Temperature control and noise reduction method and temperature control and noise reduction system for analog optical module

Abstract

The invention discloses a temperature control and noise reduction method for an analog optical module, and relates to the technical field of data processing. The method comprises the following steps: collecting and preprocessing an original temperature signal, environment temperature time sequence data and an optical wavelength original detection value of an optical module; improving the LSTM model, constructing a junction temperature dynamic prediction model, and outputting prediction data; a pre-trained BP neural network is utilized to calculate the temperature compensation amount according to the real-time wavelength offset, and a PPO reinforcement learning algorithm is adopted to optimize with the temperature stability, the wavelength offset and the TEC power consumption as multiple targets to obtain an optimal PID parameter mapping table; and finally, predicting, compensating and optimizing parameters are fused, and control output for driving the TEC is generated through a multi-mode intelligent decision. The problems of insufficient control precision and response lag are effectively solved, cooperative high-precision control over the junction temperature and the emission wavelength of the laser is achieved, transmission noise is remarkably reduced, and meanwhile system energy consumption and the self-adaptive capacity are considered.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and specifically relates to a method for temperature control and noise reduction of an analog optical module. Background Technology

[0002] In traditional analog optical modules, the heat dissipation measure for the optical components involves designing the component housing to be in full contact with the overall product housing, relying on passive heat dissipation through the housing. While passive heat dissipation can control the operating temperature of the optical components within a certain range, this range still deviates from the optimal operating temperature. This deviation causes a shift in the wavelength of the emitted light from the optical components, thus affecting the overall signal transmission. Especially when the receiver has strict requirements for the optical wavelength, this wavelength shift can generate significant noise.

[0003] Existing technologies suffer from insufficient control precision and response lag, single control dimension ignoring optical performance indicators, rigid models lacking adaptability and intelligence, contradictions between algorithm complexity and embedded deployment, and the challenge of balancing multi-objective optimization. Summary of the Invention

[0004] (a) Technical problems to be solved To address the problems in related technologies, this invention provides a method for temperature control and noise reduction of an analog optical module, thereby overcoming the aforementioned technical problems existing in the prior art.

[0005] (II) Technical Solution To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: S1. Collect the raw temperature signal of the optical module, the time-series data of the ambient temperature, and the raw detection value of the optical wavelength, and perform preprocessing to obtain the processed real-time multi-source data; S2. Improve the LSTM to obtain the initial LSTM junction temperature dynamic prediction model; use historical optical module feature data combined with optimization algorithm to train the LSTM junction temperature dynamic prediction model to obtain the final LSTM junction temperature dynamic prediction model. Based on the processed real-time multi-source data, the junction temperature prediction data is obtained through the final LSTM junction temperature dynamic prediction model. S3. Pre-train the BP neural network model to obtain the pre-trained BP neural network model; calculate the real-time wavelength offset based on the real-time wavelength; and obtain the real-time temperature compensation amount based on the real-time wavelength offset and the pre-trained BP neural network model. Based on multi-dimensional optimization objectives, the optimal PID parameter mapping table is obtained by training with the PPO reinforcement learning algorithm. S4. Based on the temperature compensation amount, junction temperature prediction data, and optimal PID parameter mapping table, a control output is generated through a PID algorithm; the control output is then converted into a precise current output by driving the TEC. This invention integrates multi-source sensor data, uses an improved LSTM network to achieve dynamic prediction of junction temperature, combines a BP neural network to generate temperature compensation based on wavelength offset, and utilizes PPO reinforcement learning to optimize PID parameter mapping. Finally, it outputs a precise control signal to drive the TEC through an intelligent PID algorithm. This effectively solves the problems of response lag and single control dimension in traditional temperature control methods, significantly improves the wavelength stability and temperature control accuracy of optical modules, and achieves power consumption optimization and greatly reduces transmission noise.

[0006] Preferably, step S1 includes the following steps: S11. Collect raw temperature signals, ambient temperature time-series data, and raw optical wavelength detection values ​​to obtain real-time multi-source data; S12. Preprocess the real-time multi-source data to obtain processed real-time multi-source data; This invention provides high-quality, low-noise standardized input data for subsequent intelligent control models by synchronously acquiring data from multiple sources, fusing and filtering the data, and standardizing the preprocessing. This lays a solid foundation for high-precision control.

[0007] Preferably, step S11 includes the following steps: S111. Using a high-precision thermistor mounted on the back of the laser chip, the laser junction temperature data is collected through a sampling circuit to obtain the original temperature signal. S112. Use the ambient temperature sensor integrated inside the module to collect the ambient temperature of the module's working environment through the digital interface to obtain ambient temperature time-series data. S113. Using an integrated optical power meter and wavelength detection unit, the real-time emission wavelength is collected through a built-in spectrometer and grating spectrometer to obtain the original detection value of the optical wavelength; This invention achieves multi-dimensional direct measurement of the working status of the optical module by using a high-precision thermistor, environmental sensor and spectral detection unit to collect junction temperature, ambient temperature and light wavelength data, thus providing a comprehensive and reliable data foundation for precise temperature control.

