Optical fiber array coupling optical system with self-calibration function

Abstract

The application discloses a fiber array coupling optical system with a self-calibration function and belongs to the technical field of optical precision alignment. The system comprises a fiber array substrate, a coupling optical unit, a monitoring light source unit, a self-calibration unit and an execution unit; a reverse injection light path of a heteroband monitoring light is adopted, a Fresnel back reflection light is utilized, a multi-dimensional position and posture deviation is extracted through a Hartmann-Shack wavefront sensor, a built-in second-order polynomial temperature-drift response model is relied on to eliminate quasi-static drift caused by temperature, a time-domain convolution network is combined to predict the deviation evolution trend, a model predictive control algorithm is adopted to solve the optimal compensation amount, a piezoelectric ceramic and a thermo-optic micro-lens driver are matched to compensate for the angle deviation in the lateral and axial directions. The application realizes full-optical direct monitoring of a coupling state, temperature drift and dynamic disturbance distinction, deviation trend prediction and multi-degree-of-freedom decoupling and parallel compensation, and significantly improves the long-term stability and anti-interference ability of the system.

Description

Technical Field

[0001] This invention relates to the field of optical precision alignment technology, and more specifically, to a fiber array coupled optical system with self-calibration function. Background Technology

[0002] Fiber optic array coupled optical systems are core components for achieving efficient optical signal transmission in fields such as optical communication and lidar. Their coupling efficiency hinges on high-precision spatial alignment between the fiber optic array and the target device. In practical applications, factors such as ambient temperature fluctuations, mechanical vibrations, packaging stress release, and material aging can cause slow drift or sudden shifts in alignment, leading to decreased coupling efficiency or even link interruption. Existing technologies mainly fall into three categories: first, purely passive rigid packaging, relying on high-precision mechanical positioning and adhesive fixation, lacking active adjustment capabilities and exhibiting poor long-term stability; second, manual fine-tuning frames or electric displacement stages, requiring manual optimization during shutdown, resulting in slow response, poor repeatability, and difficulty adapting to dynamic operating conditions; and third, integrated position sensor closed-loop feedback, but only detecting mechanical displacement rather than directly measuring the optical path, leading to nonlinear errors and limited calibration accuracy. Furthermore, existing technologies often employ single-degree-of-freedom polling adjustment, neglecting the strong coupling characteristics of axial, lateral, and angular deviations, and lacking dedicated temperature desensitization and deviation trend prediction mechanisms, resulting in slow convergence, oscillation, and difficulty in quickly maintaining the optimal coupling operating point.

[0003] Therefore, there is an urgent need to design a fiber array coupling optical system that can monitor its own true coupling state in real time, automatically decouple the combined effects of multi-degree-of-freedom errors, distinguish between quasi-static temperature drift and dynamic disturbances, and has the ability to predict deviation trends and provide feedforward compensation. This is to solve the performance degradation problem caused by environmental disturbances in the long-term operation of existing systems and improve the stability, robustness and intelligent self-calibration level of the system. Summary of the Invention

[0004] This invention proposes a fiber array coupling optical system with self-calibration function, aiming to solve the problems of existing fiber array coupling systems lacking real-time self-calibration capability, being unable to distinguish between temperature drift and dynamic disturbance, being unable to predict the evolution trend of deviation, and being difficult to quickly and accurately compensate for multi-degree-of-freedom coupling errors. By constructing a closed-loop self-calibration architecture integrating wavefront sensing, inverse optical field desensitization, temperature desensitization, time-series prediction, and multi-axis parallel compensation, it achieves full optical direct monitoring of coupling state, decoupling and separation of multi-degree-of-freedom errors, quasi-static thermal drift elimination, dynamic deviation trend prediction, and high-precision autonomous alignment recovery.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] This invention provides a fiber array coupling optical system with self-calibration function. The system includes: a fiber array substrate, a coupling optical unit, a monitoring light source unit, a self-calibration unit, and an execution unit. The system is used in conjunction with a target photonic integrated circuit, which has an input waveguide port.

[0007] The fiber array substrate fixes and positions the input fiber array. The output end face of the input fiber array is optically polished and used to transmit the main optical signal.

[0008] A coupling optical unit is disposed on the emitting side of the fiber array substrate along the main optical path direction, including at least one focusing lens or lens group, which converges the main optical signal and couples it to the input waveguide port of the target photonic integrated circuit.

[0009] The monitoring light source unit is integrated inside or beside the fiber array substrate. It outputs a monitoring light signal with a wavelength band different from that of the main optical signal and injects the monitoring light signal into the main optical path in reverse isolation through wavelength division multiplexing coupling. After the monitoring light signal propagates along the main optical path to the input waveguide port of the target photonic integrated circuit, it is reflected back along the original path after Fresnel reflection at the end face of the port.

