Light beam focal length controllable system of BOSA optical device and implementation method of light beam focal length controllable system

Abstract

The invention discloses a light beam focal length controllable system of a BOSA optical device and an implementation method thereof, relates to the technical field of optical communication, and solves the problems of focal length drift, light beam quality reduction and poor eye pattern quality caused by ambient temperature fluctuation in the prior art. Comprising a first focal length quality analysis module, a second focal length quality analysis module, an intelligent focal length adjustment module and a signal optimization module, the first focal length quality analysis module predicts future environment temperature change through a long-short term memory network, and plans an optimal focal length path in advance; the second focal length quality analysis module identifies light beam characteristic parameters by using a depth image identification algorithm, and performs multi-objective optimization in combination with an environment temperature change trend; the intelligent focal length adjustment module dynamically adjusts the focal length according to the optimization suggestion; the signal optimization module carries out waveform shaping and noise suppression by using a deep learning network, so that the eye pattern quality is remarkably improved; according to the invention, the focal length stability of the system is greatly improved, and the performance and reliability of the optical communication system are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of optical communication technology, and more specifically to a beam focal length controllable system of a BOSA optical device and an implementation method thereof. Background Art

[0002] As global communication networks develop towards higher bandwidth, faster speed and larger capacity, especially the popularization of 5G technology and the continuous expansion of data centers, unprecedented requirements are placed on the performance of optical communication components. As a key component for achieving single-fiber bidirectional transmission, bidirectional photonic subassemblies (BOSA) play an increasingly important role in this context. BOSA not only needs to have high-efficiency data transmission capabilities, but also needs to maintain good stability and reliability while being miniaturized and integrated. In particular, the requirements for focal length control are increasing to ensure high-quality signal transmission even in complex and changing working environments. Therefore, how to effectively manage the focal length of the light beam in BOSA has become a technical problem that needs to be solved urgently.

[0003] In order to meet the above challenges, the industry has developed a variety of technologies to achieve precise control of the focal length of the light beam in BOSA. One of the widely used methods is based on the thermal regulation mechanism, that is, adjusting the position or shape of the lens by changing the material temperature to achieve the purpose of controlling the focal length. In the prior art, Chinese patent CN 118011572A proposes a parallel light BOSA device with a three-port circulator and its assembly method. This design achieves efficient isolation and beam collimation between the transmitting end and the receiving end by introducing an optical prism combination placed at a specific angle and elements such as a Faraday rotator; secondly, the document "Research on Beam Control Technology Based on Phase-Type Spatial Light Modulator" proposes the use of a non-mechanical servo-controlled liquid crystal spatial light modulator (LCSLM) to modulate the wavefront phase in real time, thereby achieving beam deflection control; these innovations provide new ideas for improving the focusing characteristics of the beam in BOSA;

[0004] However, in actual use, since thermal regulation depends on the physical property changes of the material with temperature, this mechanism makes the entire system extremely sensitive to the ambient temperature. Once the ambient temperature fluctuates, the originally set focal length may drift, affecting the quality of data transmission. Secondly, thermal effects may also cause the quality of the beam to deteriorate. Especially in application scenarios where frequent focal length adjustment is required, long-term heating may cause problems such as uneven light emission or deformation of the light spot, thereby reducing the clarity of the signal. Finally, poor eye diagram quality is also a prominent problem. Due to inaccurate beam quality control, the eye diagram margin is low in high-speed signal transmission, the signal level line is thick and has double eyelids, the rising and falling edges are exceeded, and the random jitter is also large. These problems directly affect the performance and reliability of the optical communication system. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention discloses a beam focal length controllable system for a BOSA optical device and an implementation method thereof, aiming to solve the problems in the background technology.

[0006] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0007] A beam focal length controllable system for a BOSA optical device, comprising: a first focal length quality analysis module, a second focal length quality analysis module, an intelligent focal length adjustment module and a signal optimization module;

[0008] The first focal length quality analysis module is used to collect ambient temperature data in real time, build a thermal prediction model through a long short-term memory network, predict the future ambient temperature change trend and output it to the second focal length quality analysis module;

[0009] The second focal length quality analysis module is used to collect the beam brightness distribution and spot shape data in real time through the photoelectric detector, and identify and classify the focus position accuracy, beam divergence and light intensity distribution uniformity parameters of the collected data through the deep image recognition algorithm; based on the recognition result, the second focal length quality analysis module is also used to construct a multi-objective optimization model by taking into account the beam quality, the ambient temperature change trend output by the first focal length quality analysis module and the focal length stability through the beam quality optimization algorithm, and output the optimal focal length adjustment plan;

[0010] The intelligent focus adjustment module is used to perform a focus adjustment operation according to the output result of the second focus quality analysis module;

[0011] The signal optimization module is used to apply a deep learning network to perform waveform shaping and noise suppression. When poor eye diagram quality or large jitter is detected, correction is performed through pre-distortion processing and adaptive equalization algorithm to optimize the signal morphology of the transmitting end and the receiving end.

[0012] As a further technical solution of the present invention, the thermal prediction model includes a data input layer, a feature layer, a training layer, a trend prediction layer and an output layer; the data input layer is used to collect temperature data of the environment where the BOSA optical device is located in real time, and the collected data is used to identify periodic components and non-periodic trends through a time series analysis method, and input into the feature layer in the form of a time series; the feature layer is used to extract feature vectors from the output temperature data of the data input layer through principal component analysis and autoencoder, and transmit the extracted feature data to the training layer for modeling; the training layer is used to control the information flow through a gating mechanism based on the extracted data of the feature layer, learn the long-term dependencies in the time series, and train the thermal prediction model; the trend prediction layer is used to predict the future trend of ambient temperature changes based on the trained model; during the prediction process, a sliding window is used to improve the continuity and accuracy of the prediction results; the output layer is used to output the predicted ambient temperature change trend to the second focal length quality analysis module.

[0013] As a further technical solution of the present invention, the working method of the depth image recognition algorithm is:

[0014] S301. Based on the beam image data captured by the photodetector and the image sensor, the deep image recognition algorithm removes noise and maintains edge information through adaptive median filtering and bilateral filtering, and adjusts all images to the same scale through a normalization method; wherein the adaptive median filtering is based on each pixel point P, calculates the median M(P) in the neighborhood of the pixel point P, and determines the pixel value according to a local statistical method; the bilateral filtering performs smoothing by combining spatial distance and pixel intensity difference, and the formula for bilateral filtering smoothing is:

[0015]

[0016] In formula (1), x and y represent the current pixel position and the neighboring pixel position respectively; l(x) and l(y) are the pixel intensities at positions x and y respectively; G as (x, y) represents the spatial Gaussian kernel, which is used to measure the spatial distance weight between two pixels; as represents the standard deviation of the spatial distance, which is used to control the influence of the spatial distance; G ar (l(x),l(y)) represents the range Gaussian kernel, which is used to weigh the intensity difference between two pixels; ar is the intensity standard deviation, which is used to control the influence of intensity difference;

[0017] S302, the deep image recognition algorithm strengthens the image features through histogram equalization and adaptive gamma correction, and identifies the edge and texture image features in the enhanced beam image through a multi-layer convolutional neural network; wherein the histogram equalization increases the image contrast through grayscale value frequency distribution, and the adaptive gamma correction enhances the dark details through image brightness curve adjustment; during the learning process of the multi-layer convolutional neural network, the deep image recognition algorithm converges and classifies different types of beam images using the pre-trained ResNet architecture through a transfer learning mechanism, and dynamically allocates weights to different image areas through an attention mechanism; wherein the formula expression for the recognition of the multi-layer convolutional neural network is:

[0018]

[0019] In formula (2), a i represents the attention weight of the i-th region; r i represents the unnormalized attention score of the i-th region; represents the weight matrix used to calculate the attention score; exp represents the exponential function, which is used for Softmax normalization; b a represents the bias term; W x and W y They are used to process the input feature vector x i and context information y i The weight matrix of

[0020] S303. Based on the identified data, the normal and abnormal beam characteristics are distinguished by an abnormality detection algorithm based on a variational autoencoder. The formula is:

[0021]

[0022] In formula (3), D KL represents KL divergence; q(z|x) represents encoder distribution; p(z) represents prior distribution; p(x|z) represents the data distribution generated by the decoder; N represents the number of image pixels; c i Represents the i-th pixel value of the original image; e i Represents the i-th pixel value of the reconstructed image; Indicates the reconstruction error. If the reconstruction error is detected to exceed the preset threshold, the alarm mechanism is immediately triggered and the closed-loop control system is notified to make adjustments. Otherwise, monitoring continues;

[0023] S304, accumulate new sample data through the online learning mechanism. If the quality assessment score S(θ) of the new sample data is higher than the preset threshold, update the model parameters; otherwise, maintain the current parameters unchanged. The calculation formula of the quality assessment score is:

[0024]

[0025] In formula (4), S(θ) represents the quality assessment score of the sample data; f(θ; x i ) represents the model for the i-th sample G i The predicted value of H i represents the true label of the i-th sample; σ i represents the standard deviation of the i-th sample; θ represents the model parameter; n represents the number of samples.

[0026] As a further technical solution of the present invention, the working method of the beam quality optimization algorithm is as follows: first, the beam characteristic parameters of the deep image recognition algorithm, including beam intensity, spot size and edge clarity, are received, and the future ambient temperature change prediction value output by the first focal length quality analysis module is combined, and the parameters are mapped to the beam quality and focal length stability indicators through a multi-objective optimization framework. If the current ambient temperature is stable and the beam quality and focal length stability indicators are higher than or equal to the preset threshold, the beam quality and focal length stability indicators are not triggered to continue to be evaluated; otherwise, the dynamic adjustment mechanism is triggered; the dynamic adjustment mechanism uses the Pareto optimal solution, and when the beam quality score Q(ω) under the current focal length setting is higher than the preset threshold and the focal length drift risk R(ω) caused by the ambient temperature change is lower than the preset threshold, the setting parameters are recorded and regarded as a potential optimal solution; otherwise, other configurations are continued to be searched; wherein the beam quality score Q(ω) and the focal length drift risk R(ω) are quantified by the comprehensive evaluation function F(ω); the formula expression of the comprehensive evaluation function F(ω) is:

[0027] F(ω)=w 1 Q(ω)+w 2 (1-R(ω)) 2 (5)

[0028] In formula (5), ω represents the focal length adjustment parameter; w 1 and w 2 are weight coefficients, which represent the priorities of beam quality and focal length stability. When the predicted value of future ambient temperature change output by the first focal length quality analysis module is higher than the preset threshold, w is increased through a dynamic weighting mechanism. 2 ; When the beam quality shows a downward trend, increase w 1 ; After dynamic adjustment, the beam quality optimization algorithm also quickly converges to the optimal solution in the parameter space through genetic algorithm and particle swarm optimization.

