Adaptive quality management system for large size photovoltaic cell doping / oxidation / annealing process

Abstract

The application discloses a large-size photovoltaic cell doping / oxidation / annealing process adaptive quality management system, which comprises a data acquisition layer, an intelligent decision-making layer and a control execution layer; the data acquisition layer collects process parameters through doping, oxidation and annealing PLC systems, constructs a time-space aligned process data cube, and outputs a multi-dimensional feature matrix containing time x process x parameter; the intelligent decision-making layer analyzes the multi-dimensional feature matrix by using a random forest hidden crack classification model and outputs a hidden crack risk grade; the control execution layer comprises a dynamic compensation strategy composed of a diffusion furnace temperature control module and an annealing cooling compensator, is used for generating an OPC UA instruction according to the hidden crack risk grade, and triggers an annealing furnace PID regulator through the OPC UA instruction to dynamically adjust process parameters of the annealing furnace; the application solves the data island problem by integrating multi-process PLC data and the random forest model, realizes high-precision hidden crack prediction, and significantly improves the quality detection efficiency.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology for photovoltaic cells, and particularly relates to an adaptive quality management system for the doping / oxidation / annealing process of large-size photovoltaic cells. Background Technology

[0002] With the global energy structure shifting towards cleaner and lower-carbon energy, the photovoltaic industry is experiencing rapid development. Large-size, high-efficiency photovoltaic cells are gradually becoming the market mainstream due to their higher power output and lower cost per kilowatt-hour. Photovoltaic cell manufacturing involves multiple key processes, including doping, oxidation, annealing, coating, ribbon bonding, and cell assembly and positioning. The process is highly complex and requires strict consistency and yield standards. Meanwhile, advancements in intelligent manufacturing technology are driving photovoltaic manufacturing towards digitalization and intelligence. Achieving precise control of process parameters and real-time quality management has become a key focus for the industry.

[0003] Currently, process control in the manufacturing of large-size photovoltaic cells mainly relies on programmable logic controller (PLC) systems to independently monitor and adjust each process step. This involves collecting key parameters through sensors such as temperature, gas flow, and pressure, and then using preset process thresholds for single-point control. For example, patent CN110506258A discloses a "Method and System for Equipment-Based Engineering of Programmable Logic Controller (PLC)" that employs a dual-core design. One core of the equipment processor runs the equipment firmware to handle data acquisition from underlying hardware (such as temperature and flow sensors), while the other core runs an engineering server integrated with a web server to receive user commands. A communication path between the two cores is established through a management program to achieve data interaction.

[0004] However, the above methods have significant limitations. Sensor data from each process is stored in separate PLC systems, forming data silos and making it impossible to achieve cross-process data correlation analysis and real-time control. This fragmented data management approach severely restricts the efficiency of process parameter optimization and quality traceability, making it difficult to meet the complex needs of multi-process coupling in the manufacturing of large-size photovoltaic cells. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide an adaptive quality management system for the doping / oxidation / annealing process of large-size photovoltaic cells that integrates multi-source sensor data and machine learning, thereby solving the problems of data silos and process coupling optimization.

[0006] Technical solution: The quality management system of the present invention includes a data acquisition layer, an intelligent decision-making layer and a control execution layer connected in sequence;

[0007] The data acquisition layer includes a doping PLC system, an oxidation PLC system, and an annealing PLC system, which are used to acquire process parameters during the manufacturing process of photovoltaic cells, construct a spatiotemporally aligned process data cube, and output a multidimensional feature matrix containing time × process × parameter.

[0008] The intelligent decision-making layer includes a process data cube, a random forest hidden crack classification model, and a collaborative analysis engine, which are used to receive the multidimensional feature matrix, calculate the hidden crack risk level through the random forest hidden crack classification model, and output the hidden crack risk level.

[0009] The control execution layer includes a dynamic compensation strategy consisting of a diffusion furnace temperature control module and an annealing cooling compensator. This strategy generates an OPC UA command based on the microcrack risk level and triggers the annealing furnace PID controller via the OPC UA command to dynamically adjust the process parameters of the annealing furnace.

[0010] This invention acquires multi-source process parameters in real time during photovoltaic cell manufacturing through a data acquisition layer, constructing a spatiotemporally aligned process data cube to achieve precise integration and multi-dimensional feature extraction of the entire process data. The intelligent decision-making layer, based on a random forest microcrack classification model and a collaborative analysis engine, performs in-depth mining and risk modeling of process parameters, significantly improving the early identification and classification accuracy of microcrack defects. The control execution layer, through dynamic compensation strategies and OPC UA commands, achieves closed-loop dynamic control of annealing process parameters, effectively reducing microcrack risk and improving process stability. This three-layer collaborative mechanism forms a complete control loop from data perception and intelligent analysis to precise execution, significantly improving the manufacturing yield and quality consistency of photovoltaic cells while enhancing the adaptive adjustment capability of the process.

[0011] Preferably, the doping PLC system, oxidation PLC system and annealing PLC system in the data acquisition layer are respectively integrated with multi-process sensing modules through industrial Internet of Things gateways to form a data acquisition network for three process chains;

[0012] The doping PLC system connects two types of sensors: K-type thermocouples and online refractometers. The K-type thermocouples are arranged in a 3×3 matrix in the upper, middle, and lower layers of the diffusion furnace cavity, synchronously collecting temperature data at preset intervals and transmitting temperature gradient values ​​via the 4-20mA+HART protocol. The online refractometers are deployed at the inlet and outlet of the phosphorus source delivery pipeline to monitor concentration fluctuations in real time and output 3σ values ​​via the RS485 Modbus-RTU protocol.

