Search

CN-121980353-A - Glass seal welding process optimization and defect prediction method based on deep learning

CN121980353ACN 121980353 ACN121980353 ACN 121980353ACN-121980353-A

Abstract

The invention relates to the technical field of glass seal welding process optimization and defect prediction, in particular to a glass seal welding process optimization and defect prediction method based on deep learning. The method comprises the steps of firstly collecting electric signal data of a multi-source sensor of a welding furnace in real time, and synchronizing and preprocessing the electric signal data to form a standard time sequence data sequence. And then inputting the data sequence into a pre-trained multi-task deep learning model, and synchronously realizing dynamic process optimization and early defect prediction by a shared feature extraction network and two parallel task branch networks. After the process is finished, the system correlates the actual quality result with the process data to form an incremental sample, and performs on-line fine adjustment on the model based on an intelligent trigger mechanism and an anti-forgetting algorithm, so that the system can adapt to the change of equipment and materials. The invention realizes the conversion from fixed parameter control to real-time closed loop intelligent optimization of the glass seal welding process, and effectively improves the product yield, process stability and production intelligent level.

Inventors

  • YAN TINGFU

Assignees

  • 青岛福润德微电子设备有限公司

Dates

Publication Date
20260505
Application Date
20260126

Claims (10)

  1. 1. The glass seal welding process optimization and defect prediction method based on deep learning is characterized by comprising the following steps of: s1, acquiring multi-source sensor electric signal data in real time when a glass-sealed welding furnace runs, wherein the multi-source sensor electric signal data at least comprise a temperature sensor signal, a process gas flowmeter signal and a push-pull boat motor driving signal, and performing time synchronization alignment and preprocessing on all acquired electric signal data to form a sensor data sequence with uniform time stamps; s2, inputting the sensor data sequence into a pre-trained multi-task deep learning model, wherein the multi-task deep learning model comprises a shared feature extraction network and two parallel task branch networks; S3, executing two tasks in parallel through the multi-task deep learning model, wherein a first task branch network dynamically outputs an optimization adjustment suggestion aiming at the current process parameter according to the sensor data sequence; S4, after a glass seal welding process of a batch is completed, acquiring an actual quality detection result of the batch of products, and correlating the actual quality detection result with the sensor data sequence and the risk probability and the optimization suggestion output by the model at the time to form an incremental training sample for carrying out online fine adjustment updating on the multi-task deep learning model; And S5, carrying out real-time visual presentation on the optimization adjustment suggestion and the defect risk probability through a man-machine interaction interface, and triggering an early warning signal when the risk probability exceeds a preset threshold.
  2. 2. The method according to claim 1, wherein in step S1, the "time synchronization alignment" is specifically: The method comprises the steps of marking uniform high-precision time marks for electric signal data streams from different physical sensors, unifying the data streams to data points with the same time interval through an interpolation algorithm for sequence length differences caused by different acquisition frequencies, and forming a multidimensional sensor data sequence with a strict time sequence corresponding relation.
  3. 3. The method according to claim 2, wherein in step S2, the shared feature extraction network is a one-dimensional convolutional neural network including an attention mechanism configured to automatically extract deep timing features related to process states and defect formation from the sensor data sequence, the first task branch network is a fully connected regression network for outputting continuous process parameter adjustment amounts, and the second task branch network is a classification network for outputting occurrence probabilities of different defect categories.
  4. 4. The method according to claim 1, wherein in step S4, the "online trimming update" is not performed immediately after each process is completed, but an update trigger mechanism is set: And when the deviation between the predicted defect risk probability of the model and the subsequent actual quality detection result continuously exceeds a dynamic threshold value or the difference between the characteristic distribution of the sensor data sequence of the new batch and the characteristic distribution of the model training data exceeds a preset range in the process of continuous N batches, starting to finely tune the model by utilizing the recently accumulated incremental training samples.
  5. 5. The method of claim 4, further comprising the step of verifying the actual quality test results prior to initiating an online trim update: If the difference between the input actual quality detection result and the prediction result made by the model based on the corresponding sensor data sequence is larger than a preset value and the result is seriously inconsistent with the data mode of the historical normal batch, the system marks the batch data as to-be-confirmed, pauses the batch data to be used for model updating, and prompts an operator to review through a human-computer interaction interface.
  6. 6. The method of claim 4 or 5, wherein the online trimming update employs an elastic weight consolidation algorithm to impose constraints on important, learned historical process knowledge related parameters in the model when updating model parameters with incremental training samples to prevent catastrophic forgetfulness of previously learned key process rules when adapting to new data.
  7. 7. The method according to claim 1, wherein in step S3, the "optimal adjustment suggestion" specifically includes one or more of an adjustment amount of a heating zone temperature set value, an adjustment amount of a process gas flow, and an adjustment amount of a push-pull boat movement speed, and the suggestion is presented in a natural language description and a parameter curve overlapping highlighting manner on a man-machine interaction interface.
  8. 8. The method of claim 7, wherein the human-machine interface further provides an "optimization simulation" function, and wherein upon receiving the optimization adjustment suggestion, an operator authorizes the system to simulate the operation result of the future process step based on the suggestion parameters, and wherein the multi-task deep learning model predicts the quality index and the defect risk change after performing the optimization suggestion based on the sensor data generated by the simulation, and compares the simulation prediction result with the original process prediction result.
  9. 9. The method of claim 1, further comprising the step of creating and maintaining a process-quality case library for storing the sensor data sequences, model outputs, actual quality inspection results and final determined process states associated in step S4, and labeling each case with material lot, equipment number, environmental conditions for supporting rapid retrieval and comparative analysis of historical cases.
  10. 10. The method of claim 9, wherein the process-quality case library supports interactive queries with the multi-task deep learning model, automatically retrieving historical cases with most similar sensor data sequence patterns in the case library when the model has low confidence in the predictions of the current process, and pushing complete data, processing, and final results of the historical cases to a human-machine interaction interface as reference information for aid decisions.