[0029] As a further technical solution of the present invention, the multi-objective optimization framework includes a weight allocation module, a multi-objective collaborative optimization module and a dynamic adjustment module; the weight allocation module is used to analyze the changing trend of the beam characteristic parameters through a deep learning algorithm, and calculate the relative weights of the beam quality and focal stability in the optimization process in combination with the predicted value of the ambient temperature; the multi-objective collaborative optimization module is used to receive the weight calculation results output by the weight allocation module, as well as the objective functions of the beam quality and focal stability, and iteratively search the objective functions of the beam quality and focal stability under the guidance of the co-evolutionary algorithm to find the optimal solution, and after each iteration, adjust the search direction and step size based on the improvement of the objective function; the dynamic adjustment module is used to receive the optimal solution output by the multi-objective collaborative optimization module, and use a closed-loop control system to monitor the changes in the beam quality and focal stability indicators in real time. If the real-time monitoring value deviates from the preset range, the adjustment amount is calculated and the focal control parameters of the optical device are adjusted.

[0030] As a further technical solution of the present invention, the working principle of the intelligent focal length adjustment module is as follows: first, the optimization suggestion of the second focal length quality analysis module is received, including the optimal focal length setting parameter ω and the predicted value of the ambient temperature change, and the current actual state is collected once every t seconds through the real-time monitoring module, including the current beam focal length quality state and the ambient temperature data, and compared with the optimization suggestion; the current beam focal length quality state includes the focus position accuracy, beam divergence and light intensity distribution uniformity parameters. If the deviation between the current actual state and the optimization suggestion is higher than a preset threshold, the focal length adjustment mechanism is triggered. When the ambient temperature change value in a preset time window is less than or equal to a preset threshold, the focal length adjustment mechanism adopts a PID control algorithm to control the internal lens position or curvature of the optical device through a piezoelectric ceramic driver; when the ambient temperature change predicted value in a preset time window is higher than a preset threshold, the focal length adjustment mechanism plans the focal length path in advance through a model predictive control method to adjust the focal length.

[0031] As a further technical solution of the present invention, the working method of the signal optimization module is: receiving the original signal data of the photoelectric detector, and collecting the current eye diagram quality and jitter situation once every t seconds through the real-time monitoring module; if the eye diagram closure is detected to be lower than the set threshold E tereshold Or the jitter exceeds the allowable range J max , the deep learning network is triggered to perform waveform shaping and noise suppression processing; otherwise, the existing signal state is maintained; during the deep learning network processing, a real-time monitoring module is used to continuously monitor changes in signal quality, and the parameters and strategies of the deep learning network are continuously fine-tuned through a closed-loop feedback mechanism to achieve dynamic optimization.

[0032] As a further technical solution of the present invention, the waveform shaping method of the deep learning network is: the collected signal waveform is extracted by a convolutional neural network, and the waveform contour, jitter mode and noise distribution in the signal are gradually abstracted by using multiple convolutional layers and pooling layers. The extracted features are input to the fully connected layer, and nonlinear transformation is performed by a deep learning algorithm to eliminate waveform distortion caused by device nonlinearity, dispersion effect or environmental factors. During the waveform shaping process, the deep learning network also dynamically adjusts the equalizer parameters according to the current signal quality through an adaptive equalization algorithm; the method for noise suppression processing of the deep learning network is: based on the original signal, amplitude adjustment parameters, time-frequency analysis parameters, frequency domain signal inverse transformation and filtering parameters and deep learning network parameters, the inverse distortion adjustment amount of the original signal is calculated by an inverse distortion function, which is used to correct the original signal; the formula expression of the inverse distortion function is:

[0033]

[0034] In formula (6), z(k) represents the signal after inverse distortion processing; g(k) represents the original signal captured by the receiving end; Represents the set of training parameters of the deep learning network, including weights, biases, and any regularization parameters, used to describe the inverse distortion function I -1 The nonlinear mapping relationship of A s,β g(k) represents the amplitude adjustment of the signal g(k). Parameters s and β control the amplitude and mode of adjustment, which are used to compensate for the amplitude distortion caused by the nonlinear factors of the device during the transmission process. γ,δ (STEFg(k)) represents the weighted processing of the short-time Fourier transform result of the signal g(k), and the parameters γ and δ control the weight and method of weighting, which is used to compensate the signal distortion in the frequency domain; B Q,T (IFFT(W γ,δ )) represents the amount of filtering processing of the frequency domain signal after the weighted inverse Fourier transform; the parameters Q and T control the parameters and methods of the filter to further suppress noise and interference.

[0035] As a further technical solution of the present invention, the working principle of the adaptive equalization algorithm is as follows: first, the input signal is analyzed using the hierarchical feature extraction mechanism of the convolutional neural network to extract signal features, including the intensity distribution of frequency components and slight shifts in phase relationships, and the extracted features are integrated into a signal quality vector through a fully connected layer; based on the current signal quality vector, a nonlinear transformation is performed through a multilayer perceptron to output an action vector containing an equalizer parameter adjustment strategy, and the filter parameters of the equalizer are adjusted according to the action vector to compensate for frequency distortion and phase deviation in the signal; during the equalizer parameter adjustment process, the output signal quality is continuously monitored, and if the signal quality improvement is lower than a preset threshold, the gradient is calculated based on the error signal using the minimum mean square error method, the equalizer parameters are updated, and the amplitude and direction of the parameter update are limited by the regularization term; after the equalizer parameter adjustment process, the monitored signal quality information is fed back to the strategy network through a feedback mechanism to optimize the network parameters and strategy prediction accuracy.

[0036] As a further technical solution of the present invention, a method for realizing controllable focal length of a BOSA optical device beam comprises the following steps:

[0037] Step 1: Use a temperature sensor to collect temperature data of the BOSA optical device and the environment in which the BOSA optical device is located in real time, remove noise, detect and eliminate outliers, and perform data smoothing on the collected temperature data, and output the preprocessed temperature data set;

[0038] Step 2: Based on the temperature data set output in step 1, a thermal prediction model is constructed through a long short-term memory network to predict the future trend of ambient temperature changes;

[0039] Step 3: Use a photodetector to monitor the beam quality emitted by the BOSA optical device in real time, including brightness distribution and spot shape, and pre-process the collected beam quality data. Use a deep image recognition algorithm to identify and classify the focus position accuracy, beam divergence, and light intensity distribution uniformity parameters of the collected data, and output the beam quality data set;

[0040] Step 4: Based on the beam quality data set outputted in step 3, a multi-objective optimization model is constructed by using a beam quality optimization algorithm taking into account the beam quality, the ambient temperature variation trend outputted in step 2, and the focal length stability, and an optimal focal length adjustment scheme is outputted;

[0041] Step 5: According to the focal length pre-adjustment strategy and optimization strategy outputted in step 4, the focal length is initially adjusted by a stepper motor. During the focal length adjustment process, the adjustment error is monitored in real time by a closed-loop feedback control mechanism, and the PID algorithm is used for real-time compensation.

[0042] Step 6: After the initial adjustment of the focal length, the expansion and contraction of the piezoelectric ceramic is controlled by the piezoelectric ceramic fine-tuning mechanism to perform a secondary adjustment of the focal length;

[0043] Step 7. According to the eye diagram quality and jitter detection results, the signal emitted by the BOSA optical device is subjected to waveform shaping, noise suppression and adaptive equalization through a deep learning network. When poor eye diagram quality or large jitter is detected, correction is performed through pre-distortion processing and adaptive equalization algorithm to optimize the signal shape at the transmitter and receiver.

[0044] Based on the above technical solutions, the positive and beneficial effects of the present invention are:

[0045] 1. The first focal length quality analysis module uses the long short-term memory network (LSTM) to build a thermal prediction model, collect and predict future ambient temperature change trends in real time, and plan the optimal focal length path in advance. The intelligent focal length adjustment module uses the PID control algorithm or model predictive control method to dynamically adjust the focal length according to the optimal focal length adjustment scheme output by the second focal length quality analysis module to ensure focal length stability and response speed. The multi-objective optimization framework combines the beam quality and ambient temperature change trends for comprehensive evaluation, and quickly converges to the optimal solution through the Pareto optimal solution and genetic algorithm, further improving the focal length adjustment accuracy. These measures work together to enable the system to maintain high focal length stability when the external ambient temperature fluctuates, significantly improving the quality and reliability of data transmission.

[0046] 2. The deep image recognition algorithm removes noise and enhances image features through methods such as adaptive median filtering, bilateral filtering, histogram equalization and adaptive gamma correction, and accurately identifies the focus position, divergence and uniformity of light intensity distribution of the beam. The beam quality optimization algorithm combines the trend of ambient temperature changes and focal length stability, and dynamically adjusts the beam parameters through a multi-objective optimization framework to ensure that the beam quality is always in the best state. The variational autoencoder anomaly detection algorithm distinguishes between normal and abnormal beam characteristics, and promptly triggers the closed-loop control system for adjustment to prevent the degradation of beam quality. These measures enable the system to maintain the uniformity of the light source and the stability of the spot shape during high-frequency focal length adjustment, significantly improving the clarity and transmission quality of the signal.

[0047] 3. The signal optimization module applies deep learning networks to perform waveform shaping and noise suppression, extracts signal features through convolutional neural networks, and eliminates waveform distortion caused by device nonlinearity, dispersion effects, or environmental factors. The adaptive equalization algorithm dynamically adjusts the equalizer parameters according to the current signal quality, compensates for the frequency distortion and phase deviation in the signal, and improves the eye diagram quality. The pre-distortion processing and adaptive equalization algorithm work together to ensure the optimization of the signal morphology at the transmitter and receiver, significantly improve the eye diagram quality, increase the eye diagram margin, reduce the double eyelid phenomenon and edge over-standard problems, and reduce random jitter. These measures enable the system to exhibit excellent eye diagram quality in high-speed signal transmission, improve the clarity and stability of the signal, and enhance the performance and reliability of the optical communication system. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0049] Figure 1 This is a schematic diagram of a beam focal length controllable system of a BOSA optical device according to the present invention;

[0050] Figure 2 This is a working principle architecture diagram of the intelligent focal length adjustment module of the present invention;

[0051] Figure 3 A diagram showing the working method of the beam quality optimization algorithm of the present invention;

[0052] Figure 4 A step diagram of a method for controlling the focal length of a beam of a BOSA optical device according to the present invention;

[0053] Figure 5 It is a principle framework diagram of the thermal prediction model of the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] like Figure 1As shown, a beam focal length controllable system of a BOSA optical device includes: a first focal length quality analysis module, a second focal length quality analysis module, an intelligent focal length adjustment module and a signal optimization module;

[0056] The first focal length quality analysis module is used to collect ambient temperature data in real time, build a thermal prediction model through a long short-term memory network, predict future ambient temperature change trends and output them to the second focal length quality analysis module; wherein the thermal prediction model includes a data input layer, a feature layer, a training layer, a trend prediction layer and an output layer; the data input layer is used to collect temperature data of the environment where the BOSA optical device is located in real time, and the collected data is used to identify periodic components and non-periodic trends through a time series analysis method, and input into the feature layer in the form of a time series; the feature layer is used to extract feature vectors from the output temperature data of the data input layer through principal component analysis and an autoencoder, and transmit the extracted feature data to the training layer for modeling; the training layer is used to control the information flow through a gating mechanism based on the extracted data of the feature layer, learn the long-term dependencies in the time series, and train the thermal prediction model; the trend prediction layer is used to predict future ambient temperature change trends based on the trained model; during the prediction process, a sliding window is used to improve the continuity and accuracy of the prediction results; the output layer is used to output the predicted ambient temperature change trend to the second focal length quality analysis module.

