test station is a dedicated test unit used for accelerated aging tests on optical modules under conditions such as high temperature, high humidity, and high current stress. During the production of optical communication devices, optical modules need to undergo aging tests in an environment simulating long-term operating stress to screen for early failures and evaluate their lifespan performance. This station includes an aging furnace, a multi-channel test board, and a data acquisition and control unit. The aging furnace provides a controllable high-temperature environment, such as 50℃~85℃; the multi-channel test board simultaneously supports multiple optical modules, providing electrical connections and optical interfaces; and the data acquisition and control unit monitors the operating parameters of each module in real time and issues adjustment commands.

[0028] It is understandable that the bias voltage signal of the driver chip refers to the DC bias voltage applied to the laser diode by the output terminal of the laser driver IC, which is a direct control quantity for controlling the laser injection current. During the aging process, as the laser threshold current gradually increases, the bias voltage needs to be appropriately increased to maintain the target optical power. According to the current-voltage characteristics of the laser, the relationship between the bias voltage and the injection current is as follows: in, The bias current injected into the laser, This is the bias voltage output by the driver chip. This is the threshold voltage of the laser diode. For series resistance, This is the dynamic resistance of the laser diode.

[0029] By monitoring the bias voltage output by the driver chip, the actual bias current injected into the laser can be indirectly estimated.

[0030] The thermoelectric cooler's operating current signal refers to the current value flowing through the TEC inside the optical module. The thermoelectric cooler achieves temperature control of the laser chip through the Peltier effect: forward current for cooling and reverse current for heating. The magnitude and direction of this current determine the cooling / heating power of the thermoelectric cooler, directly affecting the stability of the laser junction temperature.

[0031] The thermoelectric cooler temperature signal refers to the temperature monitoring value of the cold end of the thermoelectric cooler, that is, the side close to the heat sink of the laser. It is usually collected by a negative temperature coefficient thermistor or thermocouple.

[0032] The backlight monitoring current signal refers to the photocurrent generated by the photodiode integrated inside the laser package. This photodiode receives the leakage light from the back of the laser, and its output current is proportional to the forward output light power of the laser.

[0033] It is understandable that a multi-source electrothermal signal sequence refers to a structured time-series data unit formed after time synchronization, preprocessing, and non-overlapping segmentation of the aforementioned four types of signals.

[0034] In one feasible implementation, step S10 may include: determining the sampling frequency and sampling duration according to the optical module model and aging process requirements, and deploying a multi-channel synchronous data acquisition unit at the aging test station; synchronously acquiring the driver chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal, and backlight monitoring current signal at the sampling frequency through the multi-channel synchronous data acquisition unit, and recording the hardware timestamp in each sampling cycle to obtain the original signal matrix; obtaining the data packet reception time of each channel recorded by the software layer as the software reception timestamp, and determining the software reception timestamp based on the hardware timestamp and the software reception timestamp. The timestamp determines the transmission delay of each channel, and a delay compensation matrix is ​​constructed. Phase compensation is performed on the driver chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal, and backlight monitoring current signal based on the delay compensation matrix to generate a time-aligned signal set. The time-aligned signal set is then subjected to sliding window mid-range filtering to generate a filtered signal set. Each channel signal in the filtered signal set is normalized to generate a standardized electrothermal signal sequence. The standardized electrothermal signal sequence is then non-overlapped according to a preset duration to generate a multi-source electrothermal signal sequence.

[0035] It should be noted that the sampling frequency must satisfy the Nyquist sampling theorem, meaning the sampling frequency should be greater than twice the highest frequency component of the measured signal. In optical module aging tests, the bandwidth of the driver chip bias voltage signal is mainly determined by power supply noise and modulation response, typically in the kilohertz range; the bandwidth of the thermoelectric cooler operating current signal is determined by the thermoelectric cooler's response speed, typically in the hertz range; the bandwidth of the thermoelectric cooler temperature signal is determined by the thermal time constant, typically in the millihertz range; and the bandwidth of the backlight monitoring current signal is determined by the laser's direct modulation response, typically in the megahertz range.

[0036] The highest frequency component of each channel signal is determined based on the optical module model, and four times the maximum value is taken as the unified sampling frequency. For example, for a 10G optical module, the bandwidth of the backlight monitoring current signal is approximately 10MHz, so the sampling frequency is determined to be 40MHz; for low-speed optical modules, the sampling frequency can be appropriately reduced to save storage resources. The sampling duration is determined according to the temperature cycle in the aging process formula, usually covering a complete temperature cycle to ensure the capture of transient and steady-state characteristics. The aging process typically includes multiple stages such as heating, holding, and cooling, with each stage lasting from several seconds to tens of seconds. To ensure that the acquired signal can reflect the complete aging process, this embodiment preferably uses a sampling duration of 10 seconds, which includes sufficient timing information for state estimation while avoiding excessive data volume that could lead to processing delays. The sampling frequency can be adjusted within the range of 1kHz to 10kHz, and the sampling duration can be adjusted within the range of 5 seconds to 30 seconds, depending on the specific optical module model and aging process requirements, to achieve a balance between signal fidelity and processing efficiency.

[0037] Understandably, the multi-channel synchronous data acquisition unit includes a four-channel analog front-end, an analog-to-digital converter array, a field-programmable gate array clock synchronization module, and a high-speed data buffer. The four-channel analog front-end is designed specifically for the characteristics of four different signals. Specifically, the driver chip bias voltage signal channel is a high-input-impedance voltage follower with a bandwidth of DC-1MHz and an input range of ±5V; the thermoelectric cooler operating current signal channel is a Hall-effect current sensor or a precision sampling resistor with a bandwidth of DC-10kHz and a range of ±3A; the thermoelectric cooler temperature signal channel is an NTC thermistor or thermocouple excited by a constant current source, amplified after signal conditioning, with a bandwidth of DC-1Hz and a range of -10℃ to +100℃; and the backlight monitoring current signal channel is a transimpedance amplifier with an adjustable gain of 10. 3 -10 6 V / A, bandwidth DC-100MHz, noise equivalent power less than -80dBm. The analog-to-digital converter (ADC) array employs a four-channel synchronous sampling ADC, with each channel sharing a sampling clock to ensure strict synchronization at the hardware level. A field-programmable gate array (FPGA) is used to construct a global clock distribution network, generating a 10MHz synchronous clock, which is distributed to each ADC channel via a fan-out buffer. The ADC conversion is triggered on the rising edge of each sampling cycle, and a 64-bit hardware counter value is latched as a hardware timestamp. The hardware timestamp accuracy is determined by the counter clock frequency; for example, a 100MHz clock corresponds to a 10ns time resolution. The original signal matrix is ​​stored in a two-dimensional array, with row indices representing sampling times and column indices representing signal channels. It includes bias voltage, TEC current, TEC temperature, and backlight monitoring current, with each element containing a sampled value and its corresponding hardware timestamp.

[0038] The software receive timestamp refers to the time recorded by the operating system when a data packet arrives at the host computer software. Due to the multi-level buffering involved in the data transmission path from the analog-to-digital converter through the field-programmable gate array, direct memory access controller, bus interface to the software layer, there is a fixed or random delay difference between the software receive timestamp and the hardware timestamp.

[0039] For each sampling point, the difference between the software receiving timestamp and the hardware timestamp is determined. The delay distribution of multiple sampling points is statistically analyzed, and the median or mean is taken as the transmission delay estimate of the channel to form a delay compensation matrix. This matrix is ​​used to correct the time offset caused by different transmission paths in each channel. The delay compensation matrix is ​​a diagonal matrix, and the diagonal elements are the complex exponential phase factors of the transmission delay of each channel.

[0040] Each channel signal is converted to the frequency domain by Fast Fourier Transform, multiplied by a delay compensation matrix for phase rotation, and then restored to the time domain by Inverse Fast Fourier Transform, achieving precise time alignment of each channel signal with an alignment error of less than one sampling period.

[0041] Sliding window median filtering is used to suppress high-frequency switching noise and electromagnetic interference, preserve the baseband characteristics of the signal, and avoid filtering distortion affecting subsequent feature extraction. The window width is determined based on the signal characteristics and noise frequency, typically an integer multiple of the signal period or noise period. For thermoelectric cooler temperature signals, the window width is set to 50ms to cover the typical response time of the thermoelectric cooler; for driver chip bias voltage signals, the window width is set to 1ms to cover the switching cycle of the power supply. The sampling points within the window are sorted by amplitude, and the median is taken as the filtered output at the window center. Compared to mean filtering, median filtering has better robustness to impulse noise and outliers, preventing outliers from spreading to adjacent sampling points. The beginning and end of the window are padded with data using symmetrical or repeated extension methods to ensure that the length of the filtered signal is consistent with the original signal.

[0042] Normalization aims to eliminate differences in the dimensions and numerical ranges of signals from different channels, making signals of different physical quantities comparable and facilitating neural network processing and feature fusion. This implementation uses minimum-maximum normalization, linearly mapping each channel signal to the [0,1] interval. Alternatively, a 10% margin can be reserved, i.e., the target mapping interval is [0.05, 0.95], to prevent truncation distortion when subsequent signals exceed historical ranges. For mass production scenarios, the maximum and minimum values ​​of the signal are determined using statistics from all optical modules within a batch, rather than the historical values ​​of a single optical module, to ensure the consistency of data distribution within the batch.

[0043] Non-overlapping segmentation refers to dividing a continuous long time-series signal into short sequences of fixed length, facilitating batch processing and neural network input. The preset duration is determined based on the stage characteristics and control cycle of the aging process. For example, for a temperature cycle containing three stages—heating, isothermal, and cooling—the preset duration can be set to 10 seconds, covering either a rapid temperature change stage or a steady-state maintenance stage. There is no data overlap between adjacent segments to avoid data redundancy, but the preset duration must contain complete feature information. The multi-source electrothermal signal sequence is a three-dimensional tensor with dimensions [batch size, sequence length, number of channels], where the sequence length equals the preset duration multiplied by the sampling frequency, and the number of channels is 4.