[0010] The self-calibration unit is located on the return path of the monitoring optical signal and is connected to the optical path of the fiber array substrate and the coupling optical unit. It extracts the pitch and yaw components from the returned monitoring optical signal. The unit integrates a wavefront sensing module, a state analysis module, a temperature desensitization module, a prediction module, and a control command generation module. An embedded temperature sensor is set inside the self-calibration unit or on the side of the system optical path. The sensor is electrically connected to the self-calibration unit and collects the current ambient temperature value.

[0011] The execution unit, electrically connected to the self-calibration unit, drives the focusing lens or lens group in the coupling optical unit to perform coordinated lateral translation and axial defocusing movements. In order to achieve multi-degree-of-freedom precision displacement compensation, the pitch and yaw components are equivalently canceled by lateral translation and axial defocusing when the system has no independent angle adjustment mechanism.

[0012] Furthermore, in the self-calibration unit:

[0013] The wavefront sensing module receives the monitoring optical signal reflected back from the input waveguide port of the target photonic integrated circuit, and generates a two-dimensional interference pattern containing the wavefront phase distribution based on the Hartmann-Shack wavefront sensing principle, and outputs the original wavefront sensing data.

[0014] The state analysis module, electrically connected to the wavefront sensing module, performs Zernike polynomial fitting and decoupling operations on the raw wavefront sensing data to extract the raw deviation state vector characterizing the relative spatial pose between the output end face of the current input fiber array and the input waveguide port of the target photonic integrated circuit. The raw deviation state vector includes axial defocus, lateral X-axis offset, lateral Y-axis offset, yaw angle component, and pitch angle component.

[0015] The temperature desensitization module is electrically connected to the state analysis module and the embedded temperature sensor. It receives the original deviation state vector and the current ambient temperature value collected by the embedded temperature sensor. This module has a built-in independent second-order polynomial temperature-drift response model for each deviation component. By fitting the quasi-static drift relationship between temperature and each deviation component, it removes the quasi-static drift component caused by temperature changes from the original deviation state vector component by component, and outputs the residual deviation state vector after temperature desensitization. The residual deviation state vector only represents the rapid dynamic deviation caused by mechanical vibration, packaging stress release, and random disturbance.

[0016] Furthermore, the temperature-drift response model pre-stored within the temperature desensitization module is a second-order polynomial response model independently constructed for each deviation component; the temperature drift estimates for axial defocus, lateral offset, yaw angle component, and pitch angle component are independently calculated based on the current ambient temperature value; the polynomial response model includes at least the second-order coefficient, first-order coefficient, and reference bias coefficient of the temperature term, accurately fitting the quasi-static drift component caused by temperature;

[0017] The prediction module, electrically connected to the temperature desensitization module, receives a continuous time-series residual deviation state vector sequence and, based on a pre-trained temporal convolutional network model, predicts the evolution trend of the residual deviation state vector within a preset time window, outputs the predicted residual trajectory, and provides a basis for feedforward compensation for the control system.

[0018] The control command generation module is electrically connected to the prediction module and the state analysis module. It fuses multi-source information such as the original deviation state vector, residual deviation state vector, predicted residual trajectory, and current ambient temperature value at the current moment. It constructs a cost function with the optimization objective of maximizing coupling efficiency, solves it using a rolling time-domain optimization algorithm based on model predictive control, and outputs a three-channel compensation control signal that optimizes the expected coupling efficiency. In the three-channel compensation control signal, the X-axis control signal and the Y-axis control signal contain high-frequency dynamic components for real-time dynamic compensation, and the axial control signal contains low-frequency quasi-static drift components for temperature drift compensation.

[0019] Furthermore, the execution unit includes a piezoelectric ceramic actuator array and a thermo-optical effect microlens actuator. The piezoelectric ceramic actuator array includes two independently driven piezoelectric ceramic actuators, which respond to the high-frequency dynamic components in the compensation control signal, respectively, and perform sub-millisecond-level lateral X-axis displacement and lateral Y-axis displacement compensation. The operating bandwidth of the piezoelectric ceramic actuator is higher than the closed-loop control bandwidth of the system. The piezoelectric ceramic actuator array has a built-in anti-aliasing filter and order reduction smoothing processing unit to suppress high-frequency resonance and oscillation interference. The thermo-optical effect microlens actuator responds to the low-frequency quasi-static drift components in the compensation control signal. By locally heating, it changes the refractive index distribution of a specific lens in the coupled optical unit, finely adjusts the equivalent focal length of the optical system, and achieves fine compensation for axial defocus. The pitch angle and yaw angle deviations are mapped to a combined control quantity of lateral translation and axial defocus through a control decoupling algorithm. The two types of actuators work together to achieve equivalent compensation for angle deviations.