[0013] The oxidation PLC system connects two types of detection devices, including a quartz crystal microbalance and a linear CCD scanning system. The quartz crystal microbalance is suspended above the conveyor belt at the outlet of the oxidation furnace. It triggers sampling when each cell passes through and transmits film thickness uniformity data via TCP / IP protocol. The linear CCD scanning system is installed on a horizontal track, and its scanning rate matches the production line rate. It sends the film thickness distribution variance in real time via Profinet protocol.

[0014] The annealing PLC system connects two types of monitoring devices, including a thermopile sensor and a thermal conductivity gas sensor. The thermopile sensor is distributed at four temperature measurement points at the inlet and four at the outlet of the annealing furnace cooling zone. It polls and collects temperature data at preset intervals and calculates the cooling rate via a CAN bus. The thermal conductivity gas sensor is embedded in the exhaust pipe. It extracts gas samples at preset intervals to analyze the N2 / H2 ratio and transmits the concentration value via a 4-20mA analog signal.

[0015] This data acquisition layer integrates multi-process sensor networks through an industrial IoT gateway to achieve high-precision data acquisition throughout the entire process: the doping PLC system uses a distributed thermocouple matrix and an online refractometer to accurately monitor the temperature gradient of the diffusion furnace and changes in phosphorus source concentration, ensuring the stability of the doping process; the oxidation PLC system combines a quartz crystal microbalance and high-speed CCD scanning to capture the uniformity and distribution characteristics of oxide film thickness in real time, improving the ability to monitor oxidation quality; the annealing PLC system deploys a multi-channel thermopile and gas sensors to dynamically track the cooling rate and atmosphere composition, enhancing the reliability of annealing process control. The entire system, through multi-protocol fusion transmission and spatialized sensor layout, constructs a high spatiotemporal resolution process parameter acquisition system, providing comprehensive and accurate data support for subsequent intelligent decision-making, while also enhancing the early detection capability of abnormal process conditions.

[0016] Preferably, the intelligent decision-making layer includes a process data cube, a random forest hidden split classification model, and a collaborative analysis engine;

[0017] The process data cube integrates a multi-process PLC system through an industrial IoT gateway, collects process data from the data acquisition layer in real time, constructs a spatiotemporally aligned process dataset, and outputs a multi-dimensional feature matrix containing time, process, and parameter dimensions.

[0018] The random forest split classification model employs an ensemble learning approach, combining multiple decision trees for classification prediction. Specifically: training data is selected using random sampling with replacement; each decision tree randomly selects some features as candidate features when its nodes split; and the final classification result is determined through a majority voting mechanism.

[0019] The collaborative analysis engine employs a process parameter correlation prediction metric method to quantify the correlation between features. This method includes: assessing the correlation strength between features through proxy splitting in a decision tree; and measuring the correlation based on the substitutability of features in decision rules.

[0020] The association prediction measurement method is based on a decision tree proxy splitting mechanism to establish a feature association evaluation model, which is used to handle the nonlinear relationship between process parameters.

[0021] This intelligent decision-making layer achieves spatiotemporal alignment and structured integration of multi-source heterogeneous process data through a process data cube, providing a high-dimensional feature matrix for quality prediction. The random forest microcrack classification model effectively improves the model's generalization ability and microcrack identification accuracy by utilizing ensemble learning and random feature selection mechanisms, while reducing the risk of overfitting. The collaborative analysis engine, based on surrogate splitting and feature substitutability analysis, deeply mines the nonlinear correlations between process parameters, enhancing the explanatory power of complex process interactions. Through multi-level collaboration of data fusion, ensemble learning, and correlation analysis, the entire system achieves accurate prediction of microcrack defects and intelligent analysis of process influencing factors in photovoltaic cell manufacturing, providing a reliable decision-making basis for process optimization and improving the timeliness and accuracy of quality risk warnings.

[0022] Preferably, the input to the random forest cryptic classification model is a cube of key physical quantity data, and the input training data is... Including N samples; the construction of the random forest split classification model includes the following steps:

[0023] Bootstrap sampling: Sampling with replacement is performed on the training set to generate J independent bootstrap sample sets;

[0024] Decision tree construction: Train an unpruned classification and regression decision tree for each bootstrap sample set;

[0025] Forest generation: J decision trees are constructed by repeating the bootstrap sampling and decision tree construction steps to form a random forest;

[0026] Feature importance analysis: Feature importance is assessed by calculating the average Gini decrease of each feature across all decision trees.

[0027] This random forest microcrack classification model enhances the coverage of the sample space through a bootstrap sampling strategy, ensuring the diversity and robustness of model training. The unpruned decision tree construction method fully captures the complex nonlinear relationship between process parameters and microcrack defects, improving classification accuracy. Integrating multiple decision trees to form a forest structure effectively reduces the risk of overfitting from a single model and improves prediction stability. Feature importance analysis based on Gini descent values ​​objectively identifies key process parameters, providing clear directions for process optimization. This modeling method significantly improves the classification accuracy and generalization ability of microcrack defects, while enhancing the model's interpretability for high-dimensional process features, providing a reliable data-driven solution for predicting the manufacturing quality of photovoltaic cells.

[0028] Preferably, the decision tree construction includes:

[0029] When a node splits, m candidate features are randomly selected from M features;

[0030] The optimal split point is selected using the Gini coefficient, where the Gini coefficient is... The calculation method is as follows:

[0031] ;in, This represents the total number of samples in the node. Let be the number of samples belonging to class d in the node, where D is the number of classes. For nodes The proportion of samples in category d;

[0032] Calculate the Gini coefficient after splitting: ; in, This represents the node currently being evaluated in the decision tree, which contains a set of samples and data labels; This represents a specific splitting rule that produces two subsets. and ; and These are the left and right child nodes after the split. , , The number of samples;

[0033] Choosing to lower the Gini The most significant feature and the split point:

[0034] .