[0044] In the specific implementation, a multi-level clock tree structure is constructed inside the field-programmable gate array (FPGA). The root node is connected to an external crystal oscillator, and the leaf nodes are connected to the clock input terminals of the analog-to-digital converters (ADCs) of each channel through fan-out buffers. The transmission delay difference from each leaf node to the root node is measured to generate a clock skew compensation table. The trigger time of each ADC is adjusted using the clock skew compensation table to ensure that the sampling time deviation of the four channels is less than 50ns. At the same trigger time, the sampled values ​​of each channel and the corresponding hardware timestamp are latched to generate a set of sampled values ​​after trigger alignment. The transmission delay of each channel is determined by comparing the hardware timestamp with the software receive timestamp, and channels with abnormal delays are identified. When the transmission delay of a channel exceeds a preset delay threshold, the channel is marked as an abnormal delay channel and a resynchronization process is triggered. For channels determined to be normal, a delay compensation matrix is ​​constructed, and the original time-aligned data is converted using a fast Fourier transform. The signal matrix is ​​converted to the frequency domain, and phase rotation compensation is performed using a delay compensation matrix. Then, it is restored to the time domain through inverse fast Fourier transform to generate a phase-compensated signal matrix. Based on the thermoelectric cooler's operating status indicator, a pre-stored filter parameter mapping table is retrieved. When the thermoelectric cooler's operating status indicator is in cooling mode, a low-pass characteristic parameter set with a filter window width of 50ms and a cutoff frequency of 10Hz is selected. When the thermoelectric cooler's operating status indicator is in heating mode, a faster response parameter set with a filter window width of 30ms and a cutoff frequency of 20Hz is selected. The selected parameter set is used to perform sliding window mid-range filtering on the phase-compensated signal matrix. During the sliding window mid-range filtering process, abnormal sampling points that exceed the absolute deviation range of ±3 times the median are detected. The abnormal sampling points are replaced by linear interpolation of adjacent normal sampling points, thereby generating a filtered signal set.

[0045] Step S20: Construct a batch association graph structure with multiple optical modules in the same batch as nodes.

[0046] It should be noted that optical modules within the same batch undergo the same wafer growth, chip dicing, packaging coupling, and aging testing processes during manufacturing, exhibiting similar physical characteristics and aging behaviors. By constructing a batch association graph structure, optical modules within the same batch are modeled as graph network nodes. The message passing mechanism of the graph neural network is used to achieve cross-module information aggregation, improving the accuracy and robustness of state estimation for individual optical modules.

[0047] Traditional methods treat each optical module as an independent sample, failing to utilize the correlation between devices within a batch. This implementation explicitly models the similarity relationships between devices within a batch using a graph structure, enabling state estimation to gain information gain from historical or real-time data of similar devices, making it suitable for state inference of early or anomalous samples.

[0048] In one feasible implementation, step S20 may include: obtaining the identification code of each optical module and identifying optical modules in the same batch based on the identification code; obtaining the aging process recipe parameters of the optical modules in the same batch and encoding the aging process recipe parameters into a process fingerprint vector, wherein the aging process recipe parameters include temperature cycling curve setpoint, current stress level, heating rate, and holding time, and the process fingerprint vector is used to characterize the current thermal stress environment and electrical stress environment of the optical module; determining the real-time thermal state characteristics of each optical module based on the thermoelectric cooler temperature signal and thermoelectric cooler operating current signal in the standardized electrothermal signal sequence, wherein the real-time thermal state characteristics include the current junction temperature estimate and heat flux density; correcting the real-time thermal state characteristics based on the temperature cycling curve setpoint in the aging process recipe parameters to generate thermal state correction features; taking each optical module in the same batch as a graph node, using the thermal state correction features as the node initial features, and concatenating the process fingerprint vector with the node initial features to form a node composite feature. The process involves setting node importance weights based on the current stress level in the aging process formula parameters, generating a weighted graph node set based on the node composite features and the node importance weights; determining the similarity of node composite features between any target nodes in the graph node set, and obtaining the physical distance and process time difference of the target nodes in the aging furnace, where the physical distance is determined based on the aging furnace compartment number and the process time difference is the difference in the time of entry into the aging furnace; when the node composite feature similarity is greater than a preset similarity threshold, the physical distance is less than a preset distance threshold, and the process time difference is less than a preset time difference threshold, establishing graph connection edges between target nodes and generating a graph connection edge set; constructing an adjacency matrix based on the graph connection edge set, and performing symmetry processing and self-connection enhancement processing on the adjacency matrix to generate an enhanced adjacency matrix; performing row normalization processing on the enhanced adjacency matrix to obtain a message passing weight matrix; and constructing a batch association graph structure with multiple optical modules in the same batch as nodes based on the message passing weight matrix and the weighted graph node set.

[0049] It should be noted that the optical module identification code is a unique code, typically containing information such as the production batch number, serial number, and production date. The identification code is affixed to the optical module casing in the form of a QR code or barcode, or laser-etched onto a metal base. The process involves reading the optical module identification code using a scanning device; parsing the batch number field within the identification code; querying the Manufacturing Execution System (MES) database to verify the validity of the batch number; and grouping optical modules with the same batch number into the same batch set.

[0050] Understandably, the aging process formulation parameters are obtained from the manufacturing execution system or the aging test equipment control system, and preset by process engineers according to product specifications and reliability requirements. These parameters include temperature cycling curve setpoints, current stress levels, heating rates, and holding times. The temperature cycling curve setpoints are the set trajectory of the aging furnace temperature over time, for example, 25℃→85℃→25℃, with a cycle time of 4 hours. The current stress level is the percentage of the bias current relative to the rated operating current, which can be divided into three levels: 80%, 100%, and 120%. The heating rate is the rate at which the temperature rises from a low temperature to a high temperature, such as 1℃ / min, 2℃ / min, and 5℃ / min. The holding time is the time to maintain the temperature at the set point, such as 30min, 60min, and 120min. The temperature cycling curve setpoints determine the thermal stress history, the current stress level determines the electrical stress intensity, and the heating rate and holding time together determine the cumulative thermal shock effect.

[0051] The process fingerprint vector is obtained by encoding the aging process formula parameters, specifically including: parsing the temperature cycling curve setpoints in the aging process formula parameters, extracting the target isothermal temperature, the slope of the heating segment, and the temperature fluctuation tolerance, determining the temperature acceleration factor (calculated based on the Arrhenius equation, reflecting the degree of temperature influence on the aging process), combining the temperature acceleration factor with the current stress level to generate the process stress index; normalizing the target isothermal temperature, the slope of the heating segment, the temperature fluctuation tolerance, the current stress level, and the process stress index, and concatenating them to form a five-dimensional process fingerprint vector; and associating and storing the process fingerprint vector with the unique identifier of the optical module. The formula for calculating the temperature acceleration factor is: in, As a temperature acceleration factor, It is the activation energy, measured in electron volts (eV), ranging from 0.3 to 1.0 eV. Different optical module materials and processes will have different activation energy values. k is the Boltzmann constant, with a value of approximately 8.617 × 10⁻⁶. -5 eV / K, This is a reference temperature, selected from the normal operating temperature of the optical module, for example, 25℃, which translates to 298K in Kelvin. This is the current junction temperature, expressed in Kelvin.

[0052] Based on the thermoelectric cooler hot-junction temperature signal and thermoelectric cooler operating current signal in the standardized electrothermal signal sequence, the laser junction temperature is estimated as the current junction temperature estimate, and the heat flux per unit time is calculated as the heat flux density to generate real-time thermal state characteristics. The formulas for calculating the current junction temperature estimate and heat flux density are as follows: in, This is the current estimated junction temperature. This is the temperature signal for the thermoelectric cooler. For thermal resistance parameters, The electrical power of the thermoelectric cooler. This refers to the cooling capacity of the thermoelectric cooler. For heat flux density, Let A be the optical power and A be the area of ​​the laser chip.

[0053] Real-time thermal state characteristics are estimated based on current measurements and do not account for deviations between process settings and actual execution. When the actual temperature of the aging furnace deviates from the set value, the estimated junction temperature contains systematic errors, thus requiring correction. The real-time thermal state characteristics are corrected based on the temperature cycling curve settings in the aging process recipe parameters. For example, if the aging furnace is set to 85℃ but the actual temperature is 82℃, the current estimated junction temperature is adjusted based on the temperature deviation. Assuming a linear relationship between the temperature deviation and the estimated junction temperature, the estimated junction temperature can be corrected by calculating the proportion of the temperature deviation. The corrected thermal state characteristics more accurately reflect the thermal state of the optical module under actual process conditions, eliminating the impact of process execution deviations on state estimation and improving the consistency and comparability of the characteristics.

[0054] The node composite feature is a comprehensive feature vector formed by concatenating the thermal state correction feature and the process fingerprint vector. It is used to characterize the complete state of the optical module, including the current thermal state and long-term stress environment, as shown in the following formula: in, For the composite feature vector of nodes, This is the thermal state correction feature vector. The process fingerprint vector is a nine-dimensional vector, and the node composite feature vector is a nine-dimensional vector.

[0055] After splicing, batch-level normalization is performed on each dimension to eliminate dimensional differences and make different feature dimensions have similar numerical ranges.

[0056] Based on the current stress level in the aging process formula parameters, an importance weight is assigned to each optical module node. A higher current stress level indicates a faster aging process, more drastic state changes, and greater informational value to neighboring nodes, thus receiving a higher weight. During graph message passing, messages from neighboring nodes are aggregated according to their importance weight, with high-weight nodes having a greater impact on the central node.