[0020] Furthermore, the system also includes a host computer management platform that communicates with the self-calibration unit. The host computer management platform is deployed on a remote private cloud server, receiving and visualizing the original deviation state vector, residual deviation state vector, predicted residual trajectory, current ambient temperature value, and coupling efficiency data. At the same time, the host computer management platform defines the data labeling rules for calibration intervention periods and steady-state non-intervention periods. Using historically accumulated steady-state deviation state data and corresponding temperature data, it performs periodic incremental retraining and version updates on the temperature-drift multinomial models of each component in the temperature desensitization module and the temporal convolutional network model inside the prediction module. The updated model parameters are sent to the self-calibration unit through the nighttime maintenance window. The self-calibration unit uses a non-disruptive smooth switching mechanism to complete the model parameter replacement, avoiding transient disturbances during calibration.

[0021] Furthermore, the temperature-drift response model pre-stored within the temperature desensitization module is a second-order polynomial response model independently constructed for each deviation component; the temperature drift estimates for axial defocus, lateral offset, yaw angle component, and pitch angle component are independently calculated based on the current ambient temperature value; the polynomial response model includes at least the second-order coefficient, first-order coefficient, and reference bias coefficient of the temperature term, accurately fitting the quasi-static drift component caused by temperature.

[0022] Furthermore, the control command generation module contains a rolling time-domain optimization solver, which runs synchronously with the system's fixed control cycle. Each control cycle performs the following operations: acquiring the current original deviation state vector, residual deviation state vector, predicted residual trajectory for future steps, and current ambient temperature; embedding this information into the coupled optical system forward model within the module. The forward model is based on Gaussian beam propagation theory and aberration diffraction theory. The mode field radius, Rayleigh length, and angle deviation tolerance parameters in the model are fixed according to the system's optical structure and can be slightly modified during model retraining; the forward model predicts the expected coupling efficiency within a future time window given the current deviation state and compensation control signal; constructing a cost function, which includes a tracking loss term based on the expected coupling efficiency, a compensation control signal amplitude penalty term, and a compensation control signal change rate penalty term; using a constrained quadratic programming optimization algorithm to solve for the optimal compensation control signal sequence that minimizes the cost function, and taking its first step as the three-channel compensation control signal output for this cycle; since the angle deviation has no independent control channel, it is integrated into the three-channel control signal synchronously through a decoupling algorithm.

[0023] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0024] This invention achieves direct, fully optical monitoring of coupled states and real-time decoupling of multi-degree-of-freedom errors by constructing a closed-loop self-calibration architecture integrating wavefront sensing, inverse optical field desensitization, temperature desensitization, temporal prediction, and multi-axis parallel compensation. Compared with existing schemes that rely on mechanical displacement sensors or indirect optical power detection, this invention directly extracts multi-dimensional spatial pose errors such as axial defocus, lateral offset, yaw angle, and pitch angle from the wavefront phase of the monitored optical signal, avoiding the influence of nonlinear errors and mechanical backlash. Furthermore, it eliminates the need for an independent angle drive mechanism, effectively offsetting angle deviations through synergistic lateral and axial compensation. This simplifies the hardware structure, reduces costs, and overcomes the shortcomings of traditional single-degree-of-freedom polling adjustment, which suffers from slow convergence and oscillation.

[0025] This invention introduces a temperature desensitization module, which constructs a temperature-drift response model using independent second-order polynomials for each deviation component. This model removes low-frequency quasi-static drift components caused by ambient temperature changes from the original deviation state vector in real time, component by component. This design effectively solves the problem of distinguishing between thermal drift and dynamic disturbances such as mechanical vibration and stress release, ensuring that the residual deviation state vector only reflects rapid dynamic deviations. This significantly improves the prediction accuracy of the prediction module and the anti-interference capability of the closed-loop control.

[0026] This invention embeds a prediction module based on a temporal convolutional network model into the control loop. It uses historical residual deviation time series as input to predict future deviation evolution trajectories and incorporates this as feedforward information into the rolling temporal optimization solver. The system can predict impending offset trends in advance and achieve pre-compensation. Especially in scenarios involving sudden thermal shocks, mechanical vibrations, and other dynamic disturbances, it can output compensation control signals before the coupling efficiency significantly deteriorates, effectively suppressing instantaneous drops and significantly enhancing the system's active robustness and adaptability to different operating conditions.