[0035] This decision tree construction method effectively enhances model diversity by randomly selecting candidate features, avoiding overfitting and improving generalization ability. The Gini coefficient-based split point selection mechanism accurately evaluates feature partitioning performance, ensuring that each node maximizes the differentiation of different categories of hidden crack defects. By calculating the Gini coefficient decrease value before and after splitting, the feature selection and splitting strategy are dynamically optimized, thereby improving the classification accuracy and interpretability of the decision tree. This method not only strengthens the model's ability to capture the nonlinear relationships of complex process parameters but also improves the stability and reliability of hidden crack prediction, providing efficient data-driven decision support for quality control in photovoltaic cell manufacturing.

[0036] Preferably, the feature importance analysis includes an unbiased feature importance assessment method:

[0037] For each tree The original classification accuracy was calculated using the out-of-bag (OOB) samples. ;

[0038] Permutation features The value in the unbiased feature importance value disrupts the association between features and labels;

[0039] Calculate the accuracy after permutation ;

[0040] The unbiased feature importance value is calculated by comparing the difference between the original accuracy and the accuracy after permutation. The calculation formula is as follows:

[0041]

[0042] The larger the unbiased feature importance value, the more important the feature.

[0043] This unbiased feature importance assessment method effectively eliminates bias interference in traditional feature assessment through an innovative out-of-bag (OOB) sample permutation strategy, significantly improving the objectivity and reliability of feature importance analysis. Utilizing OOB samples unique to random forests for dual validation maintains the independence of the assessment process while ensuring computational efficiency. Through a mechanism comparing eigenvalue permutation with classification accuracy, the actual contribution of each process parameter to microcrack prediction is precisely quantified, making the identification of key process parameters more scientifically based. This method not only enhances the model's interpretability but also provides precise directions for process optimization. Furthermore, by eliminating interference from irrelevant features, it further improves the generalization ability and stability of the microcrack prediction model.

[0044] Preferably, the feature importance analysis further includes a Gini gain importance assessment method:

[0045] For each tree Each node Record splitting characteristics The decline in the Gini coefficient ;

[0046] For each feature, accumulate the Gini decrease when it is used as a splitting feature across all decision trees;

[0047] The Gini gain importance value is calculated using the following formula:

[0048]

[0049] Among them, the Gini gain importance value indicates the stronger the discriminative power of the feature in the decision rule. for The total number of times it was used for splitting.

[0050] This Gini gain importance assessment method achieves dynamic quantitative evaluation of the discriminative ability of process parameters by deeply mining the Gini decrease during the splitting process of decision tree nodes. The importance value is calculated based on the cumulative Gini decrease of features across all decision trees, accurately reflecting the actual contribution of each process parameter in the microcrack classification decision and effectively identifying key process variables that significantly impact quality. By capturing the discriminative ability of features in multi-level decision rules, this method not only enhances the reliability of model feature selection but also improves the accuracy of process parameter importance ranking, providing data support for process optimization. Simultaneously, it strengthens the analytical ability of the random forest model for key influencing factors in photovoltaic cell manufacturing quality.

[0051] Preferably, the collaborative analysis engine uses a process parameter correlation prediction metric method to quantify the correlation between characteristic physical quantities, including the following steps:

[0052] Determine the optimal split: at the decision tree node Selective features And the split point u is taken as the optimal splitting rule;

[0053] Search agent splitting: for candidate features Find the split point v to make the agent split rule Maximize the consistency with the optimal split, i.e., for candidate features Select all possible split points v, and select those that make Maximize v;

[0054] Calculate the sample proportion: Statistically calculate the proportion of samples that satisfy the optimal splitting and surrogate splitting rules, including: sample proportion of And proportion of If the optimal split is satisfied, then proceed to the left child node; proportion In and When the proportion of samples satisfies the optimal splitting and surrogate right rule; when the proportion In and At that time, the proportion of samples that satisfy the optimal splitting and surrogate right rule;

[0055] Calculate the node-level correlation metric: Based on the optimal surrogate split point v, substitute the proportions into the following formula to obtain the process parameter correlation prediction metric. :

[0056]

[0057] Forest-level aggregation: Traverse all decision tree nodes, for each pair of features The average of the process parameter correlation prediction metrics is taken as the final correlation metric.

[0058] This collaborative analysis engine achieves precise quantification of nonlinear correlations among process parameters through an innovative proxy splitting mechanism. Based on consistency analysis of optimal splitting of decision tree nodes and proxy splitting rules, it can deeply mine potential correlation patterns among features; through sample proportion statistics and node-level correlation measurement calculations, it effectively captures the interaction of process parameters in complex decision paths; and finally, forest-level aggregation improves the stability of evaluation results. This method overcomes the limitations of traditional linear correlation analysis, significantly improves the analytical accuracy of correlations among multi-process parameters, provides a reliable basis for process collaborative optimization, and enhances the ability of quality prediction models to explain complex coupling effects in manufacturing processes.

[0059] Preferably, the collaborative analysis engine also includes a risk quantification mapping mechanism, which: classifies the hidden crack risk into levels 0-5; compares the changes in the hidden crack risk level between the current round and the previous round; generates control instructions based on the latest risk level and its changing trend; and transmits the control instructions to the control execution layer for dynamic adjustment of process parameters.

[0060] This risk quantification and mapping mechanism achieves real-time quantification and precise early warning of manufacturing process risks by dynamically assessing the risk level and changing trends of microcracks. Based on a multi-level risk classification system, it can accurately identify abnormal process states and determine their severity; through trend analysis of risk level changes between rounds, it can predict the direction of quality deterioration in advance; combined with intelligently generated closed-loop control commands, it makes process parameter adjustments more predictable and targeted. This mechanism significantly improves the risk control capabilities of the manufacturing process, realizing full-chain management from risk monitoring and trend prediction to proactive intervention, effectively reducing the probability of microcrack defects, and enhancing the adaptive adjustment capability of the process system to quality fluctuations.