[0057] The composite feature similarity of nodes is the cosine similarity of the composite features of any two nodes i and j. The similarity value ranges from -1 to 1, with a larger value indicating greater feature similarity. The preset similarity threshold is preferably 0.85, meaning nodes are considered similar when the included angle is less than approximately 32 degrees. Physical distance is the Euclidean or Manhattan distance determined by the spatial coordinates of the aging furnace compartment number. Aging furnaces are typically multi-layered rack structures, with compartment numbers encoding the layer, row, and column numbers. For example, 3-2-5 represents the 3rd layer, 2nd row, and 5th column. The preset distance threshold is usually set to 1 or 2, meaning optical modules in adjacent compartments or separated by one compartment are considered physically close. The process time difference is the absolute value of the difference between the times two optical modules enter the aging furnace. Optical modules in the same batch should enter the aging furnace as synchronously as possible to ensure consistency in the aging process. In actual production, loading and unloading operations result in time dispersion. The preset time difference threshold is usually set to 300 seconds; optical modules entering within this time window are considered to be in process synchronization. Meeting all three conditions simultaneously ensures that connected nodes are related in three dimensions: similar state, spatial proximity, and time synchronization, thus avoiding spurious connections.

[0058] Symmetry processing involves adding the adjacency matrix to its transpose and then dividing by 2 to make the adjacency matrix symmetric, ensuring that the edge connections are undirected, and making the connection relationships between nodes reciprocal, which conforms to the actual physical meaning.

[0059] Self-join enhancement, following symmetry processing, adds a positive-zero constant to the diagonal elements of the adjacency matrix to strengthen the connectivity of each node. This process allows each node to retain more of its own characteristic information during message passing, preventing excessive dilution of node features after multiple message passes.

[0060] Row normalization is performed on each row of the enhanced adjacency matrix to normalize its elements. By adjusting the sum of each row to 1, the numerical stability of message passing is preserved. The message passing weight matrix and the weighted set of graph nodes are combined into a batch association graph structure. In the graph-augmented long short-term memory network, each node aggregates the hidden states of its neighbors through the message passing weight matrix, achieving cross-module information aggregation and improving state estimation accuracy.

[0061] Step S30: Input the multi-source electrothermal signal sequence and the batch association graph structure into the graph-enhanced long short-term memory network for cross-module information aggregation to obtain the laser equivalent junction temperature estimate and threshold current aging rate estimate.

[0062] It should be noted that the Graph-Enhanced Long Short-Term Memory (LSTM) network is a deep fusion architecture of Graph Neural Networks (GNNs) and Long Short-Term Memory Networks (LSTMs) for processing temporal data with graph topologies. In this embodiment, the Graph-Enhanced LSTM network includes an encoding layer, a graph message passing layer, a LSTM layer, and a dual decoder head structure. Traditional LSTM networks only process single-sample temporal sequences, while the Graph-Enhanced LSTM network introduces spatial relationships between samples through the graph message passing layer and dynamically adjusts the neighbor influence weights through the graph attention layer to achieve cross-module information aggregation.

[0063] The equivalent junction temperature (EJT) of a laser refers to the actual operating temperature of the active region of the laser, which determines the laser's output optical power, threshold current, and lifetime. Since the temperature of the active region cannot be directly measured, the hidden state is estimated from the multi-source electrothermal signals using a graph-enhanced long short-term memory (LSTM) network. The threshold current aging rate refers to the rate of change of the laser's threshold current over time, reflecting the speed of device aging and degradation.

[0064] In one feasible implementation, step S30 may include: constructing a graph-enhanced long short-term memory network, wherein the graph-enhanced long short-term memory network includes an encoding layer, a graph message passing layer, a long short-term memory layer, and a dual decoder head structure; inputting the multi-source electrothermal signal sequence into the encoding layer, extracting temporal features from the multi-source electrothermal signal sequence through a one-dimensional convolutional network to generate encoded feature sequences corresponding to each node; inputting the message passing weight matrix in the batch association graph structure and the encoded feature sequences together into the graph message passing layer, aggregating the hidden states of neighboring nodes of each node according to the message passing weight matrix through the graph message passing layer to generate an aggregated node representation vector; inputting the aggregated node representation vector into the long short-term memory layer, embedding a laser during the cell state update process of the long short-term memory layer. The heat conduction physics equation is used as a bias term to generate updated cell states and hidden states. This equation describes the dynamic relationship between the laser junction temperature and the changes in electrical power input and ambient temperature. A graph attention mechanism is introduced at the output of the long short-term memory layer. Based on the neighbor relationships of each node in the batch association graph structure, the attention weights of different neighbor nodes for updating the current node state are determined. These attention weights are then weighted and fused with the hidden states to generate an attention-enhanced node state vector. This enhanced node state vector is input to the dual decoding head structure. The equivalent junction temperature estimate of the laser is output through the fully connected network of the first decoding head, and the threshold current drift is output through the fully connected network of the second decoding head. The threshold current drift is numerically differentiated to obtain an estimate of the threshold current aging rate.

[0065] It should be noted that a graph-enhanced long short-term memory network is constructed, comprising an encoding layer, a graph message passing layer, a long short-term memory layer, and a dual-decoder structure. The encoding layer is used to extract temporal features from multi-source electrothermal signal sequences, converting the original signals into a high-dimensional latent space representation. The graph message passing layer aggregates neighbor node information based on the message passing weight matrix in the batch-associative graph structure. The long short-term memory layer embeds the laser heat conduction physics equation as a bias term in the cell state update, realizing physical prior-guided temporal modeling. The graph attention mechanism layer dynamically learns the importance weights of different neighbor nodes to the current node, realizing adaptive information aggregation. The dual-decoder structure outputs the laser's equivalent junction temperature and threshold current drift through two independent fully connected network branches.

[0066] Before being used, graph-enhanced long short-term memory (LSSM) networks require pre-training. The training of these networks employs a strategy of supervised learning combined with physical constraint optimization. The training objective is to minimize the composite loss function, enabling the network to accurately estimate the equivalent junction temperature and threshold current aging rate of the laser from a multi-source electrothermal signal sequence. The training process comprises three stages: a pre-training stage using high-fidelity simulation data and a small amount of real data for initial training; a physical alignment stage introducing residual constraints from physical equations to enhance the physical rationality of the network's output; and an online incremental learning stage continuously updating the network parameters during mass production to adapt to batch-to-batch differences in devices.

[0067] It's worth noting that during the pre-training phase, a high-fidelity simulation dataset containing 100,000 samples was loaded, network parameters were initialized, the Adam optimizer was used, the learning rate was 1e-3, and training was conducted for 200 epochs. The model was saved every 5 epochs, and the validation set loss was monitored. Training stopped when the validation loss did not decrease for 10 consecutive epochs. During the physical alignment phase, an experimental calibration dataset containing 5,000 samples, including measured thermocouple junction temperatures, was loaded. The pre-trained model parameters were loaded, the learning rate was reduced to 1e-4, and the physical constraint weights λ were increased. phy =0.5, trained for 50 epochs, with the residual of the physical equation evaluated after each epoch, and the model with the smallest residual of the physical equation saved. In the online incremental learning phase, the model is deployed to the production line, data is collected in real time, and an experience replay buffer with a capacity of 10,000 samples is maintained. An incremental update is performed every 500 new samples accumulated, and the EWC regularization term is determined to prevent catastrophic forgetting. The learning rate is 1e-5, and updates are performed for 3 epochs. After each update, the model performance is evaluated, and if it declines, it is rolled back.

[0068] After training, the multi-source electrothermal signal sequence is input into the encoding layer. The encoding layer converts the original multi-source electrothermal signal sequence into a high-dimensional feature representation, extracting temporal dynamic patterns. It consists of a one-dimensional convolutional network containing multiple convolutional blocks. Each convolutional block includes: a one-dimensional convolutional layer with a kernel size of 3 or 5 to extract local temporal patterns; a batch normalization layer to stabilize the training process; an activation function, ReLU or LeakyReLU, introducing nonlinearity; and a max-pooling layer for downsampling and extracting salient features. The encoded feature sequence is a three-dimensional tensor [batch size, sequence length, feature dimension], typically with a feature dimension of 64 or 128. The one-dimensional convolutional network can extract local patterns in the signal, such as the slope of temperature rise and the step response of bias current. Pooling operations reduce the temporal dimension while preserving important features. For example, the temperature rise slope feature in the TEC temperature signal is captured by the convolutional kernel, serving as an important basis for subsequent junction temperature estimation.

[0069] The graph message passing layer, based on the message passing weight matrix in the batch-related graph structure, weights and aggregates the hidden states of each node's neighbor nodes, enabling information propagation within the graph structure. This layer achieves cross-module information aggregation, i.e., information sharing between different optical modules within the same batch. The message passing mechanism is as follows: in, Indicates that node i is in l +1 layer hidden state, The activation function is either Sigmoid or ReLU, used to introduce a nonlinear transformation. Let be the set of neighboring nodes of node i. It is the first l The message passing weights from node j to node i in each layer are derived from the message passing weight matrix. Is node y at the th l The hidden state of the layer It is the first l The bias vector of the layer.

[0070] Through this message-passing mechanism, each node can integrate information from its neighbors to update its hidden state, improving the robustness of state estimation. In graph-reinforced long short-term memory networks, the graph message-passing layer enables information exchange between different optical modules, breaking the limitation of traditional models that only consider a single module, thereby improving the estimation accuracy of the laser's equivalent junction temperature and threshold current aging rate.

[0071] The Long Short-Term Memory (LSTM) layer embeds the laser heat conduction physics equation as a bias term during cell state updates, ensuring that the network's temporal evolution conforms to thermodynamic laws. This embedding method makes the network output both data-driven and physically constrained, improving the model's generalization ability and physical interpretability. The physical constraint bias term is the calculation result of the embedded laser heat conduction physics equation, guiding the network to learn the junction temperature evolution law that satisfies energy conservation.

[0072] Graph attention is an adaptive weighting mechanism introduced at the output of the Long Short-Term Memory (LSTM) layer. It dynamically calculates the importance weights of different neighboring nodes for the current node's state update. Unlike a fixed message passing weight matrix W, graph attention weights are dynamically calculated based on the current node state, adapting to information aggregation needs under different operating conditions. Graph attention enables the network to automatically identify the most relevant neighbors to the current node. For example, when an optical module is in an accelerated aging phase, such as when the threshold current drifts rapidly, the attention mechanism automatically assigns higher weights to neighboring nodes also in the accelerated phase, while suppressing the influence of neighboring nodes that have not yet entered the accelerated phase. This dynamic weighting mechanism is superior to fixed-weight aggregation.