[0008] Preferably, step S12 includes the following steps: S121. Using a sliding window filtering algorithm, all raw sensor data in real-time multi-source data are processed by convolution operation to obtain a filtered smooth signal. S122. The filtered smooth signal is organized and sampled using a memory circular buffer to obtain a structured time series dataset. S123. By standardizing with z-score, the sensor dimensions of the structured time-series dataset are transformed to obtain processed real-time multi-source data. This invention effectively eliminates high-frequency noise through sliding window filtering, constructs regular time-series data using a circular buffer, and adopts z-score standardization to unify the units, significantly improving the quality and consistency of multi-source data and providing a stable and reliable input foundation for subsequent intelligent models.

[0009] Preferably, step S2 includes the following steps: S21. Construct an initial LSTM model, improve the initial LSTM model, and obtain an initial LSTM junction temperature dynamic prediction model. S22. Collect historical optical module temperature time-series data, ambient temperature data, and wavelength data to obtain historical optical module characteristic data; S23. Use historical optical module feature data to train the initial LSTM junction temperature dynamic prediction model. During the training process, combine the optimization algorithm to find the learning rate of the initial LSTM junction temperature dynamic prediction model and obtain the optimal solution. Use the optimal solution as the learning rate of the initial LSTM junction temperature dynamic prediction model to obtain the final LSTM junction temperature dynamic prediction model. S24. Extract the features of the processed real-time multi-source data to obtain real-time optical module feature data; input the real-time optical module feature data into the final LSTM junction temperature dynamic prediction model to obtain junction temperature prediction data; This invention achieves high-precision dynamic prediction of the junction temperature of optical modules by constructing an improved LSTM model, employing a composite loss function and model compression technology, and combining optimization algorithms to determine the optimal learning rate, thus significantly improving the anticipation and accuracy of temperature control.

[0010] Preferably, step S21 includes the following steps: S211. Based on the historical temperature single-dimensional time series data, add ambient temperature and driving current as joint input features to the input of the LSTM model, add real-time wavelength detection value as supervision signal, and delete redundant signals unrelated to temperature control to form a three-dimensional feature tensor and establish an electro-thermal-optical coupling model. S212. To minimize the temperature prediction error of the LSTM model, add a wavelength stability penalty term and modify the loss function to a composite form, directly embedding optical indicators into the machine learning optimization objective. S213. Modify the fixed step size temperature prediction of LSTM to predict the critical control window as the target. S214. Remove the complex gating structure of the LSTM model, adopt a single-layer LSTM, and add a knowledge distillation compression process to modify the calculation accuracy and model size to meet the real-time requirements of the embedded MCU. S215. Change the offline training of the LSTM model with fixed model parameters to an online incremental learning module, and modify the weight update strategy to selective update. This invention significantly improves the accuracy and adaptability of the junction temperature prediction model by constructing electro-thermal-optical coupled input features, embedding a wavelength stability penalty term in the loss function, focusing on the prediction of key control windows, using knowledge distillation to achieve lightweight model compression, and introducing an online incremental learning mechanism, while ensuring the real-time performance of the algorithm on embedded hardware.

[0011] Preferably, the step S23, in which the optimization algorithm is used to find the learning rate of the initial LSTM junction temperature dynamic prediction model during training to obtain the optimal solution, includes the following steps: S231. Set the maximum number of iterations and construct a composite evaluation function; S232. Perform a three-layer screening process in the discrete learning rate space, which includes coarse screening, fine-tuning, and validation. S233. Initialize the search interval with the fine-tuned optimal value as the center, and obtain the optimal solution by iteratively shrinking the interval using the golden section method. This invention uses a composite evaluation function combined with three-layer screening and the golden ratio method to automatically and accurately determine the optimal learning rate of the LSTM model, significantly improving the model's convergence speed and prediction accuracy, and effectively avoiding the blindness and inefficiency of traditional manual parameter tuning.

