Process feedback system for contact lens production lines

Abstract

The present application relates to the technical field of production line data management, in particular to a process feedback system of a contact lens production line, which acquires real-time operation data of each section of the contact lens production line in real time, extracts multi-time window features from the real-time operation data, uses a defect prediction model to predict the multi-time window features, and obtains a defect prediction result. At the same time, the image is detected for defects by using machine vision technology, and a defect detection result is obtained. Finally, the defect prediction result and the defect detection result are sliced by batch, the data samples obtained by each slice are analyzed for deviation, and a production line adjustment scheme is generated based on the result of the deviation analysis. Based on the operation data of each device in the whole line, the prediction model predicts potential defects in advance based on real-time data, and the actual detection result calibrates the prediction algorithm in reverse, forming a dynamic optimization mechanism that verifies each other. The optimization system breaks through the data fragmentation of traditional systems and truly realizes fine quality control of the production process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of production line data management, in particular to a process feedback system for contact lens production line. BACKGROUND

[0002] The production process of contact lenses includes multiple key process steps such as pad printing, mold assembly, mold disassembly, and hydration, each of which has a significant impact on the quality of the final product. Due to the complexity of production conditions and process parameters, production defects of contact lenses are often difficult to detect in advance, and are usually revealed during final product inspection. Traditional quality control methods usually rely on final product inspection and cannot provide effective early warning of defects in the production process. Without early intervention, it may lead to the production of a large number of unqualified products, increasing production costs and waste. In recent years, with the development of data collection and machine learning technology, a large amount of process parameters and production data can be collected and analyzed during the production process, which provides the possibility for defect prediction based on data. In order to improve the production efficiency and quality of contact lenses, it is necessary to use prediction technologies such as machine learning to identify potential production defects in advance by analyzing production data and provide real-time adjustment suggestions during the production process.

[0003] In the prior art, only process parameters are used for prediction, and there is a lack of effective correlation with actual defect detection results; another part only focuses on the final inspection link of finished products, ignoring the key role of early warning in the production process, resulting in the inability to form a collaborative optimization mechanism throughout the production line. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a process feedback system for contact lens production line to solve the problems in the background art.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] The process feedback system for contact lens production line of the present application comprises:

[0007] The data acquisition module is used to acquire real-time running data of each section of the contact lens production line and acquire contact lens images output by each process equipment of the contact lens production line, wherein the real-time running data includes real-time environmental data, real-time process data and real-time equipment data, the real-time environmental data includes time sequence value sequences of multiple environmental parameters, the real-time process data includes time sequence value sequences of multiple process parameters, and the real-time equipment data includes time sequence value sequences of multiple equipment running parameters;

[0008] a prediction module configured to extract feature vectors of multiple time windows from the real-time operation data, and input the feature vectors into a pre-constructed defect prediction model to obtain a defect prediction result, wherein the defect prediction result comprises a defect type and a defect quantity;

[0009] a detection module configured to perform defect detection on the contact lens image to obtain a defect detection result;

[0010] an analysis module configured to slice the defect detection results and the defect prediction results of multiple batches of contact lenses to obtain multiple data samples, perform deviation analysis on the defect prediction results and the defect detection results of the contact lenses in the data samples to obtain a deviation analysis result, and select the data samples to obtain target sample data, wherein the target sample data is used to adjust and train the defect prediction model;

[0011] an adjustment module configured to generate a production line adjustment scheme based on the deviation analysis result, and feed back the production line adjustment scheme and the trend analysis result to a target object.

[0012] The contact lens production line process feedback system of the present application has the following advantages: the real-time operation data of each section of the contact lens production line is collected in real time, and the multiple time window features in the real-time operation data are extracted, the defect prediction model is used to predict the multiple time window features to obtain the defect prediction result. At the same time, the machine vision technology is used to detect the defects of the image to obtain the defect detection result. Finally, the defect prediction result and the defect detection result are sliced by batch, the deviation analysis is performed on the data samples obtained from each slice, and the production line adjustment scheme is generated based on the deviation analysis result. Based on the process data and environmental data of each device in the whole line, the prediction model predicts potential defects in advance based on real-time data, the actual detection result is used to calibrate the prediction algorithm in reverse based on the self-developed optical detection system, and a dynamic optimization mechanism is formed for mutual verification. The optimization system breaks through the data fragmentation of the traditional system and truly realizes the fine quality control of the production process. BRIEF DESCRIPTION OF DRAWINGS

[0013] The present application will be further described below in conjunction with the drawings and examples:

[0014] Figure 1 is a structure diagram of the contact lens production line process feedback system in an embodiment of the present application;

[0015] Figure 2 is a whole process schematic diagram in an embodiment of the present application;

[0016] Figure 3 is a prediction process schematic diagram in an embodiment of the present application;

[0017] Figure 4 a detection module in an embodiment of the present application runs a flow chart;

[0018] Figure 5 a schematic diagram of a disc image in an embodiment of the present application;

[0019] Figure 6 a schematic diagram of a parameter adjustment flow in an embodiment of the present application;

[0020] Figure 7 a schematic diagram of a scheme issuing flow in an embodiment of the present application;

[0021] Figure 8 a schematic diagram of an effect tracking and archiving flow in an embodiment of the present application. DETAILED DESCRIPTION

[0022] The present application is described in more detail by the specific embodiments below, and other advantages and effects of the present application can be easily understood by those skilled in the art from the content disclosed in the specification. The present application can also be implemented or applied by different specific embodiments, and various modifications or changes can be made to the details in the specification based on different views and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0023] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and only the layers related to the present application are shown in the diagrams, not the number of layers, shapes and sizes when actually implemented. The actual implementation of each layer pattern, quantity and proportion can be a random change, and the layer layout pattern can also be more complex.

[0024] In the following description, a large number of details are discussed to provide a more thorough explanation of the embodiments of the present application, however, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details.

[0025] Figure 1 is a structural diagram of a process feedback system of a contact lens production line shown in an embodiment of the present application, as shown in Figure 1 The process feedback system of the contact lens production line of the present embodiment includes data acquisition module 110, prediction module 120, detection module, analysis module 140, adjustment module 150, the functions and principles of each module are described as follows:

[0026] The data acquisition module 110 is configured to acquire real-time running data of each section of the contact lens production line and acquire contact lens images output by each process equipment of the contact lens production line, wherein the real-time running data comprises real-time environment data, real-time process data and real-time equipment data, the real-time environment data comprises time sequence value sequences of a plurality of environment parameters, the real-time process data comprises time sequence value sequences of a plurality of process parameters, and the real-time equipment data comprises time sequence value sequences of a plurality of equipment running parameters.

[0027] The prediction module 120 is configured to extract feature vectors of a plurality of time windows from the real-time running data, input the feature vectors into a pre-constructed defect prediction model, and obtain a defect prediction result, wherein the defect prediction result comprises a defect type and a defect product quantity.

[0028] The detection module 130 is configured to perform defect detection on the contact lens images to obtain a defect detection result.

[0029] The analysis module 140 is configured to slice the defect detection results and the defect prediction results of a plurality of batches of contact lenses to obtain a plurality of data samples, perform deviation analysis on the defect prediction results and the defect detection results of the contact lenses in the data samples to obtain a deviation analysis result, and select the data samples to obtain target sample data, wherein the target sample data is used to adjust and train the defect prediction model.