[0027] This invention employs an execution unit architecture that combines a piezoelectric ceramic actuator array with a thermo-optical microlens actuator, with clear division of labor between high and low frequencies and complementary advantages. Meanwhile, the host computer management platform defines data labeling rules, incrementally retrains the model, and uses non-disruptive parameter switching, enabling the system to have self-evolution and self-optimization capabilities throughout its entire life cycle. It continuously adapts to characteristic drift caused by material aging and packaging creep, and maintains high-precision self-calibration performance in a long-term stable manner. Attached Figure Description

[0028] Figure 1 This is a system structure block diagram of a fiber array coupled optical system with self-calibration function proposed in this invention;

[0029] Figure 2 This is a schematic diagram of the data flow for residual deviation state prediction using the temporal convolutional network model in Embodiment 3 of the present invention. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] Example 1:

[0032] Please see Figure 1 This invention provides a fiber array coupling optical system with self-calibration function. The system includes: a fiber array substrate, a coupling optical unit, a monitoring light source unit, a self-calibration unit, and an execution unit. This embodiment aims to clearly illustrate the core components, optical path architecture, and basic workflow of the system.

[0033] The fiber optic array substrate is made of Kovar alloy, which has a coefficient of thermal expansion of 5.1 × 10⁻⁶. -6 / K, matching the thermal expansion coefficient of adjacent ceramic ferrules; a V-groove array is formed on the substrate by precision laser processing, with an angle error of less than ±0.5 degrees and a groove depth deviation of less than ±0.5 μm, and a row of input fiber arrays composed of 12 single-mode fibers is fixed with high precision; the output end face of the fiber array substrate is ground and polished, and its flatness is better than λ / 30 (λ is 1550 nm), and the surface is coated with a dual-band anti-reflection coating covering 1310 nm and 1550 nm; the input fiber array is used to transmit the main optical signal, which is a dense wavelength division multiplexing communication light in the C-band (1530-1565 nm);

[0034] The coupling optics unit is positioned on the output side of the fiber array substrate along the main optical path and specifically includes a collimating lens and an aspherical focusing objective. The collimating lens has a focal length of 8 mm and a numerical aperture of 0.25, converting the diverging light emitted from the fiber array into a collimated beam with a beam waist diameter of 1.2 mm. The aspherical focusing objective has a focal length of 2.75 mm and a numerical aperture of 0.55, focusing the collimated beam onto the input waveguide port end face of the target photonic integrated circuit. The aspherical focusing objective is mounted via a flexible hinged displacement stage, which has motion decoupling characteristics in the X and Y directions. Furthermore, a compensation lens made of SF11 optical glass with a refractive index temperature coefficient of -6 × 10⁻⁶ is embedded in the coupling optics unit. -6 / K, with a micro-thin film heater surrounding it, can independently control the temperature of the compensation lens and achieve fine compensation of axial defocus;

[0035] The monitoring light source unit is integrated into a micro-package on the side of the fiber array substrate, specifically a distributed feedback laser diode; its output wavelength is 1310 nm (isolated from the wavelength of the main C-band optical signal), the linewidth is less than 1 MHz, and the output optical power is stable at 1 mW; the monitoring optical signal output by the monitoring light source unit is reverse-coupled to the output end face of the fiber array substrate through a 1×2 fused tapered wavelength division multiplexer; the 1310 nm port of the wavelength division multiplexer is connected to the monitoring light source unit, the 1550 nm port is connected to the main communication light source, and the common end is connected to the fiber tail of the fiber array substrate;

[0036] The monitoring optical signal is emitted in the forward direction along the main optical path, passes through the coupling optical unit and illuminates the end face of the input waveguide port of the target photonic integrated circuit; due to the abrupt change in the refractive index of air and silicon waveguide, about 4% Fresnel reflection occurs, and the reflected monitoring optical signal is transmitted back along the original path;

[0037] The self-calibration unit completes optical path splitting and isolation through a four-port polarization-maintaining optical circulator and a wavelength division multiplexer: the first port of the optical circulator is connected to the monitoring light source to output forward light, the second port is connected to the 1310 nm port of the 1×2 wavelength division multiplexer to realize the reverse injection of monitoring light into the main optical path, the third port receives the back-reflected monitoring light and sends it to the wavefront sensing module, and the fourth port is used for monitoring light power equalization and crosstalk suppression; the optical circulator ensures unidirectional isolation of the forward injection and reverse return optical paths to avoid crosstalk between the monitoring light and the main communication light;

[0038] The self-calibration unit integrates a wavefront sensing module, a state analysis module, a temperature desensitization module, a prediction module, and a control command generation module. Each module is integrated on a heterogeneous computing board based on Zynq UltraScale+FPGA. The ARM core is responsible for process control, communication, and parameter scheduling, while the FPGA logic is responsible for the parallel pipeline real-time processing of wavefront sensing data.