[0012] Preferably, step S3 includes the following steps: S31. Collect historical optical module full life cycle data, including temperature scan data, current disturbance data, and aging dataset; An input vector is constructed based on the historical optical module lifecycle data, and the label is the temperature compensation amount, resulting in an input vector with a temperature compensation label. The BP neural network model uses a three-layer fully connected network. The input layer receives real-time wavelength deviation and aging factor, the hidden layer uses ReLU activation, and a dual-path loss function is designed. A pre-trained BP neural network model is obtained by combining a progressive training strategy with a three-stage verification mechanism; the progressive training strategy includes basic training. S32. Based on the real-time wavelength and preset wavelength in the processed real-time multi-source data, the real-time wavelength offset is obtained by subtraction. The wavelength offset and sensor aging data are input into a pre-trained BP neural network model to obtain the real-time temperature compensation. S33. Collect historical optical module temperature deviation, deviation change rate, ambient temperature and drive current data to obtain historical optical module parameter data; use historical optical module parameter data to construct state space and obtain optical module environmental feature vector; With temperature stability, wavelength shift, and TEC power consumption as optimization objectives, a weighted reward function is constructed. Based on the optical module environmental feature vector and the weighted reward function, the PPO algorithm is used for reinforcement training to obtain the optimal PID parameter mapping table. This invention constructs a BP neural network compensation model based on full lifecycle data, combines a dual-path loss function and a progressive training strategy to accurately calculate real-time temperature compensation. Simultaneously, it utilizes the PPO reinforcement learning algorithm, with temperature stability, wavelength shift, and TEC power consumption as multi-objective optimization functions, to autonomously train and generate an optimal PID parameter mapping table, thereby achieving adaptive and precise control and multi-objective performance optimization of the optical module under complex operating conditions.

[0013] Preferably, step S4 includes the following steps: S41. Add the original target temperature to the temperature compensation amount to obtain the corrected target temperature; By subtraction, the temperature deviation between the target temperature and the predicted junction temperature is calculated and the core control variable is obtained. S42. Set the absolute value threshold for temperature deviation; calculate the real-time absolute value of temperature deviation based on the core control variables; obtain the working status of the optical module based on the real-time absolute value of temperature deviation and the absolute value threshold for temperature deviation. The dynamic PID parameter mode is selected based on the optical module's operating state to obtain the final dynamic PID parameters. When the optical module is in a large deviation mode, a high-gain PID parameter is obtained by hard-coding parameter settings and loading a strong response parameter. This high-gain PID parameter is then used as the final dynamic PID parameter. When the optical module is in a steady-state mode, a scenario-based scaling factor is obtained by querying a PID mapping table generated by machine learning. Based on the real-time wavelength, an integral term correction factor is calculated to obtain an adaptive Ki value. The scenario-based scaling factor and the adaptive Ki value are then used as the final dynamic PID parameters. S43. Based on the final PID dynamic parameters, the control quantity is calculated using the PID algorithm to obtain the control output quantity; the control output quantity is then converted into a precise current output by driving the TEC. This invention generates a correction target by fusing temperature compensation and junction temperature prediction data, and adopts a dual-mode PID parameter dynamic switching mechanism. When there is a large deviation, high-gain parameters are activated for rapid convergence, and an optimized mapping table is called for fine adjustment in steady state. Ultimately, it achieves accurate output of TEC current, which significantly improves the system's control accuracy, response speed and wavelength stability.

[0014] A temperature control and noise reduction system for an analog optical module, used to implement the aforementioned temperature control and noise reduction method for an analog optical module, includes a multi-source data acquisition and preprocessing module, an improved LSTM junction temperature dynamic prediction module, a temperature compensation and intelligent optimization module, and a multi-modal intelligent decision-making and control output module. The multi-source data acquisition and preprocessing module is used to synchronously acquire raw laser junction temperature, ambient temperature and operating wavelength data from multiple sensors of the optical module; a series of preprocessing steps are performed on this set of real-time multi-source data, including using a sliding window filtering algorithm to eliminate high-frequency noise, organizing the data into a structured time-series dataset through a circular buffer, and finally performing z-score standardization to unify the units, so as to provide high-quality and standardized input data for subsequent models. The improved LSTM junction temperature dynamic prediction module is used to deeply improve the LSTM neural network model. The improvement measures include introducing ambient temperature and driving current as joint input features, adding a wavelength stability penalty term to the loss function, focusing on the prediction of key control windows, lightweighting and compressing the model to meet embedded deployment requirements, and integrating online incremental learning capabilities to adapt to device aging. It uses historical data and combines optimization algorithms to automatically find the optimal learning rate for training, and finally uses preprocessed real-time data to predict future junction temperature change trends with high accuracy. The temperature compensation and intelligent optimization module includes a pre-trained BP neural network and an optimal PID parameter mapping table based on the PPO reinforcement learning algorithm. The BP network predicts the required real-time temperature compensation based on the wavelength offset calculated in real time and aging data. The PPO algorithm uses temperature stability, wavelength offset and system power consumption as multi-objective optimization functions. Through reinforcement training with interaction with the environment, it learns and generates an optimal PID parameter lookup table that can adapt to different working states, providing intelligent parameters for the final control decision. The multimodal intelligent decision and control output module serves as the decision-making and execution center of the entire system. It calculates the deviation between the final target temperature and the real-time temperature for control through junction temperature prediction and temperature compensation. Based on the magnitude of the deviation, it determines whether the system is in a "large deviation mode" or a "steady-state mode," and intelligently switches control strategies accordingly. In the case of large deviations, it uses preset strong response parameters for rapid convergence; in the case of steady-state operation, it queries scenario-based parameters from the mapping table generated by the PPO and performs fine-tuning using machine learning functions. These dynamic PID parameters are used to calculate precise control quantities and drive the TEC to perform corresponding current operations, thereby achieving high-precision and high-stability temperature and wavelength control.

