Optical module coupling positioning intelligent calibration method based on ai and production big data

Abstract

The application discloses an AI and production big data-based intelligent calibration method for optical module coupling positioning, and relates to the technical fields of artificial intelligence and optical communication devices; by acquiring the batch identification and the environment temperature of the optical module to be coupled, the historical production big data base is used to extract the tolerance distribution of the same batch, and a batch initial bias compensation vector is generated; the geometric centroid is acquired by performing semantic segmentation on the end face image, and high-precision coarse positioning is realized in combination with the compensation vector; a multi-dimensional state feature vector fusing real-time optical power, power gradient and six-degree-of-freedom coordinates is constructed, a non-fixed step instruction is output through a reinforcement learning model, intelligent fine movement optimization is performed until the target threshold is reached; the coordinate drift slope in continuous production is monitored in real time, and the compensation vector is updated in an online closed loop; the application realizes adaptive batch tolerance, dynamic compensation of thermal drift and efficient intelligent optimization, and significantly improves the coupling efficiency and yield of the optical module.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and optical communication device technology, and specifically relates to an intelligent calibration method for optical module coupling and positioning based on AI and production big data. Background Technology

[0002] In the field of optical communication device manufacturing technology, optical modules are the core components for realizing photoelectric signal conversion. With the evolution of 5G / 6G networks and computing centers towards 400G, 800G, and even higher bandwidths, the coupling precision requirements between the optical chips and optical fibers or lenses within optical modules have entered the sub-micron level. Coupling efficiency directly determines the insertion loss, transmission distance, and final product yield of the optical module, making it the most critical and time-consuming step in the optical module manufacturing process.

[0003] Currently, existing optical module coupling positioning mainly relies on machine vision coarse positioning combined with mechanical fixed-path blind search, such as hill climbing and spiral search methods. In specific application scenarios, the system first uses a camera to identify the contours of the chip and optical fiber for initial alignment. Then, the displacement stage moves point by point according to a preset fixed step size and fixed trajectory, and the maximum power point is found by using the real-time feedback value from the optical power meter. This method is feasible for dealing with low-speed, high-tolerance optical modules. However, with the increase in packaging density of high-frequency optical modules, materials have inherent systematic geometric deviations under different wafer batches and different cutting processes. At the same time, continuous production in industrial settings can cause equipment motors to heat up and jigs to wear, resulting in micron-level mechanical thermal drift. Traditional visual alignment cannot identify these implicit physical deviations, resulting in the coarse positioning point being far from the true optical peak point.

[0004] This shows that the existing technology has the following problems: 1. Fixed step size search is inefficient and is prone to getting stuck in local optima or oscillating around the peak value; 2. It lacks the ability to dynamically compensate for production batch tolerances and equipment thermal drift, which causes the coupling time to gradually increase with production time, requiring frequent shutdowns for manual calibration, which seriously restricts the mass production efficiency and yield of high-end optical modules. Summary of the Invention

[0005] (a) Technical problems to be solved To address the problems in related technologies, this invention provides an intelligent calibration method for optical module coupling positioning based on AI and production big data, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] (II) Technical Solution To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: S1. Input the production batch identifier of the current optical module to be coupled and the ambient temperature data into the historical production database for cluster matching, extract the distribution pattern of the historical best coupling coordinates of the same batch, and generate the initial bias compensation vector of the batch accordingly. S2. Perform edge enhancement and feature extraction on the initial end face image stream of the current optical module to be coupled, and locate the initial geometric centroid coordinates of the light-emitting area of ​​the optical chip and the receiving area of ​​the optical fiber. The initial geometric centroid coordinates and the initial batch offset compensation vector are added together to generate coarse positioning target coordinates, and the six-degree-of-freedom high-precision displacement stage is driven to move to the coarse positioning target coordinates. S3. At the coarsely located target coordinates, collect real-time optical power signals and real-time three-dimensional spatial coordinates, perform first-order differential processing on the real-time optical power signals to obtain the optical power change gradient; align the real-time optical power signals, optical power change gradients and real-time three-dimensional spatial coordinates with timestamps to construct a multi-dimensional coupled state feature vector. S4. Input the multidimensional coupling state feature vector into the reinforcement learning optimization model in real time to obtain a non-fixed step displacement increment command, drive the displacement stage to perform micro-displacement detection, and calculate the action reward value based on the change in optical power before and after displacement; repeat this step until the current optical power exceeds the target threshold, and lock the current coordinates as the optimal coupling coordinates; S5. Extract the optimal coupling coordinates of the most recent consecutive preset number of optical modules, calculate the coordinate drift slope, and when the coordinate drift slope exceeds the mechanical thermal drift safety threshold, update the batch initial offset compensation vector in the historical production database using the coordinate drift slope. By integrating historical production big data with real-time environmental information to generate batch offset compensation vectors, and using these vectors to correct visual coarse positioning coordinates, the unknown search space is compressed into an extremely small high-probability interval, effectively eliminating material batch tolerances and equipment systematic deviations. By introducing a deep reinforcement learning model to replace the traditional fixed-step search, and by perceiving multi-dimensional coupled state characteristics, non-fixed-step displacement commands are output, achieving intelligent, fast, and accurate optimization from the middle to the peak, solving the problems of low efficiency, easy getting trapped in local optima, and frequent shutdowns for calibration in existing technologies. Through monitoring and automatic updating of the drift slope of continuous production data, the system has the ability to continuously optimize, dynamically offsetting equipment thermal drift, and ensuring long-term production line stability and high yield.

[0007] Preferably, step S1 includes the following steps: S11. Synchronously acquire the production batch identifier, ambient temperature data, and initial end-face image stream of the current optical module to be coupled; S12. In the historical production database, select historical successful coupling records that have the same material supplier code and the same production batch identifier as the current optical module to be coupled, and form a batch sample set; if there are no historical successful coupling records with the same material supplier code and the same production batch identifier, select the historical successful coupling record with the highest similarity as the batch sample set. Extract the initial geometric centroid coordinates and the final optimal coupling coordinates of all samples in the batch sample set, and calculate the three-dimensional spatial difference vector between the two. S13. After removing outlier data from the three-dimensional spatial difference vector that deviate from the average by more than three standard deviations, assign increasing weights to the remaining three-dimensional spatial difference vector according to the time distance, and calculate the weighted average vector. S14. The weighted average vector is calibrated by combining the current ambient temperature data with the temperature difference compensation coefficient of the standard room temperature, and the initial bias compensation vector of the batch is output. The temperature difference compensation coefficient is obtained by multiplying the preset thermal expansion coefficient by the temperature difference. By extracting the inherent spatial deviation pattern between the visual geometric center and the optical optimal point from the historical data of the same batch, and introducing environmental temperature compensation, a high-precision batch initial bias compensation vector is generated. This vector with prior knowledge can correct the physical geometric center of the optical chip and optical fiber to a position close to the actual optical peak after the vision system identifies the physical geometric center of the optical chip and optical fiber. This changes the mode of purely blind search for unknown materials and provides a precise data foundation for subsequent coarse positioning.