[0039] The wavefront sensing module is specifically a Hartmann-Shack wavefront sensor based on a microlens array. The reflected monitoring light signal is amplified by a relay optical system with a magnification of 2× before entering the module. It is then divided into 32×32 sub-aperture spots by a microlens array with a spacing of 300 μm and a focal length of 15 mm. The focal plane detector is an InGaAs area array detector with a response band of 900-1700 nm, 640×512 pixels, and a pixel size of 15 μm. It is stably cooled to -10°C to reduce dark current noise. The module's FPGA logic calculates the offset between the centroid coordinates of the sub-aperture spots and the ideal reference coordinates in real time, reconstructs the wavefront phase distribution, and outputs the raw wavefront sensing data. The data update rate is 200 Hz.

[0040] The state analysis module is electrically connected to the wavefront sensing module and is implemented by a dedicated FPGA hard core logic and an ARM core. The module receives raw wavefront sensing data, performs fast fitting of Zernike polynomials based on Gram-Schmidt orthogonality, and decomposes the wavefront phase distribution into the first nine Zernike coefficients. Through a pre-calibrated calibration decoupling matrix (introducing standard displacement and tilt at the reference temperature and recording the changes in Zernike coefficients to complete the calibration), the changes in Zernike coefficients are linearly mapped to the changes in the original deviation state vector. The final output of the original deviation state vector contains five components: lateral X-axis offset (ΔX), lateral Y-axis offset (ΔY), axial defocus (ΔZ), pitch angle around the X-axis (Δθx), and yaw angle around the Y-axis (Δθy).

[0041] The execution unit and the self-calibration unit are electrically connected to a control command generation module, specifically including a dual-channel piezoelectric ceramic actuator array and a thermo-optical effect microlens actuator. The dual-channel piezoelectric ceramic actuator array responds to the high-frequency dynamic components in the X and Y directions of the compensation control signal, performing sub-millisecond-level lateral translation compensation to drive the aspherical focusing objective of the coupling optical unit to move in the X and Y directions, offsetting the lateral X-axis and lateral Y-axis offset deviations. The driving voltage range is -120 V to +120 V, with a hardware output bandwidth of 2 kHz. The piezoelectric ceramic actuator array incorporates an anti-aliasing filter and order reduction smoothing processing unit, matched with a 200 Hz closed-loop control cycle, suppressing high-frequency resonance. The thermo-optical effect microlens actuator responds to the low-frequency quasi-static drift component in the axial control signal, finely adjusting the refractive index by changing the local temperature of the compensation lens, and altering the equivalent back focal length of the coupling optical unit to achieve fine compensation for axial defocus deviation, with a compensation range of ±5 μm and a resolution better than 10. nm; The pitch and yaw components have no independent angle drive mechanism. The control system uses a decoupling algorithm to map the angle deviation into a combination of X and Y lateral translation and axial defocus control signals. The pose alignment is achieved by the coordinated compensation of the two types of actuators.

[0042] Example 2:

[0043] Based on Example 1, this embodiment defines the specific structure, model form, and operating logic of the temperature desensitization module and the prediction module, and also fully defines the closed-loop self-calibration standard process of the self-calibration unit.

[0044] The temperature desensitization module pre-stores a temperature-drift response model, which consists of five independent second-order polynomial response models for each of the five deviation components. The model is constructed as follows: under the control of a host computer management platform, the system is placed in a temperature-controlled chamber, and temperature cycling is performed in 2°C steps within the range of -10°C to +60°C. At each steady-state temperature point, the original deviation state vector output by the state analysis module is collected. Under steady-state conditions without external disturbances, the original deviation only reflects the quasi-static temperature drift. After binning and averaging the data, a quadratic function is independently fitted to each deviation component using the least squares method.

[0045] Taking the axial defocusing temperature drift model as an example:

[0046] ;

[0047] in, This is an estimated value of axial defocusing caused solely by temperature. This is the current ambient temperature value; Reference temperature (take) ); and These are the second-order and first-order temperature coefficients, respectively. The residual defocusing amount is offset at the reference temperature; the lateral offset, pitch angle, and yaw angle components all adopt the same form of second-order polynomial, only replacing the corresponding independent fitting coefficients;

[0048] When the system is working normally, the temperature desensitization module reads the real-time temperature of the high-precision embedded temperature sensor at ±0.1-0.1°C next to the fiber array substrate, and independently calculates the estimated temperature drift values ​​for the axial defocusing, lateral X-axis offset, lateral Y-axis offset, yaw angle component, and pitch angle component. The temperature drift component is subtracted from the original deviation state vector component by component to obtain the residual deviation state vector. The residual deviation only includes the dynamic deviation caused by mechanical vibration, stress release, and random disturbance, thus achieving effective separation of thermal drift and dynamic disturbance.