[0057] In a specific embodiment, Figure 5As shown in the figure, the data input layer extracts periodic components through Fourier transform and eliminates short-term fluctuations through moving average filter, thereby separating long-term trends. This method not only improves the signal-to-noise ratio of the data, but also makes subsequent feature extraction more effective. The feature layer combines two technologies: principal component analysis (PCA) and autoencoder. PCA is used to reduce dimensionality, reduce redundant information and retain the main variation direction; autoencoder automatically learns the potential representation of data through unsupervised learning. Through PCA, the feature layer can extract the most representative principal components from high-dimensional data, simplifying the subsequent modeling process. Through autoencoder, the feature layer can capture nonlinear relationships and hidden patterns in the data, providing richer feature representation. This dual mechanism greatly enhances the accuracy and robustness of feature extraction. The training layer uses a long short-term memory network (LSTM), a recurrent neural network (RNN) specifically designed for processing sequence data. LSTM controls the flow of information through a gating mechanism, allowing the model to remember long-term dependencies while ignoring irrelevant information. Specifically, through the forget gate, input gate, and output gate, LSTM can selectively update or maintain the cell state at each time step, thereby effectively learning complex dynamic behaviors. This method not only improves the model's ability to learn long sequence data, but also enhances its flexibility in responding to environmental changes. The trend prediction layer introduces the sliding window technology. Specifically, by setting a fixed-size time window, the model can use the latest historical data as input at each prediction, gradually move forward, and form continuous prediction results. This method not only ensures a smooth transition of the prediction results, but also improves the accuracy of short-term predictions. In addition, by integrating multiple LSTM models, the trend prediction layer can integrate data features from different time periods to further improve the prediction performance. The output layer adjusts the prediction results to a unified scale through standardization and normalization, which is convenient for other modules to use directly. In addition, by introducing a feedback mechanism, the output layer can dynamically adjust the parameters of the prediction model according to the actual temperature changes and continuously optimize the prediction effect. This method not only improves the system's adaptability, but also enhances the stability and reliability of the prediction.

[0058] In actual applications, the thermal prediction model collects temperature data at fixed intervals (such as 1 second) and performs real-time predictions through the above process. When it is detected that the ambient temperature is about to change significantly, the system will plan the optimal focal path in advance to avoid focal drift caused by temperature fluctuations. At the same time, the dynamic thermal management module adjusts the operating temperature of the BOSA optical device according to the prediction results to ensure that it operates stably within the optimal temperature range. This closed-loop control system not only improves the accuracy and response speed of focal control, but also enhances the overall performance and reliability of the system.

[0059] Compared with the prior art, the thermal prediction model of the present invention combines multiple advanced technologies such as time series analysis, PCA, autoencoder and LSTM, and the present invention can more comprehensively capture and process the complex features in the ambient temperature data, thereby improving the accuracy and stability of the prediction. Secondly, the application of sliding window technology makes the prediction results more continuous and smooth, reduces the impact of short-term fluctuations, and improves the adaptability and flexibility of the system. In addition, the introduction of the feedback mechanism enables the model to dynamically adjust parameters according to the actual temperature changes, realizes adaptive optimization, and further enhances the prediction effect. More importantly, the thermal prediction model of the present invention is not only an independent prediction tool, but is closely integrated into the entire BOSA optical device beam focal length controllable system. By working in conjunction with the second focal length quality analysis module, the thermal prediction model not only improves the focal length stability and beam quality of the system, but also, when the beam quality is guaranteed, the signal received by the receiving end is clearer and more stable, the rising and falling edges in the eye diagram are sharper, and the random jitter is also reduced, thereby significantly improving the eye diagram quality. This all-round optimization measure enables the present invention to show excellent performance and reliability in practical applications, and is suitable for various complex application scenarios.

[0060] The second focal length quality analysis module is used to collect the beam brightness distribution and spot shape data in real time through the photoelectric detector, and identify and classify the focus position accuracy, beam divergence and light intensity distribution uniformity parameters of the collected data through the deep image recognition algorithm; based on the recognition result, the second focal length quality analysis module is also used to consider the beam quality, the ambient temperature change trend output by the first focal length quality analysis module and the focal length stability through the beam quality optimization algorithm to construct a multi-objective optimization model and output the optimal focal length adjustment plan; wherein, the working method of the deep image recognition algorithm is: S301, based on the beam image data captured by the photoelectric detector and the image sensor, the deep image recognition algorithm removes noise and maintains edge information through adaptive median filtering and bilateral filtering, and adjusts all images to the same scale through the normalization method; wherein, the adaptive median filtering is based on each pixel point P, calculates the median M(P) in the neighborhood of the pixel point P, and determines the pixel value according to the local statistical method; the bilateral filtering performs smoothing by combining the spatial distance and the pixel intensity difference, and the formula expression of the bilateral filtering smoothing is:

[0061]

[0062] In formula (1), x and y represent the current pixel position and the neighboring pixel position respectively; l(x) and l(y) are the pixel intensities at positions x and y respectively; G as (x, y) represents the spatial Gaussian kernel, which is used to measure the spatial distance weight between two pixels; as represents the standard deviation of the spatial distance, which is used to control the influence of the spatial distance; G ar (l(x), l(y)) represents the range Gaussian kernel, which is used to measure the weight of the intensity difference between two pixels; ar is the intensity standard deviation, which is used to control the influence of the intensity difference; S302, the deep image recognition algorithm strengthens the image features through histogram equalization and adaptive gamma correction, and identifies the edge and texture image features in the enhanced beam image through a multi-layer convolutional neural network; wherein, the histogram equalization increases the image contrast through gray value frequency allocation, and the adaptive gamma correction enhances the dark details through image brightness curve adjustment; during the learning process of the multi-layer convolutional neural network, the deep image recognition algorithm uses the pre-trained ResNet architecture through the transfer learning mechanism to converge and classify different types of beam images, and dynamically allocates weights to different image areas through the attention mechanism; wherein, the formula expression of the recognition of the multi-layer convolutional neural network is:

[0063]

[0064] In formula (2), a i represents the attention weight of the i-th region; r i represents the unnormalized attention score of the i-th region; represents the weight matrix used to calculate the attention score; exp represents the exponential function, which is used for Softmax normalization; b a represents the bias term; W x and W y They are used to process the input feature vector x i and context information y i S303, based on the identified data, distinguish the normal and abnormal beam characteristics through the abnormality detection algorithm based on the variational autoencoder, the formula expression is:

[0065]

[0066] In formula (3), D KL represents KL divergence; q(z|x) represents encoder distribution; p(z) represents prior distribution; p(x|z) represents the data distribution generated by the decoder; N represents the number of image pixels; c i Represents the i-th pixel value of the original image; e i Represents the i-th pixel value of the reconstructed image; Represents the reconstruction error. If the reconstruction error is detected to exceed the preset threshold, the alarm mechanism is immediately triggered and the closed-loop control system is notified to make adjustments. Otherwise, monitoring continues. S304: accumulate new sample data through the online learning mechanism. If the quality assessment score S(θ) of the new sample data is higher than the preset threshold, the model parameters are updated. Otherwise, the current parameters are maintained unchanged. The calculation formula of the quality assessment score is:

[0067]

[0068] In formula (4), S(θ) represents the quality assessment score of the sample data; f(θ; x i ) represents the model for the i-th sample G i The predicted value of H i represents the true label of the i-th sample; σ i represents the standard deviation of the i-th sample; θ represents the model parameter; n represents the number of samples.

[0069] Among them, adaptive median filtering dynamically adjusts the size of the filter window and calculates the median of each pixel in its neighborhood, thereby effectively removing random noise such as salt and pepper noise. The core of this method is to adaptively select the filter window according to local statistical characteristics, avoiding the loss of image details that may be caused by traditional median filtering. Bilateral filtering combines spatial distance and pixel intensity difference for smoothing. It uses spatial Gaussian kernel to measure the spatial distance between pixels and uses range Gaussian kernel to measure the intensity difference between pixels, thereby achieving smoothing effect while ensuring edge information. Histogram equalization increases the contrast of the image by adjusting the frequency distribution of grayscale values, making the details in the image clearer. It redistributes grayscale values ​​based on the probability distribution function of grayscale values ​​so that grayscale values ​​are evenly distributed in the image. Adaptive gamma correction dynamically adjusts according to the brightness curve of the image, enhances dark details by adjusting the brightness curve of the image, and makes the overall brightness of the image more uniform. Multi-layer convolutional neural network extracts and classifies input images through structures such as convolutional layers, pooling layers, and fully connected layers. The convolution layer extracts local features by sliding the convolution kernel on the image; the pooling layer reduces the dimension of the feature map through downsampling operations to improve computational efficiency; the fully connected layer combines and classifies the extracted features. The transfer learning mechanism enables CNN to converge quickly using pre-trained models, while the attention mechanism can dynamically assign weights to different image regions to improve recognition accuracy and robustness. The variational autoencoder is a generative model that reconstructs input data through the structure of the encoder and decoder. The encoder maps the input data to a distribution in the latent space, and the decoder samples and generates data from the latent space. The KL divergence is used to measure the difference between the encoder distribution and the prior distribution to ensure the stability and reliability of the VAE model when generating data.