[0008] Preferably, the similarity calculation in S12 includes the following steps: S121. Extract the physical dimensions of the optical chip, lens focal length, nominal wavelength and thermal expansion coefficient of the substrate material from the historical records, and construct a normalized historical material feature vector. S122. Obtain the physical dimensions of the optical chip, lens focal length, nominal wavelength, and thermal expansion coefficient of the substrate material of the current optical module to be coupled, and construct the current material feature vector. S123. The Euclidean distance method is used to calculate the spatial distance between the current material feature vector and each historical material feature vector. The record with the smallest distance is taken as the historical coupling success record with the highest similarity. When there are no identical batch records in the database, the system does not simply use the global average value. Instead, it constructs a feature vector and performs Euclidean distance matching by quantifying the physical parameters of the material, such as chip size, lens focal length, and coefficient of thermal expansion. This enables the system to intelligently retrieve the historical batch data with the closest physical characteristics as a reference, ensuring the physical rationality of the offset compensation benchmark and improving the initial positioning accuracy under cold start or new material batches.

[0009] Preferably, the initial centroid coordinates for positioning geometry in S2 include the following steps: S21. Obtain the original grayscale image of the optical module end face through a coaxial infrared CCD camera, and apply adaptive histogram equalization to enhance the contrast between the light-emitting point of the optical chip waveguide and the fiber core. S22. Use the final lightweight convolutional neural network to perform semantic segmentation on the enhanced image and output a binary mask image containing the waveguide region and the fiber core region. S23. Calculate the pixel geometric center of the waveguide region and the fiber core region in the binarized mask image, and convert the pixel coordinates into physical space absolute coordinates through the camera calibration matrix to obtain the initial geometric centroid coordinates. By deeply integrating big data intelligent compensation with real-time visual positioning, and through adaptive image enhancement and lightweight neural networks, the waveguide and fiber core are accurately segmented. The nine-point calibration method is used to achieve accurate coordinate transformation from pixels to physical space. The calculated geometric centroid coordinates are vector-summed with the big data offset vector of S1, enabling the mechanical displacement stage to directly cross the blind search area and quickly reach the high-probability starting point. This significantly reduces the travel range for subsequent AI micro-motion optimization and significantly improves overall efficiency.

[0010] Preferably, the construction of the multidimensional coupled state feature vector in S3 includes the following steps: S31. The voltage signal output by the photodetector is acquired using a high-speed analog-to-digital converter, and high-frequency circuit noise is eliminated by a moving average filtering algorithm to obtain the real-time optical power signal. S32. Read the real-time feedback data of the grating ruler inside the six-degree-of-freedom high-precision displacement stage to obtain real-time three-dimensional spatial coordinates containing three translational dimensions and three rotational dimensions. S33. Calculate the difference between the optical power in the current sampling period and the optical power in the previous sampling period, divide it by the sampling time interval, and obtain the optical power change gradient. S34. Using the hardware clock of the displacement stage controller as a reference, the real-time optical power signal, optical power change gradient and real-time three-dimensional spatial coordinates within the same time window are spliced ​​together to generate a multi-dimensional coupled state feature vector. By synchronously acquiring and highly accurate spatiotemporal alignment, spatial location information, optical power energy information, and their changing trends are fused into a complete state vector. This multidimensional feature description enables the reinforcement learning model to not only perceive location but also energy intensity and the trend of energy change, providing comprehensive and real-time sensory input for making the correct next decision, which is the foundation of intelligent optimization.

[0011] Preferably, the construction and training of the deep reinforcement learning optimization model in S4 includes the following steps: S411. Construct an initial deep reinforcement learning optimization model that includes an Actor network and a Critic network; obtain multi-dimensional state data sequences and corresponding displacement operation sequences from historical manual coupling processes, and construct a demonstration dataset; S412. Supervised learning pre-training of the Actor network is performed using the demonstration dataset to make the initial output behavior of the Actor network approximate the human optimization trajectory, thus obtaining the pre-trained Actor network. S413. Build a physical simulation environment coupled with optical modules, and connect the pre-trained Actor network and Critic network into the physical simulation environment for reinforcement learning training. S414. Set the training reward function; multiply the increase in optical power compared to the previous moment by the positive gain coefficient as the positive reward, multiply the time consumption of a single displacement by the time penalty coefficient as the negative reward, and give a finite negative reward value if a physical collision occurs. S415. When the average cumulative reward value of the model in the maximum number of rounds in the physical simulation environment converges and no longer increases, stop training, solidify the current network parameters, and obtain the pre-trained deep reinforcement learning optimization model. By using expert data to pre-train the Actor network through supervised learning, the model's initial behavior closely approximates the optimal human trajectory, avoiding the risk of collisions caused by random exploration in the early training stages. In the physical simulation environment, the reward function guides the model to learn how to maximize light power while minimizing the number of steps and time consumption, and actively avoid dangerous areas. This pre-training combined with reinforcement learning mode balances safety, learning efficiency, and the optimality of the optimization strategy, solving the problem of high risk and low efficiency in the zero-to-one exploration of reinforcement learning models in precision manufacturing.

[0012] Preferably, the training reward function in S414 includes: a positive gain coefficient, a time penalty coefficient, and a fixed negative reward; The ratio between the positive gain coefficient and the time penalty coefficient is determined based on historical test data to ensure that the model prioritizes maximizing optical power while tending to select the path with the fewest steps. The fixed negative reward is used to force the model to avoid dangerous areas that exceed mechanical limits or cause hard contact between the optical fiber and the chip when outputting displacement increment commands. Positive gain encourages the model to find the point of maximum power, time penalty prompts it to choose the shortest path, and huge negative reward for collision sets an absolute safety red line. The combination of these three factors forces the model to find the optimal coupling path in a complex multidimensional space that can rapidly increase the light power to the target, fully guarantee the physical safety of the precision device, and has the shortest time consumption.