[0061] Preferably, the dynamic compensation strategy includes a diffusion furnace temperature control module and an annealing cooling compensator, wherein:

[0062] The diffusion furnace temperature control module includes a temperature sensing system, a heating execution unit, and a cascade PID control system, wherein:

[0063] The temperature sensing system includes an internal thermocouple for detecting the actual temperature of the silicon wafer and an external thermocouple for monitoring the temperature gradient of the furnace body.

[0064] The heating execution unit adopts a segmented independent heating structure, and the heating power of each segment can be adjusted independently.

[0065] The cascaded PID control system includes a main PID controller and a secondary PID controller. The main PID controller outputs a target temperature setpoint based on the actual temperature of the silicon wafer, and the secondary PID controller adjusts the heating power in real time based on the furnace wall temperature as feedback.

[0066] The cascaded PID control system also includes a temperature simulation module, which is used to predict the drift of the constant temperature zone and dynamically adjust the heating power at both ends of the furnace tube to compensate for heat loss and keep the temperature difference on the silicon wafer surface within the set range.

[0067] The annealing cooling compensator includes a temperature sensing system, a gradient cooling system, and an airflow control system, wherein:

[0068] The temperature sensing system is used to monitor the temperature distribution of the silicon wafer in real time.

[0069] The gradient cooling system includes an air cooling unit, a water cooling unit, and a heat exchange unit. The air cooling unit uses adjustable flow rate gas spray cooling, and the water cooling unit uses liquid cooling medium circulation cooling.

[0070] The airflow control system includes a fan and a flow equalization device for optimizing the distribution of cooling airflow;

[0071] The annealing cooling compensator adopts a multi-stage gradient cooling strategy, using different cooling methods in different temperature ranges, and can dynamically adjust the cooling parameters according to the detected temperature difference of the silicon wafer.

[0072] This dynamic compensation strategy, through cascaded PID control and temperature simulation prediction in the diffusion furnace temperature control module, achieves precise temperature regulation and optimized thermal field uniformity of the silicon wafer, effectively suppressing the effects of isothermal drift and heat loss. The annealing cooling compensator, combined with multi-mode gradient cooling and intelligent airflow control, dynamically adjusts cooling parameters based on the real-time temperature distribution of the silicon wafer, significantly improving the stability and uniformity of the cooling process. The entire system, through the coordinated operation of a temperature sensing network, multi-level control algorithms, and segmented actuators, forms a closed-loop temperature control system from heating to cooling. This not only significantly reduces the risk of microcracks caused by thermal stress but also enhances the process system's adaptive compensation capability for temperature fluctuations, providing a reliable guarantee for the high-quality manufacturing of photovoltaic cells.

[0073] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. By integrating multi-process PLC data with a random forest model, the problem of data silos is solved, high-precision microcrack prediction is achieved, and the efficiency of quality inspection is significantly improved; 2. Based on real-time risk level feedback, the diffusion furnace temperature and annealing parameters are automatically adjusted to form a closed-loop control, reducing microcracks and improving process stability; 3. The correlation of process parameters is quantified using a proxy splitting method, providing analysis of key influencing factors and optimizing production decisions; 4. By combining a multi-source sensor network, the accurate acquisition of key process parameters is achieved, providing reliable data support for the model. Attached Figure Description

[0074] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0075] Figure 2 This is a schematic diagram of the furnace body and heating module of the present invention;

[0076] Figure 3 This is a schematic diagram of the overall random forest classification model of the present invention;

[0077] Figure 4 This is a schematic diagram of the random forest classification process of the present invention. Detailed Implementation

[0078] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0079] This invention provides an adaptive quality management system for the doping / oxidation / annealing process of large-size photovoltaic cells. It constructs a digital manufacturing control system integrating "sensor acquisition - microcrack prediction - correlation quantification." By integrating a PLC system, it collects key physical quantities (including diffusion temperature gradient, oxide film thickness uniformity, and annealing cooling rate) of processes such as doping, oxidation, annealing, coating, soldering, and cell assembly positioning in real time. A high-precision microcrack classification classifier is established based on a random forest model, and an innovative correlation prediction metric method is applied to quantify the nonlinear dependence strength between process parameters, revealing the cross-process parameter coupling mechanism. For example... Figure 1 As shown, the system includes a data acquisition layer, an intelligent decision-making layer, and a control execution layer.

[0080] like Figure 2As shown, the data acquisition layer includes a doping PLC system, an oxidation PLC system, and an annealing PLC system. These systems are integrated with a multi-process PLC system via the Industrial Internet of Things (IIoT) and can acquire data from the doping, oxidation, and annealing processes in real time. This data includes diffusion temperature gradient, phosphorus source concentration fluctuation, film thickness uniformity, growth rate drift, annealing cooling rate, and N2 / H2 ratio. After acquiring the above data, the data acquisition layer calculates the edge gateway, i.e., constructs a spatiotemporally aligned process data cube. The timestamp synchronization error of this dataset is less than 50ms, and it can output a multi-dimensional feature matrix of time × process × parameter based on the process data cube.

[0081] The intelligent decision-making layer includes a process data cube, a random forest microcrack classification model, and a collaborative analysis engine. This layer analyzes the collected data, calculates the microcrack risk level, deploys the random forest classification model, receives a multi-dimensional feature matrix, and outputs microcrack risk levels of 0-5. The fragmentation rates for microcrack risk levels 0-5 can be manually set to be greater than 0.1%, 0.5%, 1.5%, 2.1%, 5.7%, and 18.2% respectively (0-level fragmentation rate 0.1%, 5-level fragmentation rate 18.2%). When the system identifies a risk level greater than 3, a compensation command will be triggered.

[0082] The control execution layer includes a dynamic compensation strategy, a diffusion furnace temperature control module, and an annealing cooling compensator. This layer generates OPC UA instructions based on the risk level to trigger the annealing furnace PID controller.