[0049] The prediction module receives the time-series sequence of continuous residual deviation state vectors output by the temperature desensitization module. Internally, it integrates a temporal convolutional network (TCN) model, which, based on dilated convolutional structures, predicts the residual trajectory for the next 10 time steps based on the residual deviation sequence of the past 40 time steps. The system sampling rate is 200 Hz, meaning the input tensor dimension for the past 0.2 seconds is 40×5, and the output predicted trajectory tensor dimension for the next 0.05 seconds is 10×5. The specific structure of this network includes dilated convolutional block parameters, fully connected layer configuration, and Dropout strategy; see Example 3 and its appendix for details. Figure 2 Description;

[0050] The control command generation module operates with a fixed 5 ms control cycle (200 Hz), which is strictly matched with the wavefront sensing sampling frequency; it integrates a rolling time-domain optimization solver, which executes the following every cycle: obtain the current original deviation state vector, the current residual deviation state vector, the predicted residual trajectory for the next 10 steps, and the current ambient temperature;

[0051] The input coupling optical system is a forward model, which is based on the theory of Gaussian beam propagation and aberration diffraction. The mode field radius, Rayleigh length, and angular deviation tolerance are the system calibration parameters.

[0052] The formula for predicting the expected coupling efficiency is as follows:

[0053] ;

[0054] in, For the future Predicted value of expected coupling efficiency at each prediction step size; To predict the timing step number, This corresponds to the next 10 consecutive control cycles; This represents the maximum theoretical coupling efficiency under ideal alignment conditions. Based on the natural constant An exponential function with base 0; The k-th prediction step is equivalent to the horizontal Towards the overall offset; No. Prediction step equivalent lateral Towards the overall offset; , To focus the light spot on the target waveguide end face , Radius of the model field; No. Prediction step equivalent comprehensive angle deviation; This refers to the angular deviation tolerance parameter. No. Predicting the equivalent integrated axial defocusing amount in the step; The Rayleigh length of the Gaussian beam;

[0055] Based on this positive model, the control command generation module constructs the following cost function:

[0056] ;

[0057] in, To minimize the objective function, For the first Prediction step time series weighting coefficients; This represents the current control voltage amplitude for channel m. To control the amplitude penalty coefficient; This represents the change in control voltage between adjacent cycles. To control the rate of change penalty coefficient; To control the channel sequence number, The control signals correspond sequentially to the X-axis, Y-axis, and axial control channels; angle deviations are incorporated into the three-channel optimization solution.

[0058] The solver uses an effective set quadratic programming algorithm with Box constraints to solve for the optimal control sequence, and takes the first step as the compensation control signal output for the current cycle.

[0059] The self-calibration unit performs a complete self-calibration process, which includes the following steps:

[0060] Step S1: The wavefront sensing module receives the reflected monitoring light at a frequency of 200 Hz and generates raw wavefront sensing data in real time.

[0061] Step S2: The state analysis module performs Zernike polynomial fitting and decoupling frame by frame, and outputs the current original deviation state vector;

[0062] Step S3: The temperature desensitization module reads the real-time temperature, calculates the temperature drift component through an independent second-order polynomial model for each component, removes the quasi-static drift component from the original deviation, and outputs the residual deviation state vector.

[0063] Step S4: The prediction module concatenates the residual sequences of 40 historical time steps, inputs them into the temporal convolutional network, and outputs the predicted residual trajectory for the next 10 steps.

[0064] Step S5: The control command generation module integrates multi-source information and outputs a three-channel compensated control signal through rolling time-domain optimization.

[0065] Step S6: The execution unit receives the control signal, the piezoelectric ceramic performs high-frequency dynamic component compensation in the X / Y direction, the thermo-optical microlens performs low-frequency quasi-static drift component compensation in the axial direction, and the angle deviation is compensated by a decoupling algorithm.

[0066] Step S7: Repeat S1~S6 to form a closed loop until the magnitude of the original deviation state vector converges to the preset dead zone threshold: axial defocus <15 nm, lateral offset <8 nm, angular deviation <0.3 μrad; continuously and stably maintain for 50 control cycles (0.25 s), determine that the self-calibration is complete and lock the working point.

[0067] Example 3:

[0068] Please see Figure 2 Based on Example 2, this embodiment defines the collaborative mechanism of the host computer management platform, the model retraining rules, and the logic for switching parameters without disturbance. It also fully describes the system response process under sudden thermal shock conditions and verifies robustness and predictive compensation capabilities.

[0069] This system is equipped with a host computer management platform, deployed on an enterprise private cloud server, and communicates bidirectionally and asynchronously with each field self-calibration unit via the MQTT protocol; the platform receives and visualizes raw deviation, residual deviation, predicted trajectory, temperature, coupling efficiency, and control command data in real time;

[0070] The platform maintains a historical operating condition database and a model parameter library, and also has built-in data time period marking rules: human intervention calibration and power-on initial calibration are defined as calibration intervention periods, and the steady-state autonomous closed-loop operation period of the system is defined as non-intervention periods. Only valid data from non-intervention periods are used for model retraining.