[0070] In a specific embodiment, the deep image recognition algorithm collects beam image data once every fixed time interval (such as 1 second) and processes it in real time through the above process. When a significant change in the beam characteristics is detected, the system will immediately trigger the closed-loop control system to make adjustments to ensure that the beam quality is always in the best state. In addition, through the online learning mechanism, the system can continuously accumulate new sample data, continuously optimize model parameters, and adapt to changing environmental conditions. This closed-loop control system not only improves the accuracy and response speed of beam quality analysis, but also enhances the overall performance and reliability of the system.

[0071] Compared with the prior art, the deep image recognition algorithm of the present invention has significant advantages and characteristics. First, by combining adaptive median filtering and bilateral filtering, the present invention can more effectively remove noise and maintain edge information, thereby improving the quality of image data and recognition accuracy. Secondly, the application of histogram equalization and adaptive gamma correction makes the image features more prominent, providing a high-quality data foundation for subsequent multi-layer convolutional neural network recognition. In addition, the introduction of transfer learning mechanism and attention mechanism greatly improves the generalization ability and robustness of the model, shortens training time and resource consumption. More importantly, the application of anomaly detection algorithm based on variational autoencoder and online learning mechanism enables the system to have higher adaptability and response speed, and can respond to emergencies in a timely manner and continuously optimize model parameters.

[0072] Further, such as Figure 3 As shown, the working method of the beam quality optimization algorithm is as follows: first, the beam characteristic parameters of the deep image recognition algorithm, including beam intensity, spot size and edge clarity, are received, and the predicted value of the future ambient temperature change output by the first focal length quality analysis module is combined, and the parameters are mapped to the beam quality and focal length stability indicators through a multi-objective optimization framework. If the current ambient temperature is stable and the beam quality and focal length stability indicators are higher than or equal to the preset threshold, the beam quality and focal length stability indicators are not triggered to continue to be evaluated; otherwise, the dynamic adjustment mechanism is triggered; the dynamic adjustment mechanism uses the Pareto optimal solution, and when the beam quality score Q(ω) under the current focal length setting is higher than the preset threshold and the focal length drift risk R(ω) caused by the ambient temperature change is lower than the preset threshold, the setting parameters are recorded and regarded as a potential optimal solution; otherwise, other configurations are continued to be searched; wherein the beam quality score Q(ω) and the focal length drift risk R(ω) are quantified by the comprehensive evaluation function F(ω); the formula expression of the comprehensive evaluation function F(ω) is:

[0073] F(ω)=w 1 Q(ω)+w 2 (1-R(ω)) 2 (5)

[0074] In formula (5), ω represents the focal length adjustment parameter; w 1 and w 2 are weight coefficients, which represent the priorities of beam quality and focal length stability. When the predicted value of future ambient temperature change output by the first focal length quality analysis module is higher than the preset threshold, w is increased through a dynamic weighting mechanism. 2 ; When the beam quality shows a downward trend, increase w 1 ; After dynamic adjustment, the beam quality optimization algorithm also quickly converges to the optimal solution in the parameter space through genetic algorithm and particle swarm optimization.

[0075] In a specific embodiment, through the Pareto optimal solution method, the algorithm can find the best balance between different goals to ensure that the beam quality and focal length stability are optimal at the same time. Through the dynamic weighting mechanism, when the predicted value of the ambient temperature change is higher than the preset threshold, the focal length stability weight is appropriately increased; when the beam quality shows a downward trend, the beam quality weight is increased. This dual mechanism not only improves the flexibility of the system, but also enhances its ability to cope with complex environmental changes. Based on the above parameter mapping and evaluation results, the beam quality optimization algorithm triggers or does not trigger the dynamic adjustment mechanism. Specifically, if the current ambient temperature is stable and the beam quality and focal length stability indicators are higher than or equal to the preset threshold, it will not be triggered, and the beam quality and focal length stability indicators will continue to be evaluated; otherwise, the dynamic adjustment mechanism will be triggered. The dynamic adjustment mechanism uses the Pareto optimal solution. When the beam quality score under the current focal length setting is higher than the preset threshold and the focal length drift risk caused by the ambient temperature change is lower than the preset threshold, the setting parameters are recorded and regarded as a potential optimal solution; otherwise, other configurations are continued to be searched. This method not only improves the response speed of the system, but also ensures that each adjustment can reach the optimal state, avoiding unnecessary frequent adjustments. In order to quantify the beam quality and focal stability, the beam quality optimization algorithm introduces a comprehensive evaluation function. This function prioritizes the focal adjustment parameters and the weight coefficients of beam quality and focal stability through formula (5). In order to quickly converge to the optimal solution in the parameter space, the beam quality optimization algorithm combines the genetic algorithm GA and the particle swarm optimization PSO. GA simulates the natural selection process through selection, crossover and mutation operations to explore the global optimal solution; PSO uses swarm intelligence to find the local optimal solution. The combination of these two methods enables the algorithm to have both global exploration capabilities and the advantages of local fine adjustment. Specifically, if GA finds a new potential optimal solution, it will be passed to PSO for further refinement; conversely, if PSO finds a better solution, it will be fed back to GA to guide the subsequent search direction. This method not only improves the convergence speed, but also ensures the quality and reliability of the final solution.

[0076] In practical applications, the beam quality optimization algorithm collects beam characteristic parameters once every fixed time interval (such as 1 second) and processes them in real time through the above process. When a significant change in beam quality or focal length stability is detected, the system will immediately trigger the intelligent focal length adjustment module to adjust the focal length in advance to ensure that the beam quality is always in the best state. In addition, through the dynamic weighting mechanism, the system can adjust the weight coefficient in real time according to the changes in ambient temperature to ensure that each adjustment can achieve the optimal state and avoid unnecessary frequent adjustments. This closed-loop control system not only improves the accuracy and response speed of beam quality optimization, but also enhances the overall performance and reliability of the system.

[0077] Compared with the existing technology, the introduction of the dynamic adjustment mechanism enables the system to monitor and evaluate the beam quality score and focal drift risk in real time, and dynamically adjust the focal length according to the evaluation results. This feature enables the system to quickly adapt to changes in different environmental conditions and maintain high beam quality and focal length stability. In addition, the application of genetic algorithms and particle swarm optimization enables the system to efficiently search for the optimal solution in the parameter space. This advantage not only improves the system's optimization accuracy for beam focal length adjustment parameters, but also shortens the time of the optimization process and improves the overall efficiency of the system. The beam quality optimization algorithm comprehensively considers multiple factors such as beam quality, focal length stability, and ambient temperature changes, and balances the priority between these factors by setting weight coefficients and thresholds. This feature enables the system to maintain a high overall performance under different environmental conditions and meet the diverse needs in practical applications.

[0078] In specific implementation, the hardware working environment of the second focal length quality analysis module mainly includes: photodetectors, high-resolution image sensors, high-performance computing platforms: including high-performance CPUs and GPUs, which are used to run deep image recognition algorithms to realize functions such as beam image preprocessing, feature extraction, classification, and anomaly detection. Closed-loop control system: According to the output results of the deep image recognition algorithm, the beam focal length of the BOSA optical device is automatically adjusted to ensure the stability and accuracy of the beam quality. Display and recording equipment: used to display the beam image and recognition results in real time, and record experimental data.

[0079] Based on the above hardware, a comparative experiment was designed, and the second focal length quality analysis module (Group A) and the traditional method (Group B) were used for comparative experiments; among them, the traditional method mainly relies on manual adjustment and simple image processing algorithms, and adopts a fixed focal length setting, and does not dynamically adjust according to the beam quality or ambient temperature changes. First, the beam focal length of the BOSA optical device is manually adjusted by manually observing the beam image; secondly, the beam image is preliminarily processed using a simple image processing algorithm (such as threshold segmentation, edge detection, etc.) to assist manual adjustment. Each group of experiments was carried out under the same experimental conditions, including the same light source, initial value of the beam focal length, image acquisition parameters, etc. Each group conducted five experiments, and the beam focal length adjustment accuracy, adjustment time and beam quality evaluation results of each experiment were recorded. The data recording table is shown in Table 1:

[0080] Table 1 Comparison of the second focal length quality analysis module

[0081]

[0082]

[0083] From the data table 1, it can be seen that the application of the second focal length quality analysis module (Group A) is superior to the traditional method (Group B) in terms of beam focus adjustment accuracy, adjustment time and beam quality evaluation. The average beam focus adjustment accuracy of Group A is 2.0μm, which is much lower than the 5.0μm of Group B; the average adjustment time is 5.2s, which is much lower than the 20.0s of Group B; the average beam quality evaluation score is 9.55 points, which is much higher than the 7.56 points of Group B. Therefore, the application of the second focal length quality analysis module that applies the deep image recognition algorithm and the beam quality optimization algorithm in the BOSA optical device beam focus controllable system has the positive and beneficial effects of significantly improving the beam quality, shortening the adjustment time and improving the adjustment accuracy.

[0084] Furthermore, the multi-objective optimization framework includes a weight allocation module, a multi-objective collaborative optimization module and a dynamic adjustment module; the weight allocation module is used to analyze the changing trend of the beam characteristic parameters through a deep learning algorithm, and calculate the relative weights of the beam quality and focal stability in the optimization process in combination with the predicted value of the ambient temperature; the multi-objective collaborative optimization module is used to receive the weight calculation results output by the weight allocation module, as well as the objective functions of the beam quality and focal stability, and perform iterative search on the objective functions of the beam quality and focal stability under the guidance of the co-evolutionary algorithm to find the optimal solution, and after each iteration, adjust the search direction and step size based on the improvement of the objective function; the dynamic adjustment module is used to receive the optimal solution output by the multi-objective collaborative optimization module, and use a closed-loop control system to monitor the changes in the beam quality and focal stability indicators in real time. If the real-time monitoring value deviates from the preset range, the adjustment amount is calculated and the focal control parameters of the optical device are adjusted.

