Workpiece sealing detection method, model training method, electronic equipment and storage medium

Abstract

The invention discloses a workpiece sealing detection method, a model training method, electronic equipment and a storage medium. The method comprises the following steps: obtaining sealing parameters related to a to-be-detected workpiece in a sealing process; inputting the sealing parameters into the target sealing quality prediction model, and predicting the sealing qualification probability corresponding to the to-be-detected workpiece; and performing sealing detection on the to-be-detected workpiece based on the sealing qualification probability to obtain a sealing detection result corresponding to the to-be-detected workpiece.

Description

Technical Field

[0001] This application relates to the field of sealing detection technology, specifically to a workpiece sealing detection method, a model training method, electronic equipment, and a storage medium. Background Technology

[0002] In the manufacturing of key components such as power battery covers and casings, the sealing process, such as laser sealing welding, is a core quality indicator that determines the safety and service life of the product.

[0003] Currently, the industry generally adopts a quality control model that combines "sampling metallographic analysis during welding" with "full inspection using helium mass spectrometry leak detection equipment". However, this model has systematic shortcomings, as all workpieces are indiscriminately subjected to the high-precision but limited-throughput helium mass spectrometry leak detection process, resulting in low efficiency in sealing inspection of the workpieces. Summary of the Invention

[0004] This invention provides a workpiece sealing inspection method, a model training method, an electronic device, and a storage medium, which can improve the efficiency of sealing inspection.

[0005] Firstly, a method for detecting the sealing of a workpiece is provided, comprising the following steps: Obtain the sealing parameters involved in the sealing process of the workpiece under test; Input the sealing parameters into the target sealing quality prediction model to predict the sealing qualification probability of the workpiece under test. The sealing test is performed on the workpiece to be tested based on the probability of successful sealing, and the sealing test result corresponding to the workpiece to be tested is obtained.

[0006] In one exemplary embodiment, a seal test is performed on the workpiece to be tested based on the seal pass probability to obtain the seal test result corresponding to the workpiece to be tested, including: Based on the probability of successful sealing, determine the type of sealing test corresponding to the workpiece to be tested; According to the type of sealing test, the sealing test is performed on the workpiece to be tested, and the corresponding sealing test result of the workpiece to be tested is obtained.

[0007] In this embodiment, by determining the sealing test type based on the sealing qualification probability of the workpiece to be tested, and performing the corresponding sealing test processing according to the sealing test type, it is possible to achieve graded processing of the workpiece to be tested, thereby achieving accurate sealing test of the workpiece and improving the sealing test efficiency and accuracy of the workpiece.

[0008] In one exemplary embodiment, the seal inspection type corresponding to the workpiece under test is determined based on the seal pass probability, including: When the probability of a seal passing inspection exceeds the preset passing range, the seal inspection type corresponding to the workpiece to be tested is determined to be a non-standard seal inspection. When the probability of a successful seal is within the predicted success range, the seal inspection type for the workpiece to be tested is determined to be standard seal inspection. If the sealing pass probability does not reach the preset pass range, the sealing test type for the workpiece to be tested is determined to be terminated.

[0009] In this embodiment, by determining the sealing test type based on the sealing qualification probability of the workpiece to be tested, and performing the corresponding sealing test processing according to the sealing test type, it is possible to achieve graded processing of the workpiece to be tested, thereby achieving accurate sealing test of the workpiece and improving the sealing test efficiency and accuracy of the workpiece.

[0010] In one exemplary embodiment, the sealing process is a welding process, and the sealing test result is obtained by performing an airtightness test on the workpiece under test in the airtightness test process.

[0011] In an exemplary embodiment, the target seal quality prediction model is obtained by using the second predicted seal pass probability output by the trained first seal quality prediction model as the second training label, and training the model based on the input first training seal features and the second training label. The first seal quality prediction model is obtained by training the model based on the input first training seal features and the first training seal detection features.

[0012] In this embodiment, a first sealing quality prediction model is trained using sealing features and sealing detection features, enabling the first sealing quality prediction model to output an accurate sealing pass probability based on these features. The predicted sealing pass probability output by the first sealing quality prediction model is then used to train a target sealing quality prediction model, allowing the target sealing quality prediction model to indirectly learn the relationship between sealing features, sealing detection features, and sealing pass probability. Furthermore, when the target sealing quality prediction model is applied online, it can accurately predict the sealing pass probability of the workpiece even with only sealing features as input, ensuring the accuracy of the sealing pass probability prediction.

[0013] In one exemplary embodiment, the workpiece seal detection method further includes: Acquire the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label; Based on the first training sealing features, the first training sealing detection features, and the first training label, the first sealing quality prediction model to be trained is iteratively optimized to obtain the first sealing quality prediction model after training. The second predicted seal pass probability output by the first trained seal quality prediction model based on the first trained seal features and the first trained seal detection features is used as the second training label. Based on the first training sealing features, the first training label, and the second training label, the second sealing quality prediction model to be trained is iteratively optimized to obtain the target sealing quality prediction model after training.

