Glasses lens distributed collaborative manufacturing system based on federated learning and closed-loop control

Abstract

The application belongs to the technical field of industrial internet of things, and discloses a glasses lens distributed collaborative manufacturing system based on federal learning and closed-loop control; comprising: obtaining an encrypted order, and calculating corresponding processing difficulty coefficients, a theoretical lens model and initial processing parameters; dynamically evaluating the fitness scores of each intelligent processing node, and screening out a target intelligent processing node; inputting the processing difficulty coefficients and the initial processing parameters into a pre-constructed local process prediction model to obtain optimized processing parameters, and combining the theoretical lens model to execute a closed-loop grinding process to obtain closed-loop processing data; locally training the local process prediction model according to the closed-loop processing data to generate encrypted model parameter update amounts; aggregating the encrypted model parameter update amounts to generate a global updated process model; and the application can realize high-precision, high-efficiency, intelligent and safe collaborative manufacturing of the lens processing process.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) technology, and more specifically, to a distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control. Background Technology

[0002] Eyeglass lens processing falls under the field of precision optical manufacturing. Its processing quality directly determines the assembly accuracy of the lens and frame, wearing comfort, and the final visual correction effect. With the continuous expansion of eyewear retail networks and the sustained growth in personalized customization demands, lens processing is gradually evolving from the traditional centralized processing center model to a distributed, networked manufacturing model. This involves collecting three-dimensional geometric data of the frame and optometry prescription data from retail outlets, and then distributing orders to different processing nodes for manufacturing via the network. While existing remote processing technologies have solved the problem of cross-regional order transfer to some extent, they are essentially still based on a "remote order transmission + independent node processing" model. The quality of lens processing remains highly dependent on the equipment status of individual processing nodes at the time of order execution. Especially after long-term operation of edging machines and the accumulation of tool wear, systematic dimensional deviations can easily occur, leading to out-of-tolerance lens dimensions, rework, or even scrapping. It is difficult to guarantee stable and consistent processing quality across the entire network.

[0003] Furthermore, the scheduling logic of existing distributed processing systems is usually quite simple, mostly allocating resources based solely on geographical distance or order sequence. This fails to comprehensively consider the differences in processing difficulty determined by multi-dimensional process parameters such as lens material, base curvature, and design type. It also fails to fully integrate and optimize factors such as historical yield rates, real-time load, and logistics costs at each processing node. This results in high-difficulty orders potentially being assigned to processing nodes that are not suited for them, leading to increased scrap rates and higher operating costs. Simultaneously, each processing node continuously accumulates a large amount of valuable process data and compensation experience during processing. However, this data is usually confined to the node itself, making cross-node sharing difficult while protecting privacy and trade secrets, creating information silos and limiting the continuous improvement of the entire manufacturing network's process capabilities. Therefore, there is an urgent need for a distributed collaborative manufacturing system for eyeglass lenses that can simultaneously achieve local closed-loop processing control, multi-dimensional intelligent scheduling in the cloud, and secure evolution of cross-node process knowledge.

[0004] In view of this, the present invention proposes a distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control, comprising: The order acquisition module is used by the data acquisition terminal to obtain the three-dimensional geometric data of the eyeglass frame, the optometry prescription data, and the order selection parameters. It then generates an encrypted order by summarizing the data and uploads it to the cloud platform. The difficulty calculation module is used by the cloud platform to dynamically calculate the processing difficulty coefficient of the encrypted order based on the optometry prescription data and order selection parameters in the encrypted order. The intelligent scheduling module is used by the cloud platform to generate corresponding theoretical lens models and initial processing parameters based on encrypted orders, obtain the comprehensive status data of each intelligent processing node, dynamically evaluate the adaptability score of each intelligent processing node, select target intelligent processing nodes, and issue processing instructions containing theoretical lens models, initial processing parameters and processing difficulty coefficients to the target intelligent processing nodes. The closed-loop machining module is used by the target intelligent machining node to input the machining difficulty coefficient and initial machining parameters into the pre-built local process prediction model to obtain optimized machining parameters, and then execute the closed-loop grinding process according to the optimized machining parameters and the theoretical lens model to obtain closed-loop machining data. The local training module is used by the target intelligent processing node to train the local process prediction model locally based on closed-loop processing data, generate encrypted model parameter update data, and upload it to the cloud platform. The federated aggregation module is used by the cloud platform to aggregate the encrypted model parameter updates uploaded by each intelligent processing node, generate a globally updated process model, and distribute it to each intelligent processing node.

[0006] Furthermore, methods for generating encrypted orders include: A unique order code is generated for the current order, and the terminal identification code of the data acquisition terminal is obtained; the three-dimensional geometric data of the eyeglass frame, the optometry prescription data, the order selection parameters, the order code and the terminal identification code are integrated to form the original order data; the original order data is subjected to data integrity verification, and the original order data that passes the integrity verification is encrypted to obtain the encrypted order; The three-dimensional geometric data of the eyeglass frame includes the lens outline parameters, lens size parameters, and eyeglass frame structure parameters; the prescription data includes prescription records for the left and right eyes, and each prescription record includes spherical power, cylindrical power, cylindrical axis, pupillary distance, and pupillary height; the order selection parameters include lens material type, lens refractive index grade, lens design type, and lens coating type.

[0007] Furthermore, methods for dynamically calculating the processing difficulty coefficient of encrypted orders include: The three-dimensional geometric data of the eyeglass frame, the optometry prescription data, and the order selection parameters in the encrypted order are decrypted to obtain the decrypted three-dimensional geometric data of the eyeglass frame, the optometry prescription data, and the order selection parameters. Based on the spherical and cylindrical power values ​​in the optometry prescription data, the prescription complexity level is determined according to the prescription complexity dimension; based on the lens material type and refractive index level in the order selection parameters, the material processing difficulty level is determined according to the material processing dimension; based on the lens design type in the order selection parameters, the design complexity level is determined according to the design complexity dimension. The system presets the base difficulty values ​​for each of the following levels: prescription complexity level, material processing difficulty level, and design complexity level. It also presets the dimension weights for each of the prescription complexity dimension, material processing dimension, and design complexity dimension. The system obtains the base difficulty values ​​corresponding to the judgment results of each dimension, and calculates the weighted sum of each base difficulty value based on the dimension weights of each dimension to obtain the processing difficulty coefficient.

[0008] Furthermore, methods for generating theoretical lens models and initial processing parameters include: Based on the decrypted optometry prescription data, order selection parameters, and three-dimensional geometric data of the frame, the optical surface parameters, edge contour curves, and structural assembly parameters of the lens are calculated to form a theoretical lens model; among them, the structural assembly parameters include the lens center thickness and positioning offset. Based on the lens material type in the order's optional parameters, the corresponding recommended grinding parameter range is searched from the pre-built material and process parameter library; based on the edge contour curve in the theoretical lens model, the maximum contour curvature and curvature fluctuation are extracted. Determine the grinding feed rate within the recommended grinding parameter range based on the maximum contour curvature; determine the grinding wheel speed within the recommended grinding parameter range based on the curvature fluctuation; determine the roughing allowance and fine grinding allowance based on the lens material type, edge contour curve, and lens center thickness. The grinding feed rate, grinding wheel speed, rough grinding allowance and finish grinding allowance are integrated to form the initial machining parameters.

[0009] Furthermore, the methods for evaluating the adaptability score of each intelligent processing node include: For each intelligent processing node, an evaluation is conducted from four evaluation dimensions based on the processing difficulty coefficient, order selection parameters, terminal identification code, and corresponding node comprehensive status data, and an evaluation score is obtained for each evaluation dimension. The four evaluation dimensions are difficulty matching dimension, load balancing dimension, inventory adaptation dimension, and logistics timeliness dimension. For the dimensions of difficulty matching, load balancing, inventory adaptation, and logistics timeliness, corresponding evaluation weights are set. Based on the evaluation weights corresponding to each evaluation dimension, the evaluation scores of each evaluation dimension under the same intelligent processing node are weighted and summed to obtain the adaptation score of the corresponding intelligent processing node.

