Contactless transmission path coordinated control system

Abstract

The present application belongs to the technical field of cooperative control, and particularly relates to a contactless transmission path cooperative control system, which comprises a monitoring module, a risk coupling module, a cooperative control module and a feedback module; the monitoring module collects control parameters, evaluation indexes and process parameters and quality data of a processing process of a contactless transmission process in real time; the risk coupling module constructs a transmission and processing risk matrix by using the data, and generates a risk coupling matrix and a production coupling chain based on an algorithm such as a hidden Markov algorithm; the cooperative control module generates transmission and processing control instructions according to the coupling risk information of each node in the chain; and the feedback module realizes real-time adjustment through reverse abnormality evaluation positioning, so that the production process continuously conforms to the standard process chain; the present application realizes risk cooperative control and closed-loop optimization of the transmission and processing processes, and significantly improves the production yield and the intelligent level of the system.

Description

Technical Field

[0001] This invention belongs to the field of collaborative control technology, and particularly relates to a contactless transmission path collaborative control system. Background Technology

[0002] In the high-end packaging and testing of semiconductor discrete components, especially automotive-grade devices, the cleanliness, stability, and coordination with processing equipment during material transfer across multiple processes such as expansion, assembly, soldering, cleaning, packaging, and testing are crucial. Existing production lines mostly employ mechanical contact transfer methods, which are prone to introducing particulate contamination, mechanical vibration, and surface scratches. Furthermore, the lack of an intelligent coupling mechanism based on real-time quality data between transfer and processing equipment hinders proactive risk perception and closed-loop control, thus restricting further improvements in product yield and production efficiency. Therefore, there is an urgent need for a system that can achieve contactless, low-disturbance transfer throughout the entire process and enable real-time collaborative risk assessment and dynamic control of the transfer and processing processes. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a contactless transmission path collaborative control system. This system includes a monitoring module, a risk coupling module, a collaborative control module, and a feedback module. The monitoring module collects control parameters and evaluation indicators of the contactless transmission process, as well as process parameters and quality data of the processing. The risk coupling module uses this data to construct a transmission and processing risk matrix and generates a risk coupling matrix and production coupling chain based on algorithms such as Hidden Markov Models. The collaborative control module generates transmission and processing control commands based on the risk information coupled at each node in this chain. The feedback module achieves real-time adjustments through reverse anomaly assessment and location, ensuring that the production process continuously conforms to the standard process chain. This invention realizes collaborative risk management and closed-loop optimization of the transmission and processing processes, significantly improving production yield and system intelligence.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] The contactless transmission path collaborative control system includes: a monitoring module, a risk coupling module, a collaborative control module, and a feedback module;

[0006] The monitoring module is used to acquire first state data and second state data of the contactless production process of semiconductor devices according to the configured intelligent sensor network.

[0007] The risk coupling module is used to construct a first transmission risk matrix based on the first state data, and simultaneously construct a second processing risk matrix using the second state data. It also combines a coupling algorithm and a risk transfer algorithm according to the priority of the contactless semiconductor device production process to obtain the risk coupling matrix and the production coupling chain.

[0008] The collaborative control module is used to respond to the standard manufacturing process chain of semiconductor devices and combine the coupling risk triggering information corresponding to each processing control node and adjacent transmission connection relationship of the production coupling chain to generate coupling risk control instructions; the coupling risk control instructions include at least one transmission control instruction and one processing control instruction.

[0009] The feedback module is used to acquire the transmission and processing status information of the production coupling chain in real time, and combine it with the preset distributed evaluation algorithm and standard production process chain to perform reverse anomaly evaluation and location along the production coupling chain, and feed back the location anomaly information to the production coupling chain to adjust the risk coupling matrix and the production coupling chain in real time until the contactless production process meets the standard production process chain in real time.

[0010] Specifically, the first state data is associated with the contactless transfer control process of the semiconductor device, including at least contactless transfer control parameters and transfer evaluation indicators; the second state data is associated with the contactless processing control process of the semiconductor device, including at least processing process control parameters, processing process evaluation indicators, processing equipment operating status parameters and processing contamination evaluation indicators.

[0011] Each processing control node in the production coupling chain corresponds one-to-one with each processing technology node in the contactless production process; each connection in the production coupling chain corresponds one-to-one with the transmission process between two adjacent processing technology nodes in the contactless production process; the standard production process chain includes standard processing technology evaluation indicators and standard transmission evaluation indicators, which correspond one-to-one with each processing control node and transmission stage in the production coupling chain, and are used to evaluate the real-time processing technology and transmission errors corresponding to each processing control node and transmission stage in the production coupling chain.

[0012] Specifically, the risk coupling module includes a first risk unit and a second risk unit; the transmission evaluation indicators include at least a transmission pollution indicator, a transmission delay indicator, and a transmission location accuracy indicator; the workflow of the first risk unit includes:

[0013] Based on the transmission evaluation index parameters corresponding to each transmission stage and the evaluation algorithm, the transmission evaluation index score and estimated confidence level of the corresponding transmission stage are obtained.

[0014] Based on the score of each transmission evaluation index in the current transmission stage and the corresponding transmission evaluation index threshold, the transmission evaluation error set of the current transmission stage is obtained.

[0015] Based on the transmission evaluation error set of the current transmission stage and the processing pass probability of the corresponding processing technology of the next processing control node, the contribution of each transmission evaluation error to the processing pass probability of the corresponding processing technology of the next processing control node is obtained through gradient boosting tree.

[0016] Based on the contribution and estimated confidence of each transmission evaluation error to the processing qualification probability of the corresponding processing technology of the next processing control node, the first processing failure risk probability of the current transmission stage to the next processing control node is obtained through a Bayesian neural network.

[0017] Based on the probability of the first processing failure risk corresponding to each transmission stage and the preset first risk threshold, a processing risk triggering information set is constructed; the processing risk triggering information corresponds to the processing control command.

[0018] Specifically, the workflow of the second risk unit includes:

[0019] The system collects the operating status parameters of the processing equipment, the processing technology control parameters, and the processing pollution assessment index parameters corresponding to each processing control node of the production coupling chain in real time. The collected data is then preprocessed by timestamp alignment and standardization to obtain the preprocessed operating status parameters of the processing equipment, the processing technology index parameters, and the processing pollution assessment index parameters.

[0020] Based on the operating status parameters of the processing equipment and the historical processing pass rate change rate, obtain the performance score and change trend of the processing equipment corresponding to the current processing control node;

[0021] Based on the performance score and trend of the processing equipment corresponding to the current processing control node, and combined with the corresponding processing technology index parameters, the probability of processing success of the current processing control node is obtained through the first evaluation algorithm with built-in Bayesian function.

[0022] Using the processing control node's processing pass risk probability as a condition and combining it with processing pollution assessment index parameters, a first evaluation algorithm with a built-in Bayesian function is used to obtain the first transmission trigger risk probability corresponding to the current processing control node. At the same time, a transmission risk trigger information set is constructed by combining the first transmission trigger risk probability corresponding to each processing control node with the corresponding first transmission trigger risk probability threshold; the transmission risk trigger information corresponds to the transmission control command.

[0023] Specifically, the risk coupling module further includes a risk coupling unit; the risk coupling unit is used to construct a first transmission risk matrix according to the first processing failure risk probability of all transmission stages and the order of transmission stage timestamps, and simultaneously construct a second processing risk matrix according to the first transmission trigger risk probability of all processing control nodes and the processing technology priority, obtain a risk coupling matrix based on the first transmission risk matrix and the second processing risk matrix combined with a hidden Markov model, construct a production coupling chain based on all processing control nodes combined with the risk coupling matrix and a loose coupling algorithm according to the order priority of transmission and processing in the production process, map the transmission evaluation index to the transmission coupling connection in the production coupling chain, map the processing technology control parameters, processing technology evaluation index and processing pollution evaluation index to the corresponding processing control nodes, and embed the processing risk trigger information set into the input end of each processing control node and the transmission risk trigger information set into the output end of each processing control node.