[0015] (III) Beneficial Effects The present invention has the following beneficial effects: This invention fundamentally solves the limitations of traditional passive heat dissipation and single temperature feedback control by constructing a multi-source data fusion and intelligent predictive control system. It collects multi-dimensional data such as junction temperature, ambient temperature and light wavelength in real time through a high-precision sensor array, and provides high-quality, low-noise input signals for the model through sliding window filtering and standardized preprocessing, laying the data foundation for high-precision control.

[0016] The core innovation of this invention lies in the multi-dimensional targeted improvements to the LSTM model, which achieves advanced and accurate prediction of junction temperature changes. By introducing electro-thermal-optical coupling input features, embedding wavelength stability penalty terms in the loss function, focusing on the prediction of key control windows, and using knowledge distillation to lighten the model, the prediction model is not only highly accurate and fast in response, but can also directly optimize optical performance indicators. It has been successfully deployed in resource-constrained embedded hardware, meeting the stringent requirements of real-time control. This invention innovatively introduces a wavelength feedback compensation mechanism based on a BP neural network and a multi-objective parameter optimization strategy based on PPO reinforcement learning. The former analyzes the real-time wavelength offset and uses a pre-trained neural network to dynamically calculate the accurate temperature compensation amount, effectively offsetting the temperature drift coefficient change caused by device aging and achieving long-term stability. The latter optimizes the target with multiple dimensions such as temperature, wavelength, and power consumption, and automatically learns the PID parameter mapping table with the best performance under different operating conditions, thereby intelligently achieving the best balance between control performance and system energy consumption.

[0017] This invention ultimately achieves adaptive, high-performance control output by organically combining prediction, compensation, and optimization parameters through a multimodal intelligent decision-making mechanism. Based on the magnitude of real-time temperature deviation, it intelligently switches between a "large deviation mode" and a "steady-state mode," employing high-gain parameters to quickly suppress disturbances or querying an optimized mapping table for fine-tuning, thus combining rapid dynamic response and extremely high steady-state control accuracy. Finally, by driving the TEC to output precise current, it ensures the dual stability of the laser junction temperature and emission wavelength, significantly reducing transmission noise caused by wavelength drift, improving the overall performance and reliability of the optical module, and extending device lifespan.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart of a method for temperature control and noise reduction of an analog optical module according to the present invention. Figure 2 This is a schematic diagram of a temperature control and noise reduction system for an analog optical module according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0022] In the description of this invention, it should be understood that the terms "opening", "upper", "lower", "top", "middle", "inner", etc., which indicate orientation or positional relationship, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the components or elements referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the invention.