[0085] In the specific implementation, the weight distribution module of the multi-objective optimization framework is based on a deep learning algorithm, and receives the beam characteristic parameters (including beam intensity, spot size, edge clarity) from the deep image recognition algorithm and the predicted value of future ambient temperature changes provided by the first focal length quality analysis module. First, the module preprocesses these input data to ensure the consistency and comparability of the data. Then, a deep learning model (such as LSTM or GRU network) is used to perform time series analysis on the beam characteristic parameters to identify their changing trends. The key to this step is to accurately capture the dynamic changes of the beam characteristic parameters and provide a basis for subsequent weight calculations. Finally, combined with the predicted value of ambient temperature, a multi-objective optimization algorithm (such as NSGA-II) is used to dynamically calculate the relative weights of beam quality and focal length stability in the optimization process. When calculating the weights, the module comprehensively considers the degree of deviation of the current beam characteristic parameters and the potential impact of future ambient temperature changes on beam performance, ensuring that the weight distribution can reflect both the current state and foresee future trends. The calculated weight values ​​are then output to the multi-objective collaborative optimization module to provide guidance for the subsequent optimization process. The multi-objective collaborative optimization module receives the weight values ​​and the objective functions of beam quality and focal stability output by the adaptive weight allocation module, and performs iterative search based on the co-evolution algorithm to find the optimal solution. The module first generates a set of random focal control parameters as the initial population, and then uses the weight value and objective function to evaluate the fitness of each individual. In the selection stage, the module adopts strategies such as roulette selection and tournament selection to select individuals with higher fitness from the current population as parents. Then, crossover and mutation operations are performed on the parent individuals to generate new offspring individuals. After the offspring individuals are merged with the parent individuals, the optimal individuals are selected according to the fitness value to form a new population. This process is iterated continuously until the preset number of iterations or convergence conditions are reached. Finally, the module outputs the optimal solution, that is, the optimal focal control parameters, which are passed to the real-time feedback and dynamic adjustment module to provide a basis for subsequent real-time monitoring and predictive adjustment. The dynamic adjustment module is based on a closed-loop control system, receives the optimal solution output by the multi-objective collaborative optimization module, and monitors the changes in beam quality and focal stability indicators in real time. The module first uses sensors and signal processing technology to monitor the beam quality and focal stability indicators in real time to ensure the accuracy and real-time nature of the data. Then, the real-time monitoring value is compared with the preset range and the error value is calculated to evaluate the deviation between the current beam performance and the expected performance. Based on the error value and the optimal solution, the module calculates the adjustment amount. The adjustment strategy can be based on algorithms such as PID control and fuzzy control to ensure that the adjustment process is both fast and accurate. After the adjustment amount is calculated, the module applies it to the focal control system of the optical device to adjust the focal control parameters to improve the beam quality and focal stability. The adjusted real-time monitoring value is used as a reference for the next adjustment, forming a closed-loop feedback mechanism to ensure the continuous optimization of beam performance.

[0086] The intelligent focus adjustment module is used to perform a focus adjustment operation according to the output result of the second focus quality analysis module; Figure 2 As shown, the working principle of the intelligent focal length adjustment module is as follows: first, the optimization suggestion of the second focal length quality analysis module is received, including the optimal focal length setting parameter ω and the predicted value of the ambient temperature change, and the current actual state is collected once every t seconds through the real-time monitoring module, including the current beam focal length quality state and the ambient temperature data, and compared with the optimization suggestion; the current beam focal length quality state includes the focus position accuracy, beam divergence and light intensity distribution uniformity parameters, if the deviation between the current actual state and the optimization suggestion is higher than the preset threshold, the focal length adjustment mechanism is triggered, and when the ambient temperature change value in the preset time window is less than or equal to the preset threshold, the focal length adjustment mechanism adopts the PID control algorithm to control the internal lens position or curvature of the optical device through the piezoelectric ceramic driver; when the ambient temperature change prediction value in the preset time window is higher than the preset threshold, the focal length adjustment mechanism plans the focal length path in advance through the model predictive control method to adjust the focal length.

[0087] In the specific implementation, the intelligent focus adjustment module first receives optimization suggestions from the second focus quality analysis module. These suggestions are based on in-depth analysis of the beam focus quality, including optimal focus setting parameters and predicted values ​​of ambient temperature changes. At the same time, the module uses the built-in real-time monitoring module to collect the actual status of the current optical device at fixed intervals, including key parameters such as focus position accuracy, beam divergence and uniformity of light intensity distribution, as well as ambient temperature data. This integrated data processing method enables the module to compare the current status with the optimization suggestions in real time, so as to make accurate adjustment decisions. By combining optimization suggestions with real-time monitoring data, the intelligent focus adjustment module can achieve real-time tracking and dynamic adjustment of the focal state of the optical device, effectively improving the accuracy and response speed of focal control.

[0088] When the deviation between the current actual state and the optimization suggestion is detected to be higher than the preset threshold, the intelligent focal length adjustment module triggers the focal length adjustment mechanism. When the ambient temperature change is relatively stable (that is, the change value within the preset time window is less than or equal to the preset threshold), the module adopts the PID (proportional-integral-differential) control algorithm. The PID algorithm calculates the error between the current focal length and the target focal length, and adjusts the output of the piezoelectric ceramic driver according to the proportion, integral and differential of the error, thereby accurately controlling the position or curvature of the lens inside the optical device and achieving fine-tuning of the focal length. The application of the PID algorithm enables the intelligent focal length adjustment module to achieve precise control of the focal length when the environmental changes are small, ensuring the stability of the beam quality.

[0089] When the predicted value of the ambient temperature change is large (i.e., the change value in the preset time window is higher than the preset threshold), the intelligent focus adjustment module adopts a method based on model predictive control. This method predicts the change trend of the focal length in the future by constructing a mathematical model of the focal length adjustment of the optical device, and plans the focus adjustment path in advance accordingly. Through the optimization algorithm, the module calculates the optimal focus adjustment strategy and implements the pre-adjustment of the focus through the piezoelectric ceramic driver. The focus path planning based on model predictive control enables the intelligent focus adjustment module to make adjustments in advance when the ambient temperature changes greatly, effectively avoiding drastic fluctuations in the focal length and improving the predictability and stability of the beam focus control.

[0090] Compared with the prior art, the intelligent focal length adjustment module flexibly selects the PID control algorithm or the method based on model predictive control to adjust the focal length according to different changes in ambient temperature. This flexible and changeable control strategy enables the module to adapt to various complex working environments and ensure the stability of the beam focal length. In addition, the focal length path planning based on model predictive control enables the module to make adjustments in advance, effectively avoiding the drastic fluctuations in the focal length. This predictive adjustment method greatly improves the stability of the beam focal length control and ensures the long-term stable operation of the optical device. Secondly, the module continuously accumulates experience and knowledge by recording the records of each adjustment and optimizing and analyzing them. This continuous optimization and self-learning ability enables the module to continuously improve the accuracy and efficiency of focal length adjustment and adapt to the ever-changing working environment. At the same time, the intelligent focal length adjustment module is not only suitable for the beam focal length controllable system of BOSA optical devices, but can also be widely used in other optical systems that require high-precision focal length control. Its highly integrated, intelligent, flexible and changeable control strategy, predictability and stability make the module have broad application prospects and market potential.

[0091] The signal optimization module is used to apply deep learning network to perform waveform shaping and noise suppression. When poor eye quality or large jitter is detected, the signal morphology of the transmitting and receiving ends is corrected through pre-distortion processing and adaptive equalization algorithm. The working method of the signal optimization module is as follows: receiving the original signal data of the photoelectric detector, and collecting the current eye quality and jitter every t seconds through the real-time monitoring module; if the eye closure is detected to be lower than the set threshold E tereshold Or the jitter exceeds the allowable range J max , the deep learning network is triggered to perform waveform shaping and noise suppression processing; otherwise, the existing signal state is maintained; during the deep learning network processing, a real-time monitoring module is used to continuously monitor changes in signal quality, and the parameters and strategies of the deep learning network are continuously fine-tuned through a closed-loop feedback mechanism to achieve dynamic optimization.

[0092] The waveform shaping method of the deep learning network is as follows: the collected signal waveform is feature extracted by a convolutional neural network, and the waveform contour, jitter mode and noise distribution in the signal are gradually abstracted by using multiple convolutional layers and pooling layers. The extracted features are input to the fully connected layer, and nonlinear transformation is performed by a deep learning algorithm to eliminate waveform distortion caused by device nonlinearity, dispersion effect or environmental factors. During the waveform shaping process, the deep learning network also dynamically adjusts the equalizer parameters according to the current signal quality through an adaptive equalization algorithm; the noise suppression processing method of the deep learning network is as follows: based on the original signal, amplitude adjustment parameters, time-frequency analysis parameters, frequency domain signal inverse transformation and filtering parameters and deep learning network parameters, the inverse distortion adjustment amount of the original signal is calculated by an inverse distortion function, which is used to correct the original signal; the formula expression of the inverse distortion function is:

[0093]

[0094] In formula (6), z(k) represents the signal after inverse distortion processing; g(k) represents the original signal captured by the receiving end; Represents the set of training parameters of the deep learning network, including weights, biases, and any regularization parameters, used to describe the inverse distortion function I -1 The nonlinear mapping relationship of A s,β g(k) represents the amplitude adjustment of the signal g(k). Parameters s and β control the amplitude and mode of adjustment, which are used to compensate for the amplitude distortion caused by the nonlinear factors of the device during the transmission process. γ,δ (STEFg(k)) represents the weighted processing of the short-time Fourier transform result of the signal g(k), and the parameters γ and δ control the weight and method of weighting, which is used to compensate the signal distortion in the frequency domain; B Q,T (IFFT(W γ,δ )) represents the amount of filtering processing of the frequency domain signal after the weighted inverse Fourier transform; the parameters Q and T control the parameters and methods of the filter to further suppress noise and interference.

[0095] In a specific example, a convolutional neural network is used to extract meaningful features from complex signals. Its working principle is based on local perceptual field, weight sharing and pooling operations. The local perceptual field allows the network to focus on the local area of ​​the input data, while weight sharing reduces the number of parameters and enhances the generalization ability of the model. The pooling layer further reduces the feature dimension by downsampling while retaining key information. In the present invention, CNN gradually abstracts the waveform contour, jitter pattern and noise distribution in the signal through multiple convolutional layers, and inputs these features into the fully connected layer for nonlinear transformation, thereby eliminating waveform distortion caused by device nonlinearity, dispersion effects or environmental factors. This hierarchical feature extraction method significantly improves the effect of waveform shaping.

[0096] The adaptive equalization algorithm aims to compensate for the frequency distortion and phase deviation introduced during the transmission process. Its core lies in dynamically adjusting the equalizer parameters according to the current signal quality to maintain the best signal shape. Specifically, the adaptive equalization algorithm uses the feature vector extracted by the convolutional neural network to evaluate the frequency distortion and phase deviation in the signal in real time, and dynamically adjusts the equalizer parameters through optimization algorithms such as minimum mean square error (LMS) or recursive least squares (RLS). This method not only improves the effect of waveform shaping, but also enhances the adaptability and robustness of the system, ensuring that each adjustment can reach the optimal state.

[0097] The inverse distortion function is a mathematical mapping relationship used to perform inverse distortion adjustment on the original signal. Its working principle is based on amplitude adjustment, time-frequency analysis, inverse frequency domain transformation and filtering of the signal. Specifically, the inverse distortion function performs amplitude adjustment \(A(t)\) on the original signal \(s(t)\) captured by the receiving end, and combines the weighted processing \(W(f)\) of the short-time Fourier transform result, and finally achieves noise suppression by filtering the frequency domain signal after the weighted inverse Fourier transform \(F(s)\). The training parameter set \(\theta\) in formula (6) describes the nonlinear mapping relationship of the inverse distortion function, which is used to compensate for various distortion phenomena in the transmission process. This method not only improves the effect of noise suppression, but also enhances the clarity and stability of the signal.