[0024] Specifically, the risk coupling matrix is ​​obtained by combining the first transmission risk matrix and the second processing risk matrix with a hidden Markov model, including:

[0025] Based on the timestamp sequence of each transmission stage in the first transmission risk matrix and the corresponding first processing failure risk probability value, and the process priority sequence of each processing control node in the second processing risk matrix and the corresponding first transmission trigger risk probability value, the joint multi-dimensional observation sequence of the hidden Markov model is obtained according to the priority order of each transmission and processing process in the production process.

[0026] Based on the hidden state sequence and joint observation sequence labeled in historical production data, the correspondence between the hidden state and the joint observation value is extracted. The Baum-Welch algorithm is used to perform maximum likelihood estimation iterative training on the initial state probability vector, state transition probability matrix and observation probability density function represented by Gaussian mixture model of the hidden Markov model until the rate of change of the model likelihood function is lower than the preset convergence threshold.

[0027] Based on the trained Hidden Markov Model, the transition probability values ​​between all states are extracted from its state transition probability matrix to construct a risk coupling matrix.

[0028] Specifically, obtaining the risk coupling matrix based on the first transmission risk matrix and the second processing risk matrix using a hidden Markov model also includes:

[0029] Based on the real-time acquired joint observation sequence and the risk coupling matrix, the Viterbi dynamic programming algorithm is used to decode the optimal hidden state path at the current moment, and the risk state evolution trend of the production process is predicted based on the transition probability distribution of the terminal state in the optimal hidden state path. Based on the risk state evolution trend of the production process and combined with the radial kernel function, the backward risk distribution prediction function is fitted.

[0030] Based on the backward risk distribution prediction function, the trained Hidden Markov Model is fine-tuned a second time using the Markov Chain Monte Carlo sampling algorithm. This updates the posterior distribution of the model parameters and the backward risk distribution prediction deviation, maximizing the likelihood of the posterior distribution with the sample data and minimizing the backward risk distribution prediction deviation. This completes the model training and yields the fine-tuned risk coupling matrix and the backward risk distribution prediction function.

[0031] Specifically, the process of obtaining the hidden state sequence and the joint observation sequence includes:

[0032] Based on the probability of the first transmission trigger risk corresponding to the previous processing control node in the current transmission stage, and combined with the real-time collected transmission delay index parameters and transmission position accuracy index parameters, confidence analysis is conducted to obtain the confidence level of delay risk and transmission positioning risk.

[0033] Meanwhile, the transmission pollution confidence level can be obtained by combining the transmission pollution index parameters of the current transmission stage with the processing pollution risk probability corresponding to the previous processing control node.

[0034] By combining the confidence levels of delay risk, transmission location risk, and transmission contamination with a comprehensive fuzzy algorithm and the evaluation score level range, the transmission evaluation index score and the processing risk level corresponding to the first processing failure risk probability are obtained.

[0035] At the same time, based on the probability of the first transmission trigger risk of the corresponding processing control node after the corresponding transmission stage, combined with the preset transmission risk level range, the corresponding transmission risk level is obtained.

[0036] Based on the risk coupling matrix corresponding to the processing risk level and transmission risk level between adjacent transmission stages and processing control nodes, the hidden state sequence and joint observation sequence of the processing control node after the corresponding transmission stage are obtained.

[0037] Specifically, the feedback module includes a two-way monitoring unit;

[0038] The bidirectional monitoring unit is used to monitor the first state data and second state data corresponding to each processing control node in real time along the production coupling chain, starting from the current processing control node or transmission stage. At the same time, based on the first state data and second state data corresponding to the transmission stage or processing control node, combined with the risk coupling matrix and the backward risk distribution prediction function, it predicts the first processing failure risk probability or the first transmission trigger risk probability corresponding to the next transmission stage or processing control node. It also monitors the first trigger result information between the first processing failure risk probability and the processing risk trigger information, or the second trigger result information between the first transmission trigger risk probability and the transmission risk trigger information.

[0039] Specifically, the feedback module also includes an evaluation and early warning unit and a feedback adjustment unit; the evaluation and early warning unit is used to combine the first and second state data of forward monitoring, the first processing failure risk probability or the first transmission trigger risk probability of backward prediction and the corresponding first trigger result information or second trigger result information, combined with the corresponding preset causal reasoning model and the standard processing technology evaluation index built into each processing control node to perform forward anomaly evaluation and correlation anomaly analysis, and obtain the correlation anomaly risk transmission chain and corresponding anomaly characteristic information.

[0040] The feedback adjustment unit is used to obtain an anomaly resolution strategy based on the associated anomaly risk transmission chain and corresponding anomaly characteristic information, combined with a preset anomaly strategy library and a fast index constructed based on historical index information, to perform real-time forward anomaly processing control parameter correction and backward processing control parameter pre-adjustment on the associated anomaly risk transmission chain, until the production process meets the standard production process chain in real time.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention addresses existing shortcomings by constructing a contactless transmission path collaborative control system, achieving deep risk coupling and intelligent closed-loop control of the transmission and processing stages in semiconductor manufacturing. The system dynamically generates a transmission and processing risk matrix using multi-source monitoring data, accurately predicts risk transfer paths based on a Hidden Markov Model, and implements forward-looking decision-making through a collaborative control module. This significantly improves production line yield and equipment utilization, fundamentally eliminating quality defects caused by transmission contamination, vibration, and processing parameter drift. Furthermore, through reverse anomaly localization and real-time adjustment mechanisms, it greatly reduces the intensity of manual intervention, achieving autonomous optimization of the production process. Attached Figure Description

[0043] Figure 1 This is a simplified flow chart of the contactless production process of the present invention;

[0044] Figure 2This is a block diagram of the contactless transmission path collaborative control system of the present invention; Detailed Implementation

[0045] Example 1

[0046] In current contactless semiconductor production lines, the transport and processing systems typically employ independent control strategies, lacking a real-time risk linkage mechanism across stages. Risk factors such as contamination, vibration, and positioning deviations generated during transport cannot be dynamically predicted and proactively addressed, leading to decreased yields in the processing stages and difficulties in real-time traceability and closed-loop control of production line anomalies. Please refer to [link / reference]. Figure 2 The present invention provides an embodiment of a contactless transmission path collaborative control system, comprising: a monitoring module, a risk coupling module, a collaborative control module, and a feedback module;

[0047] The monitoring module is used to acquire first state data and second state data of the contactless production process of semiconductor devices according to the configured intelligent sensor network.

[0048] The risk coupling module is used to construct a first transmission risk matrix based on the first state data, and simultaneously construct a second processing risk matrix using the second state data. It also combines a coupling algorithm and a risk transfer algorithm according to the priority of the contactless semiconductor device production process to obtain the risk coupling matrix and the production coupling chain.

[0049] The collaborative control module is used to respond to the standard manufacturing process chain of semiconductor devices and combine the coupling risk triggering information corresponding to each processing control node and adjacent transmission connection relationship of the production coupling chain to generate coupling risk control instructions; the coupling risk control instructions include at least one transmission control instruction and one processing control instruction.

[0050] The feedback module is used to acquire the transmission and processing status information of the production coupling chain in real time, and combine it with the preset distributed evaluation algorithm and standard production process chain to perform reverse anomaly evaluation and location along the production coupling chain, and feed back the location anomaly information to the production coupling chain to adjust the risk coupling matrix and the production coupling chain in real time until the contactless production process meets the standard production process chain in real time.

[0051] It should be further explained that the first state data associated with the contactless transmission control process of the semiconductor device in this embodiment includes at least contactless transmission control parameters and transmission evaluation indicators; the second state data associated with the contactless processing control process of the semiconductor device includes at least processing process control parameters, processing process evaluation indicators, processing equipment operating status parameters, and processing contamination evaluation indicators; wherein, the contactless transmission control parameters in the first state data include at least transmission speed setpoints, acceleration curve parameters, and suspension height setpoints; the transmission evaluation indicators include at least measured values ​​of suspended particulate matter concentration, the difference between standard transmission time and actual transmission time, and the X-axis deviation, Y-axis deviation, and rotation angle deviation of the semiconductor device in the target processing node coordinate system; the second state data includes at least temperature setpoints, pressure control values, process durations for each processing node, welding qualification rate and cleaning cleanliness score obtained through the vision inspection system, vibration frequency and temperature deviation values ​​of the processing equipment spindle, and particle concentration monitoring values ​​in the processing cavity. Since the processing control parameters corresponding to each processing node are different, some parameter examples are given here. The specific parameters should be set by those skilled in the art according to the parameters required by the equipment corresponding to the specific processing process, and will not be elaborated further here.