[0083] Secondly, a digital detection module for the risk of microcracks in photovoltaic cells is provided, based on a large-size photovoltaic cell manufacturing quality prediction and process dynamic control system as described in any one of the first aspects, wherein the PLC system in the data acquisition layer integrates three process chains through a multi-process sensing module via an industrial Internet of Things gateway.

[0084] The PLC for the doping process connects to two types of sensors: nine K-type thermocouples and two online refractometers. The nine Omega KMQSS-125G-12 K-type thermocouples are arranged in a 3×3 matrix in the upper / middle / lower layers of the diffusion furnace cavity, synchronously acquiring temperature data every 10 seconds with an accuracy of ±0.5℃. Temperature gradient values ​​are transmitted via 4-20mA+HART protocol. The two Anton Paar L-Rix 5100 online refractometers are deployed at the inlet and outlet of the phosphorus source delivery pipeline, monitoring concentration fluctuations in real time at a frequency of 10Hz. They output a 3σ value via RS485 Modbus-RTU protocol. This value measures the dispersion or fluctuation range of a set of data, indicating the degree of dispersion of data points relative to their average value. A small 3σ value indicates that the measured value is very stable with minimal fluctuation; conversely, a large 3σ value indicates a large fluctuation range, relatively poor stability, or significant fluctuations in the process itself, suggesting anomalies.

[0085] The oxidation process PLC connects to two types of detection devices, including two quartz crystal microbalances and one linear CCD scanning system. The two INFICON SQM-160 quartz crystal microbalances are suspended 1.5m above the conveyor belt at the oxidation furnace exit. Sampling is triggered when each cell passes through at a speed of ≤5 seconds / cell, and film thickness uniformity data is transmitted via TCP / IP protocol. The Banner P4CM-080Q linear CCD scanning system is installed on a transverse track. Because the production line speed is 180mm / s, it scans the cell surface at a speed of 200mm / s to match the production line speed, and transmits the film thickness distribution variance in real time via Profinet protocol.

[0086] The annealing process PLC connects to two types of monitoring equipment, including eight thermopile sensors and one thermal conductivity gas sensor. The eight thermopile sensors (Tyco TS318-1B0814) are distributed at four temperature measurement points each at the inlet and outlet of the annealing furnace cooling zone. They poll and collect temperature data every 0.5 seconds and calculate the cooling rate via the CAN bus. The one thermal conductivity gas sensor (SIK GMS800) is embedded in the exhaust pipe. It extracts a gas sample every 30 seconds to analyze the N2 / H2 ratio and transmits the concentration value via a 4-20mA analog signal.

[0087] Thirdly, a digital intelligent decision-making layer for the risk of microcracks in photovoltaic cells is provided, based on a large-size photovoltaic cell manufacturing quality prediction and process dynamic control system as described in any of the first aspects. The intelligent decision-making layer includes a process data cube, a random forest microcrack classification model, and a collaborative analysis engine.

[0088] The process data cube integrates a multi-process PLC system through an industrial IoT gateway, collects data from the PLC system in real time, and constructs a process dataset of time × process × parameter. The collected raw process dataset is processed to output a spatiotemporally aligned multidimensional feature matrix that can characterize the entire multi-process integrated manufacturing process. The raw data includes, but is not limited to, diffusion temperature gradient, phosphorus source concentration fluctuation, film thickness uniformity, growth rate drift, cooling rate, N2 / H2 ratio, and welding temperature.

[0089] The random forest split classification model combines multiple decision trees. Each time, the dataset is randomly selected with replacement, and some features are randomly selected as input. In the classification problem, the combiner selects the majority classification result as the final result, which can effectively improve the accuracy of quantization mapping.

[0090] The collaborative analysis engine refers to a method for predicting the correlation of process parameters. It's an innovative indicator used to quantify the correlation between features through surrogate splitting similarity calculation. Its core idea is to assess the strength of the correlation between parameter features through surrogate splitting in a decision tree. Unlike traditional correlation coefficients, this method directly measures the substitutability of features in decision rules, making it more suitable for high-dimensional nonlinear data.

[0091] Fourthly, a digital random forest microcrack classification model for photovoltaic cell microcrack risk is provided, based on the digital intelligent decision-making layer for photovoltaic cell microcrack risk described in any of the third aspects. The input of the random forest microcrack classification model is the aforementioned output process data cube, and the input training data is... The training dataset contains N samples and M features, where J represents the number of decision trees (DTs) and m represents the number of features randomly selected when a node splits. ;

[0092] like Figure 3 As shown, the random forest cryptic classification model mainly includes the following steps:

[0093] S1: Collect data from the data acquisition layer: diffusion temperature gradient, phosphorus source concentration fluctuation, film thickness uniformity, growth rate drift, annealing cooling rate, and N2 / H2 ratio. Optimize the data by removing obvious outliers from the original data, assigning appropriate class labels to the output, and then processing the data according to... Figure 4 The steps involved in training all decision trees (dt) within the random forest model using the preprocessed input and output are as follows:

[0094] Step a: Bootstrapping from the original training set TD ( N A random sample with replacement was drawn from a sample of [number] samples. N Next, a self-service sample set is formed. Dj Mathematically, it is represented as: ;

[0095] Step b: Constructing a decision tree - random feature selection (column sampling). When splitting a node, m candidate features are randomly selected from all M features. Mathematically, this is expressed as: ;

[0096] Step c: Constructing a decision tree - finding the optimal split point: Calculating the Gini coefficient of each node:

[0097]

[0098] Where D is the number of categories, For nodes The proportion of samples in category d. For nodes The number of samples in category d, where n is the total number of samples in the nodes;

[0099] Calculate the Gini coefficient after splitting:

[0100]

[0101] in, This represents the node currently being evaluated in the decision tree, which contains a set of samples and data labels; This represents a specific splitting rule that produces two subsets. and , and These are the left and right child nodes after the split. The number of samples;

[0102] Choosing to lower the Gini The most significant feature and the split point:

[0103]

[0104] Finally, the optimal split is selected:

[0105]

[0106] Step d: Perform the split when When, it is assigned to the left node. Otherwise, it will be assigned to the right node. ,right and Repeat the above steps until the conditions are met;

[0107] Step e: Output the decision tree.