[0073] The dual-decoder structure refers to starting from the attention-enhanced node state vector and outputting the laser's equivalent junction temperature and threshold current drift through two independent fully connected network branches. This structure decouples two physically different output quantities, facilitating independent optimization and physical constraints.

[0074] The first decoding head is used for junction temperature estimation and consists of a three-layer fully connected network: the first layer has 64 neurons, the second layer has 32 neurons, and the third layer has 1 neuron. Each layer is followed by a ReLU activation function, except for the output layer, which outputs the estimated value of the equivalent junction temperature of the laser. The typical output range is 20℃~100℃.

[0075] The second decoding head, used for threshold current drift, also consists of a three-layer fully connected network: the first layer has 64 neurons, the second layer has 32 neurons, and the third layer has 1 neuron. Each layer is followed by a ReLU activation function to output the threshold current drift amount, with a typical output range of 0~10mA.

[0076] The two decoders share the features extracted from the preceding layers, namely the encoding layer, graph message passing layer, long short-term memory layer, and graph attention layer, but perform backpropagation optimization independently, and their respective loss functions can be weighted separately.

[0077] The threshold current drift output by the second decoder is numerically differentiated to obtain an estimate of the threshold current aging rate.

[0078] Step S40: Based on the estimated equivalent junction temperature of the laser, the estimated threshold current aging rate, and the current cumulative aging time, predict the optical power deviation at the future target time using the optical performance prediction model.

[0079] It should be noted that the current cumulative aging time is the effective aging time that the optical module has accumulated under electrothermal stress since entering the aging test station. The aging rate of the laser is exponentially related to the junction temperature, following the Arrhenius equation. Aging for 1 hour at a higher temperature has the same aging effect as aging for several hours or even tens of hours at a reference temperature. Therefore, the cumulative aging time must be recalculated based on the actual junction temperature history.

[0080] Understandably, the optical performance prediction model is a mathematical model based on the physical mechanism of lasers, used to describe the mapping relationship between optical power and junction temperature and aging accumulation. The model contains two core parts: the exponential decay effect of junction temperature on optical power, and the linear decay effect of aging accumulation on optical power.

[0081] Optical power deviation refers to the difference between the predicted optical power and the target optical power, reflecting the gap between the current state and the expected performance.

[0082] In one feasible implementation, step S40 may include: performing a moving average filtering process on the estimated equivalent junction temperature of the laser to generate a filtered junction temperature sequence; constructing a temperature-accelerated aging factor based on the Arrhenius equation according to the junction temperature values ​​at each time point in the filtered junction temperature sequence, generating a temperature-accelerated aging factor sequence corresponding to each time point, wherein the temperature-accelerated aging factor is used to characterize the exponential acceleration effect of junction temperature on the aging rate; obtaining the current cumulative aging time, determining the equivalent aging rate according to the estimated threshold current aging rate and the factor value at the corresponding time point in the temperature-accelerated aging factor sequence, and integrating the equivalent aging rate over time in the prediction time domain to obtain the predicted cumulative aging amount; inputting the junction temperature value at the predicted start time in the filtered junction temperature sequence and the predicted cumulative aging amount into the optical performance prediction model to obtain the predicted optical power at the future target time, wherein the optical performance prediction model includes an exponential decay relationship between laser optical power and junction temperature and a linear decay relationship between optical power and cumulative aging amount; and determining the optical power deviation at the future target time based on the predicted optical power and the target optical power.

[0083] It should be noted that the estimated equivalent junction temperature (JT) of the laser may contain high-frequency random noise or instantaneous fluctuations. Moving average filtering is a low-pass filtering method that suppresses high-frequency noise and extracts the trend of junction temperature changes by averaging the estimated JT values ​​over multiple consecutive time points. The preset filter window length is 5 sampling points, corresponding to 5 seconds, i.e., a sampling frequency of 1Hz. This window length is chosen based on the following considerations: junction temperature changes are jointly determined by the electrical power input and TEC regulation, and their time constants are typically on the order of several seconds to tens of seconds. A 5-second window effectively suppresses noise without excessively smoothing the dynamic changes in junction temperature. For the first 4 points of the time series, a symmetrical window truncation is used; that is, if the number of valid points within the window is less than 5, the average of the valid points is taken. The filtered junction temperature sequence eliminates the influence of high-frequency noise and can more realistically reflect the physical trend of junction temperature changes. For example, when the bias current of the driver chip undergoes a step change, the junction temperature should show an exponential upward trend, rather than an instantaneous jump; filtering can suppress glitches caused by sensor noise or estimation errors, making the subsequent calculation of the temperature-accelerated aging factor more stable.

[0084] The temperature-accelerated aging factor is a physical quantity describing the effect of junction temperature on the aging rate. The aging process of a laser follows a thermal activation mechanism, and its aging rate is related to temperature by the Arrhenius equation: for every certain increase in temperature, the aging rate increases exponentially. This factor is used to convert the aging rate at the actual junction temperature to the equivalent aging rate at a reference temperature. The formula for the temperature-accelerated aging factor is the same as that mentioned earlier and will not be repeated here. Based on the junction temperature values ​​at each moment in the filtered junction temperature sequence, the corresponding temperature-accelerated aging factor can be calculated using this formula, thus generating a temperature-accelerated aging factor sequence. This sequence clearly reflects the influence of junction temperature at different moments on the aging rate.

[0085] The equivalent aging rate refers to the equivalent aging rate at the current junction temperature after being adjusted for the temperature acceleration factor, i.e., the equivalent aging rate at the reference temperature. The equivalent aging rate is obtained by multiplying the estimated threshold current aging rate by the temperature acceleration factor. Integrating the equivalent aging rate over the prediction time domain yields the predicted cumulative aging, which is the expected drift of the threshold current over a future period.

[0086] The relationship between laser output power and junction temperature can be derived from the rate equation. The slope efficiency decreases exponentially with increasing junction temperature. Under constant bias current and threshold current, the optical power decreases exponentially with increasing junction temperature. This exponential decay relationship stems from the physical properties of semiconductor materials: increased junction temperature leads to a decrease in the active region gain coefficient and a reduction in internal quantum efficiency. The threshold current increases linearly with aging accumulation. Under constant junction temperature, the optical power has a linear relationship with aging accumulation, and the decay slope is the slope efficiency η. Since η also slowly decays during aging, γ is slightly smaller than η when using a linear model in practice.

[0087] The optical power deviation at the future target time refers to the predicted future optical power P. pre d and target optical power P target The difference between the two values. This deviation signal reflects the degree of deviation between the optical power at the future target time and the expected value without control intervention, and is the object that the subsequent collaborative compensation controller needs to eliminate.

[0088] Step S50: Based on the optical power deviation, solve the collaborative compensation trajectory of bias current and thermoelectric cooler temperature, and generate bias current compensation sequence and thermoelectric cooler temperature setpoint correction sequence.

[0089] It should be noted that the optical power output of the optical module is determined by both the bias current and the laser junction temperature. The junction temperature is further influenced by both the bias current (electrothermal coupling) and the thermoelectric cooler temperature (heat transfer). Therefore, a complex coupling relationship exists between the bias current and the thermoelectric cooler temperature: increasing the bias current can improve optical power, but it also leads to an increase in junction temperature, which in turn inhibits optical power; decreasing the thermoelectric cooler temperature can lower the junction temperature and improve optical power, but it increases power consumption. Based on the predicted optical power deviation, by solving the collaborative compensation trajectory of the bias current and thermoelectric cooler temperature, a set of coordinated control command sequences on the time axis is generated. This allows the two actuators to cooperate and compensate for each other's weaknesses, quickly and accurately eliminating the optical power deviation at the future target time with minimal energy consumption and thermal stress.

[0090] Traditional methods treat bias current regulation and thermoelectric cooler temperature regulation as independent channels and perform feedback control separately. This can easily lead to mutually canceling control actions, such as one channel increasing current and the other channel decreasing temperature. The effects of the two are superimposed, but energy consumption is multiplied. This implementation method uses a collaborative compensation trajectory to solve the control commands of the two channels together under a unified optimization framework, achieving true collaborative control and avoiding the waste of control energy and unnecessary accumulation of thermal stress.

[0091] A collaborative compensation trajectory refers to a sequence of control commands arranged in the predicted time domain along the time axis. It includes both bias current compensation commands and thermoelectric cooler temperature setpoint correction commands, with the commands from the two channels coordinating in time and complementing each other in effect. The bias current compensation sequence is formed by arranging all bias current compensation commands in the collaborative compensation trajectory in chronological order, and the thermoelectric cooler temperature setpoint correction sequence is formed by arranging all thermoelectric cooler temperature setpoint correction commands in the collaborative compensation trajectory in chronological order.

[0092] In its implementation, this method employs a model-based predictive control approach to solve for the collaborative compensation trajectory. The core idea is to use an electrothermal coupled dynamic model to predict future optical power changes, and then use rolling optimization to find the optimal control sequence that minimizes the deviation between the predicted and target optical power. The electrothermal coupled dynamic model is a set of mathematical equations describing the dynamic relationship between bias current, thermoelectric cooler temperature, laser junction temperature, and optical power. This model forms the basis for solving the collaborative compensation trajectory and is used to predict future trends.