[0013] Preferably, the loop execution mechanism in S4 includes the following steps: S421. After receiving a non-fixed step displacement increment command, the displacement stage performs physical movement and triggers a synchronization signal after moving into position. S422. After receiving the synchronization signal, the system re-executes S3 to generate a new multidimensional coupled state feature vector. S423. Determine whether the current real-time optical power signal is greater than or equal to the target threshold. If not, input the new multidimensional coupled state feature vector into the deep reinforcement learning optimization model and continue to output the next instruction. S424. If yes, start the stabilization timer. If the optical power does not drop below the target threshold within the continuous detection period, the coupling is considered successful, the current coordinate is locked as the optimal coupling coordinate, and the loop is terminated. By synchronizing the signal and re-acquiring the state after displacement execution, a real-time closed loop between AI decision-making and changes in the physical world is ensured. The stable timer and continuous drop-free verification mechanism introduced after the target threshold is reached can effectively filter out false peaks caused by accidental factors such as platform micro-vibration and instantaneous changes in electrical signals. Only the optical power high point that is verified to be continuously stable will be finally locked as the optimal coupling coordinate, thus improving the reliability and consistency of the coupling results.

[0014] Preferably, triggering the global compensation mechanism in S5 includes the following steps: S51. Set the sliding window length and extract the optimal coupling coordinates of the most recent consecutive optical modules in the same batch; S52. Extract the X-axis, Y-axis and Z-axis data of the optimal coupled coordinates in S51 respectively. Use the least squares method to perform linear fitting on the data of each axis as the production serial number changes, and extract the slope of the fitted line of each axis as the coordinate drift slope. S53. Set a mechanical thermal drift safety threshold; if the absolute value of the coordinate drift slope of any axis is greater than the mechanical thermal drift safety threshold, it is determined that the production equipment has undergone thermal deformation or the jig has worn. S54. Multiply the coordinate drift slope that exceeds the mechanical thermal drift safety threshold by the sliding window length as the drift compensation amount, and add the drift compensation amount to the batch initial bias compensation vector corresponding to the current batch in the historical production database to complete the online update; By continuously monitoring the optimal coupling coordinate sequence of continuous production, the system can keenly detect micron-level trend drift caused by equipment thermal deformation or jig wear. Once the drift is determined to exceed the limit, the system automatically calculates the compensation amount and updates the batch offset compensation vector, so that the coarse positioning of subsequent modules can follow the physical state changes of the equipment in real time. This ensures that the initial positioning accuracy is always kept at the optimal level in large-scale uninterrupted production, and curbs the problem of coupling time deteriorating with the increase of production time from the source. It gives the production line the intelligent ability of self-sensing and self-evolution, changing the passive mode of relying on manual experience judgment or periodic shutdown calibration.

[0015] Preferably, an intelligent calibration system for optical module coupling positioning based on AI and production big data is provided to implement the intelligent calibration method for optical module coupling positioning based on AI and production big data. The system includes: The data perception and bias generation module is used to simultaneously acquire the optical module production identifier and image stream, and retrieve the historical database to generate the batch initial bias compensation vector; The visual calculation and coarse localization module receives the image stream and the batch initial offset compensation vector, calculates the geometric center and drives the displacement stage to the coarse localization target coordinates; The real-time state feature construction module synchronously collects high-frequency optical power and spatial coordinates after coarse localization to construct a multi-dimensional coupled state feature vector. The reinforcement learning micro-motion optimization module takes the multi-dimensional coupled state feature vector as input, uses the Actor-Critic network architecture to output continuous displacement commands, and searches for the optimal coupled coordinates in a closed loop. The drift monitoring and self-evolution module monitors the coordinate drift slope during continuous production and dynamically updates the batch initial offset compensation vector when the mechanical thermal drift safety threshold is reached.

[0016] (III) Beneficial Effects The present invention has the following beneficial effects: This invention improves the accuracy of initial positioning and the environmental robustness of the system by integrating production big data and a dynamic compensation mechanism. By extracting the distribution pattern of the historical best coupled coordinates of the same batch and combining the ambient temperature and the coordinate drift slope during equipment operation, the system can automatically generate and update the initial bias compensation vector of the batch in real time. It breaks through the static limitations of traditional visual positioning, effectively eliminates the systematic deviation caused by material batch tolerances and equipment thermal deformation, enables coarse positioning to cross the blind search area, and significantly shortens the subsequent optimization path.

[0017] This invention introduces a deep reinforcement learning algorithm to replace the traditional mechanical fixed path search, achieving efficient and accurate optimization in complex multidimensional spaces. By constructing a multidimensional coupled state feature vector containing optical power, power change gradient, and six-degree-of-freedom spatial coordinates, and using an Actor-Critic network architecture to output continuous displacement commands with non-fixed step sizes, the system can quickly approximate with large step sizes when the power gradient is large, and finely adjust with micro-step sizes when approaching the peak. This biomimetic intelligent optimization method solves the problem of traditional hill-climbing methods easily getting trapped in local optima and oscillating near the peak, greatly improving the efficiency and success rate of high-precision coupling.

[0018] This invention constructs a comprehensive safety monitoring and feature decoupling system to ensure high reliability of automated production; it improves the visual recognition anti-interference capability in complex reflective environments by using a lightweight convolutional neural network to perform adaptive histogram equalization and semantic segmentation on end face images; and it introduces a strict negative reward penalty based on physical collision critical points into the reinforcement learning reward function to ensure that the displacement stage will not suffer damage to piezoelectric ceramic impacts or optical fiber hard contacts during the pursuit of extreme speed optimization, thus balancing production efficiency and equipment physical safety.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the intelligent calibration method for optical module coupling positioning based on AI and production big data according to the present invention.

[0022] Figure 2 This is a schematic diagram of the process for obtaining the three-dimensional spatial difference vector according to the present invention;