[0052] Each processing control node in the production coupling chain corresponds one-to-one with each processing technology node in the contactless production process; each connection in the production coupling chain corresponds one-to-one with the transmission process between two adjacent processing technology nodes in the contactless production process; the standard production process chain includes standard processing technology evaluation indicators and standard transmission evaluation indicators, which correspond one-to-one with each processing control node and transmission stage in the production coupling chain, and are used to evaluate the real-time processing technology and transmission errors corresponding to each processing control node and transmission stage in the production coupling chain.

[0053] Please see Figure 1 The production process in this embodiment includes processes such as film expansion, assembly, welding, cleaning, encapsulation, post-curing, waste removal, reflow soldering, electroplating (outsourcing), lead cutting, TMTT testing, final inspection, packaging, and OQC inspection. Each of these processes involves complex processing control technology and there is a risk of contamination during processing and transportation. This greatly reduces the production quality and precision of the high-precision TVS diodes and Zener diodes corresponding to this embodiment during the production process.

[0054] The manufacturing precision of semiconductor devices such as Schottky diodes and rectifier diodes has decreased. Therefore, this embodiment further reduces the risk probability of the entire production process by coupling the risks of each process before and after it.

[0055] It should be further noted that the risk coupling module in this embodiment includes a first risk unit, a second risk unit, and a risk coupling unit; the transmission evaluation indicators include at least a transmission contamination indicator, a transmission delay indicator, and a transmission location accuracy indicator.

[0056] The workflow of the first risk unit includes:

[0057] Based on the transmission evaluation index parameters corresponding to each transmission stage and the evaluation algorithm, the transmission evaluation index score and estimated confidence level of the corresponding transmission stage are obtained.

[0058] Based on the score of each transmission evaluation metric in the current transmission stage and the corresponding transmission evaluation metric threshold, the transmission evaluation error set for the current transmission stage is obtained. An example illustrating this transmission evaluation error set for the current transmission stage is shown below:

[0059] The transmission evaluation index parameters are collected in real time during the transmission stage from the welding process to the cleaning process. The concentration of suspended particulate matter is obtained by a laser particle counter deployed on the transmission platform as a transmission pollution index. The transmission execution time is obtained by a high-precision photoelectric encoder as a transmission delay index. The positional deviation data of the material in the current processing control node coordinate system is obtained by a vision positioning system as a transmission position accuracy index.

[0060] Based on the transported pollution index, the transported pollution index score is obtained by comparing the measured pollution concentration value with the standard pollution concentration value and calculating their relative deviation percentage. At the same time, based on the measurement uncertainty level provided by the laser particle counter calibration certificate, the estimated confidence level of the transported pollution index is obtained by calculating the reciprocal of the ratio of the measurement uncertainty to the measured value.

[0061] Based on the transmission delay index, the transmission delay index score is obtained by calculating the difference between the measured transmission time and the standard transmission time and dividing it by the maximum allowable time deviation. At the same time, based on the accuracy parameters provided by the photoelectric encoder accuracy level certificate, the probability that the time measurement value falls within the standard value range is calculated by the normal distribution cumulative function to obtain the estimated confidence level of the transmission delay index.

[0062] Based on the aforementioned transmission position accuracy index, a two-dimensional coordinate system is established with the standard position point of the target processing control node as the center. The X-axis deviation, Y-axis deviation, and rotation angle deviation of the current position of the material in this coordinate system are measured. By normalizing each deviation component with the corresponding maximum permissible deviation and weighting them together, the transmission position accuracy index score is obtained. At the same time, based on the variance data of repeated measurements by the visual positioning system, the confidence level of the transmission position accuracy index is obtained by calculating the reciprocal normalized value of the variance.

[0063] Based on the transmission pollution index score, transmission delay index score, and transmission location accuracy index score, a weighted summation is performed using predefined weighting coefficients to obtain the total transmission evaluation index score. Simultaneously, based on the estimated confidence scores of each index, a confidence score synthesis calculation is performed using DS evidence theory to obtain the total estimated confidence score.

[0064] Based on the total score of the transmission evaluation index, the transmission pollution index error, transmission delay index error, and transmission position accuracy index error are obtained by calculating the absolute difference between each individual score and the corresponding index threshold and dividing by the threshold. Finally, a transmission evaluation error set containing the three error values ​​is formed, and the error set is associated with and stored with the corresponding individual estimated confidence scores.

[0065] Based on the transmission evaluation error set of the current transmission stage and the processing pass probability of the corresponding processing technology of the next processing control node, the contribution of each transmission evaluation error to the processing pass probability of the corresponding processing technology of the next processing control node is obtained through gradient boosting tree.

[0066] Based on the contribution and estimated confidence of each transmission evaluation error to the processing qualification probability of the corresponding processing technology of the next processing control node, the first processing failure risk probability of the current transmission stage to the next processing control node is obtained through a Bayesian neural network.

[0067] For example, to better describe the probability of first processing failure at the next processing control node in the current transmission stage, we take the process from welding to cleaning as an example, and the specific process is as follows:

[0068] The transmission stage from welding to cleaning is monitored to obtain a transmission assessment error set. Simultaneously, historical processing pass rate data for recent production batches in the cleaning process is acquired, and this historical pass rate data is used as the target variable. This includes standardized data across three dimensions: transmission contamination index error, transmission delay index error, and transmission position accuracy index error.

[0069] The model is trained using a gradient boosting tree algorithm with multiple regression trees. The transmission pollution index error, transmission delay index error, and transmission position accuracy index error are used as input features, and the processing qualification rate of the cleaning process is used as the supervised learning objective. A series of decision tree models are constructed sequentially using a forward stepwise algorithm. The node splitting parameters and leaf node output values ​​of each decision tree are gradually optimized by minimizing the mean square error loss function.

[0070] Based on a well-trained gradient boosting tree model, the SHAP value analysis framework is used to calculate the marginal contribution of each input feature. By constructing all possible feature subsets and calculating the output change brought about by adding each feature value to the subset, the contribution of transmission contamination index error, transmission delay index error and transmission position accuracy index error to the prediction result of the processing qualification rate of the cleaning process is quantified, and the standardized contribution values ​​corresponding to the three errors are obtained.

[0071] It should be further explained that the construction and training process of the gradient boosting tree model in this embodiment includes:

[0072] Based on the initial training dataset, which includes historical transmission evaluation error data as feature variables and the corresponding processing pass rate as target variables, the hyperparameters of the gradient boosting tree are set, including the number of trees, learning rate, tree depth and minimum number of leaf node samples, and the basic predictor is initialized as a constant model.

[0073] By calculating the negative gradient between the initial predicted value and the actual target value, the pseudo residual sequence of the first iteration is obtained. Based on the pseudo residual sequence and feature variable data, a regression decision tree is constructed for fitting. The tree node splitting criterion adopts the principle of maximizing the reduction of variance, and the leaf node output value is obtained by optimizing the loss function through linear search.

[0074] Based on the constructed regression decision tree, the update amount for the current iteration is obtained by multiplying the learning rate by the tree model prediction result, the current prediction result is added to the update amount to obtain the new prediction value, and the pseudo residual sequence is recalculated based on the new prediction value.

[0075] Repeatedly execute the pseudo residual calculation, regression tree construction and model update process to generate a series of weak learners in sequence. Each iteration builds a new regression tree based on the residual results of the previous round, and controls the contribution weight of each tree through the learning rate.

[0076] During each iteration, the loss function value on the validation set is monitored using the ten-fold cross-validation method. When the validation loss no longer decreases after several consecutive iterations, the early stopping mechanism is triggered to terminate the training process.