[0108] S2: Feature Importance Analysis: In this section, the importance of all features is quantified, and their impact on the classification performance of the parameters is analyzed. Two effective quantitative metrics are used: unbiased feature importance and gain-improving feature importance.

[0109] Gain-enhanced feature importance analysis, based on the internal structure of the decision tree and traversing the nodes of each tree, reflects the feature contribution in real time and identifies key control parameters in the manufacturing process. It is more efficient and more sensitive to data. The steps are as follows:

[0110] Step a: Extract split gain: Collect from all nodes across all trees ;

[0111] Step b: Accumulation and Counting: Calculate the total gain and number of splits ;

[0112] Step c: Calculate the average gain: .

[0113] Unbiased feature importance can improve the predictive performance of random forest models, quantify the proportion of features in the decision-making process, reveal the intrinsic mechanism of parameters on quality classification, and enhance the robustness against interference. The steps are as follows:

[0114] Step a: Out-of-Bag (OOB) Sample Prediction: Evaluate the raw accuracy of each tree using out-of-bag samples. ;

[0115] Step b: Feature permutation test: Destroy the features, associate them with labels, and then re-predict to obtain the result. ;

[0116] Step c: Importance quantification: .

[0117] The two feature importance analyses described above apply the permutation importance system to manufacturing feature analysis, providing data-driven decision-making basis for process optimization and significantly improving key parameter identification, process path optimization, and quality control.

[0118] S3: After calculating the importance of each parameter, calculate the process parameter correlation prediction metric for each feature using the following steps and formulas, and plot it as a heatmap. The correlation prediction metric method is a feature correlation quantification tool for decision tree models. It can directly map the decision substitution relationship between process parameters and directly guide parameter optimization strategies based on heatmap visualization. The steps are as follows:

[0119] Step a: Determining the optimal split refers to the splitting at the decision tree node. Above, select features And the split point u is used as the optimal splitting rule; search agent splitting refers to the splitting of decision tree nodes using features. After performing the optimal split, search for another feature. Splitting rules This aims to replicate the optimal split result as closely as possible. For candidate features... Find the split point v to make the agent split rule Maximize the consistency with the optimal split. That is, for candidate features... Iterate through all possible split points v and select the one that makes v the split point. Maximize v.

[0120] Step b: Based on the optimal surrogate split point v, substitute the proportion into the following formula to obtain the process parameter correlation prediction metric.

[0121]

[0122] Among them, the sample proportion of And proportion of If the optimal split is satisfied, then proceed to the left child node; proportion In and When the proportion of samples satisfies the optimal splitting and surrogate right rule; when the proportion In and At that time, the proportion of samples that satisfy the optimal splitting and surrogate right rule;

[0123] Step c: Perform forest-level aggregation, which refers to traversing all nodes of the decision tree and performing aggregation on each pair of features. The average of the correlation prediction metrics is used as the final correlation metric. The risk quantification mapping level based on the final result guides the control layer strategy.

[0124] Step d: Based on the risk quantification mapping mechanism of the intelligent decision-making layer, the hidden crack risk level is obtained through the collaborative analysis engine and compared with the hidden crack risk level obtained in the previous round of analysis. A new round of instructions is derived using the new hidden crack risk level and its changes, and these instructions are then transmitted to the control execution layer. The risk quantification mechanism divides the risk level into 0-5 levels; different hidden crack risk levels will cause the control execution layer to perform different operations.

[0125] S4: As Figure 1 The system flowchart shown illustrates the process: Step S1 involves acquiring process parameters; Step S2 involves calculating the microcrack risk level using a model; and Step S3 involves generating the dependency matrix and control rules. The actions of the control execution layer are determined based on the risk quantification mapping level, which is artificially divided into the following five levels:

[0126] 1. Risk level 0-1: The fragmentation rate is less than 0.5%, marked as a green safety alarm, indicating that the process parameters are in the optimal range and the current production pace can be maintained;

[0127] 2. Risk Level 2: Fragmentation rate 0.5%-1.5% (not listed in the table but actually enabled), triggering a blue observation alert. It is recommended that process engineers check the parameter fluctuation trend.

[0128] 3. Risk Level 3: The fragmentation rate rises to 2.1%, activating a yellow warning alarm. The system automatically records abnormal batches and prompts for optimization of the annealing cooling curve.

[0129] 4. Risk level 4: Fragmentation rate 5.7%, orange high-risk alarm is activated, the gradient cooling module of the annealing furnace is forcibly activated (≤2.8℃ / s) and subsequent wafer feeding of this batch is suspended;

[0130] 5. Risk Level 5: Fragmentation rate reaches 18.2%, triggering a red emergency alarm, simultaneously executing equipment emergency stop, MES system isolation of defective batches, and pushing accident diagnosis report to the quality management terminal.