[0098] The closed-loop feedback mechanism achieves dynamic optimization by continuously monitoring changes in signal quality and constantly fine-tuning the parameters and strategies of the deep learning network based on the actual results. Specifically, the closed-loop feedback mechanism uses the tiny changes captured by the real-time monitoring module and transmits this information to the deep learning network through the feedback path, enabling it to adjust parameters in a timely manner to ensure that each adjustment achieves the expected effect. This method not only improves the accuracy of the adjustment, but also enhances the response speed and stability of the system. The application of the closed-loop feedback mechanism enables the system to maintain stable performance in complex environments and ensures that each adjustment can achieve the optimal state.

[0099] In practical applications, the specific implementation method of the deep learning network is as follows: First, the module receives the original signal data from the photodetector, and collects the current eye diagram quality and jitter at fixed intervals (such as 1 second) through the real-time monitoring module. If the eye diagram closure is detected to be lower than the set threshold or the jitter exceeds the allowable range, the deep learning network is triggered to perform waveform shaping and noise suppression processing; otherwise, the existing signal state is maintained. Next, the deep learning network uses a convolutional neural network to extract features from the collected signal waveform, and uses multiple convolutional layers and pooling layers to gradually abstract the waveform contour, jitter pattern, and noise distribution in the signal. Then, the equalizer parameters are dynamically adjusted according to the current signal quality in combination with the adaptive equalization algorithm to eliminate the frequency distortion and phase deviation in the signal. Finally, the inverse distortion adjustment amount of the original signal is calculated through the inverse distortion function, which is used to correct the original signal, and the parameters and strategies of the deep learning network are continuously fine-tuned through the closed-loop feedback mechanism to achieve dynamic optimization.

[0100] Compared with the existing technology, the deep learning network extracts features and nonlinear transformation of signals through efficient algorithms such as convolutional neural networks, which can quickly and accurately realize waveform shaping and noise suppression, and improve signal quality. In addition, the deep learning network has a strong self-learning ability, and can dynamically adjust parameters and strategies according to the changes in signal quality monitored in real time to achieve dynamic optimization. This adaptability enables the system to cope with signal transmission requirements under different environments and conditions. Secondly, the deep learning network can effectively suppress the noise and interference generated during the transmission process through complex nonlinear transformation and filtering processing, and improve the robustness and stability of the system. At the same time, the structure and parameters of the deep learning network can be adjusted and optimized according to actual needs, making the system more scalable and flexible. At the same time, with the continuous development of deep learning technology, the performance of the system can be further improved. In addition, the application of deep learning networks makes the BOSA optical device beam focal length controllable system more intelligent, able to independently judge the signal quality and perform corresponding optimization processing, reducing the need and cost of manual intervention.

[0101] Furthermore, the working principle of the adaptive equalization algorithm is as follows: first, the input signal is analyzed using the hierarchical feature extraction mechanism of the convolutional neural network to extract signal features, including the intensity distribution of frequency components and slight shifts in phase relationships, and the extracted features are integrated into a signal quality vector through a fully connected layer; based on the current signal quality vector, a nonlinear transformation is performed through a multi-layer perceptron to output an action vector containing an equalizer parameter adjustment strategy, and the filter parameters of the equalizer are adjusted according to the action vector to compensate for frequency distortion and phase deviation in the signal; during the equalizer parameter adjustment process, the output signal quality is continuously monitored. If the signal quality improvement is lower than a preset threshold, the gradient is calculated based on the error signal using the minimum mean square error method, the equalizer parameters are updated, and the amplitude and direction of the parameter update are limited by the regularization term; after the equalizer parameter adjustment process, the monitored signal quality information is fed back to the strategy network through a feedback mechanism to optimize the network parameters and strategy prediction accuracy.

[0102] In a specific embodiment, the adaptive equalization algorithm first uses the hierarchical feature extraction mechanism of the convolutional neural network (CNN) to analyze the input signal and extract signal features. The convolution layer captures local features through sliding window operations, while the pooling layer reduces the feature dimension by downsampling and retains key information. This hierarchical feature extraction method can abstract the intensity distribution of frequency components and subtle shifts in phase relationships from complex signals, which not only improves the accuracy of feature extraction, but also enhances the generalization ability of the system, making subsequent processing more accurate and reliable.

[0103] Based on the current signal quality vector, the adaptive equalization algorithm performs nonlinear transformation through a multilayer perceptron (MLP) and outputs an action vector containing the equalizer parameter adjustment strategy. MLP is a feedforward neural network consisting of multiple hidden layers, each of which introduces nonlinear characteristics through an activation function; the action vector contains the specific adjustment values ​​of the filter parameters to compensate for the frequency distortion and phase deviation in the signal. This method not only improves the accuracy of parameter adjustment, but also enhances the response speed and stability of the system. During the equalizer parameter adjustment process, the adaptive equalization algorithm continuously monitors the output signal quality. If the signal quality improvement is lower than the preset threshold, the gradient is calculated according to the error signal through the minimum mean square error method (LMS) to update the equalizer parameters. LMS is a classic optimization algorithm that adjusts parameters by minimizing the sum of squared errors to ensure that each adjustment can achieve the expected effect. Specifically, by calculating the gradient through the error signal, LMS can evaluate the effectiveness of the current parameters in real time, and dynamically adjust the parameters through the gradient descent method to gradually approach the optimal solution. In addition, the amplitude and direction of parameter update are limited by the regularization term, which avoids overfitting and enhances the robustness and stability of the system.

[0104] After the equalizer parameter adjustment process, the adaptive equalization algorithm feeds back the monitored signal quality information to the policy network through the feedback mechanism to optimize the network parameters and policy prediction accuracy. The feedback mechanism is the key to the system's self-regulation. It continuously monitors changes in signal quality and continuously fine-tunes the parameters and strategies of the deep learning network according to the actual results to achieve dynamic optimization. This method not only improves the accuracy of adjustment, but also enhances the response speed and stability of the system, allowing the system to maintain stable performance in complex environments and ensure that each adjustment can reach the optimal state.

[0105] Compared with the existing technology, through the joint application of CNN and MLP, the system can extract signal features and generate equalization strategies in real time and accurately, ensuring the stability of the beam focal length and the high quality of signal transmission. In addition, the adaptive equalization algorithm can dynamically adjust the equalizer parameters according to changes in signal quality without manual intervention, which greatly improves the system's adaptability and intelligence level. Secondly, the combination of the LMS algorithm and the regularization term ensures the accuracy and stability of parameter adjustment, and the system can maintain stable performance even in complex signal environments. At the same time, since the algorithm is built on a deep learning framework, it is easy to expand and optimize by increasing the amount of data, adjusting the network structure, etc., to meet the needs of higher-performance optical devices in the future.

[0106] In a specific embodiment, the hardware working environment in which the signal optimization module is applied to the beam focus controllable system of the BOSA optical device mainly includes:

[0107] BOSA optical device: responsible for the generation, modulation and emission of light beams.

[0108] Photodetector: used to receive light signals and convert them into electrical signals.

[0109] High-speed data acquisition card: real-time acquisition of raw signal data from photodetectors.

[0110] Embedded processor: runs the deep learning algorithm of the signal optimization module and the real-time monitoring module.

[0111] Programmable power supply: Provides stable power supply for BOSA optical devices and the entire system.

[0112] Optical test platform: used to accurately measure beam focus and signal quality.

[0113] Based on the above hardware working environment, two groups of hardware were selected, and the signal optimization module (Group A) and the traditional method (Group B) were used for comparison. The traditional method mainly relies on fixed filters for waveform shaping and simple noise suppression processing, and lacks dynamic optimization capabilities. The experiments were carried out under the same experimental conditions, including the same beam focal length, signal strength and noise environment. The eye closure and jitter range of each experiment were measured and recorded by the optical test platform, and the waveform shaping and noise suppression effects were evaluated. The experimental record table is shown in Table 2:

[0114] Table 2 Signal optimization module comparison experiment record

[0115]

[0116] By comparing the experimental records in Table 2, it can be clearly seen that Group A (signal optimization module) is superior to Group B (traditional method) in terms of eye closure, jitter range, waveform shaping effect and noise suppression effect. The signal optimization module achieves dynamic optimization of signal quality through deep learning network and real-time monitoring module, effectively improving the stability and performance of the beam focus controllable system.

[0117] A method for realizing controllable focal length of a BOSA optical device beam, such as Figure 4 As shown, the following steps are included:

[0118] Step 1: Use a temperature sensor to collect temperature data of the BOSA optical device and the environment in which the BOSA optical device is located in real time, remove noise, detect and eliminate outliers, and perform data smoothing on the collected temperature data, and output the preprocessed temperature data set;

[0119] Step 2: Based on the temperature data set output in step 1, a thermal prediction model is constructed through a long short-term memory network to predict the future trend of ambient temperature changes;

[0120] Step 3: Use a photodetector to monitor the beam quality emitted by the BOSA optical device in real time, including brightness distribution and spot shape, and pre-process the collected beam quality data. Use a deep image recognition algorithm to identify and classify the focus position accuracy, beam divergence, and light intensity distribution uniformity parameters of the collected data, and output the beam quality data set;

[0121] Step 4: Based on the beam quality data set outputted in step 3, a multi-objective optimization model is constructed by using a beam quality optimization algorithm taking into account the beam quality, the ambient temperature variation trend outputted in step 2, and the focal length stability, and an optimal focal length adjustment scheme is outputted;

[0122] Step 5: According to the focal length pre-adjustment strategy and optimization strategy outputted in step 4, the focal length is initially adjusted by a stepper motor. During the focal length adjustment process, the adjustment error is monitored in real time by a closed-loop feedback control mechanism, and the PID algorithm is used for real-time compensation.

[0123] Step 6: After the initial adjustment of the focal length, the expansion and contraction of the piezoelectric ceramic is controlled by the piezoelectric ceramic fine-tuning mechanism to perform a secondary adjustment of the focal length;

[0124] Step 7. According to the eye diagram quality and jitter detection results, the signal emitted by the BOSA optical device is subjected to waveform shaping, noise suppression and adaptive equalization through a deep learning network. When poor eye diagram quality or large jitter is detected, correction is performed through pre-distortion processing and adaptive equalization algorithm to optimize the signal shape at the transmitter and receiver.