[0077] Based on the set of weak learners obtained after all training iterations, the outputs of all tree models are combined by weighted summation to obtain the final gradient boosting tree prediction model. This model can accurately reflect the nonlinear mapping relationship between transmission evaluation error and processing pass rate.

[0078] Based on the trained gradient boosting tree model, the tree structure parameters and node splitting rules are saved to a binary file using model persistence technology for subsequent real-time prediction tasks.

[0079] Based on the obtained error contribution values ​​and their corresponding measurement confidence data, the processing failure risk probability is calculated using a Bayesian neural network with a random variational inference layer. Specifically, each error contribution is first multiplied by its measurement confidence to obtain a weighted feature. Then, these weighted features are input into a three-layer neural network structure containing random weight distribution. Finally, the network output is converted into a risk probability value through the sigmoid activation function.

[0080] Based on the probabilistic generation characteristics of Bayesian neural networks, the Monte Carlo sampling method is used to perform multiple forward propagation calculations. By repeatedly sampling from the weight distribution of the neural network and performing the forward calculation process, a large number of risk probability prediction samples are collected. The arithmetic mean of these prediction samples is calculated as the final first processing failure risk probability. At the same time, the sample variance of these prediction values ​​is calculated as the confidence index of the risk probability.

[0081] Finally, the calculated first processing failure risk probability is compared with the preset first risk threshold. When the risk probability exceeds the threshold, a processing risk trigger signal for the cleaning process is generated. This signal, along with the corresponding confidence index and timestamp information, is encoded into a data packet and stored in the processing risk trigger information queue, providing real-time data input for the downstream risk management decision-making system.

[0082] Based on the probability of the first processing failure risk corresponding to each transmission stage and the preset first risk threshold, a processing risk triggering information set is constructed; the processing risk triggering information corresponds to the processing control instructions; it should be further noted that the processing risk triggering information in this embodiment includes at least a risk level identifier, a risk probability range description, a risk confidence level description, a timestamp, a target processing step number, a recommended control instruction type, and an instruction parameter description. For example: when a high cleaning failure risk is calculated in the welding to cleaning transmission stage, the triggering information includes a high risk identifier, a relatively high risk probability, a relatively high confidence level, a timestamp, a cleaning step number, and an instruction to reduce the cleaning fluid flow rate; when a test failure risk is identified in the packaging to testing transmission stage, the triggering information includes a medium risk identifier, a medium risk probability, a medium confidence level, a timestamp, a test step number, and an instruction to extend the test time; when an extremely high failure risk is detected in the reflow soldering process, the triggering information also includes an emergency pause instruction and equipment self-test parameter requirements.

[0083] It should be further explained that the workflow of the second risk unit in this embodiment includes:

[0084] The system collects the operating status parameters of the processing equipment, the processing technology control parameters, and the processing pollution assessment index parameters corresponding to each processing control node of the production coupling chain in real time. The collected data is then preprocessed by timestamp alignment and standardization to obtain the preprocessed operating status parameters of the processing equipment, the processing technology index parameters, and the processing pollution assessment index parameters.

[0085] Based on the operating status parameters of the processing equipment and the historical processing pass rate change rate, obtain the performance score and change trend of the processing equipment corresponding to the current processing control node;

[0086] It should be further explained that the process of obtaining the performance score and trend of the processing equipment corresponding to the current processing control node in this embodiment includes:

[0087] Based on the real-time acquisition of the operating status parameters of each processing control node in the production coupling chain, the operating status parameters of each processing device are synchronously acquired through a distributed data acquisition system, and the network time protocol is used to add millisecond-level timestamps to all parameter data; the operating status parameters include at least one of the following: vibration frequency, spindle speed, temperature deviation value, and pressure fluctuation value.

[0088] By using a time series alignment algorithm, the operating status parameters of the equipment with different acquisition frequencies are uniformly interpolated onto a common time axis to form a time-synchronized data sequence. Based on the preset upper and lower limits of the normal range of equipment parameters, the minimum-maximum normalization method is used to map all parameter values ​​to between zero and one, thereby obtaining standardized operating status parameters of the processing equipment.

[0089] Based on the preprocessed operating status parameters of the processing equipment, a data segment of the recent production batch is extracted through a sliding window, and the standard deviation and mean offset of each parameter in the data segment are calculated as stability features; at the same time, historical processing pass rate data for the corresponding time period is extracted from the production database, and the moving average of the pass rate change rate is calculated.

[0090] A mapping model between equipment parameter features and pass rate change rate is established using the random forest regression algorithm. Stability features are used as input variables and pass rate change rate is used as output variables for supervised learning training. Ten-fold cross-validation is used to optimize model parameters to ensure model generalization ability.

[0091] Based on the trained random forest model, the stability characteristics of the equipment parameters in the current time window are input, and the performance score of the processing equipment corresponding to the current processing control node is obtained through model inference. This score represents the ability index to maintain the current pass rate level.

[0092] Based on the performance score sequence of multiple consecutive time windows, the score change curve is fitted by linear regression, and the slope value of the curve is calculated as an indicator of the trend of equipment performance change; when the slope value is negative, it indicates that the performance is declining, and when the slope value is positive, it indicates that the performance is improving.

[0093] The real-time calculated equipment performance scores and trend indicators are stored in the equipment performance database, and an independent performance file is established for each processing control node, which is used for calculating the probability of passing the processing risk and making equipment early warning decisions.

[0094] Based on the performance score and trend of the processing equipment corresponding to the current processing control node, and combined with the corresponding processing technology index parameters, the probability of processing success of the current processing control node is obtained through the first evaluation algorithm with built-in Bayesian function.

[0095] Using the processing pass risk probability of the current processing control node as a condition, combined with processing contamination assessment index parameters, a first evaluation algorithm with a built-in Bayesian function is used to obtain the first transmission trigger risk probability corresponding to the current processing control node. Simultaneously, a transmission risk trigger information set is constructed by combining the first transmission trigger risk probability corresponding to each processing control node with the corresponding first transmission trigger risk probability threshold; the transmission risk trigger information corresponds to the transmission control command. It should be further noted that each piece of information in the transmission risk trigger information set in this embodiment corresponds one-to-one with a processing control node in the production coupling chain, and each piece of information is associated with a corresponding transmission control command, specifically including the following:

[0096] Risk level identification: determined by comparing the probability of the first transmission trigger risk corresponding to the current processing control node with the preset first transmission trigger risk probability threshold. The risk level includes at least four categories: low risk, medium risk, high risk, and extremely high risk.

[0097] Description of the first transmission trigger risk probability range: Define the range in which the probability of the first transmission trigger risk corresponding to the current processing control node is located. This range is divided based on the preset first transmission trigger risk probability threshold.

[0098] Risk confidence level description: The standard deviation of the probability distribution in the process of calculating the probability of the first transmission trigger risk is determined based on the built-in Bayesian function. The confidence level includes at least three categories: high confidence, medium confidence, and low confidence.

[0099] Timestamp: A time stamp used to collect the operating status parameters of the processing equipment and the processing pollution assessment index parameters corresponding to the current processing control node, or a millisecond-level time stamp used to calculate the probability of the first transmission trigger risk of the current processing control node.

[0100] Associated processing control node number: This is a unique identifier for the processing control node that currently triggers the transmission risk, used to identify the specific processing control node to which the risk belongs;

[0101] Recommended transmission control command type: This refers to the category of control actions proposed to address current transmission risks, including at least categories such as adjusting transmission speed, optimizing transmission positioning accuracy, pausing transmission, strengthening transmission channel cleaning, and improving transmission stability;

[0102] Recommended transmission control command parameter description: To specify the specific parameter requirements for the corresponding recommended transmission control command type, clarify the adjustment range, operation value or execution standard of the control command to ensure that the transmission control command can be executed directly.