[0023] Example 1: Please refer to Figure 1 This invention discloses a method for temperature control and noise reduction of an analog optical module, comprising the following steps: S1. Collect the raw temperature signal of the optical module, the time-series data of the ambient temperature, and the raw detection value of the optical wavelength, and perform preprocessing to obtain the processed real-time multi-source data; S1 includes the following steps: S11. Collect raw temperature signals, ambient temperature time-series data, and raw optical wavelength detection values ​​to obtain real-time multi-source data; S11 includes the following steps: S111. Using a high-precision thermistor (accuracy ±0.05℃) mounted on the back of the laser chip, the laser junction temperature data is collected through the ADC sampling circuit to obtain the raw temperature signal; S112. Use the ambient temperature sensor integrated inside the module to collect the ambient temperature of the module's working environment through the I²C digital interface to obtain ambient temperature time-series data. S113. Using an integrated optical power meter and wavelength detection unit, the real-time emission wavelength is collected through a built-in spectrometer and grating spectrometer to obtain the original detection value of the optical wavelength. S12. Preprocess the real-time multi-source data to obtain processed real-time multi-source data; S12 includes the following steps: S121. Using a sliding window filtering algorithm, all raw sensor data in real-time multi-source data are processed through convolution operations to obtain a filtered smooth signal; this is used to eliminate high-frequency noise. S122. Through a memory ring buffer, organize the sampled data of the filtered smooth signal to obtain a structured time-series dataset (optical module junction temperature data, ambient temperature data, optical wavelength data, and drive current data). S123. By standardizing with z-score, the sensor dimensions of the structured time-series dataset are transformed to obtain processed real-time multi-source data. S2. Improve the LSTM to obtain the initial LSTM junction temperature dynamic prediction model; use historical optical module feature data combined with optimization algorithm to train the LSTM junction temperature dynamic prediction model to obtain the final LSTM junction temperature dynamic prediction model. Based on the processed real-time multi-source data, the junction temperature prediction data is obtained through the final LSTM junction temperature dynamic prediction model. S2 includes the following steps: S21. Construct an initial LSTM model, improve the initial LSTM model, and obtain an initial LSTM junction temperature dynamic prediction model. S21 includes the following steps: S211. Based on the historical temperature single-dimensional time series data, add ambient temperature and driving current as joint input features to the input of the LSTM model, add real-time wavelength detection value as supervision signal, and delete redundant signals unrelated to temperature control (such as voltage fluctuation data) to form a three-dimensional feature tensor and establish an electro-thermal-optical coupling model. S212. To minimize the temperature prediction error (MSE loss) in the LSTM model, add a wavelength stability penalty term and modify the loss function to a composite form: Loss = MSE + α·wavelength offset (experiments determined α=0.3, and temperature control and optical performance were simultaneously optimized through backpropagation). The optical indexes are directly embedded into the machine learning optimization objective. S213. Modify the LSTM temperature prediction with a fixed step size (e.g., 1 second) to predict the critical control window: focusing on temperature changes over the next 5 seconds. S214. The complex gating structure of the LSTM model is removed, and a single-layer LSTM (64 units) is adopted. A knowledge distillation and compression process is added: a lightweight edge model is generated using the cloud teacher model, the calculation accuracy is modified to FP16, and the model size is reduced to 78KB to meet the real-time requirements of the embedded MCU (inference time <50ms). S215. Change the offline training fixed model parameters of the LSTM model to an online incremental learning module: automatically trigger fine-tuning when abnormal wavelength drift is detected, and modify the weight update strategy to selective update: only adjust the LSTM units that are strongly correlated with the current environment to achieve adaptability during the aging process of optical components (such as temperature-wavelength coefficient drift compensation). S22. Collect historical optical module temperature time-series data, ambient temperature data, and wavelength data to obtain historical optical module characteristic data; S23. Use historical optical module feature data to train the initial LSTM junction temperature dynamic prediction model. During the training process, combine the optimization algorithm to find the learning rate of the initial LSTM junction temperature dynamic prediction model and obtain the optimal solution. Use the optimal solution as the learning rate of the initial LSTM junction temperature dynamic prediction model to obtain the final LSTM junction temperature dynamic prediction model. The training process in S23, which combines optimization algorithms to find the learning rate of the initial LSTM junction temperature dynamic prediction model and obtain the optimal solution, includes the following steps: S231. Set the maximum number of iterations and construct a composite evaluation function, F(η) = 0.6·temperature MSE + 0.3·wavelength error + 0.1·convergence rate; S232. Perform a three-layer screening process in the discrete learning rate space {0.001, 0.003, 0.01, 0.03, 0.1}, including coarse screening, fine-tuning, and validation. Coarse screening: retain candidates with a loss reduction of >30% after 1 epoch; fine-tuning: calculate F(η) after 5 epochs of training; validation: perform a full 50 epochs of training on the optimal candidate. S233. With the fine-tuned optimal value as the center, initialize the ±50% search interval, and iteratively shrink the interval using the golden section method to obtain the optimal solution; calculate the 0.382 and 0.618 quantiles ηa and ηb, compare and evaluate the scores S(η)=1 / F(η), retain the high-score interval until the width is <0.0005, and output the optimal solution ηopt=(ηmin+ηmax) / 2; S24. Extract the features of the processed real-time multi-source data to obtain real-time optical module feature data; input the real-time optical module feature data into the final LSTM junction temperature dynamic prediction model to obtain junction temperature prediction data; S3. Pre-train the BP neural network model to obtain the pre-trained BP neural network model; calculate the real-time wavelength offset based on the real-time wavelength; and obtain the real-time temperature compensation amount based on the real-time wavelength offset and the pre-trained BP neural network model. Based on multi-dimensional optimization objectives, the optimal PID parameter mapping table is obtained by training with the PPO reinforcement learning algorithm. S3 includes the following steps: S31. Collect historical optical module full life cycle data, including temperature scan data (wavelength recorded in 0.1℃ steps in the range of -40℃ to 85℃), current disturbance data (wavelength response when the driving current changes by 10%-90%), and aging dataset (wavelength drift trajectory after 20,000 hours of continuous operation). An input vector is constructed based on the historical optical module lifecycle data, and the label is the temperature compensation amount ΔT, resulting in an input vector with a temperature compensation label. The BP neural network model employs a three-layer fully connected network. The input layer receives real-time wavelength deviation and aging factor, while the hidden layers use ReLU activation. A dual-path loss function, Loss = 0.7·|ΔT, is designed. pred -ΔT true | + 0.3·e (-SNR) , where ΔT represents pred The temperature difference predicted by the model, ΔT true The actual temperature deviation is represented by SNR, which is the measured value of the compensated signal-to-noise ratio, enabling direct optimization of optical performance. A pre-trained BP neural network model is obtained by combining a progressive training strategy with a three-stage verification mechanism. The progressive training strategy includes basic training (pre-training on room temperature data (25℃) until convergence), extended training (importing extreme temperature data and fine-tuning the output layer weights), aging adaptation (injecting device degradation data and updating hidden layer parameters), and using a dynamic learning rate throughout (initially 0.01, decaying by 15% every 10 rounds). The three-stage verification mechanism includes accuracy verification (test set compensation error ≤0.005℃), real-time verification (MCU inference time <2ms), and lifetime verification (wavelength stability remains ±0.015nm after simulating 5 years of aging). S32. Based on the real-time wavelength and preset wavelength in the processed real-time multi-source data, the real-time wavelength offset is obtained by subtraction. The wavelength offset and sensor aging data are input into a pre-trained BP neural network model to obtain the real-time temperature compensation. S33. Collect historical optical module temperature deviation, deviation change rate, ambient temperature and drive current data to obtain historical optical module parameter data; use historical optical module parameter data to construct state space and obtain optical module environmental feature vector; Temperature stability ΔT, wavelength shift λ shift and TEC (thermoelectric cooler) power consumption TEC To optimize the objective, a weighted reward function R = -|ΔT| - 0.5*|λ is constructed. shift | - 0.1*Power TEC The training direction is guided by a weighted reward function, which can suppress temperature drift and wavelength shift while minimizing energy consumption. Based on the optical module's environmental feature vector and weighted reward function, the optimal PID parameter mapping table is obtained through reinforcement training using the PPO algorithm. S1. Collect the raw temperature signal of the optical module, the time-series data of the ambient temperature, and the raw detection value of the optical wavelength, and perform preprocessing to obtain the processed real-time multi-source data; S2. Improve the LSTM to obtain the initial LSTM junction temperature dynamic prediction model; use historical optical module feature data combined with optimization algorithm to train the LSTM junction temperature dynamic prediction model to obtain the final LSTM junction temperature dynamic prediction model. Based on the processed real-time multi-source data, the junction temperature prediction data is obtained through the final LSTM junction temperature dynamic prediction model. S3. Pre-train the BP neural network model to obtain the pre-trained BP neural network model; calculate the real-time wavelength offset based on the real-time wavelength; and obtain the real-time temperature compensation amount based on the real-time wavelength offset and the pre-trained BP neural network model. Based on multi-dimensional optimization objectives, the optimal PID parameter mapping table is obtained by training with the PPO reinforcement learning algorithm. S4 includes the following steps: S41. Add the original target temperature to the temperature compensation amount to obtain the corrected target temperature; By subtraction, the temperature deviation between the target temperature and the predicted junction temperature is calculated and the core control variable is obtained. S42. Set the absolute value threshold for temperature deviation; calculate the absolute value of real-time temperature deviation based on the core control variables; obtain the working state of the optical module based on the absolute value of real-time temperature deviation and the absolute value threshold for temperature deviation (when the absolute value of real-time temperature deviation is ≥0.5, it is recorded as large deviation mode; when the absolute value of real-time temperature deviation is <0.5, it is recorded as steady state mode). The dynamic PID parameter mode is selected based on the optical module's operating state to obtain the final dynamic PID parameters. When the optical module is in a large deviation mode, a high-response parameter (Kp=8.0, Ki=0.1, Kd=2.0) is loaded through hard-coded parameter settings to obtain a high-gain PID parameter (used to quickly eliminate temperature abrupt changes), which is then used as the final dynamic PID parameter. When the optical module is in a steady-state mode, the lookup_Kp() function is used to query the PID mapping table generated by machine learning to obtain the scenario-based scaling factor (used to accurately maintain the target temperature). Based on the real-time wavelength, the ML_Adjust machine learning function is used to calculate the integral term correction factor to obtain the adaptive Ki value. The scenario-based scaling factor and the adaptive Ki value are then used as the final dynamic PID parameters. S43. Based on the final PID dynamic parameters, the control quantity is calculated through the PID algorithm to obtain the control output quantity; the control output quantity is converted into a precise current output by driving the TEC.