[0131] Fifthly, a digital control execution system for the risk of microcracks in photovoltaic cells is provided, based on the digital management and control system for the risk of microcracks in photovoltaic cells as described in any of the first aspects, wherein the control execution layer includes a dynamic compensation strategy consisting of a diffusion furnace temperature control module and an annealing cooling compensator mechanism;

[0132] The dynamic compensation strategy in the control execution layer includes a diffusion furnace temperature control module and an annealing cooling compensator. The furnace temperature diffusion module includes a temperature sensing system, a heating execution unit, and a control system. Its core compensation method is a dual-loop PID controller with end-mounted auxiliary heating. The temperature sensing system includes internal and external thermocouples. The former is used to detect the actual temperature of the silicon wafer, and the latter is used to monitor the furnace temperature gradient. The heating execution unit is made of 8mm heating wire and is divided into three independent heating sections, with the power of each section adjusted by a thyristor. The control system includes a cascaded PID loop and a temperature simulation module. The cascaded PID loop uses the actual process temperature detected by the internal thermocouple as a reference, outputs the set value through the PID1 controller, and uses the furnace wall temperature detected by the external thermocouple as feedback to quickly adjust the heating current through the PID2 controller. The temperature simulation module is based on a thermodynamic model to predict and dynamically compensate for the drift in the constant temperature zone. This module uses the actual temperature detected by the internal thermocouple as a reference, outputs the target value through a slow-response PID, and uses the furnace wall temperature detected by the external thermocouple as feedback. It uses a fast-response PID to adjust the heating power in real time to suppress the temperature disturbance caused by the ±10% fluctuation of the power grid within ±0.5℃. It also uses a dynamic compensation technology that increases the heating power by 5% at both ends of the furnace tube to compensate for the heat loss at the opening, so that the temperature difference on the silicon wafer surface is <±0.8℃.

[0133] The annealing cooling compensator includes a temperature sensing system, a gradient cooling system, and an airflow control system. Its core compensation methods are airflow optimization, dynamic electrical injection, and temperature difference closed-loop. The temperature sensing system shares a sensor with the diffusion furnace temperature control module to detect the actual temperature of the silicon wafer. The gradient cooling system includes an air-cooling unit, a water-cooling platform, and a heat exchanger. The air-cooling unit functions as a high-pressure nitrogen spray with a flow rate adjustable from 0 to 100 m / s and a cooling rate of 50 to 200 °C / s. The water-cooling platform has an embedded copper pipe channel with a flow rate of 10 L / min, using ethylene glycol aqueous solution as the cooling medium. The heat exchanger adopts a plate heat exchange structure to transfer heat from the silicon wafer. The airflow control system includes a centrifugal fan with a wind pressure of 0 to 5 kPa and an air volume of 0 to 3000 m³ / h, and a honeycomb flow equalization plate to reduce the unevenness of wind speed distribution. When a temperature difference greater than 2℃ is detected, the heating plate power and nitrogen flow rate are automatically adjusted. The cooling gradient is divided into three stages: Stage 1 is 1100℃→800℃, using high-pressure nitrogen spray with a cooling rate of 50℃ / s; Stage 2 is 800℃→500℃, using air cooling and pulsed electric injection with a cooling rate of 50℃ / s; Stage 3 is 500℃→room temperature, using water cooling and strong DC injection with a cooling rate of 5℃ / s, which can effectively reduce the fragmentation rate.

Claims

1. A large size photovoltaic cell doping / oxidation / annealing process adaptive quality management system, characterized in that, The system comprises a data acquisition layer, an intelligent decision-making layer and a control execution layer connected in sequence. The data acquisition layer comprises a doping PLC system, an oxidation PLC system and an annealing PLC system, which are used to collect process parameters in the manufacturing process of photovoltaic cells, construct a spatiotemporally aligned process data cube, and output a multi-dimensional feature matrix containing time, process and parameters; the doping PLC system, the oxidation PLC system and the annealing PLC system respectively integrate multi-process sensing modules through industrial Internet of Things gateways to form a data acquisition network of three process chains. The doping PLC system is connected with two types of sensors, including a K-type thermocouple and an online refractometer; the K-type thermocouple is arranged in a 3x3 matrix on the upper, middle and lower layers of the inner cavity of the diffusion furnace, and synchronously collects temperature data every preset time interval, and transmits the temperature gradient value through a 4-20mA+HART protocol; the online refractometer is respectively arranged at the inlet and outlet of the phosphorus source conveying pipeline, and monitors the concentration fluctuation in real time, and outputs a 3σ value through a RS485 Modbus-RTU protocol. The oxidation PLC system is connected with two types of detection devices, including a quartz crystal microbalance and a linear array CCD scanning system; the quartz crystal microbalance is suspended above the oxidation furnace outlet conveyor belt, and triggers sampling when each cell passes by, and transmits the film thickness uniformity data through a TCP / IP protocol; the linear array CCD scanning system is installed on a horizontal rail, and the scanning rate is matched with the production line rate, and the film thickness distribution variance is sent in real time through a Profinet protocol. The annealing PLC system is connected with two types of monitoring devices, including a thermopile sensor and a thermal conductivity gas sensor; the thermopile sensor is distributed at four temperature measuring points at the inlet and outlet of the cooling zone of the annealing furnace, and polls and collects temperature data every preset time interval, and calculates the cooling rate through a CAN bus; the thermal conductivity gas sensor is embedded in the exhaust pipeline, and extracts gas samples every preset time interval to analyze the N2 / H2 ratio, and transmits the concentration value through a 4-20mA analog signal; The intelligent decision-making layer comprises a process data cube, a random forest hidden crack classification model and a collaborative analysis engine, which are used to receive the multi-dimensional feature matrix, calculate the hidden crack risk level through the random forest hidden crack classification model, and output the hidden crack risk level; The control execution layer comprises a dynamic compensation strategy composed of a diffusion furnace temperature control module and an annealing cooling compensator, which is used to generate an OPC UA instruction according to the hidden crack risk level, and trigger the annealing furnace PID regulator through the OPC UA instruction to dynamically adjust the process parameters of the annealing furnace.