[0010] Furthermore, methods for obtaining optimized processing parameters include: The initial processing parameters and processing difficulty coefficient are extracted from the processing instructions, and the equipment status parameters are collected. The processing difficulty coefficient, initial processing parameters, and equipment status parameters are used as inputs to the pre-built local process prediction model, and the processing parameter adjustment amount is output. Among them, the processing parameter adjustment amount includes the feed rate adjustment amount, grinding wheel speed adjustment amount, rough grinding allowance adjustment amount, and fine grinding allowance adjustment amount. The grinding feed rate, grinding wheel speed, rough grinding allowance, and fine grinding allowance in the initial machining parameters are summed with the corresponding adjustment values ​​in the machining parameter adjustment values ​​to obtain the optimized grinding feed rate, optimized grinding wheel speed, optimized rough grinding allowance, and optimized fine grinding allowance in sequence. These are then integrated to form the optimized machining parameters.

[0011] Furthermore, the method for performing a closed-loop grinding process includes: Based on the edge contour curves in the theoretical lens model and optimized processing parameters, a rough grinding path is generated. Following this path, and with optimized grinding feed rate and wheel speed, a rough grinding process is performed on the lens blank to obtain a semi-finished lens. After the rough grinding process, the semi-finished lens undergoes its first online measurement to obtain measured contour data. This measured contour data is then compared point-by-point with the edge contour curves in the theoretical lens model to form a rough grinding deviation distribution. A compensating fine grinding path is generated based on this deviation distribution, and the fine grinding process is performed according to this path. After the fine grinding process is completed, a second online measurement is performed on the finely ground lens to obtain the finely ground measured contour data. The finely ground measured contour data is compared point by point with the edge contour curve in the theoretical lens model to form the fine grinding deviation distribution. The absolute value of the residual deviation value of all sampling points in the fine grinding deviation distribution is compared with the preset processing qualification threshold. Based on the comparison results, it is determined whether to perform compensation grinding and the corresponding processing result is marked. The processing result includes qualified and unqualified.

[0012] Furthermore, methods for locally training the local process prediction model include: The processing difficulty coefficient, initial processing parameters, and equipment status parameters in the closed-loop processing data are concatenated to form a training input vector. If the processing result is qualified, a training label vector is constructed based on the processing parameter adjustment amount in the closed-loop processing data. If the processing result is unqualified or compensatory grinding is performed, the processing parameter adjustment amount is corrected and used as the training label vector. The training input vector and the training label vector are combined to form training sample pairs and stored in the local training sample cache queue. It is determined whether the number of training sample pairs in the local training sample cache queue has reached the preset minimum training batch threshold. If so, all training sample pairs are retrieved to form a local training batch and local training is performed. Before performing local training, record the current model parameters of the local process prediction model and mark them as pre-training model parameters; train the local process prediction model with the training input vector as input and the training label vector as the expected output, calculate the mean squared error loss between the output of the local process prediction model and the training label vector, and update the model parameters; after local training is completed, record the model parameters of the local process prediction model and mark them as post-training model parameters; calculate the model parameter update amount based on the post-training model parameters and the pre-training model parameters.

[0013] Furthermore, methods for generating and uploading updated encrypted model parameters to the cloud platform include: Differential privacy processing is performed on the model parameter update amount to obtain the model parameter update amount with added differential privacy noise; homomorphic encryption processing is performed on the model parameter update amount with added differential privacy noise to obtain the encrypted model parameter update amount; the encrypted model parameter update amount is integrated with the number of training sample pairs to form an encrypted update upload package, and uploaded to the cloud platform through the network communication interface.

[0014] Furthermore, the method for aggregating the encrypted model parameter updates uploaded by each intelligent processing node includes: Set a federated aggregation cycle. When each federated aggregation cycle arrives, collect the encrypted update upload packets uploaded by all intelligent processing nodes in the current federated aggregation cycle; extract the encrypted model parameter update amount and the number of training sample pairs from each encrypted update upload packet; count the number of encrypted update upload packets received in the current federated aggregation cycle and mark them as the number of participating aggregation nodes; preset a minimum aggregation node threshold and compare the number of participating aggregation nodes with the minimum aggregation node threshold. If the number of participating aggregation nodes is less than the minimum aggregation node threshold, the current weighted aggregation operation is abandoned, and the process waits for the next federated aggregation cycle. If the number of participating aggregation nodes is greater than or equal to the minimum aggregation node threshold, the weighted aggregation operation is performed: all intelligent processing nodes participating in the weighted aggregation operation are marked as participating aggregation nodes, the sample size aggregation weight and quality aggregation weight of each participating aggregation node are calculated, and the weights are balanced based on a preset aggregation weight balance coefficient to obtain the comprehensive aggregation weight. Based on the comprehensive aggregation weight, the encrypted model parameter update amount of each participating aggregation node is weighted and summed in the ciphertext space to obtain the encrypted global parameter update amount.

[0015] The technical effects and advantages of the distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control in this invention are as follows: By comprehensively analyzing the three-dimensional geometric data of eyeglass frames, optometry prescription data, and order selection parameters, the actual processing characteristics of orders can be more accurately characterized. Based on this, the processing difficulty can be dynamically assessed, theoretical lens models and initial processing parameters can be generated, thereby improving the pertinence and rationality of front-end process decisions. During manufacturing execution, adaptability assessments can be performed by combining equipment load, historical yield rate, inventory status, and geographical location at each processing node. This enables intelligent matching between complex orders and processing resources, which not only improves node utilization and order flow efficiency but also helps shorten delivery cycles and reduce scheduling imbalance risks. Introducing online measurement, deviation analysis, and compensation grinding mechanisms during lens grinding allows for dynamic correction of processing paths and process parameters based on actual contour deviations after rough and fine grinding, enabling processing control to be more efficient and efficient. The traditional experience-based open-loop execution is transformed into closed-loop optimization control based on actual measurement feedback, thereby effectively reducing edge size errors, minimizing over- or under-grinding, and improving lens processing accuracy, dimensional consistency, and finished product yield. Simultaneously, each processing node trains its model locally based on actual processing results, and uploads parameter updates using differential privacy and homomorphic encryption. This ensures the security of order data, prescription data, and local processing data at each node, while enabling cross-node process knowledge sharing and global model aggregation updates. This not only continuously improves the process prediction model's adaptability to different materials, design types, and equipment states but also enhances the system's responsiveness to complex orders and dynamic production environments, ultimately achieving high-precision, high-efficiency, intelligent, and safe collaborative manufacturing in the lens processing process. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control, according to Embodiment 1 of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1:

[0018] Please see Figure 1 As shown in this embodiment, the distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control includes an order acquisition module, a difficulty calculation module, an intelligent scheduling module, a closed-loop processing module, a local training module, and a federated aggregation module. The modules are connected via wired and / or wireless means to realize data transmission between modules.

[0019] The order acquisition module is used by the data acquisition terminal to obtain the three-dimensional geometric data of the eyeglass frame, the optometry prescription data, and the order selection parameters. It then generates an encrypted order by summarizing the data and uploads it to the cloud platform.

[0020] Methods for obtaining the three-dimensional geometric data of the frame include: The data acquisition terminal deployed at the retail terminal collects three-dimensional geometric data of the eyeglass frame selected by the customer. The data acquisition terminal is a dedicated data acquisition device equipped with a structured light three-dimensional scanning module and an edge computing processing unit, and is deployed at the fitting station of the eyewear retail store. The structured light three-dimensional scanning module is used to perform non-contact optical scanning of the inner contour of the eyeglass frame to obtain three-dimensional point cloud data of the inner contour of the eyeglass frame. The three-dimensional point cloud data is a discrete point set composed of multiple spatial coordinate points, which is used to describe the three-dimensional spatial shape of the inner contour of the eyeglass frame. The edge computing processing unit performs point cloud denoising and contour fitting on the 3D point cloud data to obtain the 3D geometric data of the mirror frame. Point cloud denoising removes outliers caused by scanning noise or reflection interference from the 3D point cloud data. Point cloud denoising is a well-known technique in the field, and its specific implementation process will not be elaborated here. Contour fitting fits the denoised 3D point cloud data into a closed mirror frame contour curve to obtain the mirror frame contour parameters and mirror frame size parameters. The mirror frame contour parameters include the 3D coordinate sequence of each sampling point on the mirror frame contour curve and the curvature change value between each sampling point. The mirror frame size parameters include the maximum horizontal width of the mirror frame, the maximum vertical height of the mirror frame, and the mirror frame circumference. Contour fitting is also a conventional technique in the field and can be implemented using curve fitting or boundary reconstruction methods. The specific process will not be elaborated here. Simultaneously, the edge computing processing unit extracts the frame structure parameters from the 3D point cloud data. These parameters include the bridge distance and frame curvature. The bridge distance is the minimum horizontal distance between the inner contours of the left and right lenses at the bridge of the nose. The frame curvature is a quantitative description of the overall arc of the frame, reflecting the degree of curvature from the front to the side, and can be obtained through conventional methods such as curvature fitting or arc calculation. The extraction method for the frame structure parameters is a conventional technique in this field, and the specific process will not be elaborated here. The lens contour parameters, lens size parameters, and frame structure parameters are integrated to form the 3D geometric data of the frame.