[0125] In a specific embodiment, the core technology of the controllable focal length of the beam of BOSA (Bidirectional Optical Sub-Assembly) optical device is the integration of advanced concepts and technical means such as temperature compensation, machine learning prediction, beam quality monitoring and optimization, multi-objective optimization modeling, closed-loop feedback control and signal processing. This method uses intelligent temperature monitoring and compensation mechanisms to cope with the impact of environmental changes on optical performance; uses deep learning algorithms for thermal prediction and beam quality assessment, which improves the system's adaptability to future environmental changes and the accuracy of beam quality; introduces a closed-loop feedback control system based on the PID algorithm to ensure high precision and real-time performance during the focal length adjustment process; at the same time, the piezoelectric ceramic fine-tuning mechanism is used to achieve fine adjustment of the focal length, ensuring the stability and accuracy of the focal length; finally, the quality of the communication signal is improved through waveform shaping, noise suppression and adaptive equalization processing.

[0126] In the present invention, the temperature sensor is used to collect real-time temperature data, and after preprocessing, these data are input into a trained long short-term memory network (LSTM) model. LSTM is a special recursive neural network (RNN) that can learn long-term dependencies and is very suitable for processing data with time series characteristics, such as temperature change trends. The thermal prediction model constructed by LSTM can predict the changes in ambient temperature in advance, so that the system can take measures in advance to maintain the stability of the beam focal length. This technical feature not only reduces the impact of external temperature fluctuations on the focal length, but also improves the response speed and accuracy of the system.

[0127] The combination of photodetectors and deep image recognition algorithms provides real-time monitoring of beam quality. Photodetectors can capture the brightness distribution and morphology information of the beam, while deep image recognition algorithms can extract key parameters such as focal position, divergence, and intensity distribution. This allows the system to accurately determine the quality of the current beam and make necessary adjustments accordingly. With this combination, the system can maintain optimal beam quality at all times, thereby improving communication efficiency and reliability.

[0128] The multi-objective optimization model is designed to comprehensively consider three factors: beam quality, ambient temperature changes, and focal length stability. This model receives data from the beam quality monitoring module and the prediction results of the thermal prediction module as input, and outputs an optimal focal length adjustment plan. The plan aims to minimize the focal length drift caused by temperature changes while maximizing the stability of beam quality and focal length. Such a multi-objective optimization strategy ensures that the best communication performance can be maintained even in complex environments.

[0129] The use of stepper motors and PID algorithms provides the ability to mechanically execute and compensate for errors in real time for focal length adjustment. The stepper motor performs preliminary focal length adjustment according to the instructions given by the optimization model, while the PID controller is responsible for continuously monitoring the adjustment effect and correcting deviations in a timely manner throughout the process. This approach not only improves the speed and accuracy of focal length adjustment, but also enhances the robustness of the system, enabling it to work stably under various conditions.

[0130] The fine-tuning mechanism of piezoelectric ceramics further enhances the precision of focus adjustment. Piezoelectric materials have the property that they deform when a voltage is applied to them, and this deformation can be very subtle and easy to control, making them very suitable as a tool for fine-tuning focus. By precisely controlling the expansion and contraction of piezoelectric ceramics, sub-micron adjustments of focus can be achieved, which is very important for applications that require extremely high precision.

[0131] Finally, waveform shaping, noise suppression, and adaptive equalization are optimization measures for communication signals. These processing steps help improve the signal shape at the transmitter and receiver, reduce distortion and interference during transmission, and thus improve communication quality. In particular, when poor eye diagram quality or excessive jitter is detected, the system automatically starts pre-distortion processing and adaptive equalization algorithms to make corrections to ensure that the final received signal is as close to the original transmitted signal as possible.

[0132] The above technical features are integrated into a complete beam focus control system, and its specific implementation method is as follows: First, after the system is started, the temperature sensor begins to periodically collect temperature information of the BOSA optical device and its surrounding environment. After a series of preprocessing steps - including denoising, outlier removal and smoothing - this information is fed into the LSTM model to generate temperature predictions for a period of time in the future. At the same time, the photodetector is constantly monitoring the beam emitted from the BOSA optical device and passing the acquired data to the deep image recognition algorithm for analysis. Once the beam quality assessment is completed, the results are immediately fed back to the multi-objective optimization model to develop the most appropriate focus adjustment strategy. Subsequently, the stepper motor performs a rough focus adjustment action based on this strategy, and the PID controller is always active throughout the adjustment process, ready to correct any possible errors. After the coarse adjustment is completed, the system enters the fine adjustment stage dominated by piezoelectric ceramics to ensure that the ideal focus setting is achieved. In addition, for possible signal quality issues, the system will also synchronously run waveform shaping, noise suppression and adaptive equalization functions to ensure the integrity and clarity of the signals at both ends of the communication link.

[0133] Compared with the traditional BOSA optical device focus control method, the technical solution proposed in the present invention shows significant advantages. Traditional methods often rely on fixed mechanical structures or simple temperature compensation mechanisms, which are difficult to flexibly cope with complex and changeable working environments, especially in the case of large temperature fluctuations, which may cause unstable beam focus and thus affect communication quality. The present invention successfully overcomes the above problems by introducing advanced temperature prediction models, intelligent beam quality monitoring and optimization algorithms, and precise focus adjustment mechanisms.

[0134] First, the use of LSTM model for temperature prediction greatly improves the predictability and responsiveness of the system. It can know the future temperature change trend in advance, so that the system can make corresponding preparations before the actual temperature changes, avoiding focal drift caused by temperature mutation. Secondly, the application of deep image recognition algorithm makes the monitoring of beam quality more accurate and reliable. Unlike traditional monitoring methods that can only provide limited information, this method can comprehensively and meticulously analyze the various properties of the beam, providing a solid foundation for subsequent optimization and adjustment. Furthermore, the establishment of a multi-objective optimization model reflects the innovation of the present invention at the algorithm level. It not only focuses on a single goal, but also comprehensively considers the interaction between multiple important factors to ensure that each decision is part of the global optimal solution. More importantly, a fast and stable focal adjustment process is achieved by combining a stepper motor and a PID controller, and the ultra-fine control provided by the piezoelectric ceramic fine-tuning mechanism makes the focal adjustment both fast and accurate. Finally, the optimization processing measures for signal quality greatly improve the communication effect. Whether it is waveform shaping, noise suppression, or adaptive equalization, it effectively reduces the bit error rate during transmission and improves the reliability and efficiency of communication. In summary, the present invention not only has a breakthrough in technology, but also performs well in practical applications, bringing a brand-new solution to the field of BOSA optical devices.

[0135] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A beam focal length controllable system for a BOSA optical device, characterized in that: include: A first focal length quality analysis module, a second focal length quality analysis module, an intelligent focal length adjustment module and a signal optimization module; The first focal length quality analysis module is used to collect ambient temperature data in real time, build a thermal prediction model through a long short-term memory network, predict the future ambient temperature change trend and output it to the second focal length quality analysis module; The second focal length quality analysis module is used to collect the beam brightness distribution and spot shape data in real time through the photoelectric detector, and identify and classify the focus position accuracy, beam divergence and light intensity distribution uniformity parameters of the collected data through the deep image recognition algorithm; based on the recognition result, the second focal length quality analysis module is also used to construct a multi-objective optimization model by taking into account the beam quality, the ambient temperature change trend output by the first focal length quality analysis module and the focal length stability through the beam quality optimization algorithm, and output the optimal focal length adjustment plan; The intelligent focus adjustment module is used to perform a focus adjustment operation according to the output result of the second focus quality analysis module; The signal optimization module is used to apply a deep learning network to perform waveform shaping and noise suppression. When poor eye diagram quality or large jitter is detected, correction is performed through pre-distortion processing and adaptive equalization algorithm to optimize the signal morphology of the transmitting end and the receiving end.

2. The beam focal length controllable system of a BOSA optical device according to claim 1, characterized in that: The thermal prediction model includes a data input layer, a feature layer, a training layer, a trend prediction layer and an output layer; the data input layer is used to collect temperature data of the environment where the BOSA optical device is located in real time, and the collected data is used to identify periodic components and non-periodic trends through a time series analysis method, and input into the feature layer in the form of a time series; The feature layer is used to extract feature vectors from the output temperature data of the data input layer through principal component analysis and autoencoder, and transmit the extracted feature data to the training layer for modeling; the training layer is used to control the information flow through the gating mechanism based on the extracted data of the feature layer, learn the long-term dependency in the time series, and train the thermal prediction model; the trend prediction layer is used to predict the future trend of ambient temperature changes based on the trained model; During the prediction process, a sliding window is used to improve the continuity and accuracy of the prediction results; the output layer is used to output the predicted ambient temperature change trend to the second focal length quality analysis module.

3. The beam focal length controllable system of a BOSA optical device according to claim 1, characterized in that: The working method of the depth image recognition algorithm is: S301. Based on the beam image data captured by the photodetector and the image sensor, the deep image recognition algorithm removes noise and maintains edge information through adaptive median filtering and bilateral filtering, and adjusts all images to the same scale through a normalization method; wherein the adaptive median filtering is based on each pixel point P, calculates the median M(P) in the neighborhood of the pixel point P, and determines the pixel value according to a local statistical method; the bilateral filtering performs smoothing by combining spatial distance and pixel intensity difference, and the formula for bilateral filtering smoothing is: In formula (1), x and y represent the current pixel position and the neighboring pixel position respectively; l(x) and l(y) are the pixel intensities at positions x and y respectively; G as (x, y) represents the spatial Gaussian kernel, which is used to measure the spatial distance weight between two pixels; as represents the standard deviation of the spatial distance, which is used to control the influence of the spatial distance; G ar (l(x),l(y)) represents the range Gaussian kernel, which is used to weigh the intensity difference between two pixels; ar is the intensity standard deviation, which is used to control the influence of intensity difference; S302, the deep image recognition algorithm strengthens the image features through histogram equalization and adaptive gamma correction, and identifies the edge and texture image features in the enhanced beam image through a multi-layer convolutional neural network; wherein the histogram equalization increases the image contrast through grayscale value frequency distribution, and the adaptive gamma correction enhances the dark details through image brightness curve adjustment; during the learning process of the multi-layer convolutional neural network, the deep image recognition algorithm converges and classifies different types of beam images using the pre-trained ResNet architecture through a transfer learning mechanism, and dynamically allocates weights to different image areas through an attention mechanism; wherein the formula expression for the recognition of the multi-layer convolutional neural network is: In formula (2), a i represents the attention weight of the i-th region; r i represents the unnormalized attention score of the i-th region; represents the weight matrix used to calculate the attention score; exp represents the exponential function, which is used for Softmax normalization; b a represents the bias term; W x and W y They represent the functions used to process the input feature vector x i and context information y i The weight matrix of S303. Based on the identified data, the normal and abnormal beam characteristics are distinguished by an abnormality detection algorithm based on a variational autoencoder. The formula is: In formula (3), D KL represents KL divergence; q(z|x) represents encoder distribution; p(z) represents prior distribution; p(x|z) represents the data distribution generated by the decoder; N represents the number of image pixels; c i Represents the i-th pixel value of the original image; e i Represents the i-th pixel value of the reconstructed image; Indicates the reconstruction error. If the reconstruction error is detected to exceed the preset threshold, the alarm mechanism is immediately triggered and the closed-loop control system is notified to make adjustments. Otherwise, monitoring continues; S304, accumulate new sample data through the online learning mechanism. If the quality assessment score S(θ) of the new sample data is higher than the preset threshold, update the model parameters; otherwise, maintain the current parameters unchanged. The calculation formula of the quality assessment score is: In formula (4), S(θ) represents the quality assessment score of the sample data; f(θ; x i ) represents the model for the i-th sample G i The predicted value of H i represents the true label of the i-th sample; σ i represents the standard deviation of the i-th sample; θ represents the model parameter; n represents the number of samples.