2. The quality management system of claim 1, wherein, The intelligent decision-making layer comprises a process data cube, a random forest hidden crack classification model and a collaborative analysis engine; The process data cube integrates multi-process PLC systems through industrial Internet of Things gateways, collects process data of the data acquisition layer in real time, constructs a spatiotemporally aligned process data set, and outputs a multi-dimensional feature matrix containing time, process and parameter dimensions. The random forest hidden crack classification model adopts an ensemble learning method to perform classification prediction by combining multiple decision trees, wherein: the training data is selected by random sampling with replacement; when splitting the nodes of each decision tree, part of the features are randomly selected as candidate features; and the final classification result is determined by a majority voting mechanism; The synergy analysis engine adopts a process parameter correlation prediction measurement method to quantify the correlation between features, which includes: evaluating the correlation strength between features through proxy splitting in the decision tree; and measuring the replaceability of features in the decision rule. The correlation prediction measurement method establishes a feature correlation evaluation model based on the decision tree proxy splitting mechanism to process the nonlinear relationship between process parameters.

3. The quality management system of claim 1, wherein, The input of the random forest hidden crack classification model is a key physical quantity data cube, and the input training data including N samples; the construction of the random forest hidden crack classification model includes the following steps: Bootstrap sampling: sampling the training set with replacement to generate J independent bootstrap sample sets; Decision tree construction: training an unpruned classification and regression decision tree for each bootstrap sample set; Forest generation: repeating the bootstrap sampling and decision tree construction steps to construct J decision trees to form a random forest; Feature importance analysis: evaluating feature importance by calculating the average Gini reduction value of each feature in all decision trees.

4. The quality management system of claim 3, wherein, The decision tree construction includes: When splitting the nodes, m candidate features are randomly selected from M features; The optimal split point is selected using the Gini coefficient, where the Gini coefficient The calculation is: ; wherein, is the total number of samples in the node, is the number of samples in the node belonging to the dth class, D is the number of classes, is the proportion of samples of class d in the node, ; Calculate the Gini coefficient after splitting: ; wherein, represents the node in the decision tree that is currently being evaluated; represents a specific splitting rule that results in two subsets and ; and are the left and right child nodes after splitting, are the number of samples; Selecting to reduce the Gini Biggest features and split points: 。 5. The quality management system of claim 3, wherein, The feature importance analysis includes an unbiased feature importance evaluation method: For each tree Compute the raw classification accuracy with its out-of-bag samples OOB ; substitution features values in un-biased feature importance, scrambling the association of features with labels; Computing the accuracy of the permutation ; Calculate the unbiased feature importance value by comparing the difference between the original accuracy and the accuracy after replacement, and the calculation formula is: ; where a larger unbiased feature importance value indicates a more important feature.

6. The quality management system of claim 3, wherein, The feature importance analysis also includes a Gini gain importance evaluation method: for each node of each tree ;​​​ For each feature, accumulate the Gini reduction amount when it is used as a splitting feature in all decision trees; Calculate the Gini gain importance value, and the calculation formula is: ; wherein, the greater the Gini Gain importance value indicates the stronger the distinguishing ability of the feature in the decision rule, is total number of times the feature is used for splitting.

7. The quality management system of claim 1, wherein, The synergy analysis engine adopts a process parameter correlation prediction measurement method to quantify the correlation between characteristic physical quantities, including the following steps: determining an optimal split: selecting a feature and a split point u as an optimal split rule at a decision tree node The selected feature and the split point u as the optimal split rule; Search agent split: for candidate features , select all possible split points v, select v that maximizes ​ Compute sample proportions: Compute the proportions of samples that satisfy the optimal split and proxy split rules, including: sample proportions of and proportions of satisfy the optimal split, in which case go to the left child node; proportions of and satisfy the optimal split and proxy right rule; proportions of and satisfy the optimal split and proxy right rule; Compute node-level correlation metrics: Based on the optimal proxy split point v, plug the ratio into the following formula to get the process parameter correlation prediction metrics : ; forest-level aggregation: traverse all decision tree nodes, average the correlation prediction metrics for each pair of features of process parameters, as the final correlation metric.

8. The quality management system of claim 1, wherein, The synergy analysis engine also includes a risk quantification mapping mechanism, which: divides the hidden crack risk into 0-5 levels; compares the change in the hidden crack risk level of the current round with that of the previous round; Generates control instructions based on the latest risk level and its trend; and transmits the control instructions to the control execution layer for dynamic adjustment of process parameters.

9. The quality management system of claim 1, wherein, The dynamic compensation strategy includes a diffusion furnace temperature control module and an annealing cooling compensator, wherein: The diffusion furnace temperature control module includes a temperature sensing system, a heating execution unit, and a cascade PID control system, wherein: The temperature sensing system includes an internal thermocouple for detecting the actual temperature of the silicon wafer and an external thermocouple for monitoring the temperature gradient of the furnace body; The heating execution unit adopts a segmented independent heating structure, and the heating power of each segment can be independently adjusted; The cascade PID control system includes a main PID controller and a secondary PID controller, the main PID controller outputs a target temperature set value based on the actual temperature of the silicon wafer, and the secondary PID controller adjusts the heating power in real time based on the furnace wall temperature feedback; The cascade PID control system also includes a temperature simulation module for predicting the drift of the constant temperature zone and dynamically adjusting the heating power at both ends of the furnace tube to compensate for heat loss and control the temperature difference on the surface of the silicon wafer within a set range; The annealing cooling compensator includes a temperature sensing system, a gradient cooling system, and an airflow control system, wherein: The temperature sensing system is used to monitor the temperature distribution of the silicon wafer in real time; The gradient cooling system comprises a gas cooling unit, a water cooling unit and a heat exchange unit, the gas cooling unit adopts adjustable flow rate gas spray cooling, and the water cooling unit adopts liquid cooling medium circulation cooling; The air flow control system comprises a fan and a flow equalization device, which is used for optimizing the distribution of cooling air flow; The annealing temperature drop compensator adopts a multi-stage gradient cooling strategy, different cooling methods are used in different temperature intervals, and the cooling parameters are dynamically adjusted according to the detected silicon wafer temperature difference.

Images