[0103] For example, in a medium-risk scenario, the details are as follows: Risk level identifier: medium risk; First transmission trigger risk probability range description: within the preset probability range corresponding to medium risk; Risk confidence level description: high confidence (meets the standard deviation judgment criteria for the probability distribution corresponding to high confidence); Timestamp: the millisecond-level time identifier corresponding to the collection or calculation; Associated processing control node number: the unique identifier of the corresponding assembly and processing control node; Recommended transmission control command type: adjust transmission speed + optimize transmission positioning accuracy; Recommended transmission control command parameter description: adjust the transmission speed of the assembly and processing control node in subsequent transmission stages to a value suitable for the medium-risk scenario, and at the same time increase the transmission positioning calibration frequency to a frequency suitable for this risk scenario.

[0104] It should be further explained that, in this embodiment, the risk coupling unit is used to construct a first transmission risk matrix according to the first processing failure risk probability of all transmission stages and the order of transmission stage timestamps, and simultaneously construct a second processing risk matrix according to the first transmission trigger risk probability of all processing control nodes and the processing technology priority. Based on the first transmission risk matrix and the second processing risk matrix, a risk coupling matrix is ​​obtained by combining a hidden Markov model. Based on all processing control nodes, combined with the risk coupling matrix and a loose coupling algorithm, a production coupling chain is constructed according to the priority of transmission and processing in the production process. The transmission evaluation index is mapped to the transmission coupling connection in the production coupling chain, and the processing technology control parameters, processing technology evaluation index, and processing pollution evaluation index are mapped to the corresponding processing control nodes. At the same time, the processing risk trigger information set is embedded into the input end of each processing control node, and the transmission risk trigger information set is embedded into the output end of each processing control node.

[0105] It should be further explained that the construction process of the production coupling chain in this embodiment includes:

[0106] Based on all processing control nodes and their natural order in the production process, an ordered sequence of processing control nodes is obtained by sorting them according to the priority of the processing technology using a time-series sorting algorithm.

[0107] Based on the ordered sequence of processing control nodes and the risk coupling matrix, the risk transmission intensity index between each pair of adjacent nodes is obtained by querying the risk transfer probability value between adjacent processing control nodes in the risk coupling matrix.

[0108] Based on the core principles of risk transmission strength index and loosely coupled algorithm (i.e., maintaining relative independence between nodes while allowing the flow of risk information), a weighted graph construction algorithm is used to treat each processing control node as a vertex and use risk transmission strength as edge weight to obtain a preliminary weighted directed graph structure.

[0109] Based on a preliminary weighted directed graph structure, the minimum spanning tree algorithm optimizes the connection paths, selecting the edge with the highest risk transmission strength as the primary coupling connection, while retaining secondary connections as backup paths, thus obtaining the optimized coupled graph model. In this embodiment, the edge with the highest risk transmission strength is chosen as the primary coupling connection because the core task of the risk coupling unit is to reveal and control the transmission of risks in the production process, rather than simply establishing physical connections; therefore, the structure of the coupling chain must be dominated by the risk itself. The edge with the highest risk transmission strength represents the highest probability and most significant impact of risk cascading or amplification effects between two processes. By identifying these high-risk transmission paths as primary coupling connections, the system can precisely focus limited monitoring resources and control computing power on the links that pose the greatest threat to the stability of the entire production chain, thereby achieving the most efficient risk prevention and control. At the implementation level, the system converts risk transmission strength into connection cost; the higher the risk transmission strength, the greater the cost or risk of using this edge to connect two nodes. The goal of the minimum spanning tree algorithm is to find the set of paths connecting all nodes with the minimum total cost. Choosing these edges with the lowest cost corresponds precisely to selecting the edges with the highest risk transmission intensity in the original data as the backbone, because they are the most critical and indispensable links in the risk network. Through this method, the constructed backbone network minimizes overall risk exposure from a global perspective, ensuring the structural optimality of the production coupling chain.

[0110] It should be further explained that the construction and optimization process of the coupling graph model in this embodiment includes:

[0111] Based on the preliminary weighted directed graph structure, where vertices represent processing control nodes, directed edges represent the transmission process between nodes, and edge weights represent the risk transmission strength index calculated based on the risk coupling matrix, the cost weights suitable for minimum spanning tree calculation are obtained by taking the reciprocal of the risk transmission strength value and converting it into the cost weights required by the minimum spanning tree algorithm.

[0112] Based on the cost-weighted directed graph, the Prim minimum spanning tree algorithm is used to iteratively expand from the initial processing control node. In each iteration, all edges between the current set of connected nodes and the set of unconnected nodes are traversed, and the edge with the smallest cost weight is selected to be added to the spanning tree, thereby ensuring that all nodes are connected with the minimum total cost, and obtaining the minimum spanning tree backbone network covering all processing control nodes.

[0113] Based on the obtained minimum spanning tree backbone network, all edges selected into the backbone network are marked by the edge identification algorithm. These edges are identified as the main coupling connections because they have the lowest cost weight, i.e. the highest actual risk transmission strength, and constitute the core path for risk monitoring and transmission in the production coupling chain.

[0114] Based on the initial weighted directed graph, the set of edges that were not selected into the backbone network is obtained by subtracting the backbone edge set of the minimum spanning tree from all edges of the original directed graph through edge set difference calculation.

[0115] Based on the remaining edge set, the edges are processed by a risk transmission strength descending sorting algorithm, and the edges with higher risk transmission strength are selected as backup transmission paths to be used when abnormal risk events or performance bottlenecks occur on the backbone network edges, thus obtaining the sorted backup connection set.

[0116] Based on the minimum spanning tree backbone network and the sorted set of backup connections, the backup connections are added as additional edges to the backbone network through a graph structure fusion algorithm to construct an enhanced graph model that includes both a high-intensity risk transmission backbone and redundant backup paths, thus obtaining an optimized coupled graph model.

[0117] Based on the optimized coupled graph model, the processing control parameters, processing evaluation indicators and processing pollution evaluation indicators corresponding to each processing control node are loaded into the attribute set of the corresponding node in the graph model through the node mapping protocol, so as to obtain a set of nodes with complete processing attribute information.

[0118] Based on the transmission evaluation index dataset, the transmission evaluation index (including transmission pollution index, transmission delay index, and transmission location accuracy index) of each transmission stage is mapped to the edges between adjacent nodes in the coupled graph model through a data association algorithm, thereby obtaining a set of connection edges with transmission risk attributes.

[0119] Based on the processing risk trigger information set, the processing risk trigger information (including risk level identifier, risk probability range description, risk confidence level description, etc.) of each processing control node is written into the input port data field of the corresponding node through the information embedding interface, thereby obtaining processing control nodes with risk input perception capabilities.

[0120] Based on the transmission risk triggering information set, the transmission risk triggering information of each transmission stage is written into the output port data field of the corresponding edge through the information embedding interface, thereby obtaining a transmission connection with risk output awareness capability.

[0121] Based on all risk-enhanced processing control nodes and risk-enhanced transmission connections, a graph traversal algorithm is used to integrate nodes and edges according to the priority of the production process, generating a continuous production coupling chain structure. Each node contains processing attributes and input risk information, and each connection contains transmission attributes and output risk information.

[0122] It should be further explained that in this embodiment, the risk coupling matrix is ​​obtained based on the first transmission risk matrix and the second processing risk matrix combined with a hidden Markov model, including:

[0123] Based on the timestamp sequence of each transmission stage in the first transmission risk matrix and the corresponding first processing failure risk probability value, and the process priority sequence of each processing control node in the second processing risk matrix and the corresponding first transmission trigger risk probability value, the joint multi-dimensional observation sequence of the hidden Markov model is obtained according to the priority order of each transmission and processing process in the production process.

[0124] Based on the hidden state sequence and joint observation sequence labeled in historical production data, the correspondence between the hidden state and the joint observation value is extracted. The Baum-Welch algorithm is used to perform maximum likelihood estimation iterative training on the initial state probability vector, state transition probability matrix and observation probability density function represented by Gaussian mixture model of the hidden Markov model until the rate of change of the model likelihood function is lower than the preset convergence threshold.

[0125] It should be further explained that the process of obtaining the hidden state sequence and the joint observation sequence in this embodiment includes:

[0126] Based on the probability of the first transmission trigger risk corresponding to the previous processing control node in the current transmission stage, and combined with the real-time collected transmission delay index parameters and transmission position accuracy index parameters, confidence analysis is conducted to obtain the confidence level of delay risk and transmission positioning risk.