[0021] Methods for obtaining optometry prescription data include: The human-computer interface of the data acquisition terminal receives customer prescription information entered by the optometrist in the store. This prescription information consists of refractive correction parameters measured with an optometer and confirmed by the optometrist. The prescription data includes prescription records for both the left and right eyes. Each prescription record includes spherical power, cylindrical power, cylindrical axis, pupillary distance, and pupillary height. The spherical power is the spherical lens power required to correct myopia or hyperopia; the cylindrical power is the cylindrical lens power required to correct astigmatism; and the cylindrical axis is the astigmatism... The corrected axis angle value; pupillary distance is the horizontal distance between the centers of the pupils of the customer's left and right eyes; pupillary height is the vertical distance from the center of the pupil to the lower edge of the frame; if the prescription is a progressive multifocal prescription, the prescription record also includes near vision supplement and progressive path length; among them, near vision supplement is used to characterize the additional refractive power required for near vision, and is usually recorded separately for the left and right eyes with the same or similar values; progressive path length is the length of the transition area from the distance zone to the near zone in a progressive multifocal lens, and is a design parameter for the entire lens.

[0022] Methods for obtaining order configuration parameters include: The human-machine interface of the data acquisition terminal receives lens selection information entered by store operators based on customer needs. Among them, the order selection parameters include lens material type, lens refractive index grade, lens design type, and lens coating type. Lens material type is used to identify the type of substrate of the lens blank, including but not limited to resin, glass, and polycarbonate. Lens refractive index grade is used to identify the refractive index level of the lens, including but not limited to standard refractive index, medium refractive index, high refractive index, and ultra-high refractive index. Lens design type is used to identify the optical design of the lens, including but not limited to single-vision spherical lenses, single-vision aspherical lenses, bifocal lenses, and progressive multifocal lenses. Lens coating type is used to identify the type of functional coating required on the lens surface, including but not limited to anti-blue light coating, anti-ultraviolet coating, and anti-fog coating.

[0023] Methods for generating encrypted orders and uploading them to the cloud platform include: The edge computing processing unit generates a unique order code for the current order and obtains the terminal identification code of the data acquisition terminal; it integrates the three-dimensional geometric data of the eyeglass frame, the optometry prescription data, the order selection parameters, the order code, and the terminal identification code to form the original order data; Perform data integrity checks on the original order data. Specifically, check whether the lens outline parameters, lens size parameters, and lens structure parameters in the three-dimensional geometric data of the eyeglass frame are all non-empty; check whether the prescription records for the left and right eyes in the optometry prescription data both contain spherical power and pupillary distance; and check whether the lens material type and lens design type in the order selection parameters are both selected. If any check fails, a data missing prompt is returned to the human-computer interaction interface, requiring the store operator to supplement the data. If all checks pass, the original order data is considered complete. The original order data that has passed the integrity verification is encrypted to obtain an encrypted order, which is then uploaded to the cloud platform via a network communication interface. The encryption process uses an asymmetric encryption algorithm to encrypt the optometry prescription data and the three-dimensional geometric data of the eyeglass frame in the original order data to protect the security of customer privacy information during transmission. The asymmetric encryption algorithm is a well-known technology in this field, and the specific implementation process will not be described in detail here.

[0024] The difficulty calculation module is used by the cloud platform to dynamically calculate the processing difficulty coefficient corresponding to the encrypted order based on the optometry prescription data and order selection parameters in the encrypted order.

[0025] Methods for dynamically calculating the processing difficulty coefficient for encrypted orders include: After receiving an encrypted order, the cloud platform decrypts the frame's 3D geometric data, prescription data, and order selection parameters to obtain the decrypted frame's 3D geometric data, prescription data, and order selection parameters. The decryption operation employs an asymmetric decryption algorithm corresponding to the encryption processing of the data acquisition terminal. From the decrypted prescription data, the spherical and cylindrical power values ​​for the left and right eyes are obtained. From the order selection parameters, the lens material type, refractive index grade, and lens design type are obtained. For prescription complexity, a prescription complexity level is determined. Specifically, a pre-constructed prescription complexity grading rule table is used, containing multiple prescription grading rules. Each rule defines a spherical power range, a cylindrical power range, and a corresponding prescription complexity level. Prescription complexity levels include low complexity, medium complexity, and high complexity. For example, a spherical power absolute value less than or equal to four and a cylindrical power absolute value less than or equal to one is considered low complexity; a spherical power absolute value greater than four and less than or equal to eight or a cylindrical power absolute value greater than one and less than or equal to three is considered medium complexity; and a spherical power absolute value greater than eight or a cylindrical power absolute value greater than three is considered high complexity. Prescription complexity levels are determined for both left-eye and right-eye prescription records, and the higher of the two obtained levels is taken as the prescription complexity level for the encrypted order. The prescription complexity grading rule table is pre-constructed by those skilled in the art based on their technological experience in the field of eyeglass optical processing. For the material processing dimension, the difficulty level of material processing is determined. Specifically, a material difficulty mapping table is pre-constructed, which records the material processing difficulty level corresponding to each combination of lens material type and each lens refractive index level. The material processing difficulty level includes three levels: low difficulty, medium difficulty, and high difficulty. Based on the lens material type and lens refractive index level in the order selection parameters, the corresponding material processing difficulty level is looked up in the material difficulty mapping table. The material difficulty mapping table is pre-constructed by those skilled in the art based on the differences in the process difficulty of edge grinding under different combinations of materials and refractive indices. For example, the material processing difficulty level corresponding to the combination of resin material and standard refractive index is low difficulty, and the material processing difficulty level corresponding to the combination of polycarbonate material and high refractive index is high difficulty. To assess design complexity, a design complexity level is determined. Specifically, a pre-constructed design difficulty mapping table records the design complexity level for each lens design type. The design complexity levels are categorized into low, medium, and high complexity. Based on the lens design type in the order's optional parameters, the corresponding design complexity level is retrieved from the design difficulty mapping table. For example, a single-vision spherical lens corresponds to a low complexity level, while a progressive multifocal lens corresponds to a high complexity level. The design difficulty mapping table is pre-constructed by those skilled in the art based on the differences in edge grinding accuracy requirements for different design types. The system predefines the base difficulty values ​​for each of the following levels: prescription complexity, material processing difficulty, and design complexity. The base difficulty value for low complexity or low difficulty is less than that for medium complexity or medium difficulty, and the base difficulty value for medium complexity or medium difficulty is less than that for high complexity or high difficulty. Each base difficulty value is a value between zero and one. Simultaneously, the system predefines the dimension weights for each of the prescription complexity, material processing, and design complexity dimensions, with the sum of all dimension weights equal to one. Each base difficulty value is pre-set by those skilled in the art based on processing difficulty, equipment compatibility, processing time, and historical yield data during spectacle lens manufacturing. The base difficulty value for the prescription complexity level is determined based on the impact of different refractive power ranges on processing accuracy and finished product quality. The base difficulty value for the material processing difficulty level is determined based on the physical properties and processing compatibility of the lens material. The base difficulty value for the design complexity level is determined based on the complexity of the lens surface and the complexity of the processing path. The weights for each dimension are set based on the degree of impact of each dimension on the overall processing difficulty, processing cost, and yield rate, combined with the statistical analysis results of historical production data. Obtain the difficulty base corresponding to the judgment results of each dimension, and label them as prescription difficulty base, material difficulty base, and design difficulty base respectively; sum the products of prescription difficulty base and the dimension weight of prescription complexity dimension, material difficulty base and the dimension weight of material processing dimension, and design difficulty base and the dimension weight of design complexity dimension to obtain the processing difficulty coefficient; where the processing difficulty coefficient is a value between zero and one, the larger the value, the higher the processing difficulty of the order.

[0026] The intelligent scheduling module is used by the cloud platform to generate corresponding theoretical lens models and initial processing parameters based on encrypted orders, obtain the comprehensive status data of each intelligent processing node, dynamically evaluate the adaptability score of each intelligent processing node, select target intelligent processing nodes, and issue processing instructions containing theoretical lens models, initial processing parameters and processing difficulty coefficients to the target intelligent processing nodes.