[0030] The adjustment module 150 is configured to generate a production line adjustment scheme based on the deviation analysis result and feed back the production line adjustment scheme and the trend analysis result to a target object.

[0031] Figure 2 For the overall flowchart in an embodiment of the present application, please refer to Figure 2 To understand the principles of the following functional modules in the present application, specifically including:

[0032] A) Data acquisition module: self-developed data acquisition system for task-based, workflow management data acquisition of the whole process. For different data acquisition needs, it is packaged as an independent task, and double-parallel tasks are generated in a high-availability scenario to realize double-path acquisition.

[0033] The data acquisition module adopts a double-channel hot standby mechanism through an OPC UA, ModBus, TCP / IP, API and other multi-protocol adaptation engine, combines a subscription real-time push and a periodic polling hybrid mode, and high-concurrency reads PLC registers, device state bits, sensors and other values to generate device, environment and process data such as device state changes, temperature curves and pressure.

[0034] Taking a certain contact lens intelligent production line process as an example, the production procedures are: injection molding, vertical warehouse buffering, filling and molding, curing, mold separation and piece taking, hydration immersion, full inspection and packaging.

[0035] The key detection procedures are: filling and molding stage (detecting bubbles, dirt), full inspection and packaging stage (detecting scratches, burrs, edge damage and other defects).

[0036] The process feedback system is divided into two dimensions of parameter feedback: full production line dimension key process parameters, equipment dimension key process parameters. In addition, environmental parameters are collected through sensors.

[0037] In addition, the applicant artificially presets the adjustment gradient and range of each process parameter, covering the reasonable floating interval above and below the normal production threshold, typical values, etc.

[0038] According to the preset parameter adjustment scheme, the production line repeatedly runs corresponding production procedures, and synchronously triggers the data acquisition system to continuously collect full-quantity production data under different parameter combinations.

[0039] Through the cyclic operation of multiple rounds of parameter adjustment and data acquisition, it is ensured that all possible values and combinations of process parameters within the set range are covered, forming a complete parameter-data mapping library containing boundary values, critical values and regular values, providing data support for subsequent modeling.

[0040] B) Prediction module

[0041] Based on the real-time collection of multi-dimensional data of all devices in the whole line, the collected real-time data are input into a pre-trained model trained and optimized by a large amount of historical production data, and the quality condition of semi-finished products and finished products is analyzed and predicted by using a machine learning algorithm, and a prediction result containing defect type and occurrence probability (or defect product quantity) is output, which provides a prediction basis for subsequent production quality control. The prediction model in the application is constructed based on a Transformer-GCN fusion architecture, which uses Transformer to capture the time sequence dependency of data and uses a graph convolution network (GCN) to mine the correlation between devices. The multi-head attention mechanism is used to adaptively extract key features, and the graph attention mechanism is used to learn the device interaction mode. The model supports multi-task learning and can simultaneously predict multiple key quality indicators.

[0042] Figure 3 The prediction process in an embodiment of the application is shown in the schematic diagram as shown in Figure 3 The specific prediction process is as follows:

[0043] B1) pre-processing the real-time running data to obtain pre-processed data, wherein the pre-processing includes standardization and normalization;

[0044] First, the original data is standardized and normalized to construct a multi-dimensional feature matrix. This facilitates the generation of dynamic feature vectors using a sliding window mechanism in subsequent processes. The window size can be automatically or manually adjusted according to process characteristics.

[0045] The present application trains the model and predicts the data based on three dimensions: parameter dimension, equipment dimension, and process section dimension. Depending on the characteristics and feedback targets of different types of process data, the corresponding dimensions are matched for model training and data prediction. The three dimensions are B2-B4, as follows:

[0046] B2) Based on the pre-constructed sliding window, the fluctuation trend features of the target parameters in the pre-processed data in multiple time windows are extracted. The fluctuation trend features are input into the pre-constructed first defect prediction model to obtain the first defect prediction result. The fluctuation trend features include mean, rising slope, and falling slope.

[0047] The mean can be calculated by calling the average function directly. Details are omitted here.

[0048] The rising slope and falling slope can be obtained by linear fitting, for example:

[0049] A linear fitting equation is constructed

[0050]

[0051] where, represents the parameter value, is the slope, represents time, is the error term;

[0052] Substituting the target parameters in the pre-processed data into the above linear fitting equation for fitting, the rising slope or falling slope can be obtained.

[0053] Parameter dimension: Focus on the impact of dynamic changes of a single key process parameter on quality, such as real-time collection of environmental parameters or single operation parameters of equipment in a specific area. Combine historical data of the same period of the parameter and product quality, and extract the fluctuation trend features of the parameter in the continuous period through a sliding window. The model analyzes the amplitude and duration of the parameter deviation from the standard interval, predicts the possible changes in material properties or quality risks in subsequent processes, and provides a preliminary judgment for targeted adjustment of a single parameter.

[0054] For example, for the parameter of ambient temperature of the stand buffer area, the temperature data under different temperature conditions in the historical production and the quality detection results of the corresponding batches in the subsequent filling and molding stage are collected in the full range to train the correlation model of temperature and filling and molding quality. In real-time production, the temperature data of this area are continuously collected, the model analyzes the temperature fluctuation characteristics to predict the possible quality impact of filling and molding process, issues a warning and prompts the adjustment direction of the temperature parameter.

[0055] B3) Extracting the fluctuation time sequence feature matrix of the time sequence value sequence of the multiple operating parameters of the same device in the pre-processed data based on the pre-constructed sliding window, and inputting the fluctuation time sequence feature matrix into the pre-constructed second defect prediction model to obtain a second defect prediction result, wherein the fluctuation time sequence feature includes peak value, valley value, variance, range and fluctuation frequency;

[0056] The peak value and the valley value can be representative maximum value and minimum value The range is the difference between the maximum value and the minimum value The fluctuation frequency can be extracted in the frequency domain data after fast Fourier transform.

[0057] Device dimension: For the analysis of the synergistic effect of multiple parameters of a single device, such as for a specific process device, multiple core operating parameters are synchronously collected, the fluctuation time sequence features of the parameters in a certain period of time are captured through the Transformer layer, and the synergistic relationship between the parameters is analyzed through the GCN layer. When abnormal fluctuations are detected in some parameters, the model predicts the possible defect types and impact range in the subsequent production of the device, realizing the accurate prediction of the quality risk of a single device.

[0058] For example, for the filling and molding device, the torque command, liquid injection control parameters, running speed and other parameters of multiple sets of liquid injection mechanisms and the lens quality detection data corresponding to the production are collected in the full range to train the multi-parameter synergy and filling and molding quality correlation model. In real-time production, the operating parameters of each mechanism are synchronously collected, the model analyzes the synergistic relationship between the parameters, predicts the possible bubble, dirt and other defect risks, and outputs the synergistic adjustment suggestions of the device parameters.

[0059] B4) Determining the associated process parameters of different sections based on the pre-constructed multi-task learning model, constructing a parameter matrix based on the associated process parameters of different sections to obtain associated features, and inputting the associated features into the pre-constructed third defect prediction model to obtain a third defect prediction result.