[0023] Figure 3 This is a schematic diagram of the optical module coupling positioning intelligent calibration system based on AI and production big data of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, and not all embodiments. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0025] Please see Figure 1 This invention discloses an intelligent calibration method for optical module coupling positioning based on AI and production big data, comprising the following steps: S1. Input the production batch identifier of the current optical module to be coupled and the ambient temperature data into the historical production database for cluster matching, extract the distribution pattern of the historical best coupling coordinates of the same batch, and generate the initial bias compensation vector of the batch accordingly. S1 includes the following steps: S11. Synchronously acquire the production batch identifier, ambient temperature data, and initial end-face image stream of the current optical module to be coupled; S12. In the historical production database, select historical successful coupling records that have the same material supplier code and the same production batch identifier as the current optical module to be coupled, and form a batch sample set; if there are no historical successful coupling records with the same material supplier code and the same production batch identifier, select the historical successful coupling record with the highest similarity as the batch sample set. Extract the initial geometric centroid coordinates and the final optimal coupling coordinates of all samples in the batch sample set, and calculate the three-dimensional spatial difference vector between the two. Specifically, please refer to Figure 2 In embodiment S12, the method for calculating similarity is as follows: extract the physical dimensions of the optical chip, the focal length of the lens, the nominal wavelength, and the thermal expansion coefficient of the substrate material from the historical records to construct a normalized historical material feature vector; at the same time, obtain the corresponding physical parameters of the current optical module to be coupled to construct the current material feature vector; use the Euclidean distance formula to calculate the spatial distance between the current material feature vector and each historical material feature vector, and take the record with the smallest distance as the historical coupling success record with the highest similarity; Because the physical geometric center of the laser chip (LD) and lens array in the optical module often has a fixed systematic deviation from the actual optical light-emitting center under different wafer batches and different cutting processes; the system reads the QR code of the current material by scanning the barcode, parses out the supplier code and batch number, and pulls the matching past production data from the MySQL database mounted in the cloud; S13. After removing outlier data from the three-dimensional spatial difference vector that deviate from the average by more than three standard deviations, assign increasing weights to the remaining three-dimensional spatial difference vector according to the time distance, and calculate the weighted average vector. Specifically, in embodiment S13, in order to prevent extreme abnormal data caused by poor fixture clamping from polluting the overall model, the system first calculates the arithmetic mean and standard deviation of all difference vectors, and directly removes data that deviates from the mean by more than three standard deviations; considering that the production status evolves over time, such as the equipment gradually heating up, the data closer to production has greater reference value; the system assigns weights to the remaining N samples, with the oldest sample having a weight of one and the newest sample having a weight of N, and calculates the weighted average to obtain the basic compensation vector; S14. The weighted average vector is calibrated by combining the current ambient temperature data with the temperature difference compensation coefficient of the standard room temperature, and the initial bias compensation vector of the batch is output. The temperature difference compensation coefficient is obtained by multiplying the preset thermal expansion coefficient by the temperature difference. Specifically, in embodiment S14, the system reads the current ambient temperature in real time through a high-precision temperature sensor deployed on the coupling machine platform; the standard room temperature is set to 25 degrees Celsius; in this embodiment, the displacement stage and fixture are made of stainless steel and aluminum alloy components, and the comprehensive thermal expansion coefficient is set to 0.02 micrometers per degree Celsius; if the current temperature is 28 degrees Celsius, the temperature difference is 3 degrees. The system superimposes a compensation amount of 0.06 micrometers on the Z-axis of the basic compensation vector, that is, the vertical direction, where the thermal expansion is most significant, and finally outputs an accurate batch initial bias compensation vector. Through the above steps, the present invention compresses the originally completely unknown initial search space into a very small high-probability space by coupling production big data and environmental factors, providing a compensation benchmark with prior knowledge for subsequent coarse positioning. S2. Perform edge enhancement and feature extraction on the initial end face image stream of the current optical module to be coupled, and locate the initial geometric centroid coordinates of the light-emitting area of ​​the optical chip and the receiving area of ​​the optical fiber. The initial geometric centroid coordinates and the initial batch offset compensation vector are added together to generate coarse positioning target coordinates, and the six-degree-of-freedom high-precision displacement stage is driven to move to the coarse positioning target coordinates. The initial centroid coordinates for locating the geometry in S2 include the following steps: S21. Obtain the original grayscale image of the optical module end face through a coaxial infrared CCD camera, and apply adaptive histogram equalization to enhance the contrast between the light-emitting point of the optical chip waveguide and the fiber core. Specifically, in embodiment S21, the five-megapixel infrared CCD camera mounted on the coupling machine acquires images at a rate of thirty frames per second. Due to the severe reflection on the surface of the silicon photonic chip, the original image often has local overexposure or underexposure. The image processing module divides the original image into eight-by-eight pixel blocks and performs limited contrast histogram equalization on each block independently, so that the originally dark fiber core edge and waveguide end face outline are clearly visible. S22. Use the final lightweight convolutional neural network to perform semantic segmentation on the enhanced image and output a binary mask image containing the waveguide region and the fiber core region. Specifically, in embodiment S22, the system adopts a lightweight network based on the MobileNet-V3 architecture; the input is a 512×512 single-channel grayscale image, and the backbone network contains 11 depthwise separable convolutional blocks, each of which consists of a 3×3 depthwise convolution, a 1×1 pointwise convolution, and a SE (Separate Array Array). The network consists of a channel attention module, with h-swish as the activation function. The decoder uses bilinear interpolation to upsample to the original image size, and outputs a binary mask through 1×1 convolution and a sigmoid activation function. During training, the batch size is set to 16, the initial learning rate is 0.001, the Adam optimizer is used, and the training runs for 50 epochs. The total number of model parameters is approximately 1.2M, and the inference time per image is less than 10ms, meeting the real-time requirement of 30fps for industrial machines. The input layer of the network receives a 512x512 pixel single-channel grayscale image. After feature extraction through depthwise separable convolution, the decoder restores the image to a binary mask of the same size as the original image. In the mask, regions with a pixel value of 255 represent the identified waveguides and fiber cores, while regions with a pixel value of 0 represent the background. Furthermore, the training process of the final lightweight convolutional neural network includes the following steps: 221. Obtain end-face image sequences and corresponding multi-dimensional coupling state feature vectors containing temperature, initial optical power, and displacement coordinates from several historical optical module coupling processes as input samples; use optical calibration algorithms to quantize and label the geometric center offset of the image sequences, and score the waveguide edge blur, fiber core reflection interference, and initial centroid deviation with continuous values ​​between 0 and 1 respectively, and construct a historical visual feature dataset and a historical offset cause label dataset. 222. Construct an initial lightweight convolutional neural network and set a first training data ratio; divide the historical visual feature dataset and the historical offset cause label dataset according to the first training data ratio to obtain a first training dataset and a first test dataset. S223. Set a first training error threshold; input the first training dataset into the initial lightweight convolutional neural network for training; during the training process, if the training error is less than the first training error threshold, stop training and obtain the trained lightweight convolutional neural network; otherwise, continue training until the training error is less than the first training error threshold. S224. Set a first test error threshold; input the first test dataset into the trained lightweight convolutional neural network for testing; after the test is completed, obtain the first test error data; if the first test error data is less than or equal to the first test error threshold data, use the trained lightweight convolutional neural network as the final lightweight convolutional neural network; otherwise, return to S223 to continue training the trained lightweight convolutional neural network and repeat S224 until the first test error data is less than or equal to the first test error threshold. The first training data ratio, the first training error threshold, and the first test error threshold can be dynamically adjusted according to the model architecture, data volume, and actual needs during implementation. The training error and test error are both obtained by using historical offset cause label data as the reference label and solving the difference between the model output data and the reference label data through cosine similarity or Euclidean distance calculation. S23. Calculate the pixel geometric center of the waveguide region and the fiber core region in the binarized mask image, and convert the pixel coordinates into physical space absolute coordinates through the camera calibration matrix to obtain the initial geometric centroid coordinates. Specifically, in embodiment S23, the system traverses the mask image and calculates the centroid coordinates (U, V) of the white pixel cluster; it then calls a pre-calibrated nine-point affine transformation matrix to convert the pixel coordinates (U, V) into physical mechanical coordinates (X) that the displacement stage control system can recognize. geo ,Y geo Z geo ), that is, the initial geometric centroid coordinates; The process of obtaining the pre-calibrated nine-point calibration affine transformation matrix is ​​as follows: a standard nine-point grid calibration plate is placed on the displacement stage, the camera is controlled to capture the image of the calibration plate and extract the pixel coordinates of the nine grid feature points, and the corresponding nine physical mechanical coordinates fed back by the displacement stage are recorded at the same time; the affine transformation parameters from the pixel coordinate system to the physical coordinate system are solved by the least squares method to generate the nine-point calibration affine transformation matrix. Furthermore, the system will (X) geo ,Y geo Z geo The initial bias compensation vector (ΔX, ΔY, ΔZ) obtained in S14 is strictly summed to obtain the coarse positioning target coordinates (X). target ,ΔY target ,ΔZ target The system sends commands via the EtherCAT industrial Ethernet bus to drive the piezoelectric ceramic PZT six-degree-of-freedom displacement stage to move to the target coordinates at maximum speed. Through the above steps, the system not only completed the traditional visual alignment, but more importantly, it physicalized the batch inherent tolerance extracted from big data in S1, so that the displacement stage crossed the blind search area in the coarse positioning stage and fell into the halfway position very close to the real optical peak. S3. At the coarsely located target coordinates, collect real-time optical power signals and real-time three-dimensional spatial coordinates, perform first-order differential processing on the real-time optical power signals to obtain the optical power change gradient; align the real-time optical power signals, optical power change gradients and real-time three-dimensional spatial coordinates with timestamps to construct a multi-dimensional coupled state feature vector. The construction of the multidimensional coupled state feature vector in S3 includes the following steps: S31. The voltage signal output by the photodetector is acquired using a high-speed analog-to-digital converter with a sampling frequency of 10 kHz, and high-frequency circuit noise is eliminated by a moving average filtering algorithm to obtain the real-time optical power signal. Specifically, in embodiment S31, when the displacement stage reaches the coarse positioning coordinates, the laser turns on to emit light; the photodiode at the receiving end converts photons into a weak current, which is then converted into voltage by a transimpedance amplifier; the system uses a 16-bit high-precision ADC to continuously sample at a frequency of 10,000 Hz; in order to eliminate the 50 Hz power frequency interference and environmental white noise caused by fluctuations in the workshop power grid, the system sets up a sliding average filter with a window length of 100, and outputs the smoothed optical power value P(t) in real time. S32. Read the real-time feedback data of the grating ruler inside the six-degree-of-freedom high-precision displacement stage to obtain real-time three-dimensional spatial coordinates containing three translational dimensions and three rotational dimensions. Specifically, in embodiment S32, the nanoscale closed-loop grating