[0093] In one feasible implementation, step S50 may include: constructing an electrothermal coupling dynamic model of the optical module, the electrothermal coupling dynamic model including a laser junction temperature state equation, a thermoelectric cooler heat flow transport equation, and a bias current-optical power nonlinear mapping relationship, the bias current-optical power nonlinear mapping relationship being used to correct the gain coefficient in real time using the current threshold current aging rate estimate; based on the optical power deviation and the electrothermal coupling dynamic model, constructing a rolling optimization objective function in a finite time domain, the rolling optimization objective function including a quadratic term for optical power tracking error, a control increment smoothness penalty term, and a thermoelectric cooler power consumption economy weight term; using the laser equivalent junction temperature estimate as the initial value for state feedback, and the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence as optimization decision variables, combining the electrothermal coupling dynamic model to perform state deduction in the prediction time domain, generating a predicted state trajectory, the predicted... The state trajectory includes a predicted junction temperature sequence, a predicted optical power sequence, and a predicted thermoelectric cooler power consumption sequence. A time-varying constraint set is constructed based on the maximum cooling power of the thermoelectric cooler, the maximum output current of the driver chip, and the maximum allowable junction temperature of the laser. Based on the predicted state trajectory and the time-varying constraint set, a sequential quadratic programming algorithm is used to solve the rolling optimization objective function, generating an optimal control sequence. This optimal control sequence includes an optimal subsequence for bias current compensation and an optimal subsequence for thermoelectric cooler temperature setpoint correction. Feasibility pruning is performed on the optimal control sequence to generate a pruned feasible control sequence. The first-step control quantity is extracted from the pruned feasible control sequence as the bias current compensation value and the thermoelectric cooler temperature setpoint correction value at the current moment. The remaining control sequence is used as the initial value for optimized hot-start in the next sampling period until the prediction time domain ends, generating a bias current compensation sequence and a thermoelectric cooler temperature setpoint correction sequence.

[0094] It should be noted that the electrothermal coupling dynamic model is a set of mathematical equations describing the dynamic relationship between bias current, thermoelectric cooler temperature, laser junction temperature, and optical power. This model is the foundation for solving the cooperative compensation trajectory and is used to predict future state changes. The model consists of three core parts: the laser junction temperature state equation, the thermoelectric cooler heat flow transport equation, and the bias current-optical power nonlinear mapping relationship.

[0095] The change in the equivalent junction temperature (EJT) of a laser follows the law of thermal conduction. The laser junction temperature equation of state is: the rate of change of junction temperature equals the heat generated by the input electrical power minus the heat lost through the thermal resistance, divided by the heat capacity. This equation describes how the input electrical power is converted into a change in junction temperature. When the bias current increases, the electrical power increases, and the junction temperature rises; when the TEC temperature decreases, the temperature difference increases, heat dissipation is enhanced, and the junction temperature decreases.

[0096] The heat transfer equation of a thermoelectric cooler shows that the dynamic response of the actual temperature of the thermoelectric cooler to the set value can be approximated by a first-order thermal system. This equation describes the response process of the actual temperature of the thermoelectric cooler when the temperature set value changes. When the set value changes abruptly, the actual temperature approaches the set value exponentially.

[0097] The slope efficiency, or gain coefficient, in the nonlinear mapping relationship between bias current and optical power will decay as the device ages, and is corrected in real time by estimating the aging rate of the current threshold current.

[0098] The rolling optimization objective function is used to quantify the performance of control. When solving for the cooperative compensation trajectory, it is necessary to find the control sequence that minimizes this objective function. The objective function consists of three core parts: a quadratic term for optical power tracking error, a penalty term for control increment smoothness, and a weight term for the economic efficiency of thermoelectric cooler power consumption, as shown in the following equation: Where N is the prediction time domain length, This is the weighting coefficient for optical power tracking error. To predict optical power, For the target optical power, The bias current increment penalty weighting coefficient is used. This is the increment of the bias current. The weighting factor for the temperature setpoint increment penalty of the thermoelectric cooler. For the temperature setpoint increment of the thermoelectric cooler, The power consumption weighting coefficient for thermoelectric coolers. This represents the power consumption of the thermoelectric cooler.

[0099] The quadratic term for optical power tracking error is used to penalize the deviation between the predicted optical power and the target optical power. The quadratic form allows the controller to impose a higher penalty for larger deviations, prompting the rapid elimination of deviations. The smaller the value, the better the optical power tracking effect. The control increment smoothness penalty term is used to penalize drastic changes in control quantity between adjacent control cycles, preventing sudden changes in control commands from impacting the actuator. Increasing the smoothness penalty can extend the service life of the drive circuit and temperature control unit. The thermoelectric cooler power consumption economy weight term is used to optimize the power consumption of the thermoelectric cooler and reduce energy consumption.

[0100] Using the estimated equivalent junction temperature of the laser as the initial value for state feedback, and the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence as optimization decision variables, a state deduction is performed in the prediction time domain using an electrothermal coupled dynamic model to generate a predicted state trajectory. This trajectory includes the predicted junction temperature sequence, the predicted optical power sequence, and the predicted thermoelectric cooler power consumption sequence. The state deduction is a process of progressively calculating future states based on the electrothermal coupled dynamic model, using the initial state and control variables. In this process, the state at each moment is calculated based on the state at the previous moment and the current control variables.

[0101] The predicted junction temperature sequence is obtained by predicting the laser junction temperature at various future moments based on the laser junction temperature state equation, combined with the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence. It reflects the trend of laser junction temperature change over time under different control strategies. The predicted optical power sequence is obtained by predicting the optical power at various future moments based on the nonlinear mapping relationship between bias current and optical power, combined with the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence. This helps to understand the output performance of the optical module in advance. The predicted thermoelectric cooler power consumption sequence is obtained by predicting the thermoelectric cooler power consumption at various future moments based on the thermoelectric cooler heat flow transfer equation and the thermoelectric cooler temperature setpoint correction sequence. It can be used to evaluate the energy consumption of the control strategy.

[0102] It is worth noting that, based on the physical limits of the actuator and the controlled object, a time-varying set of constraints is constructed to ensure that the solved control commands are executable in the actual system without damaging the devices. The upper limit of the junction temperature in the time-varying constraint set is dynamically adjusted downward as the estimated threshold current aging rate increases.

[0103] In one embodiment, constructing a time-varying constraint set may include: acquiring accelerated aging test data of the laser material system; establishing a negative correlation mapping relationship between the upper limit of junction temperature and the threshold current aging rate, wherein the negative correlation mapping relationship indicates that the higher the aging rate, the lower the maximum allowable junction temperature, so as to suppress thermal accelerated aging from entering a positive feedback loop; determining the upper limit of dynamic junction temperature based on the negative correlation mapping relationship and the estimated threshold current aging rate at the current moment, and embedding the upper limit of dynamic junction temperature as a hard constraint into the time-varying constraint set; acquiring the maximum cooling capacity curve of the thermoelectric cooler at the current ambient temperature, and based on... The difference between the estimated equivalent junction temperature of the laser and the target junction temperature determines the required cooling power, and a dynamic upper bound constraint on the thermoelectric cooler drive current is constructed by combining the maximum cooling capacity curve. Based on the power supply voltage margin and bias current adjustment resolution of the driver chip, amplitude and rate constraints on the bias current compensation are constructed. The rate constraint is relaxed as the optical power deviation increases to allow for faster response speeds when there are large deviations. The dynamic junction temperature upper limit, the dynamic upper bound constraint on the thermoelectric cooler drive current, and the amplitude and rate constraints on the bias current compensation are spatiotemporally tensor encoded to generate a time-varying constraint set.

[0104] Sequential quadratic programming (SQP) is an iterative algorithm for solving constrained nonlinear optimization problems. In each iteration, the original problem is approximated as a quadratic programming subproblem, gradually approaching the optimal solution. When solving the rolling optimization objective function, the SQP algorithm continuously updates the control sequence to satisfy the time-varying constraint set, while simultaneously reducing the objective function value. In each iteration, the algorithm calculates the gradient and Hessian matrix of the objective function based on the current control sequence and the predicted state trajectory, thus constructing a quadratic programming subproblem. During iteration, the algorithm checks whether the control sequence satisfies the time-varying constraints. If not, it adjusts the control sequence to meet the constraints. As iteration progresses, the control sequence gradually converges to an optimal solution that satisfies the constraints and minimizes the objective function. When the iteration reaches a preset convergence condition, such as the change in the objective function value being less than a certain threshold, or the number of iterations reaching its upper limit, the algorithm stops iterating and outputs the optimal control sequence. This optimal control sequence includes the optimal subsequence for bias current compensation and the optimal subsequence for thermoelectric cooler temperature setpoint correction.

[0105] Feasibility pruning involves post-processing the optimal control sequence to remove instructions that may violate constraints, ensuring that the final control sequence is executable in the actual system. The control quantity of the first control cycle is extracted from the pruned feasible control sequence as the compensation instruction for the current moment. The remaining control sequence is used as the initial value for the optimized warm-start in the next sampling cycle, reducing the computational load and improving optimization efficiency. The compensation instruction for the current moment is applied to the optical module, adjusting the bias current and thermoelectric cooler temperature setpoint to achieve real-time control of the optical module's performance. In the next sampling cycle, the above steps are repeated. Based on the new optical power deviation and the latest laser equivalent junction temperature estimate, the rolling optimization objective function is reconstructed, and state deduction, solving for the optimal control sequence, and feasibility pruning are performed again. The control strategy is continuously adjusted to adapt to the dynamic changes in the optical module's performance until the prediction time domain ends, generating the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence.

[0106] Step S60: The bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence are sent to the drive circuit and temperature control unit of the optical module in steps to perform timing-coordinated feedforward compensation in order to achieve consistent control of optical module performance.

[0107] It should be noted that the optical module's drive circuit controls the bias current, while the temperature control unit controls the thermoelectric cooler's temperature. These two components have drastically different response times: the bias current adjustment response time is typically in the microsecond range (e.g., 10-100 μs), while the thermoelectric cooler temperature adjustment response time is in the millisecond to second range (e.g., 100 ms-10 s). If the two compensation sequences are sent simultaneously, due to the response time difference, the bias current compensation will take effect before the thermoelectric cooler temperature compensation, resulting in asynchrony on the time axis and potentially causing mutual interference or cancellation of control actions. This implementation uses a step-by-step sending and timing coordination mechanism to stagger and dynamically synchronize the compensation sequences on the time axis based on the hardware response time difference between the two actuators. This ensures that the actual adjustment effects of the two channels take effect simultaneously at the target time, achieving true collaborative feedforward compensation.

[0108] Traditional methods ignore the difference in actuator response time and issue commands simultaneously, resulting in the bias current acting first and the thermoelectric cooler acting later. This time asynchrony between the two can reduce compensation effectiveness or even cause oscillations. This implementation uses a timing coordination mechanism to precisely synchronize the adjustment effects of the two channels at the target time, significantly improving the accuracy and stability of compensation.