[0060] ​Process section dimension: Integrate the parameter correlation of multiple consecutive processes, such as collecting key process parameters of each process for multiple consecutive processes, and mining the coupling relationship between parameters in different process sections through multi-task learning. The model predicts the comprehensive quality risk that may be caused by the superposition of abnormal parameters in each process, and clearly defines the influence weight of each process section on the final quality, providing a basis for multi-process linkage adjustment.

[0061] For example, for the continuous process sections of curing, mold separation, hydration immersion, and full inspection packaging, collect the core process parameters of each section (such as curing process parameters, mold separation operation parameters, hydration treatment parameters, and packaging operation parameters) and the quality data of the corresponding batches, and train a model that associates cross-process parameter coupling with product quality. In real-time production, real-time parameters of each section are collected, the model analyzes the cross-process influence between parameters, predicts possible comprehensive quality risks, clearly defines the influence weight of each section, and outputs multi-process coordination adjustment suggestions.

[0062] In the above process, the application uses a multi-task model to determine the coupling relationship between parameters in different process sections. The specific process of determining the associated process parameters of different process sections based on a pre-constructed multi-task learning model includes:

[0063] B41) Obtain a historical sample data set of multiple process sections, wherein the historical sample data set includes historical sample data of multiple batches, and the historical sample data includes values of core parameters and quality labels;

[0064] The historical sample data set needs to cover multiple process sections (such as injection molding, pressing, curing, etc.), and each batch contains complete production records. The data content includes:

[0065] Key control variables of the process section (such as injection volume, pressing speed, environmental temperature, and pressure value), which are usually recorded in time series form, and each parameter includes a numerical value, a collection timestamp, and a process section identifier.

[0066] Quality label: The final quality result of each batch (such as defect type, defect probability, or number of defective products) as a label for supervised learning.

[0067] B42) Align the timestamps of the historical sample data of the same batch product of different process sections based on the chronological relationship of different process sections, to obtain aligned data;

[0068] Extract the start and end timestamps of each process section (such as injection molding start / end time) in batches, and calculate the time interval between process sections (such as the delay from the end of injection molding to the start of pressing).

[0069] Avoiding errors caused by time misalignment (such as misassociating parameters of injection molding with outputs of pressing).

[0070] B43) normalizing and standardizing the alignment data to obtain pre-processed sample data;

[0071] The core parameters (such as the liquid injection amount) are normalized, and the parameters with large noise (such as device vibration) are standardized, so as to avoid information distortion caused by a single method.

[0072] B44) inputting the pre-processed sample data into a pre-constructed multi-task learning model to obtain branch outputs of multiple prediction tasks, wherein each process corresponds to a prediction task;

[0073] In the present application, the multi-task model uses a deep neural network Transformer to extract cross-process general features (such as time series patterns and parameter interaction patterns); each process is configured with an independent output layer (such as outputting the bubble defect rate prediction of the injection molding process and the size qualification rate prediction of the compression process); the shared layer is used to capture the commonality between processes (such as “temperature” affecting the quality in both injection molding and curing), and the branch layer focuses on the specificity of the process.

[0074] B45) calculating the mutual information of the branch outputs of any two prediction tasks, and determining the associated processes and the associated process parameters of different processes based on the mutual information.

[0075] The mathematical expression of the mutual information in the present application is:

[0076]

[0077] In the formula, denotes the mutual information of the prediction task and the prediction task , denotes the joint probability density function of the prediction task and the prediction task , denotes the marginal probability distribution of the prediction task , and denotes the marginal probability distribution of the prediction task .

[0078] A high MI value (such as >0.5) indicates a strong correlation between processes (such as a high correlation between the injection defect rate and the compression qualification rate), and a low MI value (such as <0.1) indicates a weak correlation. Correspondingly, the core parameters of the processes with strong correlation also have strong correlation. Therefore, when extracting the correlation features, the values of the core parameters of the processes with strong correlation can be normalized to form a feature vector matrix.

[0079] B5) aggregating the first flaw prediction result, the second flaw prediction result, and the third flaw prediction result to obtain multiple flaw results and the average number of flaws of the multiple flaw results. ​​​

[0080] Finally, the three prediction results are summarized, and the average number of defective products corresponding to each defect result is calculated, for example:

[0081] Bubble defects: 2.6

[0082] Scratch defects: 3.8

[0083] Multi-piece defects: 0.2

[0084] The final results are stored in the cache pool in chronological order, waiting for calls.

[0085] C) Detection module

[0086] Relying on the independently developed high-precision contact lens detection system, the semi-finished product and finished product are comprehensively detected. The system automatically identifies lens defects through image recognition algorithm, obtains accurate and reliable semi-finished product and finished product detection results, and forms standardized detection results containing defect position, size, type and detection conclusion.

[0087] Figure 4 For the detection module running flowchart in an embodiment of the application, as shown in Figure 4 The detection module in the application performs standardized detection by preprocessing and calling algorithms, and stores the detection results in the cache pool. The standard algorithm includes rough detection and fine detection, and the specific process is described below.

[0088] Figure 5 For the schematic diagram of the carrier disc image in an embodiment of the application, the lenses in the application are placed in the acupoints of the carrier disc, please understand the visual detection scheme in the application in combination with Figure 5 comprises the following steps:

[0089] C1) Obtain a multi-light field image of the carrier disc, wherein the multi-light field image includes a point light source image, a bright field image and a dark field image;

[0090] C2) Preprocess the multi-light field image to obtain a preprocessed image, wherein preprocessing includes grayscale conversion and high-pass filtering;

[0091] Firstly, the color image is converted into a single-channel grayscale image, which simplifies the subsequent processing by reducing color information and reduces the computational complexity. High-pass filtering is used to enhance high-frequency information (such as edges and textures) in the image and suppress low-frequency noise (such as background smooth areas), thereby highlighting the contour features.

[0092] C3) Extract the contour features in the preprocessed image, and screen the contour features based on pre-configured screening conditions to obtain the carrier disc contour;

[0093] Since the carrier disk is rectangular, this application uses the rectangularity and size of multiple contour features as the basis for screening and identification. Contours whose rectangularity and size both meet preset conditions are then selected as the carrier disk contours.

[0094] C4) Extract the image of the carrier region based on the carrier region contour, align the image of the carrier region with a pre-constructed carrier mask, and extract multiple lens acupoint region images from the image of the carrier region based on the aligned carrier mask.

[0095] Specifically, alignment is achieved by aligning the center of the target contour with the positioning points of the mask image. The aligned mask image is then multiplied with the loading area to obtain multiple lens acupoint region images.

[0096] C5) Extract the average gray level of the multiple lens acupoint region images, and take the lens acupoint region image with the average gray level in the preset gray level range as the target lens acupoint region image where the lens exists.

[0097] To avoid extracting invalid features when there are no semi-finished or finished contact lenses in the acupoint area, this application also needs to extract the grayscale value of the acupoint area. When the average grayscale value of the acupoint area falls within a preset grayscale value range, it indicates that there are semi-finished or finished contact lenses.

[0098] C6) Extract the image of the acupoint region of the target lens. grayscale features The grayscale feature Including average grayscale Maximum contrast and gray variance The maximum contrast It is the ratio of the minimum gray value to the maximum gray value;

[0099] C7) Calculate multiple target lens acupoint region images grayscale features average And calculate the grayscale features of the acupoint region image for each target lens; Deviation rate from the mean , ;

[0100] C8) Deviation rate of any grayscale feature in the image of any target lens acupoint region. If the deviation rate exceeds the set threshold, the image of the acupoint area of ​​the target lens is determined to have a risk of defects; otherwise, the image of the acupoint area of ​​the target lens is determined to have no risk of defects.