ruler built into the displacement stage broadcasts the current position in real time at a frequency of 1 kilohertz; the system extracts the absolute coordinates of the three translation axes X, Y, and Z, as well as the yaw angle, pitch angle, and roll angle data of the rotation around the three axes; S33. Calculate the difference between the optical power in the current sampling period and the optical power in the previous sampling period, divide it by the sampling time interval, and obtain the optical power change gradient. Specifically, in embodiment S33, the system records the optical power P(t) at the current time t and the optical power P(t-1) at the previous millisecond time (t-1), and subtracts the two to obtain the power difference ΔP; this difference is divided by the time interval of one millisecond, which is the optical power change gradient dP / dt at the current spatial position; this gradient data is a physical quantity for determining whether the current direction of movement is correct. S34. Using the hardware clock of the displacement stage controller as a reference, the real-time optical power signal, optical power change gradient and real-time three-dimensional spatial coordinates within the same microsecond time window are spliced ​​together to generate a multi-dimensional coupled state feature vector with an eight-dimensional length. Specifically, in embodiment S34, to solve the data misalignment problem caused by inconsistent sensor sampling rates, the system uses the FPGA hardware clock of the displacement stage controller as the absolute time axis; whenever the grating ruler updates the coordinate data, the system immediately latches the filtered optical power P and gradient dP / dt at the current moment; [X,Y,Z,Theta] X Theta Y Theta Z The concatenation of [P, dP / dt] into a one-dimensional array of length eight constitutes the multidimensional coupled state feature vector; Theta X Theta represents the pitch angle of the stage rotating about the X-axis. Y Theta represents the yaw angle of the displacement stage rotating about the Y-axis. Z This indicates the roll angle of the displacement stage about the Z-axis; By constructing a multidimensional coupled state feature vector, which serves as the sole sensory input for subsequent AI decision-making, it not only includes location information but also integrates energy distribution information such as power and energy change trends of gradients, ensuring the completeness of the state description. S4. Input the multidimensional coupling state feature vector into the reinforcement learning optimization model in real time to obtain a non-fixed step displacement increment command, drive the displacement stage to perform micro-displacement detection, and calculate the action reward value based on the change in optical power before and after displacement; repeat this step until the current optical power exceeds the target threshold, and lock the current coordinates as the optimal coupling coordinates; The construction and training of the deep reinforcement learning optimization model in S4 includes the following steps: S411. Construct an initial deep reinforcement learning optimization model that includes an Actor network and a Critic network; obtain multi-dimensional state data sequences and corresponding displacement operation sequences from historical manual coupling processes, and construct a demonstration dataset; Specifically, in embodiment S411, the system records the complete process of a senior technician manually using a handle to adjust the light, and pairs the [state vector] of each millisecond during the process with the [joystick displacement command] given by the technician to form an expert dataset; S412. Supervised learning pre-training of the Actor network is performed using the demonstration dataset to make the initial output behavior of the Actor network approximate the human optimization trajectory, thus obtaining the pre-trained Actor network. Specifically, in embodiment S412, since reinforcement learning is prone to damage to the piezoelectric ceramic displacement stage due to exploration from scratch, pre-training is necessary; mean square error (MSE) is used as the loss function, and the weights of the Actor network are adjusted through the backpropagation algorithm so that the error between the displacement command output by the network and the historical expert operation is minimized when the same state vector is given. S413. Build a physical simulation environment coupled with optical modules, and connect the pre-trained Actor network and Critic network into the physical simulation environment for reinforcement learning training. Specifically, in embodiment S413, based on the Gaussian beam coupling theory, the system constructs a virtual Gaussian light field distribution model in computer memory and puts the pre-trained Actor network into this virtual environment; S414. Set the training reward function; multiply the increase in optical power compared to the previous moment by the positive gain coefficient as the positive reward, multiply the time consumption of a single displacement by the time penalty coefficient as the negative reward, and give a finite negative reward value if a physical collision occurs. Specifically, in embodiment S414, the training reward function includes: a positive gain coefficient, a time penalty coefficient, and a fixed negative reward; The ratio between the positive gain coefficient and the time penalty coefficient is determined based on historical test data to ensure that the model prioritizes maximizing optical power while tending to select the path with the fewest steps. The fixed negative reward is used to force the model to avoid dangerous areas that exceed mechanical limits or may cause hard contact between the optical fiber and the chip when outputting displacement increment commands. The quantization rules for the reward function are as follows: the positive gain coefficient is set to 100, and the time penalty coefficient is set to 0.1; the reward value R for the current step is 100 × (current optical power - previous optical power) - 0.1 × single action time; if the displacement of the model output causes the Z-axis coordinate to be lower than the safe plane, i.e. the critical point where the fiber end face and the chip make physical contact, then the penalty is triggered immediately, the reward value R for this step is forcibly set to negative 1000, and the current training round is terminated. Through the above quantization rules, the model is forced to find an optimal path that can rapidly increase the optical power, avoid taking a detour, and prevent collisions. S415. When the average cumulative reward value of the model in the physical simulation environment converges and no longer increases after a maximum number of rounds in one hundred consecutive rounds, training is stopped, the current network parameters are fixed, and the pre-trained deep reinforcement learning optimization model is obtained. Specifically, the internal structures of the Actor network and the Critic network are as follows: The Actor network comprises an input layer, three fully connected hidden layers, and an output layer. The input layer receives the multidimensional coupled state feature vector. The three hidden layers contain 256, 128, and 64 neurons, respectively, and the activation function is a linear rectified function. The output layer contains six neurons and uses a hyperbolic tangent activation function to map the output value to the range of negative one to positive one. The output value is then multiplied by the maximum allowable single step size of each physical axis to obtain the actual non-fixed step size displacement increment command. The Critic network shares the first two hidden layers with the Actor network. Its third hidden layer contains sixty-four neurons, and the output layer is a single neuron that uses a linear activation function to output the expected cumulative reward value of the current state. The loop execution mechanism in S4 includes the following steps: S421. After receiving a non-fixed step displacement increment command, the displacement stage performs physical movement and triggers a synchronization signal after moving into position. S422. After receiving the synchronization signal, the system re-executes S3 to generate a new multidimensional coupled state feature vector. S423. Determine whether the current real-time optical power signal is greater than or equal to the target threshold. If not, input the new multidimensional coupled state feature vector into the deep reinforcement learning optimization model and continue to output the next instruction. S424. If yes, start the stabilization timer. If the optical power does not drop below the target threshold within three consecutive detection cycles, the coupling is considered successful, the current coordinates are locked as the optimal coupling coordinates, and the loop is terminated. Specifically, in steps S421-S424, the eight-dimensional state feature vector generated in step S3 is input into the solidified Actor network, outputting six continuous values ​​between -1 and +1. The system multiplies these six values ​​by the maximum permissible single-step size for each physical axis, setting them to 0.5 micrometers and 0.1 milliradians, respectively, and converts them into actual non-fixed-step displacement increment commands, such as moving the X-axis by 0.2 micrometers and moving the Y-axis by -0.05 micrometers. The displacement stage receives the command and performs fine-tuning. After moving to the correct position, the system re-executes step S3 to re-acquire the state. It determines whether the current optical power is greater than or equal to the set factory target threshold, such as -1 dBm. If it is not reached, the new state is fed back to the Actor network to output the next command. If the target threshold is reached, the system does not stop immediately but starts a stabilization timer to continue hovering at the current position. If the optical power remains stable above -1 dBm for three consecutive detection cycles, it proves that the current position is not a pseudo-peak caused by vibration. At this point, the system locks the current six-degree-of-freedom coordinates and marks them as the optimal coupling coordinates. Furthermore, the target threshold is calculated by extracting the nominal luminous power of the current batch of optical chips, subtracting the theoretical insertion loss of the lens array and the fiber end face, and then multiplying it by the average coupling efficiency coefficient of similar materials in history at this temperature, to obtain the theoretical optimal optical power value under the current physical environment, and using the theoretical optimal optical power value as the target threshold. S5. Extract the optimal coupling coordinates of the most recent consecutive preset number of optical modules, calculate the coordinate drift slope, and when the coordinate drift slope exceeds the mechanical thermal drift safety threshold, update the batch initial offset compensation vector in the historical production database using the coordinate drift slope. The steps to trigger the global compensation mechanism in S5 are as follows: S51. Set the sliding window length to fifty, and extract the optimal coupling coordinates of the fifty most recent consecutive optical modules in the same batch; Specifically, the mechanical thermal drift safety threshold is based on the pre-calibrated optical module coupling tolerance curve. The critical value of spatial offset corresponding to the optical power decrease of 0.1dB is extracted, and the critical value is divided by the sliding window length of fifty to obtain the maximum average offset allowed for a single module, which is set as the mechanical thermal drift safety threshold. S52. Extract the X-axis, Y-axis and Z-axis data of these fifty optimal coupled coordinates respectively, and use the least squares method to perform linear fitting on the data of each axis as the production serial number changes, and extract the slope of the fitted line of each axis as the coordinate drift slope. S53. Set the mechanical thermal drift safety threshold to 0.05 micrometers for each module produced; if the absolute value of the coordinate drift slope of any axis is greater than the mechanical thermal drift safety threshold, it is determined that the production equipment has undergone thermal deformation or the fixture has worn. S54. Multiply the coordinate drift slope that exceeds the mechanical thermal drift safety threshold by the sliding window length as the drift compensation amount, and add the drift compensation amount to the batch initial bias compensation vector corresponding to the current batch in the historical production database to complete the online update; For example, consider a 400G DR4 silicon photonics module manufactured by an optical communication factory: When the system was first turned on during the early shift, the workshop temperature was 20 degrees Celsius. The system scanned material batch A, executed S1 to extract historical big data, and generated an initial bias compensation vector of [X: +1.5 micrometers, Y: -0.8 micrometers, Z: +2.0 micrometers]. In step S2, the vision camera captures an image of the optical chip and fiber array, and calculates the geometric center. The system directly adds the aforementioned offset vector to the geometric center and commands the displacement stage to move at high speed to the coarse positioning point. At this point, due to the precise compensation, the initial optical power directly reaches negative five dBm. Entering S3 and S4, the reinforcement learning Actor network takes over control. The network senses that the current power is -5 dBm and the gradient is large. In the first step, it outputs a command to slightly decrease the Z-axis by 0.3 micrometers, and the power jumps to -2 dBm. In the second step, the network senses that the gradient is becoming gentler, and outputs a command to shift the X-axis by 0.05 micrometers, and the power reaches -0.8 dBm, exceeding the target threshold of -1 dBm. After the system hovers for 30 milliseconds to confirm that there are no errors, it locks this point as the optimal coupling coordinate and performs dispensing curing. After three hours of continuous production, the internal motor of the equipment started to heat up. The S5 monitoring module found that the optimal Z-axis coordinates of the last fifty modules showed a clear monotonically increasing trend, with a slope of 0.07 micrometers per module. The system immediately triggered self-evolution and updated the Z-axis offset vector of batch A in the database to +5.5 micrometers. When the next module came online, the system directly guided the displacement stage to the new coordinates that included the thermal expansion offset, so that the initial optical power remained at a high level of -5 dBm, which offset the deterioration of the physical environment.