[0027] Methods for generating theoretical lens models and initial processing parameters include: Based on the decrypted optometry prescription data, order selection parameters, and three-dimensional geometric data of the frame, the optical surface parameters, edge contour curves, and structural assembly parameters of the lens are calculated to form a theoretical lens model. Specifically, based on the spherical power, cylindrical power, and cylindrical axis in the optometry prescription data, combined with the lens refractive index grade and lens design type in the order selection parameters, optical surface shape parameters and lens center thickness are obtained through an optical surface shape calculation algorithm. The optical surface shape parameters include the curvature of the lens's front surface and the curvature of its back surface. The optical surface shape calculation algorithm is a well-known technology in the field, and its specific implementation process will not be elaborated upon here. If the lens design type is a progressive multifocal lens, a continuous diopter variation function along the progressive channel direction is constructed using an interpolation algorithm based on the near-use additional power and the progressive channel length. This function, combined with curvature calculation and surface fitting methods, yields the progressive surface shape distribution parameters, which are then added to the optical surface shape parameters. The calculation process for the progressive surface shape distribution parameters is a conventional technique in the field, and its specific process will not be elaborated upon here. The geometric center of the lens rim is determined based on the lens rim profile parameters, and the optical center of the lens is determined by combining the pupillary distance and pupillary height. The positioning offset of the lens within the lens rim is obtained by calculating the relative positional relationship between the geometric center of the lens rim and the optical center of the lens. The positioning offset and the center thickness of the lens are used as structural assembly parameters. The lens rim profile curve is projected based on the lens rim profile parameters, and the edge profile curve is obtained through curve fitting and processing compensation. The calculation process of the positioning offset and the edge profile curve is a conventional technique in this field, and the specific process will not be elaborated here. The optical surface parameters, the edge profile curve, and the structural assembly parameters are integrated to form a theoretical lens model.

[0028] Based on the theoretical lens model and the order selection parameters, initial processing parameters are generated; these initial processing parameters include grinding feed rate, grinding wheel speed, rough grinding allowance, and fine grinding allowance. Specifically, a pre-built material and process parameter library is constructed, which includes recommended grinding parameter ranges and allowance reference values ​​for different lens material types. This library is pre-constructed by those skilled in the art based on the processing characteristics of different lens materials. Based on the lens material type in the order's optional parameters, the corresponding recommended grinding parameter range is retrieved from the material and process parameter library. The recommended grinding parameter range includes a recommended upper limit for feed rate, a recommended lower limit for feed rate, a recommended upper limit for grinding wheel speed, and a recommended lower limit for grinding wheel speed. The allowance reference values ​​include recommended rough grinding allowance reference values ​​and recommended fine grinding allowance reference values. Based on the edge contour curve in the theoretical lens model, curvature variation features are extracted. Specifically, for each sampling point on the edge contour curve, the local curvature value at each sampling point is calculated. The local curvature value is obtained by calculating the reciprocal of the radius of the circumcircle determined by three adjacent sampling points. The local curvature value with the largest value among all sampling points is selected and marked as the maximum contour curvature. The standard deviation of the local curvature values ​​of all sampling points is calculated and marked as the curvature fluctuation. The maximum contour curvature and the curvature fluctuation are integrated to form the curvature variation feature. Based on the maximum profile curvature in the curvature change characteristics, the grinding feed speed is determined within the recommended grinding parameter range. Specifically, the corresponding upper and lower limits of the recommended feed speed are obtained from the recommended grinding parameter range. The maximum profile curvature is compared with a preset curvature reference value, and the ratio of the maximum profile curvature to the curvature reference value is calculated and marked as the curvature ratio. If the curvature ratio is greater than one, the curvature ratio is limited to one. The difference between the upper and lower limits of the recommended feed speed is calculated and marked as the adjustable range of the feed speed. The upper limit of the recommended feed speed is subtracted from the product of the adjustable range of the feed speed and the curvature ratio to obtain the grinding feed speed. The curvature reference value is preset by those skilled in the art based on the motion control accuracy characteristics of the edge grinding machine. Based on the curvature fluctuation in the curvature change characteristics, the grinding wheel speed is determined within the recommended grinding parameter range. Specifically, the upper and lower limits of the recommended grinding wheel speed are obtained from the recommended grinding parameter range; the ratio of the curvature fluctuation to the fluctuation reference value is calculated to obtain the fluctuation ratio; if the fluctuation ratio is greater than one, the fluctuation ratio is limited to one; the difference between the upper and lower limits of the recommended grinding wheel speed is calculated and marked as the speed adjustment range; the upper limit of the recommended grinding wheel speed is subtracted from the product of the speed adjustment range and the fluctuation ratio to obtain the grinding wheel speed; wherein, the fluctuation reference value is preset by those skilled in the art according to the actual situation. Based on the lens material type, edge contour curve, and lens center thickness, the rough grinding allowance and fine grinding allowance are determined. Specifically, based on the lens material type in the order's optional parameters, the corresponding allowance reference value is retrieved from the material process parameter library. The estimated edge thickness at each sampling point on the edge contour curve is calculated based on the edge contour curve and lens center thickness in the theoretical lens model. The estimated edge thickness is calculated using the lens thickness distribution formula, which is based on the lens's front surface curvature, rear surface curvature, and the radial distance from the sampling point to the lens's optical center. This formula is a well-known technique in the field, and its specific implementation is not detailed here. The estimated edge thickness with the smallest value among all sampling points is obtained and marked as the thinnest edge thickness. A thin edge protection threshold is preset, and the thinnest edge thickness is compared with the thinnest edge thickness. The edge protection threshold is compared; if the thinnest edge thickness is less than the thin edge protection threshold, the current lens is determined to have a thin edge risk, and the grinding allowance needs to be reduced to prevent excessive grinding from causing the lens to break; the recommended coarse grinding allowance benchmark value is multiplied by the preset thin edge allowance reduction coefficient to obtain the coarse grinding allowance; the recommended fine grinding allowance benchmark value is multiplied by the thin edge allowance reduction coefficient to obtain the fine grinding allowance; if the thinnest edge thickness is greater than or equal to the thin edge protection threshold, the recommended coarse grinding allowance benchmark value is directly used as the coarse grinding allowance, and the recommended fine grinding allowance benchmark value is used as the fine grinding allowance; where the thin edge protection threshold is the lower limit of safe thickness to ensure that the lens edge will not be ground through due to excessive allowance during the grinding process; the thin edge allowance reduction coefficient is a value greater than zero and less than one; both the thin edge protection threshold and the thin edge allowance reduction coefficient are preset by those skilled in the art according to the actual situation.

[0029] Methods for obtaining the comprehensive status data of each intelligent processing node include: The cloud platform sends status query commands to all online intelligent processing nodes; after receiving the status query commands, each intelligent processing node reports its own comprehensive node status data to the cloud platform; among them, the intelligent processing nodes are distributed processing stations equipped with high-precision automatic edge grinding machines, online vision measurement modules and edge computing controllers, and are deployed in processing centers or large retail stores in different geographical locations; each intelligent processing node has a unique node identification code; The node's comprehensive status data includes equipment load status, historical yield rate records, blank inventory data, and node geographical location. Equipment load status includes the number of orders awaiting processing in the current processing queue and their estimated completion time. Historical yield rate records include the processing success rate of the intelligent processing node for different processing difficulty coefficient ranges within a preset historical statistical period. The historical statistical period is preset by those skilled in the art based on actual conditions. Specifically, the range of processing difficulty coefficients is divided into multiple difficulty ranges, and the total number of completed orders and the number of qualified orders are counted for each difficulty range. The yield rate for each difficulty range is obtained by calculating the ratio of qualified orders to the total number of orders, thus forming a historical yield rate record. Blank inventory data includes the current inventory quantity of lens blanks of various material types and refractive index grades held by the intelligent processing node. The node geographical location is the geographical coordinate information of the intelligent processing node.