[0101] Wherein, since the characteristics of the semi-finished product / finished product images in the carrier disc are theoretically consistent, in order to adapt to the production line scene and quickly perform rough inspection, the average value of the gray scale characteristics of all semi-finished product / finished product images in the reaction carrier disc is first calculated , and then the difference between the gray scale characteristics of each semi-finished product / finished product and the overall characteristics is calculated. If there are individual characteristics that differ greatly from the overall characteristics, it is determined that the corresponding semi-finished product / finished product may have defects.

[0102] The above detection process can quickly perform rough inspection on a large number of contact lens semi-finished products / finished products, thereby screening the detection range for subsequent post-precise detection, to provide detection efficiency and production efficiency.

[0103] C9) calling a defect type recognition algorithm from an algorithm pool to detect the target lens acupoint area image with a risk of defects, to obtain a single lens defect detection result, wherein the single lens defect detection result includes a defect type;

[0104] In this application, existing defect detection algorithms can be called to perform defect detection, such as the algorithms disclosed in the applicant's published patents: CN117269179B High-precision detection method and system for contact lens edge defects based on machine vision, CN119477817A High-precision multi-light-field contact lens defect detection method and device based on cloud-machine vision, etc.

[0105] C10) aggregating single lens defect detection results of the same batch to obtain a defect detection result.

[0106] Finally, the defect detection result is associated with the batch data based on the timestamp of the multi-light-field image of the carrier disc; then a time series data set is constructed based on the defect prediction result, the defect detection result, the batch ID, the timestamp, and the running data of the contact lenses of multiple batches, and the time series data set is stored in the cache pool, waiting for subsequent analysis process calls.

[0107] Based on the cache pool data, structured data is created, key metadata such as batch ID, timestamp, and process parameters are extracted, and the structured alignment of prediction and detection data is predicted. Form a dynamic time series data set and support multi-dimensional backtracking. Based on the sliding window mechanism, the dynamic slicing of the data cache pool is realized.

[0108] D) Analysis module

[0109] The prediction results output by the prediction module and the detection results generated by the detection module are stored in a dedicated cache pool according to time and production batch. The data in the cache pool is dynamically divided by using a sliding window data slicing technology, and a specified number of historical batch data is selected as analysis samples. By constructing a prediction value and detection value comparison analysis model, quantitative indexes such as absolute error and relative error between the two are calculated, and the data trend is analyzed. When the analysis result shows that the predicted overall batch product yield is lower than the set threshold, there is a quality risk, and the process adjustment process is automatically triggered to provide decision support for production process optimization.

[0110] D1) Data slicing

[0111] D11) Sliding along the time axis in the cache pool based on the sliding window, wherein the defect detection results and the defect prediction results of the plurality of batches of contact lenses are stored in the cache pool in time axis order;

[0112] D12) At each sliding, the defect rate of the target defect type in the defect detection results of the plurality of batches of contact lenses in the sliding window is extracted, and the width of the sliding window is adjusted based on the defect rate and the corresponding time point at the center of the sliding window, wherein the higher the defect rate, the smaller the width of the sliding window, and the window width when the center of the sliding window corresponds to a non-stable production time period is smaller than the window width when the center of the sliding window corresponds to a stable production time period;

[0113] D13) Based on the adjusted sliding window width, the defect prediction results and the defect detection results in the cache pool are sliced to obtain data samples of a plurality of time windows.

[0114] The application adopts correlation quality fluctuation division, and the key defect rate per unit time. When the key defect rate is greater than a set threshold, it is determined as high risk. The minimum window interval is used by default to strengthen the capture of instantaneous fluctuations in the time period.

[0115] At the same time, combined with the production period characteristic optimization division, the quality fluctuation high incidence period (such as shift operation period, equipment process parameter change) is used by default to strengthen the capture of instantaneous fluctuations in the time period.

[0116] D2) Sample selection

[0117] By using the sliding window data slicing technology, combined with the time period characteristics (such as continuous production period, quality fluctuation high incidence period) and the current total good product rate state of the batch, the cache pool data is dynamically divided, and representative historical batch data (including samples of different time periods and different good product rate intervals) is selected as analysis samples to continuously update and adjust the defect prediction model, which specifically includes:

[0118] D21) risk labeling based on the defect rate of the target defect type of the data samples in multiple time windows, wherein,

[0119] D211) when the defect rate of the target defect type of a single data sample is greater than a preset defect rate threshold, labeling as high risk;

[0120] D212) when the defect rate of the target defect type of N consecutive data samples is less than or equal to a preset defect rate threshold, and the defect rate of the target defect type of the N consecutive data samples shows an upward trend, labeling the N consecutive data samples as high risk;

[0121] D213) when the defect rate of the target defect type of N consecutive data samples is less than or equal to a preset defect rate threshold, and the defect rate of the target defect type of the N consecutive data samples shows no upward trend, labeling as low risk;

[0122] D214) and sample data labeled as high risk and low risk as target sample data;

[0123] The above screening conditions are based on production period dimension screening, specifically including:

[0124] Normal stable production period: the production process is stable, the process parameters and environmental parameters are stable, and the quality fluctuation is small. Such data can be used as a "benchmark" reference for comparison with other abnormal situations. The risk labeling of data samples in this case is: low risk.

[0125] High quality fluctuation period: there are quality abnormalities (such as increased product defect rate) in the production process, and the data reflects abnormal patterns, which is suitable for analyzing the causes of problems or optimizing control strategies. The risk labeling of data samples in this case is: high risk.

[0126] Special process period and time period: the process parameters may be in the exploration stage, the data fluctuation is large, and it is suitable for verifying the feasibility of new processes. The risk labeling of data samples in this case is: high risk.

[0127] Unstable production period after equipment maintenance: the equipment state is unstable, and the data may contain equipment abnormal characteristics, which is suitable for analyzing the impact of equipment health status on production. The risk labeling of data samples in this case is: high risk.

[0128] D22) obtaining process parameter samples corresponding to multiple sample data based on timestamps, calculating the similarity of the process parameter samples and a preset process parameter template, and taking sample data with a similarity greater than a preset similarity threshold as target sample data;

[0129] Pre-set parameters (such as temperature, pressure, time, etc.) before production are the core variables that affect product quality. In addition, environmental parameters, such as production environment-related variables (such as workshop temperature and humidity, raw material batches, etc.), may indirectly affect the execution effect of process parameters.

[0130] By comparing the set parameters (such as numerical differences, distribution patterns) and environmental parameters (such as environmental stability) of the current batch with those of the historical batches, historical batches with similar process conditions are screened out.

[0131] D23) Calculate the difference between the defect detection result and the defect prediction result in multiple sample data, and compare the difference with the pre-set error range. The sample data whose difference does not fall within the error range is taken as the target sample data.

[0132] Production results predicted based on models or experience (such as yield, quality indicators).

[0133] The difference between the predicted value and the actual result (such as error rate, deviation degree) reflects the accuracy of the prediction or the stability of the production process.

[0134] By analyzing the matching degree of the predicted value and the actual deviation (such as error range, trend consistency), historical batches with similar prediction patterns are screened out.