[0026] For further details, please refer to Figure 3 A smart calibration system for optical module coupling positioning based on AI and production big data is provided to implement the aforementioned smart calibration method for optical module coupling positioning based on AI and production big data. The system includes: The data perception and bias generation module is used to simultaneously acquire the optical module production identifier and image stream, and retrieve the historical database to generate the batch initial bias compensation vector; The visual calculation and coarse localization module receives the image stream and the batch initial offset compensation vector, calculates the geometric center and drives the displacement stage to the coarse localization target coordinates; The real-time state feature construction module synchronously collects high-frequency optical power and spatial coordinates after coarse localization to construct a multi-dimensional coupled state feature vector. The reinforcement learning micro-motion optimization module takes the multi-dimensional coupled state feature vector as input, uses the Actor-Critic network architecture to output continuous displacement commands, and searches for the optimal coupled coordinates in a closed loop. The drift monitoring and self-evolution module monitors the coordinate drift slope during continuous production and dynamically updates the batch initial offset compensation vector when the mechanical thermal drift safety threshold is reached. This invention breaks away from the open-loop, isolated model of static visual recognition combined with mechanical blind trial-and-error in traditional optical module coupling. By deeply integrating the time dimension, such as historical batch big data and equipment thermal drift trends, and the multimodal physical dimension, such as visual images, spatial coordinates, optical power and its gradient, into the AI ​​decision model, it utilizes big data prior knowledge and real-time thermal drift feedback to compress the unknown search space into an extremely small high-probability interval. On the other hand, it uses the nonlinear dynamic step size strategy of deep reinforcement learning to achieve the fastest convergence in complex gradient fields. This eliminates the systematic errors caused by material batch differences and equipment thermal deformation, avoiding high-frequency shutdown calibration. It also reduces the time consumed in a single coupling and effectively avoids the risk of collisions.