[0024] Example 2: Please refer to Figure 2A temperature control and noise reduction system for an analog optical module, used to implement the temperature control and noise reduction method for the aforementioned analog optical module, includes a multi-source data acquisition and preprocessing module, an improved LSTM junction temperature dynamic prediction module, a temperature compensation and intelligent optimization module, and a multi-modal intelligent decision-making and control output module. The multi-source data acquisition and preprocessing module is used to synchronously acquire raw laser junction temperature, ambient temperature and operating wavelength data from multiple sensors of the optical module; a series of preprocessing steps are performed on this set of real-time multi-source data, including using a sliding window filtering algorithm to eliminate high-frequency noise, organizing the data into a structured time-series dataset through a circular buffer, and finally performing z-score standardization to unify the units, so as to provide high-quality and standardized input data for subsequent models. The improved LSTM junction temperature dynamic prediction module is used to deeply improve the LSTM neural network model. The improvement measures include introducing ambient temperature and driving current as joint input features, adding a wavelength stability penalty term to the loss function, focusing on the prediction of key control windows, lightweighting and compressing the model to meet embedded deployment requirements, and integrating online incremental learning capabilities to adapt to device aging. It uses historical data and combines optimization algorithms to automatically find the optimal learning rate for training, and finally uses preprocessed real-time data to predict future junction temperature change trends with high accuracy. The temperature compensation and intelligent optimization module includes a pre-trained BP neural network and an optimal PID parameter mapping table based on the PPO reinforcement learning algorithm. The BP network predicts the required real-time temperature compensation based on the wavelength offset calculated in real time and aging data. The PPO algorithm uses temperature stability, wavelength offset and system power consumption as multi-objective optimization functions. Through reinforcement training with interaction with the environment, it learns and generates an optimal PID parameter lookup table that can adapt to different working states, providing intelligent parameters for the final control decision. The multimodal intelligent decision and control output module serves as the decision-making and execution center of the entire system. It calculates the deviation between the final target temperature and the real-time temperature for control through junction temperature prediction and temperature compensation. Based on the magnitude of the deviation, it determines whether the system is in a "large deviation mode" or a "steady-state mode," and intelligently switches control strategies accordingly. In the case of large deviations, it uses preset strong response parameters for rapid convergence; in the case of steady-state operation, it queries scenario-based parameters from the mapping table generated by the PPO and performs fine-tuning using machine learning functions. These dynamic PID parameters are used to calculate precise control quantities and drive the TEC to perform corresponding current operations, thereby achieving high-precision and high-stability temperature and wavelength control.