[0135] The above-mentioned screened samples can be used as representative basic data, which are updated to the sample data set for training the defect prediction model, so that the defect training model can gradually adapt to new processes and production lines.

[0136] In addition, the similarity comparison in the above process can use cosine similarity comparison.

[0137] D3) Deviation analysis

[0138] D31) Calculate the deviation between the defect prediction result and the defect detection result of the contact lens in multiple data samples, wherein the deviation includes absolute error , relative error and mean square error ;

[0139]

[0140]

[0141]

[0142] In the formula, is the detection defect rate (or the number of detected defects) of the target defect type of the i-th data sample, is the detection defect rate (or the number of detected defects) of the target defect type of the i-th data sample, is the detection defect rate (or the number of detected defects) of the target defect type of the i-th data sample, The predicted defect rate (or the predicted defect number) of the target defect type of the data sample.

[0143] D32) Constructing a multi-dimensional evaluation matrix based on the deviation and the yield rate of the plurality of data samples;

[0144] Wherein, the deviation has three dimensions, and the yield rate can be fused to construct a multi-dimensional evaluation matrix. Since the mean square error and the yield rate are for a whole batch of data (the data samples correspond to small batches, and the mean square error and the yield rate correspond to large batches, and a large batch contains multiple small batches), therefore, for a large batch, the corresponding multi-dimensional evaluation matrix is:

[0145]

[0146] If there are multiple large batches, the multi-dimensional evaluation matrices of different large batches can be spliced in chronological order.

[0147] D33) Weighting calculation of the risk parameters of the contact lenses in each batch in the multi-dimensional evaluation matrix based on the analytic hierarchy process, to obtain the comprehensive risk of the difference of each batch, wherein the risk parameters include the difference and the yield rate;

[0148] Wherein, the process of weighting calculation of the risk parameters in the matrix by the analytic hierarchy process includes:

[0149] Using 1-9 scale method (Saaty scale), the importance between indexes is subjectively judged.

[0150] Then, the importance of a plurality of risk parameters is normalized to obtain weights, such as the first weight , the second weight , the third weight , and the fourth weight .

[0151] Finally, the comprehensive risk of the contact lens batch corresponding to each data sample is:

[0152]

[0153] In the formula, represents the normalized absolute error, represents the normalized relative error, represents the normalized mean square error, represents the normalized yield rate.

[0154] D34) Analysis result

[0155] D341) When there are more than a set number of continuous batches with the comprehensive risk exceeding the threshold, it is determined that the defect prediction model has a risk of limitation;

[0156] This phenomenon may reflect that the prediction model has limitations. For example, the model may have overfitting problems or may not fully consider some key process parameters, environmental variables, and other influencing factors during the training phase, although it shows high accuracy, recall rate, and other performance indicators.

[0157] D342) When there are more than a set number of continuous batches with the comprehensive risk showing an upward trend, it is determined that the defect prediction model has a risk of overall missing;

[0158] If the batch-to-batch deviation shows a gradual increasing trend, it may be due to changes in some hidden parameters that are not being collected in real time, and these parameters have not been included in the input feature system of the prediction model, causing the prediction results to gradually deviate from the actual detection results. Therefore, it is necessary to reevaluate the comprehensiveness of data collection, check whether key parameters have been included in the model input features, and perform targeted optimization.

[0159] D343) When there are more than a set number of continuous batches with the comprehensive risk showing no upward trend, and the good product rate showing a downward trend, it is determined that the defect prediction model identifies a good product rate decline risk, wherein the comprehensive risk trend is achieved based on Monte Carlo simulation.

[0160] In this case, the prediction model can identify the possible downward trend of the good product rate caused by the current combination of process parameters in advance. When it is determined that it does not meet the pre-set quality requirements (the prediction model confidence is within the threshold range), the system will trigger an early adjustment of the process parameters to avoid further deterioration of the good product rate.

[0161] D344) In addition, if the deviation is occasional and has no obvious regularity, it usually indicates that the prediction model has failed to fully capture the key influencing factors in the production process (such as fluctuations in specific process parameter combinations, hidden variables, etc.), but the model is not completely failed at this time. It is necessary to reevaluate the model performance in combination with the actual process scenario, and if necessary, start the model retraining process.

[0162] In another embodiment of the present application, the trend judgment described above uses a sliding window mechanism to analyze historical quality fluctuation data in multiple windows and extrapolates the quality trend of the next 3-5 production cycles through Monte Carlo simulation. It is also possible to capture the long-short term dependence relationship of quality data based on a time series prediction model, combined with the dynamic change of the current total good product rate of the batch, to achieve dynamic prediction of key quality indicators and predict whether the optimization threshold will be reached. When the optimization threshold is reached, jump to the adjustment module to perform the optimization adjustment process.

[0163] Finally, the risk is classified by using the analytic hierarchy process, the influence range is determined by using the cause-effect analysis, and the confidence interval is calculated to form a structured risk assessment report.

[0164] E) Adjustment module

[0165] Based on the actual situation of the current production process, combined with the experience and knowledge of industry experts, the results of the cause-effect analysis of quality risk, and the successful cases and adjustment strategies in the historical adjustment rule library, intelligent algorithms are used to integrate and analyze various types of information. From multiple process parameter dimensions, scientific and reasonable process adjustment rules are generated, detailed process parameter optimization schemes are developed, the adjustment direction and specific adjustment value of each parameter are determined, the operability and effectiveness of the adjustment scheme are ensured, and the optimization scheme is issued to the production line for execution.

[0166] Figure 6 The parameter adjustment process in an embodiment of the present application is shown in the schematic diagram as shown in Figure 6 , which specifically includes:

[0167] E1) When the defect prediction model identifies a risk of yield reduction, locate the target device that needs to be adjusted;

[0168] The target device is located by the time point of the yield reduction batch and the timestamp of the device processing the batch of contact lenses.

[0169] E2) Obtain the influence weight of the multiple parameters corresponding to the target device on the quality factor, the adjustment sensitivity of the multiple parameters on the quality factor, the device correlation of the multiple parameters, the adjustment cost of the multiple parameters, and the safety boundary of the multiple parameters, wherein the multiple parameters include process parameters, environmental parameters, and device parameters, the influence weight and the adjustment sensitivity are obtained by regression analysis based on historical data, and the device correlation, the adjustment cost, and the safety boundary are generated based on an expert system;

[0170] In the present application, the preset rules are retrieved (based on expert experience and historical data extraction), and the association logic between device parameter changes and process parameter changes is determined (such as the device parameter corresponding to the filling and molding injection volume process parameter is the graduated cylinder drop height, and the association logic conversion is performed).

[0171] The historical data extraction method mainly uses past process adjustment closed-loop cases:

[0172] Historical records of process adjustment procedures triggered, including adjustment background (e.g., a batch bubble defect rate exceeds the threshold, yield continues to decline), adjustment content (e.g., the filling of the mold device is adjusted from 5ml to 4.8ml, the temperature of the warehouse buffer is reduced from 25°C to 23°C), adjustment tracking (e.g., whether the audit end considers it reasonable, whether there is intervention, whether it is adopted, etc.), actual adjustment effect (e.g., actual change of defect rate, actual change of yield).