[0109] The driving circuit is used to adjust the output current of the laser driver chip according to the bias current compensation command, thereby changing the optical power. It includes a digital-to-analog converter (DAC), a current source, a power amplifier, etc. The bias current adjustment response time is mainly determined by the conversion time of the DAC and the settling time of the power amplifier. The bias current compensation command is usually encapsulated as a 16-bit or 32-bit digital value and transmitted through I²C, SPI, or MDIO bus.

[0110] The temperature control unit is used to adjust the output current of the thermoelectric cooler driver according to the thermoelectric cooler temperature setpoint correction command, thereby controlling the cold junction temperature of the thermoelectric cooler. It includes a temperature sensor, analog-to-digital converter (ADC), PID controller, thermoelectric cooler driver, etc. The thermoelectric cooler temperature regulation response time is determined by the thermoelectric cooler's thermal time constant. The thermoelectric cooler temperature setpoint is usually encapsulated as a 16-bit digital value and transmitted via I²C or MDIO bus.

[0111] Step-by-step delivery refers to sending the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence to the drive circuit and temperature control unit in batches and at different times according to a preset time stagger relationship, rather than sending them simultaneously.

[0112] Timing-coordinated feedforward compensation refers to the simultaneous effect of the adjustment of the bias current and thermoelectric cooler temperature channels at the target time through time misalignment and dynamic synchronization adjustment, thereby achieving coordinated compensation in the sense of feedforward.

[0113] This embodiment provides a performance control method for optical modules based on multi-source state fusion and model prediction feedforward. By constructing a batch association graph structure and a graph-enhanced long short-term memory network, cross-module information aggregation is achieved, improving the accuracy of junction temperature estimation and aging rate estimation. Based on the collaborative compensation trajectory solution of model predictive control, the method realizes the advance prediction and compensation of future optical power deviation. By solving the electro-thermal response difference through time misalignment, the method significantly improves the consistency of optical power and control accuracy within the batch of mass-produced optical modules.

[0114] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter. Based on this, step S60 includes steps S601 to S605: Step S601: Obtain the hardware response bandwidth parameters of the current optical module driving circuit and temperature control unit. The hardware response bandwidth parameters include the bias current adjustment response time and the thermoelectric cooler temperature adjustment response time.

[0115] It should be noted that obtaining the hardware response bandwidth parameters of the current optical module driver circuit and temperature control unit includes the bias current adjustment response time and the thermoelectric cooler temperature adjustment response time. The bias current adjustment response time refers to the time required from the driver circuit receiving the bias current setpoint command to the actual bias current reaching 90% of the target value. The thermoelectric cooler temperature adjustment response time refers to the time required from the temperature control unit receiving the thermoelectric cooler temperature setpoint command to the actual thermoelectric cooler cold junction temperature reaching 90% of the target value. These response time parameters can be obtained by actually testing the driver circuit and temperature control unit using testing equipment such as an oscilloscope.

[0116] Step S602: Based on the ratio of the bias current adjustment response time to the thermoelectric cooler temperature adjustment response time, determine the time misalignment coefficient for compensation issuance, and align the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence on the time axis based on the time misalignment coefficient to generate a time misalignment compensation sequence pair. In the time misalignment compensation sequence pair, the thermoelectric cooler temperature setpoint correction sequence is issued earlier than the bias current compensation sequence, and the advance amount is equal to the difference between the thermoelectric cooler temperature adjustment response time and the bias current adjustment response time.

[0117] It should be noted that, based on the ratio of the bias current adjustment response time to the thermoelectric cooler temperature adjustment response time, a time misalignment coefficient for compensation issuance is determined. This coefficient is then used to align the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence along their time axes, generating a time misalignment compensation sequence pair. Since the thermoelectric cooler temperature adjustment response time is much longer than the bias current adjustment response time, to ensure that the adjustment effects of both channels take effect simultaneously, the thermoelectric cooler command must be issued Δτ seconds before the bias current command. This advance is equal to the difference between the thermoelectric cooler temperature adjustment response time and the bias current adjustment response time.

[0118] The thermoelectric cooler command sequence is reordered according to the time point of the thermoelectric cooler temperature correction command, generating a time-misaligned thermoelectric cooler command sequence. The bias current command sequence retains its original order, but its issuance time is based on the issuance time of the bias current compensation command. The two sequences are misaligned on the time axis, with the thermoelectric cooler command being issued Δτ earlier than the bias current command, forming a time-misaligned compensation sequence pair.

[0119] Step S603: Encapsulate the bias current compensation sequence in the time misalignment compensation sequence pair into a drive circuit instruction format to generate a first instruction frame sequence.

[0120] It should be noted that the bias current compensation sequence in the time misalignment compensation sequence pair is encapsulated into an instruction frame format recognizable by the driver circuit to generate the first instruction frame sequence. The driver circuit has its specific instruction format requirements, involving the arrangement of data bits, the addition of checksums, etc. During the encapsulation process, each compensation value in the bias current compensation sequence must be converted into a suitable data format according to the communication protocol of the driver circuit. For example, if the driver circuit supports I²C bus communication, the compensation value needs to be converted into the corresponding byte stream according to the I²C protocol specification, adding start bits, address bits, data bits, and stop bits, etc.

[0121] Meanwhile, to ensure the accuracy of data transmission, parity check bits or cyclic redundancy check codes are added for verification. The generated first instruction frame sequence is sent to the drive circuit sequentially according to a pre-set timing relationship to achieve precise adjustment of the bias current.

[0122] Step S604: Encapsulate the thermoelectric cooler temperature setpoint correction sequence in the time misalignment compensation sequence pair into a temperature control unit instruction format to generate a second instruction frame sequence.

[0123] It should be noted that the thermoelectric cooler temperature setpoint correction sequence in the time misalignment compensation sequence pair is encapsulated into a command frame format recognizable by the temperature control unit to generate the second command frame sequence. One command frame is sent at a time, and the system waits for an ACK confirmation signal from the drive circuit. The temperature control unit also has its own unique instruction format. When encapsulating the temperature setpoint correction sequence for the thermoelectric cooler, it needs to be processed according to the communication standard and data requirements of the temperature control unit. Taking the I²C or MDIO bus communication commonly used by temperature control units as an example, each parameter in the temperature setpoint correction sequence must be converted into a specified data length and format. For example, the temperature setpoint is converted into 16-bit binary data, and control bits and checksums are added according to the corresponding communication protocol. The resulting second instruction frame sequence will be sent to the temperature control unit in advance according to the time-staggered requirements to ensure that the temperature adjustment and bias current adjustment of the thermoelectric cooler can take effect simultaneously at the target time.

[0124] Step S605: Send the first instruction frame sequence and the second instruction frame sequence to the drive circuit and temperature control unit of the optical module in batches to perform timing collaborative feedforward compensation.

[0125] It should be noted that, after the preceding processing, a first instruction frame sequence and a second instruction frame sequence suitable for the drive circuit and the temperature control unit are obtained. At this point, according to the previously determined time misalignment relationship, the first instruction frame sequence is sent to the drive circuit of the optical module, and the second instruction frame sequence is sent to the temperature control unit.

[0126] After receiving the first instruction frame sequence, the driving circuit adjusts the output current of the laser driver chip according to the instruction information, thereby changing the optical power. Upon receiving the second instruction frame sequence, the temperature control unit adjusts the output current of the thermoelectric cooler driver according to the temperature setpoint in the instruction, thus controlling the cold junction temperature of the thermoelectric cooler. Through this step-by-step delivery and staggered alignment on the time axis, the adjustment effects of the bias current and thermoelectric cooler temperature channels can take effect simultaneously at the target time, achieving true time-coordinated feedforward compensation and effectively improving the consistency and control accuracy of the optical module performance.

[0127] In one feasible implementation, step S605 may include: determining a batching strategy based on the current communication link status; sending the first instruction frame sequence to the optical module's driving circuit in batches according to the batching strategy at a preset first time interval; sending the second instruction frame sequence to the optical module's temperature control unit in batches at a preset second time interval; and waiting for a corresponding hardware handshake confirmation signal after each batch is sent; after receiving the handshake confirmation signal from the driving circuit, reading the deviation between the actual execution timestamp of the bias current and the preset execution timestamp in the bias current compensation sequence, and generating a bias current timing deviation compensation factor; and after receiving the temperature control unit... After the handshake confirmation signal, the deviation between the actual execution timestamp of the thermoelectric cooler temperature and the preset execution timestamp in the thermoelectric cooler temperature setpoint correction sequence is read, and a thermoelectric cooler timing deviation compensation factor is generated. The bias current timing deviation compensation factor and the thermoelectric cooler timing deviation compensation factor are input to the timing synchronization regulator, so that the timing synchronization regulator dynamically adjusts the first time interval and the second time interval according to the difference between the bias current timing deviation compensation factor and the thermoelectric cooler timing deviation compensation factor, until the actual adjustment time of the bias current and the thermoelectric cooler temperature reaches a cooperative convergence state within the preset synchronization error threshold.

[0128] It should be noted that the batching strategy refers to determining how to send the instruction frame sequence to the executor in batches based on the current state of the communication link, such as bus load rate, data packet size, and transmission latency. A reasonable batching strategy can balance bus occupancy time and instruction transmission reliability.

[0129] The batch distribution strategy includes single-frame sequential distribution strategy, batch packet distribution strategy, and adaptive hybrid strategy. The single-frame sequential delivery strategy is suitable for bus idle conditions, i.e., when the bus load rate is less than 30%. One instruction frame is delivered at a time, and the next instruction frame is delivered only after waiting for the hardware handshake confirmation signal ACK. It has high reliability, and the failure of a single frame does not affect other instructions, but the bus occupancy time is long.

[0130] The batch packet delivery strategy is suitable for situations where the bus is busy, i.e. the bus load rate is higher than 70% or the instruction sequence length is large. Multiple instruction frames, such as 10-20 frames, are packaged into a data packet and delivered at once, which can reduce bus occupancy time and improve transmission efficiency. However, the failure of a single packet will result in the loss of multiple instructions, so a retransmission mechanism is required.