[0071] The platform performs incremental retraining every 24 hours.

[0072] (1) Extract historical data from the non-intervention period of the past week, and use the recursive least squares method with forgetting factor to update the second-order polynomial coefficients of each component of the temperature desensitization module to adapt to the slow drift of long-term creep and thermal expansion characteristics of the packaging material.

[0073] (2) Fix the updated temperature model, extract the residual time series and the subsequent true value of the deviation, and use the early stopping strategy to perform 5 rounds of fine-tuning training on the temporal convolutional network to suppress overfitting.

[0074] (3) During the nighttime maintenance window, the updated model parameters are encrypted and sent to their respective calibration units. The units adopt smooth parameter transition and seamless switching, and gradually replace the model weights and coefficients frame by frame to avoid parameter mutations that cause optical path alignment jumps and coupling efficiency fluctuations.

[0075] The data flow of the Temporal Convolutional Network (TCN) model is as follows: The model input is a 40×5 residual bias history sequence, containing 40 historical time steps, each with 5 bias components. The input data is first enlarged by convolutional layers, expanding the number of feature channels to 64. The data then flows through four dilated convolutional blocks, with dilation rates of 1, 2, 4, and 8 respectively. Each dilated convolutional block uses a 3-kernel one-dimensional convolution, and performs weight normalization, ReLU activation function operation, and residual connection processing sequentially. After convolutional feature extraction, the temporal dimension is compressed by a global average pooling layer, followed by two fully connected layers with 128 and 64 nodes respectively. To suppress overfitting, a Dropout layer is placed between the two fully connected layers with a dropout rate of 0.3. Finally, a 10×5 residual prediction trajectory is output through a linear output layer, enabling prediction of the next 10 time steps and 5 bias components.

[0076] System response to sudden thermal shock (start-up of high-power surrounding equipment, ambient temperature rises by 12°C within 10 seconds):

[0077] Thermal shock triggers a rapid rise in the temperature sensor reading, and the original axial defocusing amount shows a significant positive quasi-static drift. The temperature desensitization module calculates the synchronously growing temperature drift estimate in real time through a second-order polynomial model, accurately removes the thermal drift component from the original deviation, keeps the residual deviation within a very small range, and achieves effective isolation of thermal drift.

[0078] The prediction module captures the weak and slow growth trend of the residual sequence, and the temporal convolutional network identifies the evolution law of the deviation caused by the heat conduction lag, predicts the continuous upward trend of the residual in the next 0.05 s and outputs the predicted trajectory.

[0079] The control command generation module predicts that the coupling efficiency will soon drop in the cost function and outputs the axial pre-compensation control signal in advance; the thermo-optical microlens driver synchronously adjusts the heating power and corrects the lens refractive index and system focal length in advance to counteract the upcoming thermal defocusing drift.

[0080] The small X / Y lateral drift caused by the temperature gradient is synchronously compensated for by the piezoelectric ceramic actuator under the action of prediction feedforward; the main optical signal is uninterrupted throughout the entire process, and the closed-loop self-calibration continues to operate.

[0081] The host computer platform automatically labels the samples of this thermal shock condition and stores them in the historical database for use in the next round of model retraining sample enhancement, continuously improving the model's ability to adapt to extreme temperature disturbances.

[0082] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A fiber array coupled optical system with self-calibration function, characterized in that, include: Fiber optic array substrate, coupling optical unit, monitoring light source unit, self-calibration unit, and execution unit; The system is used in conjunction with a target photonic integrated circuit, which has an input waveguide port; The fiber array substrate fixes and positions the input fiber array, and its output end face is optically polished for transmitting the main optical signal. A coupling optical unit, located on the emitting side of the fiber array substrate, includes at least one focusing lens or lens group, which converges and couples the main optical signal to the input waveguide port of the target photonic integrated circuit. The monitoring light source unit is integrated inside or beside the fiber array substrate. It outputs a monitoring light signal with a different wavelength band than the main optical signal. It is injected into the main optical path in reverse through wavelength division multiplexing coupling. The monitoring light signal propagates along the main optical path to the input waveguide port of the target photonic integrated circuit. After being reflected by the end face of the port, it is transmitted back along the original path. The self-calibration unit is located in the return path of the monitoring optical signal and is connected to the optical path of the fiber array substrate and the coupling optical unit. It extracts the pitch angle component and yaw angle component from the returned monitoring optical signal and is equipped with an embedded temperature sensor to collect the current ambient temperature value. The execution unit is electrically connected to the self-calibration unit. This unit drives the focusing lens or lens group in the coupling optical unit to perform coordinated lateral translation and axial defocusing. In order to achieve multi-degree-of-freedom precision displacement compensation, the pitch and yaw components are equivalently canceled by lateral translation and axial defocusing under the condition that the system has no independent angle adjustment mechanism. The self-calibration unit includes: The wavefront sensing module receives the monitoring optical signal reflected back from the input waveguide port of the target photonic integrated circuit, and outputs the raw wavefront sensing data based on the Hartmann-Shack wavefront sensing principle. The state analysis module performs Zernike polynomial fitting and decoupling operations on the raw wavefront sensing data to extract the raw deviation state vector, including axial defocus, lateral X-axis offset, lateral Y-axis offset, yaw angle component and pitch angle component. The temperature desensitization module has a built-in second-order polynomial temperature-drift response model for each deviation component. It removes the quasi-static drift component caused by temperature change from the original deviation state vector component by component and outputs the residual deviation state vector. The prediction module receives the residual deviation state vector sequence and, based on a pre-trained temporal convolutional network model, predicts the evolution trend of the residual deviation state vector within a preset time window, outputting the predicted residual trajectory.