[0173] Quantitative transformation (quantitative processing: qualitative data is transformed into quantitative correlation rules, such as reducing the value of a certain process parameter in a certain situation to improve or reduce the detection yield of a certain defect) and rule structuring (converting expert experience and data correlation logic into preset rules) are used to effectively extract data and ensure that the extraction results can provide basic support for scheme development.

[0174] Quantitative processing can be obtained by regression analysis.

[0175] For example: quality impact weight: the impact of the parameter on the target defect (assuming that, for example, the impact weight of the liquid injection amount on the bubble defect is 0.8, which is higher than the temperature of 0.3), which is calculated by a historical data regression model.

[0176] Adjustment sensitivity: the improvement range of quality with a small change in the parameter (e.g., a 1% adjustment in the pressing speed within a certain parameter range can change the defect rate by 5%, so the sensitivity is high; a 3% adjustment within a certain parameter range can change the defect rate by 5%, so the sensitivity is low).

[0177] Device correlation: whether the parameter adjustment affects other devices (e.g., adjusting the environmental temperature may affect multiple stages such as injection molding and curing, so the correlation is high).

[0178] Adjustment cost: including time cost (e.g., mechanical parameter adjustment requires downtime), material cost (e.g., parameter adjustment leads to raw material loss rate).

[0179] Safety boundary: whether the parameter is close to the rated threshold of the device (e.g., when the temperature is close to the upper limit, the adjustment risk is high, and the priority is reduced).

[0180] E3) Weight the impact weight, adjustment sensitivity, device correlation, and safety boundary of each parameter to get the priority score of each parameter;

[0181] Assign weights to each constraint dimension (e.g., quality impact weight accounts for 40%, safety boundary accounts for 20%), and determine the weight value through expert experience and historical data verification. Calculate the comprehensive score of each parameter (e.g., the comprehensive score of the liquid injection amount is 85 points, and the temperature is 60 points). The higher the score, the higher the priority.

[0182] E4) output a priority sequence based on the priority score, and select the priority parameter with the highest priority and which is not currently being adjusted;

[0183] For example: filling amount of the filling machine (priority 1), pressing speed (priority 2), workshop environment temperature (priority 3), to ensure that the adjustment resources are concentrated on the most critical parameters.

[0184] E5) determine the adjustment direction of the priority parameter based on the typical value of the priority parameter, wherein the typical value of the priority parameter is calculated based on historical data;

[0185] If the value of the priority parameter at the current time point is greater than the typical value, the adjustment direction is to reduce; if the value of the priority parameter at the current time point is less than the typical value, the adjustment direction is to increase.

[0186] E6) adjust the value of the priority parameter by one unit based on the adjustment direction, to obtain adjusted running data; extract feature vectors of multiple time windows from the adjusted running data, and input the feature vectors into a pre-constructed defect prediction model to obtain an adjusted defect prediction result;

[0187] After adjusting a single priority parameter, the adjusted real-time running data is obtained, and the process returns to step B2 for re-prediction.

[0188] E7) when the yield rate in the adjusted defect prediction result is less than a preset yield rate threshold, return to adjusting the value of the priority parameter by one unit based on the adjustment direction, until the yield rate in the adjusted defect prediction result is greater than or equal to the preset yield rate threshold, or the yield rate in the adjusted defect prediction result is less than the preset yield rate threshold but reaches a maximum value, wherein the maximum value of the yield rate is calculated by an inflection point;

[0189] If the prediction result is poor and the predicted yield rate is still poor, return to step E6 for further adjustment until the predicted yield rate returns to a normal level, or an inflection point of the yield rate appears. The adjustment amount corresponding to the inflection point is retained, and the adjustment direction and the adjustment amount of the first priority parameter are saved.

[0190] E8) when the yield rate in the adjusted defect prediction result is less than the preset yield rate threshold but reaches a maximum value, return to selecting the priority parameter with the highest priority and which is not currently being adjusted, until the yield rate in the adjusted defect prediction result is greater than or equal to the preset yield rate threshold;

[0191] When the first priority parameter is adjusted, the yield appears to have a turning point but still fails to reach the yield threshold, return to step E4, select the next priority parameter, and repeat the above process until the predicted yield rises to the threshold.

[0192] E9) When the yield in the adjusted defect prediction result is greater than or equal to the preset yield threshold, generate an adjustment scheme based on the current priority parameter and the adjustment direction of the priority parameter, and the adjustment amount of the priority parameter.

[0193] Finally, all adjusted priority parameters, adjustment direction and adjustment amount of limited parameters are constructed into an adjustment matrix vector. Similarity calculation is performed with historical adjustment cases, and when the similarity threshold is reached, the final adjustment scheme is obtained and executed.

[0194] If it is not similar to the historical adjustment case, it is sent to the expert system for the expert to judge whether to issue.

[0195] Figure 7 The scheme issuing process in an embodiment of the present application is shown in the schematic diagram as shown in Figure 7 If the scheme needs to be issued, the adjustment effect prediction report is generated in combination with the self-developed production quality optimization system, and the decision-making double report is formed with the optimization scheme to submit to the management end.

[0196] A three-level confirmation mechanism is established: the management end performs feasibility review of the scheme, the system stores the parameter safety check reserved by the process engineer, and the operation layer completes the on-site applicability confirmation.

[0197] Management end review (feasibility): from the overall production safety and efficiency, check whether the electrical, environmental and mechanical adjustment scheme meets the production line safety specification, capacity demand and other feasibility confirmation.

[0198] Process engineer parameter limit check (safety interval): based on professional standards, confirm whether the electrical parameter adjustment is within the equipment rating range, whether the environmental parameter is within the process allowed interval, and whether the mechanical parameter meets the requirements.

[0199] Operation layer confirmation (on-site applicability): combined with the on-site operation conditions, confirm the operability of electrical adjustment, the equipment carrying capacity of environmental adjustment, and the on-site execution convenience of mechanical adjustment.

[0200] F) Process adjustment post-processing module

[0201] After the adjustment of the process parameters, the predicted detection result data of the new production batch is continuously collected to track and analyze the adjusted production quality. The quality data before and after the adjustment are compared to evaluate the actual effect of the process adjustment. If the product quality after the adjustment meets or exceeds the expected target, the system automatically stores the relevant information of this process adjustment (adjustment background, adjustment content, adjustment effect, etc.) in the historical adjustment rule library as a reference for subsequent production optimization, realizing self-improvement and updating of the rule library; if the expected effect is not achieved, the relevant data and information are recorded for further analysis by engineers to optimize and improve the process adjustment scheme.

[0202] Figure 8 The effect tracking and archiving process in an embodiment of the present application is shown in the schematic diagram as shown in Figure 8 The effect tracking and archiving includes:

[0203] A full-link quality tracking system is constructed, and a causal correlation analysis of the adjustment parameters and quality indicators is realized based on time series analysis technology. Continuous monitoring is performed regardless of whether the parameters are updated.

[0204] When the parameters are not updated, the prediction confidence interval verification technology is used for prediction result comparison. When the actual result and the prediction interval overlap degree is ≥95%, the prediction case is stored in the accurate prediction knowledge base to form an intelligent early warning rule.

[0205] When the parameters are updated, the double difference model is used to evaluate the adjustment effect. When the key quality indicator improvement amplitude exceeds the preset threshold and the stability indicator meets the standard, the adjustment scheme is converted into a standardized process instruction and stored in the dynamic rule library.