[0030] Methods for dynamically evaluating the adaptability score of each intelligent processing node include: For each intelligent processing node, an evaluation is conducted from four evaluation dimensions to obtain an evaluation score for each dimension. The four evaluation dimensions are difficulty matching, load balancing, inventory adaptation, and logistics timeliness. The evaluation method for the difficulty matching dimension is as follows: obtain the processing difficulty coefficient of the encrypted order, determine the difficulty range that the corresponding processing difficulty coefficient falls into; obtain the yield rate of the corresponding difficulty range from the historical yield rate records of the intelligent processing node, and use it as the difficulty matching score. The evaluation method for the load balancing dimension is as follows: count the number of orders to be processed in the load state of the corresponding equipment of the intelligent processing node to obtain the number of orders to be processed; preset the maximum load capacity threshold; calculate the difference between the number of orders to be processed and the ratio of the number of orders to be processed to the maximum load capacity threshold to obtain the load balancing score; if the load balancing score is less than zero, the load balancing score is set to zero; the maximum load capacity threshold shall be preset by those skilled in the art according to the actual situation. The evaluation method for inventory fit is as follows: based on the lens material type and lens refractive index grade in the order selection parameters, the corresponding inventory quantity is found in the raw material inventory data of the intelligent processing node; if the inventory quantity is greater than or equal to the preset inventory sufficiency threshold, the inventory fit score is set to one; if the inventory quantity is greater than zero and less than the inventory sufficiency threshold, the inventory fit score is set to the ratio of the inventory quantity to the inventory sufficiency threshold; if the inventory quantity is zero, the inventory fit score is set to zero; the inventory sufficiency threshold is preset by those skilled in the art based on the actual situation; The evaluation method for logistics timeliness is as follows: Based on the terminal identification code in the encrypted order, the corresponding retail terminal geographical location is searched in the terminal registration information table pre-stored on the cloud platform. The terminal registration information table is a data table maintained by the cloud platform, recording the terminal identification codes of all registered data collection terminals and the geographical coordinates of the corresponding deployed stores. When each data collection terminal first connects to the cloud platform, the store operations personnel enter the terminal identification code and store geographical coordinates into the terminal registration information table. Based on the geographical location of the intelligent processing node, the geographical distance between the retail terminal and the intelligent processing node is calculated. The method for calculating the geographical distance is a conventional technique in this field, and the specific implementation process will not be elaborated upon here. A maximum logistics distance threshold is preset, which is pre-set by those skilled in the art based on actual conditions. The difference between the logistics timeliness score and the ratio of the geographical distance to the maximum logistics distance threshold is calculated to obtain the logistics timeliness score. If the logistics timeliness score is less than zero, the logistics timeliness score is set to zero. For the difficulty matching dimension, load balancing dimension, inventory adaptation dimension, and logistics timeliness dimension, corresponding evaluation weights are set respectively. Among them, each evaluation weight is adaptively adjusted by those skilled in the art based on the processing difficulty coefficient. Specifically, when the processing difficulty coefficient is greater than or equal to the preset high difficulty threshold, the evaluation weight of the difficulty matching dimension is increased to the preset high difficulty weight value, and the evaluation weights of the other three evaluation dimensions are reduced proportionally. When the processing difficulty coefficient is less than the high difficulty threshold, each evaluation dimension adopts the preset default evaluation weight. The high difficulty threshold, high difficulty weight value, and default evaluation weight are all preset by those skilled in the art based on the actual situation. The product of the difficulty matching score and the evaluation weight of the difficulty matching dimension, the product of the load balancing score and the evaluation weight of the load balancing dimension, the product of the inventory adaptation score and the evaluation weight of the inventory adaptation dimension, and the product of the logistics timeliness score and the evaluation weight of the logistics timeliness dimension are summed to obtain the adaptation score of the corresponding intelligent processing node.

[0031] Methods for selecting target intelligent processing nodes and issuing processing instructions include: The adaptation scores of all intelligent processing nodes are sorted from largest to smallest, and the intelligent processing node with the largest adaptation score is selected as the target intelligent processing node. The theoretical lens model, initial processing parameters, processing difficulty coefficient and order code are integrated to form processing instructions, and the processing instructions are sent to the target intelligent processing node through the network communication interface.

[0032] The closed-loop machining module is used by the target intelligent machining node to input the machining difficulty coefficient and initial machining parameters into the pre-built local process prediction model to obtain optimized machining parameters, and then execute the closed-loop grinding process according to the optimized machining parameters and the theoretical lens model to obtain closed-loop machining data.

[0033] Methods for obtaining optimized processing parameters include: After receiving the processing command, the edge computing controller of the target intelligent processing node extracts the theoretical lens model, initial processing parameters, processing difficulty coefficient, and order code from the processing command. At the same time, the edge computing controller collects equipment status parameters, including the cumulative usage time of the grinding wheel, the current diameter of the grinding wheel, and the spindle vibration amplitude. The cumulative usage time of the grinding wheel is the cumulative grinding operation time since the last grinding wheel replacement. The current diameter of the grinding wheel is the actual diameter of the grinding wheel after wear, which is measured by the grinding wheel size sensor built into the grinding machine. The spindle vibration amplitude is the vibration amplitude of the grinding machine spindle in the running state, which is measured by the vibration sensor built into the grinding machine. The processing difficulty coefficient, initial processing parameters, and equipment status parameters are input to a pre-built local process prediction model to obtain optimized processing parameters. The local process prediction model is a lightweight neural network model deployed in an edge computing controller. It predicts the optimal adjustment amount of processing parameters based on current order characteristics and equipment status. This model is pre-built by those skilled in the art through sample construction and training based on historical processing data to establish the mapping relationship between the processing difficulty coefficient, initial processing parameters, equipment status parameters, and processing parameter adjustment amounts. Examples of suitable models include multilayer perceptron models and shallow fully connected neural networks. The input to the local process prediction model consists of the processing difficulty coefficient, the values ​​of each parameter in the initial processing parameters, and the values ​​of each parameter in the equipment status parameters. The concatenated vectors output the machining parameter adjustments. These adjustments include feed rate adjustment, grinding wheel speed adjustment, rough grinding allowance adjustment, and finish grinding allowance adjustment. The grinding feed rate from the initial machining parameters is added to the feed rate adjustment to obtain the optimized grinding feed rate. The grinding wheel speed from the initial machining parameters is added to the grinding wheel speed adjustment to obtain the optimized grinding wheel speed. The rough grinding allowance from the initial machining parameters is added to the rough grinding allowance adjustment to obtain the optimized rough grinding allowance. The finish grinding allowance from the initial machining parameters is added to the finish grinding allowance adjustment to obtain the optimized finish grinding allowance. The optimized grinding feed rate, optimized grinding wheel speed, optimized rough grinding allowance, and optimized finish grinding allowance are then integrated to form the optimized machining parameters.

[0034] Methods for performing closed-loop grinding processes include: The edge computing controller performs outward offset processing on the edge contour curve in the theoretical lens model according to the optimized rough grinding allowance, and generates a continuous rough grinding path through curve discretization and interpolation; drives the high-precision automatic edge grinding machine to perform the rough grinding process according to the rough grinding path, the optimized grinding feed speed, and the optimized grinding wheel speed, grinding the lens blank to near the edge contour curve and retaining the material allowance corresponding to the optimized rough grinding allowance, thus obtaining a semi-finished lens; After the rough grinding process is completed, the edge computing controller triggers the online vision measurement module to perform the first online measurement on the rough-ground lens semi-finished product. The online vision measurement module is a non-contact optical measurement device integrated on the edge grinding machine processing station, including a line laser sensor and a high-resolution industrial camera. The online vision measurement module scans the edge contour of the lens semi-finished product to obtain the rough grinding measured contour data. The rough grinding measured contour data includes the actual three-dimensional coordinate sequence of each sampling point of the edge contour of the lens semi-finished product. The edge computing controller compares the measured contour data of rough grinding with the edge contour curve in the theoretical lens model point by point, calculates the radial deviation value at each sampling point, and forms a rough grinding deviation distribution. Specifically, for each sampling point in the measured contour data of rough grinding, the nearest corresponding sampling point is found on the edge contour curve, and the radial distance difference between the two is calculated to obtain the radial deviation value of each sampling point. A positive radial deviation value indicates that there is more material at that point, and a negative radial deviation value indicates that there is less material at that point. The radial deviation values ​​of all sampling points are arranged in order of contour position to form a rough grinding deviation distribution. The calculation method of the radial deviation value is a conventional technique in this field and will not be described in detail here. The edge computing controller generates a compensated fine grinding path based on the rough grinding deviation distribution. Specifically, for sampling points with positive radial deviation values ​​in the rough grinding deviation distribution, the corresponding additional grinding amount is calculated. The additional grinding amount is equal to the radial deviation value minus the optimized fine grinding allowance. For sampling points with negative radial deviation values ​​or radial deviation values ​​less than the optimized fine grinding allowance, the corresponding additional grinding amount is set to zero. Based on the additional grinding amount and optimized fine grinding allowance of each sampling point, a compensated fine grinding path is generated. The edge grinding machine is then driven to perform the fine grinding process according to the compensated fine grinding path. After the fine grinding process is completed, the edge computing controller triggers the online vision measurement module to perform a second online measurement on the finely ground lens to obtain the finely ground measured contour data. The finely ground measured contour data is compared with the edge contour curve in the theoretical lens model point by point, and the radial deviation value at each sampling point is calculated again and marked as the residual deviation value to form the fine grinding deviation distribution. The system determines whether the absolute value of the residual deviation at all sampling points in the fine grinding deviation distribution is less than a preset processing qualification threshold. If all values ​​are less than the threshold, the current lens processing is deemed qualified, and the processing result is marked as qualified. If there are sampling points with an absolute value of residual deviation greater than or equal to the processing qualification threshold, compensation grinding is required. These sampling points are marked as compensation points. A local compensation path is generated based on all compensation points and their corresponding residual deviation values. The edge grinding machine is driven to perform compensation grinding according to the local compensation path, and the online visual measurement module is triggered again for measurement and deviation determination until the residual deviation values ​​at all sampling points meet the processing qualification threshold requirement or the preset maximum number of compensations is reached. If compensation points still exist after the maximum number of compensations is reached, the processing result is marked as unqualified. The processing qualification threshold and the maximum number of compensations are preset by those skilled in the art based on actual conditions. After the closed-loop grinding process is completed, the edge computing controller reports the order code and processing result to the cloud platform through the network communication interface. The cloud platform records the received order code and processing result as processing result feedback data and stores it.