[0027] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0028] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A smart calibration method for optical module coupling positioning based on AI and production big data, characterized in that, Includes the following steps: S1. Input the production batch identifier of the current optical module to be coupled and the ambient temperature data into the historical production database for cluster matching, extract the distribution pattern of the historical best coupling coordinates of the same batch, and generate the initial bias compensation vector of the batch accordingly. S2. Perform edge enhancement and feature extraction on the initial end face image stream of the current optical module to be coupled, and locate the initial geometric centroid coordinates of the light-emitting area of ​​the optical chip and the receiving area of ​​the optical fiber. The initial geometric centroid coordinates and the initial batch offset compensation vector are added together to generate coarse positioning target coordinates, and the six-degree-of-freedom high-precision displacement stage is driven to move to the coarse positioning target coordinates. S3. At the coarsely located target coordinates, collect real-time optical power signals and real-time three-dimensional spatial coordinates, perform first-order differential processing on the real-time optical power signals to obtain the optical power change gradient; align the real-time optical power signals, optical power change gradients and real-time three-dimensional spatial coordinates with timestamps to construct a multi-dimensional coupled state feature vector. S4. Input the multidimensional coupling state feature vector into the reinforcement learning optimization model in real time to obtain a non-fixed step displacement increment command, drive the displacement stage to perform micro-displacement detection, and calculate the action reward value based on the change in optical power before and after displacement; repeat this step until the current optical power exceeds the target threshold, and lock the current coordinates as the optimal coupling coordinates; S5. Extract the optimal coupling coordinates of the most recent consecutive preset number of optical modules, calculate the coordinate drift slope, and when the coordinate drift slope exceeds the mechanical thermal drift safety threshold, update the batch initial offset compensation vector in the historical production database using the coordinate drift slope.