[0206] When the expected effect is not achieved, the intelligent tracing mechanism is triggered, the optimization path is reconstructed through process mining technology, the failure node is located by combining the fault tree analysis method, and the improvement suggestions are generated by the expert system and the closed-loop optimization case is formed.

[0207] The present application relies on the professional lens detection system independently developed by the applicant. A unified quantitative algorithm is used to replace the traditional manual subjective judgment to establish a standardized contact lens quality detection benchmark. Through the fully automated detection process, the detection efficiency and consistency are improved, the problems of missed detection and misjudgment are effectively avoided, reliable data support is provided for the accurate prediction and optimization of the production process, and the accuracy of defect prediction is significantly improved.

[0208] An intelligent defect tracing and root cause analysis system is established. Based on the causal inference model, the root cause of the defect is automatically analyzed while the system is warning. By correlating multi-source data, the mapping relationship between defect characteristics and production links is established, and targeted process adjustment suggestions are generated to actively prevent the decrease of yield caused by problems such as process parameters, shorten the production optimization cycle, and improve the yield.

[0209] Break through the limitations of traditional detection and prediction and production process, realize the dynamic coordination optimization of whole line multi-dimensional data, deeply integrate the prediction model and the actual detection result, form the cooperative optimization mechanism throughout the whole production line, and achieve the fine quality control of the production process.

[0210] Change the problem of discrete and low efficiency of traditional manual sampling, rely on self-developed detection system to build a standardized intelligent detection system, replace subjective judgment with unified quantitative algorithm, and meet the process feedback demand of large-scale production with the consistency and reliability of detection results.

[0211] Make up for the defects of the existing technology, automatically locate the defect root cause through intelligent root cause analysis system, correlate multi-source data to generate accurate process adjustment suggestions, change passive early warning to active prevention, and effectively improve production efficiency through accurate process parameter feedback.

[0212] The embodiment also provides an electronic terminal, including a processor and a memory.

[0213] The memory is used for storing a computer program, and the processor is used for executing the computer program stored in the memory, so that the terminal executes any method in the embodiment.

[0214] The computer readable storage medium in the embodiment can be understood by those skilled in the art that all or part of the steps of the above-mentioned method embodiments can be completed by computer program related hardware. The foregoing computer program can be stored in a computer readable storage medium. The program executes the steps of the above-mentioned method embodiments when executed; and the foregoing storage medium includes ROM, RAM, magnetic disc or optical disc and various storage program codes.

[0215] The electronic terminal provided in the embodiment includes a processor, a memory, a transceiver and a communication interface, the memory and the communication interface are connected with the processor and the transceiver and complete communication between each other, the memory is used for storing a computer program, the communication interface is used for communication, and the processor and the transceiver are used for running the computer program, so that the electronic terminal executes each step of the method.

[0216] In the embodiment, the memory can include random access memory (RAM) and can also include non-volatile memory, for example, at least one disk memory.

[0217] The processor described above can be a general processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0218] In the above embodiments, although the present application has been described in conjunction with specific embodiments thereof, numerous alternatives, modifications and variations will be readily apparent to those of ordinary skill in the art in the light of the foregoing descriptions. The embodiments of the present application are intended to embrace all such alternatives, modifications and variations as falling within the scope of the appended claims.

[0219] The above embodiments only illustrate the principles and effects of the present application, but are not used to limit the present application. Any person skilled in the art can modify or change the above embodiments without departing from the spirit and scope of the present application. Therefore, all equivalent modifications or changes made by those skilled in the art without departing from the spirit and technical ideas of the present application should be covered by the claims of the present application.

Claims

1. A process feedback system for a contact lens production line, characterized in that, include: The data acquisition module is used to acquire real-time operating data of each section of the contact lens production line and to acquire contact lens images output by each process processing equipment of the contact lens production line. The real-time operating data includes real-time environmental data, real-time process data and real-time equipment data. The real-time environmental data includes time-series values ​​of various environmental parameters, the real-time process data includes time-series values ​​of various process parameters, and the real-time equipment data includes time-series values ​​of various equipment operating parameters. The prediction module is used to extract feature vectors of multiple time windows from the real-time running data and input the feature vectors into a pre-built defect prediction model to obtain defect prediction results, wherein the defect prediction results include defect type and defect quantity; The detection module is used to perform defect detection on the contact lens image and obtain defect detection results; The analysis module is used to slice the defect detection results and defect prediction results of multiple batches of contact lenses to obtain multiple data samples; to perform deviation analysis on the defect prediction results and defect detection results of contact lenses in the data samples to obtain deviation analysis results; and to filter the data samples to obtain target sample data, wherein the target sample data is used to adjust and train the defect prediction model. The adjustment module is used to generate a production line adjustment plan based on the deviation analysis results, and to feed back the production line adjustment plan and trend analysis results to the target object.

2. The process feedback system for a contact lens production line according to claim 1, characterized in that, Feature vectors for multiple time windows are extracted from the real-time running data, and these feature vectors are input into a pre-built defect prediction model to obtain defect prediction results, including: The real-time running data is preprocessed to obtain preprocessed data, wherein the preprocessing includes standardization and normalization; Based on a pre-built sliding window, the fluctuation trend features of the target parameter in the preprocessed data over multiple time windows are extracted. The fluctuation trend features are then input into a pre-built first defect prediction model to obtain a first defect prediction result. The fluctuation trend features include the mean, the upward slope, and the downward slope. Based on a pre-built sliding window, the fluctuation time series feature matrix of the time series values ​​of multiple operating parameters of the same device in the preprocessed data is extracted, and the fluctuation time series feature matrix is ​​input into the pre-built second defect prediction model to obtain the second defect prediction result. The fluctuation time series features include peak value, valley value, variance, range and fluctuation frequency. The associated process parameters of different sections are determined based on a pre-built multi-task learning model, and a parameter matrix is ​​constructed based on the associated process parameters of different sections to obtain associated features; the associated features are then input into a pre-built third defect prediction model to obtain the third defect prediction result. The first defect prediction result, the second defect prediction result, and the third defect prediction result are summarized to obtain multiple defect results and the average number of defective products for multiple defect results.

3. The process feedback system for the contact lens production line according to claim 2, characterized in that, Based on a pre-built multi-task learning model, the associated process parameters for different work sections are determined, including: Obtain historical sample datasets from multiple work sections, wherein the historical sample datasets include historical sample data from multiple batches, and the historical sample data includes the values ​​of core parameters and quality labels; Based on the sequential relationship between different work sections, the timestamps of historical sample data of the same batch of products in different work sections are aligned to obtain aligned data. The aligned data is normalized and standardized to obtain preprocessed sample data; The preprocessed sample data is input into a pre-built multi-task learning model to obtain branch outputs of multiple prediction tasks, wherein each work section corresponds to one prediction task. Calculate the mutual information of the branch outputs of any two prediction tasks, and determine the related work sections and the related process parameters of different work sections based on the mutual information.