[0131] The adaptive hybrid strategy is applicable to normal conditions, i.e., when the bus load rate is between 30% and 70%, and dynamically adjusts the batch size. The higher the load rate, the smaller the batch size, and the lower the load rate, the larger the batch size. The batch size B = B0 × (1 - load_rate), where B0 is the base batch size, such as 10 frames, and load_rate is the bus load rate.

[0132] According to the batch distribution strategy, the first instruction frame sequence and the second instruction frame sequence are distributed to the corresponding actuators in batches at preset time intervals, and a hardware handshake confirmation signal is waited for after each batch is distributed. The first time interval is the distribution interval between adjacent batches of bias current instructions, and the second time interval is the distribution interval between adjacent batches of thermoelectric cooler temperature instructions.

[0133] After each batch is issued, wait for the temperature control unit and drive circuit to return a hardware handshake confirmation signal. ACK indicates that the actuator has successfully received the instruction and is ready to execute it. NACK indicates that the actuator failed to receive the instruction and needs to retransmit it. After receiving NACK, retry 3 times, with an interval of 1ms between each attempt. If all 3 attempts fail, mark the instruction as abnormal, skip the instruction, and trigger an alarm.

[0134] After receiving the handshake confirmation signal from the drive circuit, the actual execution timestamp of the bias current is read, the deviation from the preset execution timestamp in the bias current compensation sequence is determined, and a bias current timing deviation compensation factor is generated. The acquisition of the actual execution timestamp includes: When the drive circuit executes the bias current command, it records the hardware counter value at the execution time. This counter is provided by the FPGA or microcontroller inside the optical module, and the frequency is usually 10MHz. The counter value is read through the I²C or SPI bus as the actual execution timestamp.

[0135] The preset execution timestamp is determined during the generation of the bias current compensation sequence and stored in the instruction frame. Subtracting the preset execution timestamp from the actual execution timestamp yields the deviation. This deviation reflects the difference between the actual and expected execution time of the bias current adjustment. An exponentially weighted moving average filter is applied to this deviation to generate the bias current timing deviation compensation factor.

[0136] Similarly, after receiving the handshake confirmation signal from the temperature control unit, the actual execution timestamp of the thermoelectric cooler temperature is read, and the deviation from the preset execution timestamp in the thermoelectric cooler temperature setpoint correction sequence is determined to generate a thermoelectric cooler timing deviation compensation factor. The method for obtaining the actual execution timestamp of the thermoelectric cooler temperature is similar to that of the actual execution timestamp of the bias current. When the temperature control unit executes the thermoelectric cooler temperature command, it records the hardware counter value at the execution time. This counter is also provided by the FPGA or microcontroller inside the optical module, and its frequency is generally 10MHz. The counter value is read via the I²C or SPI bus as the actual execution timestamp. The preset execution timestamp is determined and stored in the instruction frame when the thermoelectric cooler temperature setpoint correction sequence is generated. Subtracting the actual execution timestamp from the preset execution timestamp yields the deviation, which reflects the difference between the actual execution time and the expected time of the thermoelectric cooler temperature adjustment. An exponentially weighted moving average filter is applied to this deviation to generate the thermoelectric cooler timing deviation compensation factor.

[0137] The bias current timing deviation compensation factor and the thermoelectric cooler timing deviation compensation factor are input to the timing synchronization controller. The controller dynamically adjusts the first time interval and the second time interval based on the difference between the two until the actual adjustment time of the bias current and the thermoelectric cooler temperature reaches a coordinated convergence state within the preset synchronization error threshold. The timing synchronization controller uses a proportional controller or a proportional-integral controller to adjust the transmission interval according to the deviation difference.

[0138] In this embodiment, by using a step-by-step distribution and timing coordination mechanism, the compensation sequence is misaligned and aligned on the time axis according to the hardware response time difference of the two actuators, so that the actual adjustment effect of the two channels takes effect at the target time simultaneously, realizing true collaborative feedforward compensation and improving the accuracy and stability of compensation.

[0139] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the optical module performance control method based on multi-source state fusion and model prediction feedforward. Any simple modifications based on this technical concept are within the protection scope of this application.

[0140] This application also provides a performance control device for optical modules based on multi-source state fusion and model prediction feedforward. Please refer to [link / reference]. Figure 2 The optical module performance control device based on multi-source state fusion and model prediction feedforward includes: The acquisition module 10 is used to synchronously acquire the bias voltage signal of the driver chip, the operating current signal of the thermoelectric cooler, the temperature signal of the thermoelectric cooler, and the backlight monitoring current signal in real time at the optical module aging test station, and generate a multi-source electrothermal signal sequence.

[0141] Module 20 is used to construct a batch association graph structure with multiple optical modules in the same batch as nodes.

[0142] The aggregation module 30 is used to input the multi-source electrothermal signal sequence and the batch association graph structure into the graph-enhanced long short-term memory network for cross-module information aggregation, so as to obtain the laser equivalent junction temperature estimate and the threshold current aging rate estimate.

[0143] The prediction module 40 is used to predict the optical power deviation at a future target time based on the estimated equivalent junction temperature of the laser, the estimated threshold current aging rate, and the current cumulative aging time through an optical performance prediction model.

[0144] The generation module 50 is used to solve the collaborative compensation trajectory of bias current and thermoelectric cooler temperature based on the optical power deviation, and generate bias current compensation sequence and thermoelectric cooler temperature setpoint correction sequence.

[0145] The compensation module 60 is used to send the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence to the drive circuit and temperature control unit of the optical module in stages to perform timing-coordinated feedforward compensation in order to achieve consistent control of the optical module performance.

[0146] The optical module performance control device based on multi-source state fusion and model prediction feedforward provided in this application, employing the optical module performance control method based on multi-source state fusion and model prediction feedforward in the above embodiments, can solve the technical problem of low consistency control accuracy of optical power during the optical module aging process. Compared with the prior art, the beneficial effects of the optical module performance control device based on multi-source state fusion and model prediction feedforward provided in this application are the same as the beneficial effects of the optical module performance control method based on multi-source state fusion and model prediction feedforward provided in the above embodiments, and other technical features in the optical module performance control device based on multi-source state fusion and model prediction feedforward are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0147] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the optical module performance control method based on multi-source state fusion and model prediction feedforward as described above.

[0148] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for optical module performance control based on multi-source state fusion and model predictive feedforward, characterized in that, The method includes: In the optical module aging test station, the bias voltage signal of the driver chip, the operating current signal of the thermoelectric cooler, the temperature signal of the thermoelectric cooler, and the backlight monitoring current signal are collected in real time to generate a multi-source electrothermal signal sequence. Construct a batch association graph structure with multiple optical modules within the same batch as nodes; The multi-source electrothermal signal sequence and the batch association graph structure are input into a graph-enhanced long short-term memory network for cross-module information aggregation to obtain the laser equivalent junction temperature estimate and threshold current aging rate estimate. Based on the estimated equivalent junction temperature of the laser, the estimated threshold current aging rate, and the current cumulative aging time, the optical power deviation at the future target time is predicted using an optical performance prediction model. Based on the optical power deviation, the collaborative compensation trajectory of bias current and thermoelectric cooler temperature is solved, and a bias current compensation sequence and a thermoelectric cooler temperature setpoint correction sequence are generated. The bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence are sent step by step to the drive circuit and temperature control unit of the optical module to perform timing-coordinated feedforward compensation in order to achieve consistent control of optical module performance.

2. The method of claim 1, wherein, The process involves real-time synchronous acquisition of the driver chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal, and backlight monitoring current signal at the optical module aging test station, generating a multi-source electrothermal signal sequence, including: The sampling frequency and sampling duration are determined based on the optical module model and aging process requirements, and a multi-channel synchronous data acquisition unit is deployed at the aging test station. The multi-channel synchronous data acquisition unit synchronously acquires the driver chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal, and backlight monitoring current signal at the sampling frequency, and records the hardware timestamp in each sampling period to obtain the original signal matrix. The data packet reception time of each channel recorded by the software layer is obtained as the software reception timestamp. The transmission delay of each channel is determined based on the hardware timestamp and the software reception timestamp, and a delay compensation matrix is ​​constructed. Based on the delay compensation matrix, phase compensation is performed on the driver chip bias voltage signal, thermoelectric cooler operating current signal, thermoelectric cooler temperature signal and backlight monitoring current signal to generate a time-aligned signal set. The time-aligned signal set is subjected to sliding window mid-value filtering to generate a filtered signal set; Each channel signal in the filtered signal set is normalized to generate a standardized electrothermal signal sequence. The standardized electrothermal signal sequence is non-overlapped and segmented according to a preset duration to generate a multi-source electrothermal signal sequence.

3. The method of claim 1, wherein, The construction of the batch association graph structure with multiple optical modules within the same batch as nodes includes: Obtain the identification code of each optical module, and identify optical modules in the same batch based on the identification code; The aging process formula parameters of the optical modules in the same batch are obtained and encoded into a process fingerprint vector. The aging process formula parameters include temperature cycling curve setting value, current stress level, heating rate and heat preservation time. The process fingerprint vector is used to characterize the current thermal stress environment and electrical stress environment of the optical module. Based on the thermoelectric cooler temperature signal and thermoelectric cooler operating current signal in the standardized electrothermal signal sequence, the real-time thermal state characteristics of each optical module are determined. The real-time thermal state characteristics include the current junction temperature estimate and heat flux density. The real-time thermal state characteristics are corrected based on the temperature cycling curve set value in the aging process formula parameters to generate thermal state correction characteristics. Using each optical module in the same batch as a graph node, the thermal state correction feature is used as the initial feature of the node, and the process fingerprint vector is concatenated with the initial feature of the node to form a composite feature of the node. The importance weight of nodes is set according to the current stress level in the aging process formula parameters, and a weighted graph node set is generated based on the node composite characteristics and the node importance weight. Determine the similarity of the composite features of any target nodes in the graph node set, and obtain the physical distance and process time difference of the target nodes in the aging furnace. The physical distance is determined based on the aging furnace compartment number, and the process time difference is the difference in the time of entering the aging furnace. When the similarity of the composite features of nodes is greater than a preset similarity threshold, the physical distance is less than a preset distance threshold, and the process time difference is less than a preset time difference threshold, a graph connection edge is established between the target nodes, and a graph connection edge set is generated. An adjacency matrix is ​​constructed based on the graph edge set, and the adjacency matrix is ​​then symmetricized and self-connected to generate an enhanced adjacency matrix. The enhanced adjacency matrix is ​​row normalized to obtain the message passing weight matrix; Based on the message passing weight matrix and the weighted graph node set, a batch association graph structure is constructed with multiple optical modules in the same batch as nodes.