[0127] Meanwhile, the transmission pollution confidence level can be obtained by combining the transmission pollution index parameters of the current transmission stage with the processing pollution risk probability corresponding to the previous processing control node.

[0128] By combining the confidence levels of delay risk, transmission location risk, and transmission contamination with a comprehensive fuzzy algorithm and the evaluation score level range, the transmission evaluation index score and the processing risk level corresponding to the first processing failure risk probability are obtained.

[0129] At the same time, based on the probability of the first transmission trigger risk of the corresponding processing control node after the corresponding transmission stage, combined with the preset transmission risk level range, the corresponding transmission risk level is obtained.

[0130] Based on the risk coupling matrix corresponding to the processing risk level and transmission risk level between adjacent transmission stages and processing control nodes, the hidden state sequence and joint observation sequence of the processing control node after the corresponding transmission stage are obtained.

[0131] Based on the trained Hidden Markov Model, the transition probability values ​​between all states are extracted from its state transition probability matrix to construct the risk coupling matrix.

[0132] Based on the real-time acquired joint observation sequence and the risk coupling matrix, the Viterbi dynamic programming algorithm is used to decode the optimal hidden state path at the current moment, and the risk state evolution trend of the production process is predicted based on the transition probability distribution of the terminal state in the optimal hidden state path. Based on the risk state evolution trend of the production process and combined with the radial kernel function, the backward risk distribution prediction function is fitted.

[0133] Based on the backward risk distribution prediction function, the trained Hidden Markov Model is fine-tuned a second time using the Markov Chain Monte Carlo sampling algorithm. This updates the posterior distribution of the model parameters and the backward risk distribution prediction deviation, maximizing the likelihood of the posterior distribution with the sample data and minimizing the backward risk distribution prediction deviation. This completes the model training and yields the fine-tuned risk coupling matrix and the backward risk distribution prediction function.

[0134] This application constructs an intelligent risk management system based on Hidden Markov Models (HMMs) and loosely coupled algorithms. This system achieves precise quantification, forward-looking prediction, and dynamic control of risks throughout the entire contactless production process, ultimately significantly improving the yield and reliability of high-precision device production. Specifically, by transforming transmission and processing status data into a risk probability matrix and utilizing HMMs to mine hidden state correlations in historical data, a risk coupling matrix is ​​trained. This transforms the originally implicit and elusive risk transmission paths into quantifiable and predictable probabilistic models, enabling the system to possess a profound understanding of the cross-process transmission effects of risks. Consequently, it achieves ultimate optimization of resource allocation and a leap in management efficiency. Based on the risk coupling matrix… By utilizing the minimum spanning tree algorithm to accurately identify the path with the highest risk transmission intensity as the main coupling chain, limited monitoring and computing resources are concentrated on the critical links that pose the greatest threat to overall stability, avoiding resource dispersion and waste. An optimal control topology structure with risk as the core is constructed, ultimately achieving system adaptability and high robustness. By embedding real-time risk triggering information into each node and connection of the coupling chain and combining it with a backward risk prediction function, the system can not only locate and correct anomalies that have occurred, but also make forward-looking inferences and pre-adjustments for potential future risks. This gives the production process a strong self-healing ability to dynamically approach the standard process chain, effectively resisting various disturbances and ensuring product consistency and high precision.

[0135] It should be further noted that the feedback module in this embodiment includes a two-way monitoring unit, an evaluation and early warning unit, and a feedback adjustment unit;

[0136] The bidirectional monitoring unit is used to monitor the first state data and second state data corresponding to each processing control node in real time along the production coupling chain, starting from the current processing control node or transmission stage. At the same time, based on the first state data and second state data corresponding to the transmission stage or processing control node, combined with the risk coupling matrix and the backward risk distribution prediction function, it predicts the first processing failure risk probability or the first transmission trigger risk probability corresponding to the next transmission stage or processing control node. It also monitors the first trigger result information between the first processing failure risk probability and the processing risk trigger information, or the second trigger result information between the first transmission trigger risk probability and the transmission risk trigger information.

[0137] The assessment and early warning unit is used to combine the first and second state data of forward monitoring, the first processing failure risk probability or the first transmission trigger risk probability of backward prediction and the corresponding first trigger result information or the second trigger result information, and combine the corresponding preset causal reasoning model with the standard processing technology assessment index built into each processing control node to perform forward anomaly assessment and associated anomaly analysis, so as to obtain the associated anomaly risk transmission chain and the corresponding anomaly feature information.

[0138] The feedback adjustment unit is used to obtain an anomaly resolution strategy based on the associated anomaly risk transmission chain and corresponding anomaly characteristic information, combined with a preset anomaly strategy library and a fast index constructed based on historical index information, to perform real-time forward anomaly processing control parameter correction and backward processing control parameter pre-adjustment on the associated anomaly risk transmission chain, until the production process meets the standard production process chain in real time.

[0139] It should be further explained that the process of real-time correction of forward abnormal machining process control parameters and advance adjustment of backward machining process control parameters in this embodiment includes:

[0140] Based on the associated anomaly risk transmission chain and corresponding anomaly feature information, a multi-dimensional feature matching algorithm is used to calculate the Euclidean distance between the historical anomaly feature vectors in the preset anomaly strategy library to obtain the set of historical handling schemes with the highest matching degree. Based on the set of historical handling schemes, a weighted average algorithm is used to fuse the process parameter correction amount of each scheme, and Bayesian optimization is performed in combination with its confidence rating to obtain the optimal process parameter correction amount and comprehensive confidence of the current anomaly processing control node.

[0141] Based on the optimal process parameter correction amount, a binary parameter overwrite command is sent to the target processing control node through the CRC check protocol of the real-time control interface, and the optical measurement unit built into the node is activated simultaneously to perform online re-inspection of processing quality; based on the real-time processing evaluation index data obtained from the re-inspection, the response curve between the correction amount and the quality index is fitted by the least squares method, and the weight coefficient of the correction amount is dynamically adjusted.

[0142] Based on the risk state probability distribution output by the backward risk distribution prediction function, 10,000 sets of random parameter combinations are generated in the process parameter space using the Markov chain Monte Carlo sampling algorithm, and the expected risk probability value of subsequent processing control nodes under each set of parameters is calculated in parallel. Based on the calculation results, the optimal process parameter range with the highest process stability and risk probability below the threshold is extracted by Pareto front analysis.

[0143] Based on the optimal process parameter range, the partial derivative of the loss function with respect to the process parameters of the pre-processing control node is calculated using the inverse gradient propagation algorithm. The quantization vector of the parameter adjustment direction and adjustment magnitude is obtained by combining the learning rate decay strategy. Based on the quantization vector, the adjustment timing of each node is verified by the consensus algorithm of the distributed process parameter coordinator, and the multi-node parameter synchronous writing operation is triggered by the hardware synchronization signal.

[0144] Based on the adjusted real-time processing evaluation index data, a data sequence of 50 consecutive sampling points is collected through a sliding window, and the absolute error integral between the data and the standard value is calculated. Based on the absolute error integral result, an adjustment amount iterative optimization instruction is generated through a PID control algorithm. If the error integral value exceeds the preset tolerance, the parameter optimization loop is re-triggered until the error integral value stabilizes within the tolerance range.