4. The process feedback system for a contact lens production line according to claim 1, characterized in that, Contact lenses are placed on a carrier tray for processing, wherein the contact lens image is subjected to defect detection to obtain defect detection results, including: Acquire multi-field images of the carrier disk, wherein the multi-field images include point light source images, bright field images, and dark field images; The multi-field image is preprocessed to obtain a preprocessed image, wherein the preprocessing includes grayscale conversion and high-pass filtering; Extract contour features from the preprocessed image and filter the contour features based on pre-configured filtering conditions to obtain the disk contour. Based on the disc contour, extract the disc region image, align the disc region image with a pre-constructed disc mask, and extract multiple lens acupoint region images from the disc region image based on the aligned disc mask. Extract the average gray level of the multiple lens acupoint region images, and take the lens acupoint region image with the average gray level within a preset gray level range as the target lens acupoint region image containing the lens. Extract the image of the acupoint region of the target lens. grayscale features The grayscale feature Including average grayscale Maximum contrast and gray variance The maximum contrast It is the ratio of the minimum gray value to the maximum gray value; Calculate multiple target lens acupoint region images grayscale features average And calculate the grayscale features of the acupoint region image for each target lens. Deviation rate from the mean , ; Deviation rate of any grayscale feature in the image of any target lens acupoint region If the deviation rate exceeds the set threshold, the image of the acupoint area of ​​the target lens is determined to have a risk of defects; otherwise, the image of the acupoint area of ​​the target lens is determined to have no risk of defects. The defect type recognition algorithm is called from the algorithm pool to detect the acupoint area image of the target lens with defect risk, and the defect detection result of a single lens is obtained, wherein the defect detection result of a single lens includes the defect type; The defect detection results of individual lenses in the same batch are summarized to obtain the defect detection results.

5. The process feedback system for a contact lens production line according to claim 4, characterized in that, Also includes: The defect detection results are correlated with the batch data based on the timestamps of the multi-field images of the carrier disk; A time-series dataset is constructed based on the defect prediction results, defect detection results, batch ID, timestamp, and runtime data of multiple batches of contact lenses, and the time-series dataset is stored in a cache pool.

6. The process feedback system for a contact lens production line according to claim 1, characterized in that, Multiple data samples were obtained from slices of defect detection results and defect prediction results from multiple batches of contact lenses, including: Based on a sliding window that slides along the timeline within the cache pool, the defect detection results and defect prediction results of multiple batches of contact lenses are stored in the cache pool in chronological order. Each time the sliding window is slid, the defect rate of the target defect type in the defect detection results of multiple batches of contact lenses within the sliding window is extracted, and the width of the sliding window is adjusted based on the defect rate and the corresponding time point at the center of the sliding window. The higher the defect rate, the smaller the width of the sliding window. The window width when the time point corresponding to the center of the sliding window is in an unstable production period is smaller than the window width when the time point corresponding to the center of the sliding window is in a stable production period. Based on the adjusted sliding window width, the defect prediction results and defect detection results in the cache pool are sliced ​​to obtain data samples for multiple time windows.

7. The process feedback system for a contact lens production line according to claim 6, characterized in that, The data samples are filtered to obtain the target sample data, including: Risk labeling is performed based on the defect rate of the target defect type in data samples within multiple time windows. Specifically, if the defect rate of a single data sample for the target defect type exceeds a preset defect rate threshold, it is labeled as high-risk. If the defect rate of N consecutive data samples for the target defect type is less than or equal to the preset defect rate threshold, and the defect rate of the N consecutive data samples for the target defect type shows an upward trend, then the N consecutive data samples are labeled as high-risk. If the defect rate of N consecutive data samples for the target defect type is less than or equal to the preset defect rate threshold, and the defect rate of the N consecutive data samples for the target defect type shows no upward trend, then it is labeled as low-risk. The data samples labeled as high-risk and low-risk are used as the target sample data. Based on timestamps, process parameter samples corresponding to multiple sample data are obtained, the similarity between the process parameter samples and the preset process parameter template is calculated, and sample data with similarity greater than the preset similarity threshold is used as target sample data. The difference between the defect detection result and the defect prediction result in multiple sample data is calculated, and the difference is compared with a preset error range. The sample data whose difference does not fall within the error range is taken as the target sample data.

8. The process feedback system for a contact lens production line according to claim 7, characterized in that, A bias analysis was performed on the defect prediction results and defect detection results of contact lenses in the data sample. The bias analysis results include: The deviation between the defect prediction results and defect detection results of contact lenses in multiple data samples is calculated, wherein the deviation includes absolute error, relative error and mean square error; A multidimensional evaluation matrix is ​​constructed based on the deviation and yield rate of multiple data samples; The risk parameters of each batch of contact lenses in the multidimensional evaluation matrix are weighted and calculated based on the analytic hierarchy process (AHP) to obtain the comprehensive risk of the difference in each batch. The risk parameters include the difference and the yield rate. When the overall risk of more than a set number of consecutive batches exceeds a threshold, the defect prediction model is deemed to have a limitation risk; when the overall risk of more than a set number of consecutive batches shows an upward trend, the defect prediction model is deemed to have a comprehensive deficiency risk; when the overall risk of more than a set number of consecutive batches does not show an upward trend, and the yield rate shows a downward trend, the defect prediction model is deemed to have identified a yield rate decline risk, wherein the overall risk change trend is implemented based on Monte Carlo simulation.

9. The process feedback system for a contact lens production line according to claim 8, characterized in that, Based on the deviation analysis results, a production line adjustment plan is generated, including: When the defect prediction model identifies a risk of declining yield, it identifies the target equipment that needs adjustment. The system obtains the influence weights of various parameters corresponding to the target equipment on the quality factor, the adjustment sensitivity of various parameters on the quality factor, the equipment correlation of various parameters, the adjustment cost of various parameters, and the safety boundary of various parameters. The various parameters include process parameters, environmental parameters, and equipment parameters. The influence weights and adjustment sensitivity are obtained based on regression analysis of historical data. The equipment correlation, adjustment cost, and safety boundary are generated based on an expert system. The impact weight, adjustment sensitivity, device correlation, and safety boundary of each parameter are weighted to obtain the priority score of each parameter; Based on the priority score, output parameter priority sequence and select the priority parameter that is currently not adjusted and has the highest priority. The adjustment direction of the priority parameter is determined based on the typical value of the priority parameter, wherein the typical value of the priority parameter is calculated based on historical data; The priority parameter value is adjusted by one unit based on the adjustment direction to obtain the adjusted running data; feature vectors of multiple time windows are extracted from the adjusted running data, and the feature vectors are input into the pre-built defect prediction model to obtain the adjusted defect prediction result; When the yield rate in the adjusted defect prediction result is less than the preset yield rate threshold, the value of the priority parameter is adjusted by one unit based on the adjustment direction until the yield rate in the adjusted defect prediction result is greater than or equal to the preset yield rate threshold, or the yield rate in the adjusted defect prediction result is less than the preset yield rate threshold but reaches the maximum value, wherein the yield rate reaching the maximum value is calculated by the inflection point. If the yield rate in the adjusted defect prediction result is less than the preset yield rate threshold but reaches the maximum value, return to the selection of the currently unadjusted and highest priority parameter until the yield rate in the adjusted defect prediction result is greater than or equal to the preset yield rate threshold. When the yield rate in the adjusted defect prediction result is greater than or equal to the preset yield rate threshold, an adjustment scheme is generated based on the current priority parameters, the adjustment direction of the priority parameters, and the adjustment amount of the priority parameters.

10. The process feedback system for a contact lens production line according to claim 1, characterized in that, After feeding back the production line adjustment plan to the target entity, the process also includes: The effects of the production line adjustment plan are tracked and archived.

Images