[0035] Methods for obtaining closed-loop processing data include: The order code, processing difficulty coefficient, equipment status parameters, initial processing parameters, optimized processing parameters, processing parameter adjustment amount, rough grinding deviation distribution, fine grinding deviation distribution, and processing results are integrated to form closed-loop processing data.

[0036] The local training module is used by the target intelligent processing node to train the local process prediction model locally based on closed-loop processing data, generate encrypted model parameter update data, and upload it to the cloud platform.

[0037] Methods for locally training a local process prediction model based on closed-loop processing data include: The edge computing controller extracts data elements for training from the closed-loop processing data to construct training samples for this processing. Specifically, it concatenates the processing difficulty coefficient, initial processing parameters, and equipment status parameters in the closed-loop processing data to form a training input vector. The way the training input vector is constructed is consistent with the way the local process prediction model concatenates inputs when receiving inputs in the closed-loop processing module. Based on the fine grinding deviation distribution and processing results in the closed-loop machining data, a training label vector is constructed. Specifically, if the processing result is qualified, a training label vector is constructed based on the processing parameter adjustments in the closed-loop machining data. The training label vector includes the feed rate adjustment, grinding wheel speed adjustment, rough grinding allowance adjustment, and fine grinding allowance adjustment. If the processing result is unqualified or compensatory grinding is performed during the processing, the processing parameter adjustments are corrected and used as the training label vector. The correction method is as follows: obtain the fine grinding deviation distribution from the closed-loop machining data, and calculate the pairs of all sampling points in the fine grinding deviation distribution. The average residual deviation is obtained by taking the mean of the residual deviation values. Based on the sign and absolute value of the average residual deviation, the feed rate correction increment and the rough grinding allowance correction increment are determined. Specifically, if the average residual deviation is positive, it indicates that the overall size of the lens after fine grinding is too large, suggesting insufficient material removal during the rough grinding stage. In this case, the average residual deviation is multiplied by a preset feed rate correction ratio coefficient to obtain the feed rate correction increment. This increment is positive and is used to increase the feed depth in subsequent processing to remove more material. The average residual deviation is multiplied by a preset rough grinding allowance correction ratio coefficient to obtain the rough grinding allowance correction increment. The increment is negative, used to reduce the rough grinding allowance so that the rough grinding stage is closer to the final size; if the average residual deviation is negative, it indicates that the lens is too small after fine grinding, indicating that too much material was removed in the rough grinding stage; in this case, the absolute value of the average residual deviation is multiplied by the feed rate correction coefficient and the result is negative to obtain the feed rate correction increment, which is negative and used to reduce the feed depth in subsequent processing; the absolute value of the average residual deviation is multiplied by the rough grinding allowance correction coefficient to obtain the rough grinding allowance correction increment, which is positive and used to increase the rough grinding allowance so that more material is retained in the rough grinding stage; where, the feed rate correction increment is negative. The speed correction ratio and the rough grinding allowance correction ratio are both preset by those skilled in the art based on the empirical correspondence between the feed rate and the material removal amount of the edge grinding machine. The feed speed adjustment amount in the processing parameter adjustment amount is added to the feed speed correction increment to obtain the feed speed correction amount. The rough grinding allowance adjustment amount in the processing parameter adjustment amount is added to the rough grinding allowance correction increment to obtain the rough grinding allowance correction amount. The grinding wheel speed adjustment amount and the fine grinding allowance adjustment amount remain unchanged. The feed speed correction amount, grinding wheel speed adjustment amount, rough grinding allowance correction amount, and fine grinding allowance adjustment amount are combined to form a training label vector. The training input vector and the training label vector are combined to form a training sample pair; the edge computing controller stores the training sample pair in the local training sample cache queue; it determines whether the number of training sample pairs in the local training sample cache queue has reached the preset minimum training batch threshold; if not, local training is temporarily suspended, and more training samples are generated in subsequent order processing; if the minimum training batch threshold is reached or exceeded, all training sample pairs are retrieved from the local training sample cache queue to form a local training batch; the minimum training batch threshold is preset by those skilled in the art according to the actual situation. Before performing local training, the current model parameters of the local process prediction model are recorded and marked as pre-training model parameters. The local process prediction model is trained using stochastic gradient descent, with the training input vector as the input and the training label vector as the expected output. The mean squared error loss between the output of the local process prediction model and the training label vector is calculated, and the model parameters are updated using backpropagation. The number of training iterations is a preset number of local training rounds. Both stochastic gradient descent and backpropagation algorithms are well-known technologies in the field, and their specific implementation processes are not detailed here. The number of local training rounds is preset by those skilled in the art based on actual conditions. After local training is completed, the model parameters of the local process prediction model are recorded and marked as post-training model parameters. The difference between the post-training model parameters and the pre-training model parameters is calculated to obtain the model parameter update amount.

[0038] Methods for generating and uploading encrypted model parameter updates to the cloud platform include: Differential privacy processing is performed on the model parameter update amounts to obtain model parameter update amounts with added differential privacy noise. Specifically, for each parameter component in the model parameter update amounts, random noise following a preset distribution is added to it to obtain noisy parameter components. The magnitude of the random noise is controlled by a preset privacy budget parameter; the smaller the privacy budget parameter, the greater the added noise and the stronger the privacy protection. Differential privacy processing is a well-known technique in the art, used to prevent sensitive information of the original training data from being deduced from the model parameter update amounts. The privacy budget parameter is preset by those skilled in the art based on the balance requirements between privacy protection needs and model training accuracy. Homomorphic encryption is applied to the model parameter update values ​​with added differential privacy noise to obtain encrypted model parameter update values. Homomorphic encryption is a well-known technology in the field, which enables the cloud platform to perform aggregation operations on multiple encrypted model parameter update values ​​without decrypting the parameters. The specific implementation process of homomorphic encryption will not be elaborated here. The encrypted model parameter update quantity is integrated with the node identifier code of the target intelligent processing node and the number of training samples in the local training batch to form an encrypted update upload package, which is then uploaded to the cloud platform through the network communication interface.

[0039] The federated aggregation module is used by the cloud platform to aggregate the encrypted model parameter updates uploaded by each intelligent processing node, generate a globally updated process model, and distribute it to each intelligent processing node.

[0040] Methods for aggregating the encrypted model parameter updates uploaded by each intelligent processing node include: The cloud platform sets a federated aggregation cycle. When each federated aggregation cycle arrives, it collects the encrypted update upload packets uploaded by all intelligent processing nodes within the current federated aggregation cycle. The federated aggregation cycle is a preset fixed time interval, which is preset by those skilled in the art based on the order processing frequency and model update timeliness requirements of each intelligent processing node. Extract the encrypted model parameter update amount, node identifier code, and number of training sample pairs from each encrypted update upload packet; count the number of encrypted update upload packets received in the current federated aggregation cycle and mark them as the number of participating aggregation nodes; preset a minimum aggregation node threshold and compare the number of participating aggregation nodes with the minimum aggregation node threshold; if the number of participating aggregation nodes is less than the minimum aggregation node threshold, abandon the current weighted aggregation operation and wait for the next federated aggregation cycle; if the number of participating aggregation nodes is greater than or equal to the minimum aggregation node threshold, perform the weighted aggregation operation; wherein, the minimum aggregation node threshold is preset by those skilled in the art according to the actual situation.