4. The method of claim 1, wherein, The step of inputting the multi-source electrothermal signal sequence and the batch association graph structure into a graph-enhanced long short-term memory network for cross-module information aggregation to obtain the laser equivalent junction temperature estimate and threshold current aging rate estimate includes: A graph-enhanced long short-term memory network is constructed, wherein the graph-enhanced long short-term memory network includes an encoding layer, a graph message passing layer, a long short-term memory layer, and a dual decoder head structure; The multi-source electrothermal signal sequence is input into the coding layer, and the temporal features of the multi-source electrothermal signal sequence are extracted by a one-dimensional convolutional network to generate the encoded feature sequence corresponding to each node. The message passing weight matrix in the batch association graph structure and the encoded feature sequence are input together into the graph message passing layer. The graph message passing layer aggregates the hidden states of the neighbor nodes of each node according to the message passing weight matrix to generate an aggregated node representation vector. The aggregated node representation vector is input into the long short-term memory layer. During the cell state update process of the long short-term memory layer, the laser heat conduction physical equation is embedded as a bias term to generate the updated cell state and hidden state. The laser heat conduction physical equation is used to describe the dynamic relationship between the laser junction temperature and the changes in electrical power input and ambient temperature. A graph attention mechanism is introduced at the output of the long short-term memory layer. Based on the neighbor relationships of each node in the batch association graph structure, the attention weights of different neighbor nodes on the current node state update are determined. The attention weights are then weighted and fused with the hidden state to generate an attention-enhanced node state vector. The attention-enhanced node state vector is input into the dual-decoder structure. The laser equivalent junction temperature estimate is output through the fully connected network of the first decoder, and the threshold current drift is output through the fully connected network of the second decoder. The threshold current drift is numerically differentiated to obtain an estimated threshold current aging rate.

5. The method of claim 1, wherein, The prediction of optical power deviation at a future target time based on the estimated equivalent junction temperature of the laser, the estimated threshold current aging rate, and the current cumulative aging time using an optical performance prediction model includes: The estimated equivalent junction temperature of the laser is subjected to a moving average filter to generate a filtered junction temperature sequence. Based on the junction temperature values ​​at each time point in the filtered junction temperature sequence, a temperature-accelerated aging factor is constructed based on the Arrhenius equation, generating a temperature-accelerated aging factor sequence corresponding to each time point. The temperature-accelerated aging factor is used to characterize the exponential acceleration effect of junction temperature on the aging rate. Obtain the current cumulative aging time, determine the equivalent aging rate based on the estimated threshold current aging rate and the factor value at the corresponding time in the temperature accelerated aging factor sequence, and integrate the equivalent aging rate over the prediction time domain to obtain the predicted cumulative aging amount. The junction temperature value at the predicted start time in the filtered junction temperature sequence and the predicted aging accumulation are input into the optical performance prediction model to obtain the predicted optical power at the future target time. The optical performance prediction model includes the exponential decay relationship between laser optical power and junction temperature and the linear decay relationship between optical power and aging accumulation. The optical power deviation at the future target time is determined based on the predicted optical power and the target optical power.

6. The method of claim 1, wherein, The step of calculating the collaborative compensation trajectory between the bias current and the thermoelectric cooler temperature based on the optical power deviation, and generating the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence, includes: An electrothermal coupling dynamic model for an optical module is constructed. The electrothermal coupling dynamic model includes the laser junction temperature state equation, the thermoelectric cooler heat flow transport equation, and the bias current-optical power nonlinear mapping relationship. The bias current-optical power nonlinear mapping relationship is used to correct the gain coefficient in real time using the current threshold current aging rate estimate. Based on the optical power deviation and the electrothermal coupling dynamic model, a rolling optimization objective function in the finite time domain is constructed. The rolling optimization objective function includes a quadratic term for optical power tracking error, a control increment smoothness penalty term, and a thermoelectric cooler power consumption economy weight term. The estimated equivalent junction temperature of the laser is used as the initial value for state feedback. The bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence are used as optimization decision variables. The state is deduced in the prediction time domain in combination with the electrothermal coupling dynamic model to generate a predicted state trajectory. The predicted state trajectory includes a predicted junction temperature sequence, a predicted optical power sequence, and a predicted thermoelectric cooler power consumption sequence. Based on the maximum cooling power of the thermoelectric cooler, the maximum output current of the driver chip, and the maximum allowable junction temperature of the laser, a set of time-varying constraints is constructed. Based on the predicted state trajectory and the set of time-varying constraints, the rolling optimization objective function is solved by the sequential quadratic programming algorithm to generate the optimal control sequence. The optimal control sequence includes the optimal subsequence for bias current compensation and the optimal subsequence for thermoelectric cooler temperature setpoint correction. The optimal control sequence is subjected to feasibility pruning to generate a pruned feasible control sequence. The first control quantity is extracted from the pruned feasible control sequence as the bias current compensation value and thermoelectric cooler temperature setpoint correction value at the current moment. The remaining control sequence is used as the optimized hot start initial value for the next sampling period until the prediction time domain ends, generating the bias current compensation sequence and thermoelectric cooler temperature setpoint correction sequence.

7. The method of claim 1, wherein, The step-by-step distribution of the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence to the optical module's drive circuit and temperature control unit to perform timing-coordinated feedforward compensation includes: Obtain the hardware response bandwidth parameters of the current optical module driver circuit and temperature control unit, including the bias current adjustment response time and the thermoelectric cooler temperature adjustment response time. Based on the ratio of the bias current adjustment response time to the thermoelectric cooler temperature adjustment response time, a time misalignment coefficient for compensation is determined. Then, based on the time misalignment coefficient, the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence are aligned on the time axis to generate a time misalignment compensation sequence pair. In the time misalignment compensation sequence pair, the thermoelectric cooler temperature setpoint correction sequence is issued earlier than the bias current compensation sequence, and the advance amount is equal to the difference between the thermoelectric cooler temperature adjustment response time and the bias current adjustment response time. The bias current compensation sequence in the time misalignment compensation sequence pair is encapsulated in the driving circuit instruction format to generate a first instruction frame sequence. The temperature control unit instruction format is encapsulated in the thermoelectric cooler temperature setpoint correction sequence in the time misalignment compensation sequence pair to generate a second instruction frame sequence. The first instruction frame sequence and the second instruction frame sequence are sent to the driving circuit and temperature control unit of the optical module in batches to perform timing collaborative feedforward compensation.

8. The method of claim 7, wherein, The step of sending the first instruction frame sequence and the second instruction frame sequence to the driving circuit and temperature control unit of the optical module in batches to perform timing-coordinated feedforward compensation includes: Based on the current communication link status, a batching strategy is determined. According to the batching strategy, the first instruction frame sequence is sent to the driving circuit of the optical module in batches at a preset first time interval, and the second instruction frame sequence is sent to the temperature control unit of the optical module in batches at a preset second time interval. After each batch is sent, the corresponding hardware handshake confirmation signal is waited for. After receiving the handshake confirmation signal from the drive circuit, the deviation between the actual execution timestamp of the bias current and the preset execution timestamp in the bias current compensation sequence is read, and a bias current timing deviation compensation factor is generated. After receiving the handshake confirmation signal from the temperature control unit, the deviation between the actual execution timestamp of the thermoelectric cooler temperature and the preset execution timestamp in the thermoelectric cooler temperature setpoint correction sequence is read, and a thermoelectric cooler timing deviation compensation factor is generated. The bias current timing deviation compensation factor and the thermoelectric cooler timing deviation compensation factor are input to the timing synchronization regulator so that the timing synchronization regulator dynamically adjusts the first time interval and the second time interval according to the difference between the bias current timing deviation compensation factor and the thermoelectric cooler timing deviation compensation factor until the actual adjustment time of the bias current and the thermoelectric cooler temperature reaches a cooperative convergence state within a preset synchronization error threshold.

9. A device for controlling performance of an optical module based on multi-source state fusion and model predictive feedforward, characterized in that, The device includes: The acquisition module is used to synchronously acquire the bias voltage signal of the driver chip, the operating current signal of the thermoelectric cooler, the temperature signal of the thermoelectric cooler, and the backlight monitoring current signal in real time at the optical module aging test station, and generate a multi-source electrothermal signal sequence. The building module is used to construct a batch association graph structure with multiple optical modules within the same batch as nodes; The aggregation module is used to input the multi-source electrothermal signal sequence and the batch association graph structure into the graph-enhanced long short-term memory network for cross-module information aggregation, so as to obtain the laser equivalent junction temperature estimate and the threshold current aging rate estimate. The prediction module is used to predict the optical power deviation at a future target time based on the estimated equivalent junction temperature of the laser, the estimated threshold current aging rate, and the current cumulative aging time through an optical performance prediction model. The generation module is used to solve the collaborative compensation trajectory of bias current and thermoelectric cooler temperature based on the optical power deviation, and generate bias current compensation sequence and thermoelectric cooler temperature setpoint correction sequence. The compensation module is used to send the bias current compensation sequence and the thermoelectric cooler temperature setpoint correction sequence to the drive circuit and temperature control unit of the optical module in stages to perform timing-coordinated feedforward compensation in order to achieve consistent control of the optical module performance. 10.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the optical module performance control method based on multi-source state fusion and model prediction feedforward as described in any one of claims 1 to 8.

Images