2. The system according to claim 1, characterized in that, The self-calibration unit also includes a control command generation module, which fuses the original deviation state vector, residual deviation state vector, predicted residual trajectory, and current ambient temperature value at the current moment; constructs a cost function with maximizing coupling efficiency as the optimization objective, solves it using a rolling time-domain optimization algorithm based on model predictive control, and outputs a three-channel compensation control signal that optimizes the expected coupling efficiency; in the three-channel compensation control signal, the X-axis control signal and the Y-axis control signal contain high-frequency dynamic components for real-time dynamic compensation, and the axial control signal contains low-frequency quasi-static drift components for temperature drift compensation; The execution unit includes: a piezoelectric ceramic actuator array and a thermo-optical microlens actuator; The piezoelectric ceramic actuator array includes two independently driven piezoelectric ceramic actuators, which respond to high-frequency dynamic components respectively and perform lateral X-axis displacement and lateral Y-axis displacement compensation with sub-millisecond response. The piezoelectric ceramic actuator array has built-in anti-aliasing filtering and order reduction smoothing processing units to suppress high-frequency resonance and oscillation interference. The thermo-optical effect microlens driver responds to low-frequency quasi-static drift components and changes the refractive index distribution of specific lenses in the coupled optical unit by local heating, thereby fine-tuning the equivalent focal length of the optical system and achieving precise compensation for axial defocus.

3. The system according to claim 1, characterized in that, It also includes a host computer management platform that communicates with the self-calibration unit; The host computer management platform receives and visualizes the original deviation state vector, residual deviation state vector, predicted residual trajectory, and current ambient temperature value. Simultaneously, the host computer management platform defines data labeling rules for calibration intervention periods and steady-state non-intervention periods. Utilizing historically accumulated steady-state deviation state data and corresponding temperature data, it performs periodic incremental retraining and version updates on the second-order polynomial temperature-drift response model and the temporal convolutional network model for each deviation component. The updated model parameters are distributed to the self-calibration unit through the nighttime maintenance window, and the self-calibration unit uses a non-disruptive smooth switching mechanism to complete the model parameter replacement.

4. The system according to claim 1, characterized in that, The temperature desensitization module internally stores independent second-order polynomial temperature-drift response models for each deviation component. Based on the current ambient temperature, the temperature desensitization module independently calculates the temperature drift estimates for axial defocus, lateral X-axis offset, lateral Y-axis offset, yaw angle component, and pitch angle component. The second-order polynomial temperature-drift response model includes at least the second-order coefficients, first-order coefficients, and reference bias coefficients for temperature, used to fit the quasi-static drift components caused by temperature.

5. The system according to claim 2, characterized in that, The control command generation module contains a rolling time-domain optimization solver, which runs synchronously with the system's fixed control cycle, performing the following operations in each control cycle: Obtain the original deviation state vector, residual deviation state vector, predicted residual trajectory for multiple future steps, and current ambient temperature value at the current moment; The acquired original deviation state vector, residual deviation state vector, predicted residual trajectory for future multiple steps, and current ambient temperature value are embedded into the forward model of the coupled optical system within the module. This model is based on Gaussian beam propagation theory and aberration diffraction theory. The mode field radius, Rayleigh length, and angular deviation tolerance parameters in the model are fixed and calibrated according to the system's optical structure and can be slightly corrected during model retraining. The forward model predicts the expected coupling efficiency within a future time window given the current deviation state and compensation control signal. Construct a cost function, which includes a tracking loss term based on the expected coupling efficiency, a compensation control signal amplitude penalty term, and a compensation control signal change rate penalty term; A constrained quadratic programming optimization algorithm is used to find the optimal compensation control signal sequence that minimizes the cost function, and the first step of the sequence is taken as the three-channel compensation control signal output in this cycle.

Images