2. The intelligent calibration method for optical module coupling positioning based on AI and production big data according to claim 1, characterized in that, S1 includes the following steps: S11. Synchronously acquire the production batch identifier, ambient temperature data, and initial end-face image stream of the current optical module to be coupled; S12. In the historical production database, select historical successful coupling records that have the same material supplier code and the same production batch identifier as the current optical module to be coupled, and form a batch sample set; if there are no historical successful coupling records with the same material supplier code and the same production batch identifier, select the historical successful coupling record with the highest similarity as the batch sample set. Extract the initial geometric centroid coordinates and the final optimal coupling coordinates of all samples in the batch sample set, and calculate the three-dimensional spatial difference vector between the two. S13. After removing outlier data from the three-dimensional spatial difference vector that deviate from the average by more than three standard deviations, assign increasing weights to the remaining three-dimensional spatial difference vector according to the time distance, and calculate the weighted average vector. S14. The weighted average vector is calibrated by combining the current ambient temperature data with the temperature difference compensation coefficient of the standard room temperature, and the initial bias compensation vector of the batch is output. The temperature difference compensation coefficient is obtained by multiplying the thermal expansion coefficient and the temperature difference.

3. The intelligent calibration method for optical module coupling positioning based on AI and production big data according to claim 1, characterized in that, The similarity calculation in S12 includes the following steps: S121. Extract the physical dimensions of the optical chip, lens focal length, nominal wavelength and thermal expansion coefficient of the substrate material from the historical records, and construct a normalized historical material feature vector. S122. Obtain the physical dimensions of the optical chip, lens focal length, nominal wavelength, and thermal expansion coefficient of the substrate material of the current optical module to be coupled, and construct the current material feature vector. S123. The Euclidean distance method is used to calculate the spatial distance between the current material feature vector and each historical material feature vector. The record with the smallest distance is taken as the historical coupling success record with the highest similarity.

4. The intelligent calibration method for optical module coupling positioning based on AI and production big data according to claim 1, characterized in that, The initial centroid coordinates for locating the geometry in S2 include the following steps: S21. Obtain the original grayscale image of the optical module end face and apply adaptive histogram equalization to enhance the contrast between the light-emitting point of the optical chip waveguide and the fiber core. S22. Use the final lightweight convolutional neural network to perform semantic segmentation on the enhanced image and output a binary mask image containing the waveguide region and the fiber core region. S23. Calculate the pixel geometric center of the waveguide region and the fiber core region in the binarized mask image, and convert the pixel coordinates into physical space absolute coordinates through the camera calibration matrix to obtain the initial geometric centroid coordinates.

5. The intelligent calibration method for optical module coupling positioning based on AI and production big data according to claim 1, characterized in that, The construction of the multidimensional coupled state feature vector in S3 includes the following steps: S31. The voltage signal output by the photodetector is acquired using a high-speed analog-to-digital converter, and high-frequency circuit noise is eliminated by a moving average filtering algorithm to obtain the real-time optical power signal. S32. Read the real-time feedback data of the grating ruler inside the six-degree-of-freedom high-precision displacement stage to obtain real-time three-dimensional spatial coordinates containing three translational dimensions and three rotational dimensions. S33. Calculate the difference between the optical power in the current sampling period and the optical power in the previous sampling period, divide it by the sampling time interval, and obtain the optical power change gradient. S34. Using the hardware clock of the displacement stage controller as a reference, the real-time optical power signal, optical power change gradient and real-time three-dimensional spatial coordinates within the same time window are spliced ​​together to generate a multi-dimensional coupled state feature vector.

6. The intelligent calibration method for optical module coupling positioning based on AI and production big data according to claim 1, characterized in that, The construction and training of the deep reinforcement learning optimization model in S4 includes the following steps: S411. Construct an initial deep reinforcement learning optimization model that includes an Actor network and a Critic network; obtain multi-dimensional state data sequences and corresponding displacement operation sequences from historical manual coupling processes, and construct a demonstration dataset; S412. Supervised learning pre-training of the Actor network is performed using the demonstration dataset to make the initial output behavior of the Actor network approximate the human optimization trajectory, thus obtaining the pre-trained Actor network. S413. Build a physical simulation environment coupled with optical modules, and connect the pre-trained Actor network and Critic network into the physical simulation environment for reinforcement learning training. S414. Set the training reward function; multiply the increase in optical power compared to the previous moment by the positive gain coefficient as the positive reward, multiply the time consumption of a single displacement by the time penalty coefficient as the negative reward, and give a finite negative reward value if a physical collision occurs. S415. When the average cumulative reward value of the model in the maximum number of rounds in the physical simulation environment converges and no longer increases, stop training, solidify the current network parameters, and obtain the pre-trained deep reinforcement learning optimization model.

7. The intelligent calibration method for optical module coupling positioning based on AI and production big data according to claim 6, characterized in that, The training reward function in S414 includes: a positive gain coefficient, a time penalty coefficient, and a fixed negative reward; The ratio between the positive gain coefficient and the time penalty coefficient is determined based on historical test data to ensure that the model prioritizes maximizing optical power while tending to select the path with the fewest steps. The fixed negative reward is used to force the model to avoid dangerous areas that exceed mechanical limits or cause hard contact between the optical fiber and the chip when outputting displacement increment commands.

8. The intelligent calibration method for optical module coupling positioning based on AI and production big data according to claim 1, characterized in that, The loop execution mechanism in S4 includes the following steps: S421. After receiving a non-fixed step displacement increment command, the displacement stage performs physical movement and triggers a synchronization signal after moving into position. S422. After receiving the synchronization signal, the system re-executes S3 to generate a new multidimensional coupled state feature vector. S423. Determine whether the current real-time optical power signal is greater than or equal to the target threshold. If not, input the new multidimensional coupled state feature vector into the deep reinforcement learning optimization model and continue to output the next instruction. S424. If yes, start the stabilization timer. If the optical power does not drop below the target threshold within the continuous detection period, the coupling is considered successful, the current coordinates are locked as the optimal coupling coordinates, and the loop is terminated.

9. The intelligent calibration method for optical module coupling positioning based on AI and production big data according to claim 1, characterized in that, The steps to trigger the global compensation mechanism in S5 are as follows: S51. Set the sliding window length and extract the optimal coupling coordinates of the most recent consecutive optical modules in the same batch; S52. Extract the X-axis, Y-axis and Z-axis data of the optimal coupled coordinates in S51 respectively. Use the least squares method to perform linear fitting on the data of each axis as the production serial number changes, and extract the slope of the fitted line of each axis as the coordinate drift slope. S53. Set a mechanical thermal drift safety threshold; if the absolute value of the coordinate drift slope of any axis is greater than the mechanical thermal drift safety threshold, it is determined that the production equipment has undergone thermal deformation or the jig has worn. S54. Multiply the coordinate drift slope that exceeds the mechanical thermal drift safety threshold by the sliding window length as the drift compensation amount, and add the drift compensation amount to the batch initial bias compensation vector corresponding to the current batch in the historical production database to complete the online update.

10. A smart calibration system for optical module coupling positioning based on AI and production big data, characterized in that, The system for implementing the AI-based and production big data-driven optical module coupling positioning intelligent calibration method as described in any one of claims 1-9, the system comprising: The data perception and bias generation module is used to simultaneously acquire the optical module production identifier and image stream, and retrieve the historical database to generate the batch initial bias compensation vector; The visual calculation and coarse localization module receives the image stream and the batch initial offset compensation vector, calculates the geometric center and drives the displacement stage to the coarse localization target coordinates; The real-time state feature construction module synchronously collects high-frequency optical power and spatial coordinates after coarse localization to construct a multi-dimensional coupled state feature vector. The reinforcement learning micro-motion optimization module takes the multi-dimensional coupled state feature vector as input, uses the Actor-Critic network architecture to output continuous displacement commands, and searches for the optimal coupled coordinates in a closed loop. The drift monitoring and self-evolution module monitors the coordinate drift slope during continuous production and dynamically updates the batch initial offset compensation vector when the mechanical thermal drift safety threshold is reached.

Images