[0041] The method for performing weighted aggregation operation is as follows: all intelligent processing nodes participating in the weighted aggregation operation are marked as participating aggregation nodes, the number of training sample pairs of each participating aggregation node is obtained, and the sum of the number of all training sample pairs is calculated to obtain the total number of training samples; for each participating aggregation node, the ratio of the corresponding number of training sample pairs to the total number of training samples is calculated to obtain the sample size aggregation weight of the corresponding participating aggregation node. Simultaneously, for each participating aggregation node, the processing results of all orders completed by the corresponding participating aggregation node within the current federated aggregation cycle are statistically analyzed from the processing result feedback data; the ratio of the number of orders with qualified processing results to the total number of orders completed by the corresponding participating aggregation node within the current federated aggregation cycle is calculated to obtain the node cycle yield rate; the sum of the node cycle yield rates of all participating aggregation nodes is calculated to obtain the total yield rate; for each participating aggregation node, the ratio of the corresponding node cycle yield rate to the total yield rate is calculated to obtain the quality aggregation weight of the corresponding participating aggregation node; A pre-defined aggregation weight balance coefficient is used between the sample size aggregation weight and the quality aggregation weight. For each participating aggregation node, the product of the aggregation weight balance coefficient and the sample size aggregation weight is calculated, and the product of the difference between one and the aggregation weight balance coefficient and the quality aggregation weight is calculated. The two product results are added together to obtain the comprehensive aggregation weight of the corresponding participating aggregation node. The aggregation weight balance coefficient is a value between zero and one, which is preset by those skilled in the art according to the actual situation. For the updated encrypted model parameters of each participating aggregation node, a scalar multiplication operation is performed on it and the corresponding comprehensive aggregation weight in the ciphertext space to obtain the weighted encrypted model parameters. Then, an element-wise addition operation is performed on the weighted encrypted model parameters of all participating aggregation nodes in the ciphertext space to obtain the updated encrypted global parameters. The scalar multiplication and element-wise addition operations in the ciphertext space are guaranteed to be equivalent to the corresponding operations in the plaintext space by the mathematical properties of the homomorphic encryption scheme.

[0042] Methods for generating a globally updated process model and distributing it to each intelligent machining node include: Perform a decryption operation on the encrypted global parameter update amount to obtain the decrypted global parameter update amount; obtain the global model parameters of the global process model stored on the cloud platform; add the global model parameters to the decrypted global parameter update amount element by element to obtain the updated global model parameters; replace the original global model parameters with the updated global model parameters to form a global updated process model. The global process model is a process parameter prediction model deployed on a cloud platform. Its network architecture is exactly the same as the local process prediction model in each intelligent processing node, i.e., it has the same number of network layers, the same number of nodes per layer, and the same input and output dimensions. The global process model is used as an aggregation carrier of process knowledge across the entire network under the federated learning framework, storing model parameters that represent the collective process experience of all intelligent processing nodes after multiple rounds of federated aggregation. The initialization method of the global process model is as follows: when the system is first deployed, a person skilled in the art pre-trains a lightweight neural network model based on historical processing data to obtain initial global model parameters. The initial global model parameters are stored in the cloud platform as the initial version of the global process model. At the same time, the initial global model parameters are distributed to all intelligent processing nodes as the initial model parameters of each node's local process prediction model. Subsequently, as each intelligent processing node continues to process orders and uploads encrypted model parameter updates, the cloud platform continuously updates the parameters of the global process model through federated aggregation operations, and periodically distributes the updated parameters to each node to replace its local model parameters, so that the overall network process level continues to evolve and improve over time. Record the version number of the globally updated process model, which is an incrementing integer; integrate the updated global model parameters corresponding to the globally updated process model with the version number to form a global model distribution package, and distribute the global model distribution package to all intelligent processing nodes through the network communication interface; after receiving the global model distribution package, each intelligent processing node extracts the updated global model parameters and version number from the global model distribution package; replace the model parameters of the local process prediction model in the edge computing controller with the updated global model parameters to complete the update of the local process prediction model.

[0043] This embodiment comprehensively analyzes the three-dimensional geometric data of the eyeglass frame, optometry prescription data, and order selection parameters to more accurately characterize the actual processing characteristics of the order. Based on this, it dynamically assesses the processing difficulty, generates theoretical lens models and initial processing parameters, thereby improving the pertinence and rationality of front-end process decisions. During manufacturing execution, it combines equipment load, historical yield rate, inventory status, and geographical location of each processing node for adaptation evaluation, enabling intelligent matching between complex orders and processing resources. This not only improves node utilization and order flow efficiency but also helps shorten delivery cycles and reduce scheduling imbalance risks. Introducing online measurement, deviation analysis, and compensation grinding mechanisms during lens grinding allows for dynamic correction of the processing path and process parameters based on the actual contour deviations after rough and fine grinding, thus improving processing control. The process has shifted from traditional experience-based open-loop execution to closed-loop optimization control based on measured feedback. This effectively reduces edge size errors, minimizes over- or under-grinding, and improves lens processing accuracy, dimensional consistency, and finished product yield. Simultaneously, each processing node trains its model locally based on actual processing results, and uploads parameter updates using differential privacy and homomorphic encryption. This ensures the security of order data, prescription data, and local processing data at each node, while enabling cross-node process knowledge sharing and global model aggregation updates. This not only continuously improves the process prediction model's adaptability to different materials, design types, and equipment states but also enhances the system's responsiveness to complex orders and dynamic production environments, ultimately achieving high-precision, high-efficiency, intelligent, and safe collaborative manufacturing in the lens processing process. Example 2:

[0044] This application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code, which, when executed by the one or more processors, can perform the distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control as described above.

[0045] The system according to the embodiments of this application can also be implemented using the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. The storage device in the electronic device, such as a ROM or hard disk, may store the distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control provided in this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is merely exemplary; when implementing different devices, one or more components of the electronic device shown in this application may be omitted according to actual needs. Example 3:

[0046] One embodiment of this application discloses a computer-readable storage medium. The computer-readable storage medium stores computer-readable instructions. When the computer-readable instructions are executed by a processor, they can execute the distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control according to an embodiment of this application, as described with reference to the above figures. The storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0047] Furthermore, according to embodiments of this application, the processes described in the above-referenced flowcharts can be implemented as computer software programs. For example, this application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to perform instructions corresponding to the method steps provided in this application, such as a distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control. When this computer program is executed by a central processing unit (CPU), it performs the functions defined in the method of this application.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0049] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0050] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control, characterized in that, include: The order acquisition module is used by the data acquisition terminal to obtain the three-dimensional geometric data of the eyeglass frame, the optometry prescription data, and the order selection parameters. It then generates an encrypted order by summarizing the data and uploads it to the cloud platform. The difficulty calculation module is used by the cloud platform to dynamically calculate the processing difficulty coefficient of the encrypted order based on the optometry prescription data and order selection parameters in the encrypted order. Methods for dynamically calculating the processing difficulty coefficient of encrypted orders include: The three-dimensional geometric data of the eyeglass frame, the optometry prescription data, and the order selection parameters in the encrypted order are decrypted to obtain the decrypted three-dimensional geometric data of the eyeglass frame, the optometry prescription data, and the order selection parameters. Based on the spherical and cylindrical power values ​​in the optometry prescription data, the prescription complexity level is determined according to the prescription complexity dimension; based on the lens material type and refractive index level in the order selection parameters, the material processing difficulty level is determined according to the material processing dimension; based on the lens design type in the order selection parameters, the design complexity level is determined according to the design complexity dimension. The system presets the base difficulty values ​​for each of the following levels: prescription complexity level, material processing difficulty level, and design complexity level. It also presets the dimension weights for each of the following dimensions: prescription complexity dimension, material processing dimension, and design complexity dimension. The system obtains the base difficulty values ​​corresponding to the judgment results of each dimension, and calculates the weighted sum of each base difficulty value based on the dimension weights of each dimension to obtain the processing difficulty coefficient. The intelligent scheduling module is used by the cloud platform to generate corresponding theoretical lens models and initial processing parameters based on encrypted orders, obtain the comprehensive status data of each intelligent processing node, dynamically evaluate the adaptability score of each intelligent processing node, select target intelligent processing nodes, and issue processing instructions containing theoretical lens models, initial processing parameters and processing difficulty coefficients to the target intelligent processing nodes. The closed-loop machining module is used by the target intelligent machining node to input the machining difficulty coefficient and initial machining parameters into the pre-built local process prediction model to obtain optimized machining parameters, and then execute the closed-loop grinding process according to the optimized machining parameters and the theoretical lens model to obtain closed-loop machining data. The local training module is used by the target intelligent processing node to train the local process prediction model locally based on closed-loop processing data, generate encrypted model parameter update data, and upload it to the cloud platform. The federated aggregation module is used by the cloud platform to aggregate the encrypted model parameter updates uploaded by each intelligent processing node, generate a globally updated process model, and distribute it to each intelligent processing node.