[0145] This application constructs a closed-loop contactless production risk management system encompassing monitoring, risk coupling, collaborative control, and feedback. Addressing risks such as contamination, delay, and positioning deviation in the transmission and processing stages of high-precision diode device production, it leverages multi-module collaboration and precise algorithm modeling to achieve precise quantification of production risks, collaborative risk management, early prediction of risk trends, and dynamic optimization of the production process. This significantly improves the precision, stability, and risk controllability of contactless production. Specifically, firstly, regarding the improvement in risk quantification accuracy, the monitoring module accurately acquires the first-state data of the associated transmission and the second-state data of the associated processing. The first risk unit collects transmission evaluation index parameters, combines them with an evaluation algorithm to generate index scores and estimated confidence levels, and then calculates the transmission evaluation error set. Subsequently, a gradient boosting tree is used to capture the nonlinear mapping relationship between the transmission evaluation error and the pass rate of the next processing control node, quantifying the contribution of each error to the pass rate. Finally, a Bayesian neural network is used to fuse the contribution and confidence levels to calculate the first processing... The failure risk probability assessment process avoids the one-sidedness of single-parameter evaluation through multi-indicator collaboration and leverages the nonlinear fitting ability of gradient boosting trees and the uncertainty handling ability of Bayesian neural networks to transform transmission risk from a qualitative description to a quantitative probability, significantly improving the accuracy of the assessment of the impact of transmission on processing risk. The second risk unit collects real-time operating status parameters, process parameters, and transmission contamination indicators of processing equipment. After preprocessing, it combines the historical processing pass rate change rate with the data and uses a random forest regression model to establish the correlation between equipment parameter characteristics and pass rate change rate, generating equipment performance scores and trends. Then, it calculates the processing pass risk probability and the first transmission trigger risk probability through an evaluation algorithm with a built-in Bayesian function. This process combines the real-time status of the equipment with the historical pass rate trend and leverages the conditional probability calculation capability of Bayesian functions to make the risk assessment of subsequent transmission by the processing control node more closely match the actual performance of the equipment and the processing contamination status, further improving the pertinence and accuracy of risk quantification. Secondly, in terms of risk-related collaborative management, the risk coupling unit constructs a first transmission risk matrix based on the first processing failure risk probability of all transmission stages according to timestamps, and constructs a second processing risk matrix based on the first transmission trigger risk probability of all processing control nodes according to process priority. Then, the two matrices are fused through a Hidden Markov Model to obtain the risk coupling matrix. Finally, a production coupling chain is constructed according to the production sequence priority using a loose coupling algorithm, and transmission evaluation indicators, processing-related parameters, and risk trigger information are mapped and embedded into the transmission connection and processing control nodes respectively. This process captures the temporal correlation characteristics of transmission and processing risks through the Hidden Markov Model, transforming isolated transmission and processing risks into a collaboratively related overall risk system. At the same time, the loose coupling algorithm ensures that the production coupling chain not only conforms to the actual production process sequence but also achieves accurate correspondence of risks at each node, avoiding the isolation of risk assessment and enabling the risk management of the entire production process to form a synergistic effect, effectively reducing the overall management loopholes caused by the separation of transmission and processing risks.Furthermore, regarding the improvement of risk trend prediction capabilities, a joint multi-dimensional observation sequence of a Hidden Markov Model is constructed based on the first and second risk matrices. Combined with the hidden state sequence labeled with historical production data, the initial state probability vector, state transition probability matrix, and observation probability density function of the model are iteratively trained using the Baum-Welch algorithm until the model converges. Subsequently, the state transition probability matrix is ​​extracted to construct a risk coupling matrix. Then, the Viterbi dynamic programming algorithm is used to decode the current optimal hidden state path and predict the risk evolution trend. At the same time, the backward risk distribution prediction function and the Markov chain Monte Carlo sampling algorithm are combined to fine-tune the model. This process, through training with historical data, enables the Hidden Markov Model to accurately represent the state transition rules of transmission-processing risks. The Viterbi algorithm can quickly locate the current optimal risk state, while the backward risk distribution prediction function and sampling algorithm further optimize the model's ability to predict subsequent risks, realizing the extension from real-time risk assessment to future risk trend prediction, identifying potential risk evolution directions in advance, and providing a basis for proactive risk control. Finally, regarding dynamic optimization and stability assurance of the production process, the feedback module's bidirectional monitoring unit monitors the status data of each node in real time along the production coupling chain, and predicts subsequent risk probabilities by combining the risk coupling matrix and the backward risk distribution prediction function, while simultaneously monitoring the matching results of risks and trigger information; the evaluation and early warning unit analyzes forward monitoring data, backward predicted risks, and trigger results through a causal reasoning model combined with standard process indicators to locate the associated abnormal risk transmission chain and abnormal characteristics; the feedback adjustment unit obtains historical solutions from the abnormal strategy library through multi-dimensional feature matching, generates the optimal process parameter correction amount for the current node through Bayesian optimization, and issues it for execution, while also monitoring the backward risk prediction results. By extracting the optimal process parameter range for subsequent nodes through Pareto front analysis and combining iterative parameter optimization with backward gradient propagation and PID control, this closed-loop process not only takes into account both real-time status and future risks through bidirectional monitoring to avoid lag in anomaly detection, but also uses causal reasoning to locate related anomalies rather than single-node problems, ensuring accurate anomaly tracing. Furthermore, through historical strategy fusion and real-time parameter iterative optimization, it achieves synergy between forward anomaly correction and backward parameter pre-adjustment, continuously controlling production process errors within the standard range. Ultimately, this ensures stable production accuracy of high-precision diode devices, reduces the problem of production accuracy decline caused by insufficient risk control, and promotes the efficient and stable operation of contactless production processes.

[0146] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A contactless transmission path coordinated control system, characterized by, The application relates to a semiconductor device non-contact production process risk coupling system, which comprises a monitoring module, a risk coupling module, a cooperative control module and a feedback module. The monitoring module is used for acquiring first state data and second state data in a semiconductor device non-contact production process according to a configured intelligent sensor network; the first state data is related to a non-contact transmission control process of the semiconductor device and at least includes non-contact transmission control parameters and transmission evaluation indexes; the second state data is related to a non-contact processing control process of the semiconductor device and at least includes processing technology control parameters, processing technology evaluation indexes, processing equipment running state parameters and processing pollution evaluation indexes; wherein the non-contact transmission control parameters in the first state data at least include transmission speed setting values, acceleration curve parameters and suspension height setting values; the transmission evaluation indexes at least include a measured value of a suspended particulate matter concentration, a difference value between a standard transmission time and an actual transmission time, and X-axis deviation amount, Y-axis deviation amount and rotation angle deviation amount of the semiconductor device at a target processing node coordinate system; the second state data at least includes temperature setting values, pressure control values and process duration of each processing node, a welding qualification rate and cleaning cleanliness score acquired through a visual detection system, vibration frequency and temperature deviation value of a processing equipment main shaft, and a particulate concentration monitoring value in a processing cavity. The risk coupling module is used for constructing a first transmission risk matrix according to the first state data, simultaneously constructing a second processing risk matrix according to the second state data, and obtaining a risk coupling matrix and a production coupling chain by combining a coupling algorithm and a risk transfer algorithm according to a semiconductor device non-contact production process priority. The cooperative control module is used for generating a coupling risk control instruction in response to a standard production process chain of the semiconductor device and coupling risk trigger information corresponding to each processing control node and adjacent transmission connection relationship of the production coupling chain; the coupling risk control instruction at least includes one transmission control instruction and one processing control instruction. The feedback module is used for acquiring transmission and processing state information of the production coupling chain in real time, combining a preset distributed evaluation algorithm and a standard production process chain, performing reverse abnormal evaluation positioning along the production coupling chain, and feeding back positioning abnormal information to the production coupling chain to adjust the risk coupling matrix and the production coupling chain in real time until the non-contact production process meets the standard production process chain in real time. Each processing control node in the production coupling chain corresponds to one non-contact processing technology node in the non-contact production process; each connection relationship in the production coupling chain corresponds to transmission process between two adjacent non-contact processing technology nodes; the standard production process chain includes standard processing technology evaluation indexes and standard transmission evaluation indexes, and corresponds to each processing control node and transmission stage in the production coupling chain, and is used for evaluating errors of real-time processing technology and transmission corresponding to each processing control node and transmission stage in the production coupling chain. The risk coupling module includes a first risk unit and a second risk unit.