[0025] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0026] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for temperature control and noise reduction of an analog optical module, characterized in that, Includes the following steps: S1. Collect the raw temperature signal of the optical module, the time-series data of the ambient temperature, and the raw detection value of the optical wavelength, and perform preprocessing to obtain the processed real-time multi-source data; S2. Improve the LSTM to obtain the initial LSTM junction temperature dynamic prediction model; use historical optical module feature data combined with optimization algorithm to train the LSTM junction temperature dynamic prediction model to obtain the final LSTM junction temperature dynamic prediction model. Based on the processed real-time multi-source data, the junction temperature prediction data is obtained through the final LSTM junction temperature dynamic prediction model. S3. Pre-train the BP neural network model to obtain the pre-trained BP neural network model; Calculate the real-time wavelength offset based on the real-time wavelength. Based on the real-time wavelength offset, the real-time temperature compensation is obtained through a pre-trained BP neural network model. Based on multi-dimensional optimization objectives, the optimal PID parameter mapping table is obtained by training with the PPO reinforcement learning algorithm. S4. Based on the temperature compensation amount, junction temperature prediction data, and optimal PID parameter mapping table, the control output is generated through the PID algorithm. The control output is converted into a precise current output by driving the TEC.

2. The method for temperature control and noise reduction of an analog optical module according to claim 1, characterized in that, S1 includes the following steps: S11. Collect raw temperature signals, ambient temperature time-series data, and raw optical wavelength detection values ​​to obtain real-time multi-source data; S12. Preprocess the real-time multi-source data to obtain processed real-time multi-source data.

3. The method for temperature control and noise reduction of an analog optical module according to claim 2, characterized in that, S11 includes the following steps: S111. Collect laser junction temperature data through sampling circuit to obtain raw temperature signal; S112. Use the ambient temperature sensor integrated inside the module to collect the ambient temperature of the module's working environment through the digital interface to obtain ambient temperature time-series data. S113. Using an integrated optical power meter and wavelength detection unit, the real-time emission wavelength is collected through a built-in spectrometer and grating spectrometer to obtain the original detection value of the optical wavelength.

4. The temperature control and noise reduction method for an analog optical module according to claim 2, characterized in that, S12 includes the following steps: S121. Using a sliding window filtering algorithm, all raw sensor data in real-time multi-source data are processed by convolution operation to obtain a filtered smooth signal. S122. The filtered smooth signal is organized and sampled using a memory circular buffer to obtain a structured time series dataset. S123. By standardizing with z-score, the sensor dimensions of the structured time-series dataset are transformed to obtain processed real-time multi-source data.

5. The method for temperature control and noise reduction of an analog optical module according to claim 1, characterized in that, S2 includes the following steps: S21. Construct an initial LSTM model, improve the initial LSTM model, and obtain an initial LSTM junction temperature dynamic prediction model. S22. Collect historical optical module temperature time-series data, ambient temperature data, and wavelength data to obtain historical optical module characteristic data; S23. Use historical optical module feature data to train the initial LSTM junction temperature dynamic prediction model. During the training process, combine the optimization algorithm to find the learning rate of the initial LSTM junction temperature dynamic prediction model and obtain the optimal solution. Use the optimal solution as the learning rate of the initial LSTM junction temperature dynamic prediction model to obtain the final LSTM junction temperature dynamic prediction model. S24. Extract the features of the processed real-time multi-source data to obtain real-time optical module feature data; input the real-time optical module feature data into the final LSTM junction temperature dynamic prediction model to obtain junction temperature prediction data.