Description

Glass seal welding process optimization and defect prediction method based on deep learning Technical Field The invention relates to the technical field, in particular to a glass seal welding process optimization and defect prediction method based on deep learning. Background Glass seal welding is a key packaging technology in the fields of electronic components, photoelectric devices and special packaging, and is used for realizing airtight sealing between glass materials and metal or ceramic parts by high-temperature sintering, so that the reliability, service life and performance of products are directly affected. The conventional glass seal welding process is highly dependent on a combination of preset fixed parameters such as temperature profile, gas atmosphere, and transfer speed, which are usually determined based on experience of process personnel or preliminary process tests and remain unchanged for a long period of time in production. However, the actual production environment is complex and changeable, for example, the distribution drift of a thermal field is caused by the aging of a heating element of a welding furnace along with the use time, the slight physical and chemical characteristic difference exists between glass powder or metal pieces in different batches, and the purity of protective gas or the environmental temperature and humidity are fluctuated. These factors all cause the actual process state to deviate from the ideal setting, thereby causing defects such as oxidation, cracking, air leakage and the like of the product, and causing yield loss. At present, the industry generally adopts a quality control mode of 'post detection', namely defective products are screened through helium mass spectrum leakage detection, insulation test and other means after the process is finished. This approach is not only lag, but cannot be used to intervene in the defect formation process, and the production costs of all defective products are not recovered. Although a scheme of monitoring single parameters such as furnace temperature and the like through a sensor and alarming is available in the prior art, the method can only sense obviously overrun abnormality, can not identify tiny and gradual abnormality modes which cause defects from complex time sequence data of multi-parameter coupling, and can not dynamically give accurate suggestions for optimizing process parameters to compensate deviation. Therefore, an intelligent method capable of sensing the process state in real time, predicting the defect risk in advance, and adaptively adjusting the process parameters to realize closed-loop optimization is needed in the art, so as to improve the stability, yield and intelligent level of the glass seal welding process. Thus, the prior art is still to be further developed. Disclosure of Invention The invention aims to overcome the technical defects and provide a glass seal welding process optimization and defect prediction method based on deep learning so as to solve the problems in the prior art. In order to achieve the technical purpose, the invention provides a glass seal welding process optimization and defect prediction method based on deep learning, which comprises the following steps: s1, acquiring multi-source sensor electric signal data in real time when a glass-sealed welding furnace runs, wherein the multi-source sensor electric signal data at least comprise a temperature sensor signal, a process gas flowmeter signal and a push-pull boat motor driving signal, and performing time synchronization alignment and preprocessing on all acquired electric signal data to form a sensor data sequence with uniform time stamps; s2, inputting the sensor data sequence into a pre-trained multi-task deep learning model, wherein the multi-task deep learning model comprises a shared feature extraction network and two parallel task branch networks; S3, executing two tasks in parallel through the multi-task deep learning model, wherein a first task branch network dynamically outputs an optimization adjustment suggestion aiming at the current process parameter according to the sensor data sequence; S4, after a glass seal welding process of a batch is completed, acquiring an actual quality detection result of the batch of products, and correlating the actual quality detection result with the sensor data sequence and the risk probability and the optimization suggestion output by the model at the time to form an incremental training sample for carrying out online fine adjustment updating on the multi-task deep learning model; And S5, carrying out real-time visual presentation on the optimization adjustment suggestion and the defect risk probability through a man-machine interaction interface, and triggering an early warning signal when the risk probability exceeds a preset threshold. Specifically, in step S1, the "time synchronization alignment" specifically includes: The method comprises the steps of marking uniform h