2. The distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control according to claim 1, characterized in that, Methods for generating encrypted orders include: A unique order code is generated for the current order, and the terminal identification code of the data acquisition terminal is obtained; the three-dimensional geometric data of the eyeglass frame, the optometry prescription data, the order selection parameters, the order code and the terminal identification code are integrated to form the original order data; the original order data is subjected to data integrity verification, and the original order data that passes the integrity verification is encrypted to obtain the encrypted order; The three-dimensional geometric data of the eyeglass frame includes the lens outline parameters, lens size parameters, and eyeglass frame structure parameters; the prescription data includes prescription records for the left and right eyes, and each prescription record includes spherical power, cylindrical power, cylindrical axis, pupillary distance, and pupillary height; the order selection parameters include lens material type, lens refractive index grade, lens design type, and lens coating type.

3. The distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control according to claim 2, characterized in that, Methods for generating theoretical lens models and initial processing parameters include: Based on the decrypted optometry prescription data, order selection parameters, and three-dimensional geometric data of the frame, the optical surface parameters, edge contour curves, and structural assembly parameters of the lens are calculated to form a theoretical lens model; among them, the structural assembly parameters include the lens center thickness and positioning offset. Based on the lens material type in the order's optional parameters, the corresponding recommended grinding parameter range is searched from the pre-built material and process parameter library; based on the edge contour curve in the theoretical lens model, the maximum contour curvature and curvature fluctuation are extracted. Determine the grinding feed rate within the recommended grinding parameter range based on the maximum contour curvature; determine the grinding wheel speed within the recommended grinding parameter range based on the curvature fluctuation; determine the roughing allowance and fine grinding allowance based on the lens material type, edge contour curve, and lens center thickness. The grinding feed rate, grinding wheel speed, rough grinding allowance and finish grinding allowance are integrated to form the initial machining parameters.

4. The distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control according to claim 3, characterized in that, The methods for evaluating the adaptability score of each intelligent processing node include: For each intelligent processing node, an evaluation is conducted from four evaluation dimensions based on the processing difficulty coefficient, order selection parameters, terminal identification code, and corresponding node comprehensive status data, and an evaluation score is obtained for each evaluation dimension. The four evaluation dimensions are difficulty matching dimension, load balancing dimension, inventory adaptation dimension, and logistics timeliness dimension. For the dimensions of difficulty matching, load balancing, inventory adaptation, and logistics timeliness, corresponding evaluation weights are set. Based on the evaluation weights corresponding to each evaluation dimension, the evaluation scores of each evaluation dimension under the same intelligent processing node are weighted and summed to obtain the adaptation score of the corresponding intelligent processing node.

5. The distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control according to claim 4, characterized in that, Methods for obtaining optimized processing parameters include: The initial processing parameters and processing difficulty coefficient are extracted from the processing instructions, and the equipment status parameters are collected. The processing difficulty coefficient, initial processing parameters, and equipment status parameters are used as inputs to the pre-built local process prediction model, and the processing parameter adjustment amount is output. Among them, the processing parameter adjustment amount includes the feed rate adjustment amount, grinding wheel speed adjustment amount, rough grinding allowance adjustment amount, and fine grinding allowance adjustment amount. The grinding feed rate, grinding wheel speed, rough grinding allowance, and fine grinding allowance in the initial machining parameters are summed with the corresponding adjustment values ​​in the machining parameter adjustment values ​​to obtain the optimized grinding feed rate, optimized grinding wheel speed, optimized rough grinding allowance, and optimized fine grinding allowance in sequence. These are then integrated to form the optimized machining parameters.

6. The distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control according to claim 5, characterized in that, Methods for performing closed-loop grinding processes include: Based on the edge contour curves in the theoretical lens model and optimized processing parameters, a rough grinding path is generated. Following this path, and with optimized grinding feed rate and wheel speed, a rough grinding process is performed on the lens blank to obtain a semi-finished lens. After the rough grinding process, the semi-finished lens undergoes its first online measurement to obtain measured contour data. This measured contour data is then compared point-by-point with the edge contour curves in the theoretical lens model to form a rough grinding deviation distribution. A compensating fine grinding path is generated based on this deviation distribution, and the fine grinding process is performed according to this path. After the fine grinding process is completed, a second online measurement is performed on the finely ground lens to obtain the finely ground measured contour data. The finely ground measured contour data is compared point by point with the edge contour curve in the theoretical lens model to form the fine grinding deviation distribution. The absolute value of the residual deviation value of all sampling points in the fine grinding deviation distribution is compared with the preset processing qualification threshold. Based on the comparison results, it is determined whether to perform compensation grinding and the corresponding processing result is marked. The processing result includes qualified and unqualified.

7. The distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control according to claim 6, characterized in that, Methods for locally training local process prediction models include: The processing difficulty coefficient, initial processing parameters, and equipment status parameters in the closed-loop processing data are concatenated to form a training input vector. If the processing result is qualified, a training label vector is constructed based on the processing parameter adjustment amount in the closed-loop processing data. If the processing result is unqualified or compensatory grinding is performed, the processing parameter adjustment amount is corrected and used as the training label vector. The training input vector and the training label vector are combined to form training sample pairs and stored in the local training sample cache queue. It is determined whether the number of training sample pairs in the local training sample cache queue has reached the preset minimum training batch threshold. If so, all training sample pairs are retrieved to form a local training batch and local training is performed. Before performing local training, record the current model parameters of the local process prediction model and mark them as pre-training model parameters; train the local process prediction model with the training input vector as input and the training label vector as the expected output, calculate the mean squared error loss between the output of the local process prediction model and the training label vector, and update the model parameters; after local training is completed, record the model parameters of the local process prediction model and mark them as post-training model parameters; calculate the model parameter update amount based on the post-training model parameters and the pre-training model parameters.

8. The distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control according to claim 7, characterized in that, Methods for generating and uploading encrypted model parameter updates to the cloud platform include: Differential privacy processing is performed on the model parameter update amount to obtain the model parameter update amount with added differential privacy noise; homomorphic encryption processing is performed on the model parameter update amount with added differential privacy noise to obtain the encrypted model parameter update amount; the encrypted model parameter update amount is integrated with the number of training sample pairs to form an encrypted update upload package, and uploaded to the cloud platform through the network communication interface.

9. The distributed collaborative manufacturing system for eyeglass lenses based on federated learning and closed-loop control according to claim 8, characterized in that, Methods for aggregating the encrypted model parameter updates uploaded by each intelligent processing node include: Set a federated aggregation cycle. When each federated aggregation cycle arrives, collect the encrypted update upload packets uploaded by all intelligent processing nodes in the current federated aggregation cycle; extract the encrypted model parameter update amount and the number of training sample pairs from each encrypted update upload packet; count the number of encrypted update upload packets received in the current federated aggregation cycle and mark them as the number of participating aggregation nodes; preset a minimum aggregation node threshold and compare the number of participating aggregation nodes with the minimum aggregation node threshold. If the number of participating aggregation nodes is less than the minimum aggregation node threshold, the current weighted aggregation operation is abandoned, and the process waits for the next federated aggregation cycle. If the number of participating aggregation nodes is greater than or equal to the minimum aggregation node threshold, the weighted aggregation operation is performed: all intelligent processing nodes participating in the weighted aggregation operation are marked as participating aggregation nodes, the sample size aggregation weight and quality aggregation weight of each participating aggregation node are calculated, and the weights are balanced based on a preset aggregation weight balance coefficient to obtain the comprehensive aggregation weight. Based on the comprehensive aggregation weight, the encrypted model parameter update amount of each participating aggregation node is weighted and summed in the ciphertext space to obtain the encrypted global parameter update amount.

Images