6. The temperature control and noise reduction method for an analog optical module according to claim 5, characterized in that, S21 includes the following steps: S211. Based on the historical temperature single-dimensional time series data, add ambient temperature and driving current as joint input features to the input of the LSTM model, add real-time wavelength detection value as supervision signal, and delete redundant signals unrelated to temperature control to form a three-dimensional feature tensor and establish an electro-thermal-optical coupling model. S212. To minimize the temperature prediction error of the LSTM model, add a wavelength stability penalty term and modify the loss function to a composite form, directly embedding optical indicators into the machine learning optimization objective. S213. Modify the fixed step size temperature prediction of LSTM to predict the critical control window as the target. S214. Remove the complex gating structure of the LSTM model, adopt a single-layer LSTM, and add a knowledge distillation compression process to modify the calculation accuracy and model size to meet the real-time requirements of the embedded MCU. S215. Change the offline training of the LSTM model with fixed model parameters to an online incremental learning module, and modify the weight update strategy to selective update.

7. The method for temperature control and noise reduction of an analog optical module according to claim 5, characterized in that, The training process in S23, which combines optimization algorithms to find the learning rate of the initial LSTM junction temperature dynamic prediction model and obtain the optimal solution, includes the following steps: S231. Set the maximum number of iterations and construct a composite evaluation function; S232. Perform a three-layer screening process in the discrete learning rate space, which includes coarse screening, fine-tuning, and validation. S233. Initialize the search interval with the fine-tuned optimal value as the center, and obtain the optimal solution by iteratively shrinking the interval using the golden section method.

8. The method for temperature control and noise reduction of an analog optical module according to claim 1, characterized in that, S3 includes the following steps: S31. Collect historical optical module full life cycle data, including temperature scan data, current disturbance data, and aging dataset; An input vector is constructed based on the historical optical module lifecycle data, and the label is the temperature compensation amount, resulting in an input vector with a temperature compensation label. The BP neural network model uses a three-layer fully connected network. The input layer receives real-time wavelength deviation and aging factor, the hidden layer uses ReLU activation, and a dual-path loss function is designed. A pre-trained BP neural network model is obtained by combining a progressive training strategy with a three-stage verification mechanism; the progressive training strategy includes basic training. S32. Based on the real-time wavelength and preset wavelength in the processed real-time multi-source data, the real-time wavelength offset is obtained by subtraction. The wavelength offset and sensor aging data are input into a pre-trained BP neural network model to obtain the real-time temperature compensation. S33. Collect historical optical module temperature deviation, deviation change rate, ambient temperature and drive current data to obtain historical optical module parameter data; use historical optical module parameter data to construct state space and obtain optical module environmental feature vector; With temperature stability, wavelength shift, and TEC power consumption as optimization objectives, a weighted reward function is constructed. Based on the optical module environmental feature vector and the weighted reward function, the PPO algorithm is used for reinforcement training to obtain the optimal PID parameter mapping table.

9. The method for temperature control and noise reduction of an analog optical module according to claim 1, characterized in that, S4 includes the following steps: S41. Add the original target temperature to the temperature compensation amount to obtain the corrected target temperature; By subtraction, the temperature deviation between the target temperature and the predicted junction temperature is calculated and the core control variable is obtained. S42. Set the absolute value threshold for temperature deviation; calculate the real-time absolute value of temperature deviation based on the core control variables; obtain the working status of the optical module based on the real-time absolute value of temperature deviation and the absolute value threshold for temperature deviation. The dynamic PID parameter mode is selected based on the optical module's operating state to obtain the final dynamic PID parameters. When the optical module is in a large deviation mode, a high-gain PID parameter is obtained by hard-coding parameter settings and loading a strong response parameter. This high-gain PID parameter is then used as the final dynamic PID parameter. When the optical module is in a steady-state mode, a scenario-based scaling factor is obtained by querying a PID mapping table generated by machine learning. Based on the real-time wavelength, an integral term correction factor is calculated to obtain an adaptive Ki value. The scenario-based scaling factor and the adaptive Ki value are then used as the final dynamic PID parameters. S43. Based on the final PID dynamic parameters, the control quantity is calculated through the PID algorithm to obtain the control output quantity; the control output quantity is converted into a precise current output by driving the TEC.

10. A temperature control and noise reduction system for an analog optical module, characterized in that, The system implements a temperature control and noise reduction method for an analog optical module as described in any one of claims 1-9, comprising a multi-source data acquisition and preprocessing module, an improved LSTM junction temperature dynamic prediction module, a temperature compensation and intelligent optimization module, and a multi-modal intelligent decision-making and control output module.

Images