4. The beam focal length controllable system of a BOSA optical device according to claim 1, characterized in that: The working method of the beam quality optimization algorithm is as follows: first, the beam characteristic parameters of the deep image recognition algorithm, including beam intensity, spot size and edge clarity, are received, and the parameters are mapped to the beam quality and focal stability indicators through a multi-objective optimization framework in combination with the predicted value of the future ambient temperature change output by the first focal quality analysis module. If the current ambient temperature is stable and the beam quality and focal stability indicators are higher than or equal to the preset threshold, the beam quality and focal stability indicators are not triggered to continue to be evaluated; otherwise, the dynamic adjustment mechanism is triggered; the dynamic adjustment mechanism uses the Pareto optimal solution, and when the beam quality score Q(ω) under the current focal setting is higher than the preset threshold and the focal drift risk R(ω) caused by the ambient temperature change is lower than the preset threshold, the setting parameters are recorded and regarded as a potential optimal solution; otherwise, other configurations are continued to be searched; wherein the beam quality score Q(ω) and the focal drift risk R(ω) are quantified by the comprehensive evaluation function F(ω); the formula expression of the comprehensive evaluation function F(ω) is: F(ω)=w1Q(ω)+w2(1-R(ω)) 2 (5) In formula (5), ω represents the focal length adjustment parameter; w1 and w2 are weight coefficients, which respectively represent the priority of beam quality and focal length stability. When the predicted value of future ambient temperature change output by the first focal length quality analysis module is higher than the preset threshold, w2 is increased through a dynamic weighting mechanism; when the beam quality shows a downward trend, w1 is increased; after dynamic adjustment, the beam quality optimization algorithm also converges quickly to the optimal solution in the parameter space through genetic algorithm and particle swarm optimization.

5. The beam focal length controllable system of a BOSA optical device according to claim 4, characterized in that: The multi-objective optimization framework includes a weight allocation module, a multi-objective collaborative optimization module and a dynamic adjustment module; the weight allocation module is used to analyze the change trend of the beam characteristic parameters through a deep learning algorithm, and calculate the relative weights of the beam quality and focal length stability in the optimization process in combination with the predicted value of the ambient temperature; The multi-objective collaborative optimization module is used to receive the weight calculation results output by the weight allocation module, as well as the objective functions of beam quality and focal stability, and to iteratively search the objective functions of beam quality and focal stability under the guidance of the weights through a collaborative evolutionary algorithm to find the optimal solution, and after each iteration, adjust the search direction and step size based on the improvement of the objective function; the dynamic adjustment module is used to receive the optimal solution output by the multi-objective collaborative optimization module, and to use a closed-loop control system to monitor the changes in the beam quality and focal stability indicators in real time. If the real-time monitoring value deviates from the preset range, the adjustment amount is calculated and the focal control parameters of the optical device are adjusted.

6. The beam focal length controllable system of a BOSA optical device according to claim 1, characterized in that: The working principle of the intelligent focal length adjustment module is as follows: first, the optimization suggestion of the second focal length quality analysis module is received, including the optimal focal length setting parameter ω and the predicted value of the ambient temperature change; the current actual state is collected once every t seconds through the real-time monitoring module, including the current beam focal length quality state and the ambient temperature data, and compared with the optimization suggestion; the current beam focal length quality state includes the focus position accuracy, beam divergence and light intensity distribution uniformity parameters; if the deviation between the current actual state and the optimization suggestion is higher than the preset threshold, the focal length adjustment mechanism is triggered; when the ambient temperature change value in the preset time window is less than or equal to the preset threshold, the focal length adjustment mechanism adopts the PID control algorithm to control the internal lens position or curvature of the optical device through the piezoelectric ceramic driver; when the ambient temperature change predicted value in the preset time window is higher than the preset threshold, the focal length adjustment mechanism plans the focal length path in advance through the model predictive control method to adjust the focal length.

7. The beam focal length controllable system of a BOSA optical device according to claim 1, characterized in that: The working method of the signal optimization module is: receiving the original signal data of the photoelectric detector, and collecting the current eye diagram quality and jitter situation every t seconds through the real-time monitoring module; if the eye diagram closure is detected to be lower than the set threshold E tereshold Or the jitter exceeds the allowable range J max , the deep learning network is triggered to perform waveform shaping and noise suppression processing; otherwise, the existing signal state is maintained; during the deep learning network processing, a real-time monitoring module is used to continuously monitor changes in signal quality, and the parameters and strategies of the deep learning network are continuously fine-tuned through a closed-loop feedback mechanism to achieve dynamic optimization.

8. The beam focus controllable system of a BOSA optical device according to claim 7, characterized in that: The waveform shaping method of the deep learning network is as follows: the collected signal waveform is feature extracted by a convolutional neural network, and the waveform contour, jitter mode and noise distribution in the signal are gradually abstracted by using multiple convolutional layers and pooling layers. The extracted features are input to the fully connected layer, and nonlinear transformation is performed by a deep learning algorithm to eliminate waveform distortion caused by device nonlinearity, dispersion effect or environmental factors. During the waveform shaping process, the deep learning network also dynamically adjusts the equalizer parameters according to the current signal quality through an adaptive equalization algorithm; the noise suppression processing method of the deep learning network is as follows: based on the original signal, amplitude adjustment parameters, time-frequency analysis parameters, frequency domain signal inverse transformation and filtering parameters and deep learning network parameters, the inverse distortion adjustment amount of the original signal is calculated by an inverse distortion function, which is used to correct the original signal; the formula expression of the inverse distortion function is: In formula (6), z(k) represents the signal after inverse distortion processing; g(k) represents the original signal captured by the receiver; θ represents the training parameter set of the deep learning network, including weights, biases, and any regularization parameters, which is used to describe the inverse distortion function I -1 The nonlinear mapping relationship of A s,β g(k) represents the amplitude adjustment of the signal g(k). Parameters s and β control the amplitude and mode of adjustment, which are used to compensate for the amplitude distortion caused by the nonlinear factors of the device during the transmission process. γ,δ (STEFg(k)) represents the weighted processing of the short-time Fourier transform result of the signal g(k), and the parameters γ and δ control the weight and method of weighting, which is used to compensate the signal distortion in the frequency domain; B Q,T (IFFT(W γ,δ )) represents the amount of filtering processing of the frequency domain signal after the weighted inverse Fourier transform; the parameters Q and T control the parameters and methods of the filter to further suppress noise and interference.

9. The beam focus controllable system of a BOSA optical device according to claim 8, characterized in that: The working principle of the adaptive equalization algorithm is as follows: first, the input signal is analyzed using the hierarchical feature extraction mechanism of the convolutional neural network to extract signal features, including the intensity distribution of frequency components and slight shifts in phase relations, and the extracted features are integrated into a signal quality vector through a fully connected layer; based on the current signal quality vector, a nonlinear transformation is performed through a multi-layer perceptron to output an action vector containing an equalizer parameter adjustment strategy, and the filter parameters of the equalizer are adjusted according to the action vector to compensate for the frequency distortion and phase deviation in the signal; during the equalizer parameter adjustment process, the output signal quality is continuously monitored, and if the signal quality improvement is lower than a preset threshold, the gradient is calculated based on the error signal using the minimum mean square error method, the equalizer parameters are updated, and the amplitude and direction of the parameter update are limited by the regularization term; after the equalizer parameter adjustment process, the monitored signal quality information is fed back to the policy network through a feedback mechanism to optimize the network parameters and the accuracy of policy prediction.

10. A method for realizing controllable focal length of a BOSA optical device beam, characterized in that: A beam focus controllable system for a BOSA optical device as claimed in any one of claims 1 to 9, comprising the following steps: Step 1: Use a temperature sensor to collect temperature data of the BOSA optical device and the environment in which the BOSA optical device is located in real time, remove noise, detect and eliminate outliers, and perform data smoothing on the collected temperature data, and output the preprocessed temperature data set; Step 2: Based on the temperature data set output in step 1, a thermal prediction model is constructed through a long short-term memory network to predict the future trend of ambient temperature changes; Step 3: Use a photodetector to monitor the beam quality emitted by the BOSA optical device in real time, including brightness distribution and spot shape, and pre-process the collected beam quality data. Use a deep image recognition algorithm to identify and classify the focus position accuracy, beam divergence, and light intensity distribution uniformity parameters of the collected data, and output the beam quality data set; Step 4: Based on the beam quality data set outputted in step 3, a multi-objective optimization model is constructed by using a beam quality optimization algorithm taking into account the beam quality, the ambient temperature variation trend outputted in step 2, and the focal length stability, and an optimal focal length adjustment scheme is outputted; Step 5: According to the focal length pre-adjustment strategy and optimization strategy outputted in step 4, the focal length is initially adjusted by a stepper motor. During the focal length adjustment process, the adjustment error is monitored in real time by a closed-loop feedback control mechanism, and the PID algorithm is used for real-time compensation. Step 6: After the initial adjustment of the focal length, the expansion and contraction of the piezoelectric ceramic is controlled by the piezoelectric ceramic fine-tuning mechanism to perform a secondary adjustment of the focal length; Step 7. According to the eye diagram quality and jitter detection results, the signal emitted by the BOSA optical device is subjected to waveform shaping, noise suppression and adaptive equalization through a deep learning network. When poor eye diagram quality or large jitter is detected, correction is performed through pre-distortion processing and adaptive equalization algorithm to optimize the signal shape at the transmitter and receiver.

Images