2. The contactless transmission path coordinated control system of claim 1, wherein, ​ 3. The contactless transmission path coordinated control system of claim 2, wherein, ​ The transmission evaluation indexes at least include a transmission pollution index, a transmission delay index, and a transmission position accuracy index; and the workflow of the first risk unit includes: According to the transmission evaluation index parameters corresponding to each transmission stage and the evaluation algorithm, the transmission evaluation index scores and the estimation confidence of the corresponding transmission stage are obtained; Based on each transmission evaluation index score of the current transmission stage and the corresponding transmission evaluation index threshold, the transmission evaluation error set of the current transmission stage is obtained; Based on the transmission evaluation error set of the current transmission stage and the processing qualified probability of the corresponding processing technology of the next processing control node, the contribution degree of each transmission evaluation error to the processing qualified probability of the corresponding processing technology of the next processing control node is obtained through the gradient boosting tree; Based on the contribution degree of each transmission evaluation error to the processing qualified probability of the corresponding processing technology of the next processing control node and the estimation confidence, the first processing failure risk probability of the current transmission stage to the next processing control node is obtained through the Bayesian neural network; Based on the first processing failure risk probability corresponding to each transmission stage and the preset first risk threshold, the processing risk trigger information set is constructed; the processing risk trigger information corresponds to the processing control instruction.

4. The contactless transmission path coordinated control system of claim 3, wherein, The workflow of the second risk unit includes: The running state parameters, the processing technology control parameters, and the processing pollution evaluation index parameters of the processing equipment corresponding to each processing control node of the production coupling chain are collected in real time, and the collected data are time-stamped and standardized for pretreatment to obtain the pretreated processing equipment running state parameters, the processing technology index parameters, and the processing pollution evaluation index parameters; Based on the processing equipment running state parameters and the historical processing qualified rate change rate, the processing equipment performance score and the change trend corresponding to the current processing control node are obtained; The first evaluation algorithm with the Bayesian function is used to obtain the processing passing risk probability of the current processing control node based on the processing equipment performance score and the change trend corresponding to the current processing control node and the corresponding processing technology index parameters; The first evaluation algorithm with the Bayesian function is used to obtain the first transmission trigger risk probability corresponding to the current processing control node based on the processing passing risk probability of the current processing control node and the processing pollution evaluation index parameters; and the transmission risk trigger information set is constructed based on the first transmission trigger risk probability corresponding to each processing control node and the corresponding first transmission trigger risk probability threshold; the transmission risk trigger information corresponds to the transmission control instruction.

5. The contactless transmission path coordinated control system of claim 4, wherein, The risk coupling module further comprises a risk coupling unit; the risk coupling unit is configured to: construct a first transmission risk matrix according to the first processing failure risk probability of all transmission stages and in chronological order of transmission stage timestamps, construct a second processing risk matrix according to the first transmission trigger risk probability of all processing control nodes and in order of processing technology priorities, obtain a risk coupling matrix based on the combination of the first transmission risk matrix and the second processing risk matrix and a hidden Markov model, construct a production coupling chain based on all processing control nodes, the risk coupling matrix and a loose coupling algorithm and in order of priority of transmission and processing in a production process, map the transmission evaluation index to a transmission coupling connection in the production coupling chain, map the processing technology control parameter, the processing technology evaluation index and the processing pollution evaluation index to a corresponding processing control node, and embed the set of processing risk trigger information into an input end of each processing control node and embed the set of transmission risk trigger information into an output end of each processing control node.

6. The contactless transmission path coordinated control system of claim 5, wherein, The risk coupling matrix is obtained based on the combination of the first transmission risk matrix and the second processing risk matrix and a hidden Markov model, and comprises: based on the timestamp sequence of each transmission stage and the corresponding first processing failure risk probability value in the first transmission risk matrix and the process priority sequence of each processing control node and the corresponding first transmission trigger risk probability value in the second processing risk matrix, obtaining a joint multi-dimensional observation sequence of the hidden Markov model according to the priority order of each transmission and processing process in the production process; based on the annotated hidden state sequence and the joint observation sequence in the historical production data, extracting the corresponding relationship between the hidden state and the joint observation value, and using the Baum-Welch algorithm to perform maximum likelihood estimation iterative training on the initial state probability vector, the state transition probability matrix and the observation probability density function represented by the Gaussian mixture model of the hidden Markov model until the model likelihood function change rate is lower than a preset convergence threshold; based on the trained hidden Markov model, extracting all state transition probability values in the state transition probability matrix to construct a risk coupling matrix.

7. The contactless transmission path coordinated control system of claim 6, wherein, The risk coupling matrix is obtained based on the combination of the first transmission risk matrix and the second processing risk matrix and a hidden Markov model, and further comprises: based on the real-time collected joint observation sequence and the risk coupling matrix, using a Viterbi dynamic programming algorithm to decode the optimal hidden state path at the current time, and based on the transition probability distribution of the terminal state in the optimal hidden state path, predicting the risk state evolution trend of the production link, and based on the risk state evolution trend of the production link and a radial kernel function, fitting to obtain a backward risk distribution prediction function; The trained hidden Markov model is fine-tuned again by a Markov chain Monte Carlo sampling algorithm based on the backward risk distribution prediction function, the posterior distribution of the model parameters and the backward risk distribution prediction bias are updated, the likelihood of the posterior distribution and the sample data is maximized, and the backward risk distribution prediction bias is minimized, the model training is completed, and the fine-tuned risk coupling matrix and the backward risk distribution prediction function are obtained.

8. The contactless transmission path coordinated control system of claim 7, wherein, The obtaining process of the hidden state sequence and the joint observation sequence comprises: obtaining delay risk confidence and transmission positioning risk confidence by confidence analysis, taking the first transmission trigger risk probability of the previous processing control node corresponding to the current transmission stage as a condition, and combining the transmission delay index parameter and the transmission position accuracy index parameter collected in real time; obtaining transmission pollution confidence by the same delay risk analysis process, taking the processing pollution risk probability corresponding to the previous processing control node as a condition and combining the transmission pollution index parameter of the current transmission stage; obtaining the transmission evaluation index score and the processing risk level corresponding to the first processing failure risk probability by taking the delay risk confidence, the transmission positioning risk confidence and the transmission pollution confidence, combining the comprehensive fuzzy algorithm and the evaluation score level interval; obtaining the transmission risk level corresponding to the first transmission trigger risk probability of the processing control node corresponding to the transmission stage by combining the preset transmission risk level interval; obtaining the hidden state sequence and the joint observation sequence of the processing control node after the corresponding transmission stage based on the processing risk level and the transmission risk level between the adjacent transmission stages and the processing control nodes and combining the corresponding risk coupling matrix.

9. The contactless transmission path coordinated control system of claim 8, wherein, The feedback module comprises a bidirectional monitoring unit; The bidirectional monitoring unit is configured to, along the production coupling chain, take the current processing control node or transmission stage as a starting point, monitor the first state data and the second state data corresponding to each processing control node in real time, and predict the first processing failure risk probability or the first transmission trigger risk probability corresponding to the next transmission stage or processing control node in reverse based on the first state data and the second state data corresponding to the transmission stage or processing control node, the risk coupling matrix and the backward risk distribution prediction function, and monitor the first trigger result information between the first processing failure risk probability and the processing risk trigger information or the second trigger result information between the first transmission trigger risk probability and the transmission risk trigger information.

10. The contactless transmission path coordinated control system of claim 9, wherein, The feedback module further comprises an evaluation and warning unit and a feedback adjustment unit; the evaluation and warning unit is configured to perform forward abnormal evaluation and correlation abnormal analysis on the first state data and the second state data monitored in the forward direction, the first processing failure risk probability or the first transmission trigger risk probability predicted in the backward direction, the first trigger result information or the second trigger result information, and the preset causal reasoning model and the standard processing process evaluation index built in each processing control node, to obtain a correlation abnormal risk transmission chain and corresponding abnormal feature information. The feedback adjustment unit is configured to obtain an abnormality resolution strategy according to the associated abnormality risk transmission chain, the corresponding abnormality characteristic information, a preset abnormality strategy library, and a quick index constructed according to historical index information, to perform real-time forward abnormality processing parameter correction and backward processing parameter pre-adjustment on the associated abnormality risk transmission chain until the production process meets the standard production process chain in real time.

Images