[0014] In this embodiment, a first sealing quality prediction model is trained using sealing features and sealing detection features, enabling the first sealing quality prediction model to output an accurate sealing pass probability based on these features. The predicted sealing pass probability output by the first sealing quality prediction model is then used to train a target sealing quality prediction model, allowing the target sealing quality prediction model to indirectly learn the relationship between sealing features, sealing detection features, and sealing pass probability. Furthermore, when the target sealing quality prediction model is applied online, it can accurately predict the sealing pass probability of the workpiece even with only sealing features as input, ensuring the accuracy of the sealing pass probability prediction.

[0015] In an exemplary embodiment, the first training label includes a seal qualification classification label and an actual seal leakage rate; based on the first training seal features, the first training seal detection features, and the first training label, the first seal quality prediction model to be trained is iteratively optimized to obtain the trained first seal quality prediction model, including: The first training seal features and the first training seal detection features are input into the first main network and the second auxiliary network of the first seal quality prediction model to be trained. The first predicted seal pass probability is obtained through the first main network, and the predicted seal leakage rate is obtained through the second auxiliary network; Based on the seal qualification classification label, the first predicted seal qualification probability, the actual seal leakage rate, and the predicted seal leakage rate, the first target model loss is determined. Based on the loss of the first target model, the first sealing quality prediction model to be trained is iteratively optimized to obtain the first sealing quality prediction model after training.

[0016] In this embodiment, by using a first target model loss that integrates the first model loss and the second model loss to train the first sealing quality prediction model, the first main network and the second auxiliary network can be jointly optimized, thereby improving the training accuracy of the first sealing quality prediction model.

[0017] In an exemplary embodiment, based on the first training sealing features, the first training label, and the second training label, the second sealing quality prediction model to be trained is iteratively optimized to obtain the trained target sealing quality prediction model, including: The first training seal features are input into the second seal quality prediction model to be trained to obtain the third seal qualification probability; The loss of the second target model is determined based on the sealing qualification classification label, the second training label, and the third sealing qualification probability; Based on the loss of the second target model, the second sealing quality prediction model to be trained is iteratively optimized to obtain the trained target sealing quality prediction model.

[0018] In this embodiment, the second sealing quality prediction model is trained by using the first training sealing features, the first training label, and the second training label. This allows the second sealing quality prediction model to indirectly learn the sealing detection features, such as the patterns contained in the helium detection curve, by imitating the second training label, thereby ensuring the training accuracy and prediction accuracy of the target sealing quality prediction model.

[0019] In an exemplary embodiment, before acquiring the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label, the method further includes: Based on the sealing end time and sealing detection start time corresponding to multiple initial sample workpieces, the actual process time difference corresponding to each initial sample workpiece is determined. Based on the actual process time difference and the preset standard process time difference for each initial sample workpiece, identify the abnormal time sample workpiece among multiple initial sample workpieces; Abnormal time sample workpieces are screened out from multiple initial sample workpieces to obtain sample workpieces.

[0020] In this embodiment, by determining and filtering out abnormal time sample workpieces based on the actual process time difference and the preset standard process time difference, the accuracy of the training data corresponding to the sample workpieces can be guaranteed, thereby ensuring the training accuracy of the target sealing quality prediction model.

[0021] In an exemplary embodiment, before acquiring the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label, the method further includes: Obtain the reference sealing end time corresponding to the pre-selected workpiece. The pre-selected workpiece is the workpiece whose workpiece identifier is successfully associated with the sealing data but fails to be associated with the sealing detection data. The sealing data is used to extract sealing features, and the sealing detection data is used to extract sealing detection features. The target time search range is determined based on the difference between the reference sealing end time and the preset standard process time. Acquire multiple sealing detection data points whose start time falls within the target time search range, and determine the target sealing detection data point from among the multiple sealing detection data points; Pre-selected workpieces that are successfully associated with the target sealing test data are used as sample workpieces.

[0022] In this embodiment, by supplementing the missing sealing test data for pre-selected workpieces and using them as sample workpieces, the amount of training data corresponding to the sample workpieces can be increased, thereby ensuring the training accuracy of the target sealing quality prediction model.

[0023] Secondly, a method for training a sealing quality prediction model is provided, including the following steps: Acquire the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label; Based on the first training sealing features, the first training sealing detection features, and the first training label, the first sealing quality prediction model to be trained is iteratively optimized to obtain the first sealing quality prediction model after training. The second predicted seal pass probability output by the first trained seal quality prediction model based on the first trained seal features and the first trained seal detection features is used as the second training label. Based on the first training sealing features, the first training label, and the second training label, the second sealing quality prediction model to be trained is iteratively optimized to obtain the target sealing quality prediction model after training.

[0024] Thirdly, a sealing detection device is also provided, comprising: The data acquisition module is used to acquire the sealing parameters involved in the sealing process of the workpiece under test; The sealing quality prediction module is used to input sealing parameters into the target sealing quality prediction model and predict the sealing qualification probability of the workpiece under test. The sealing detection module is used to perform sealing detection on the workpiece under test based on the sealing pass probability, and obtain the sealing detection result corresponding to the workpiece under test.

[0025] Fourthly, a training device for a sealing quality prediction model is also provided, comprising: The data acquisition module is used to acquire the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label; The first model training module is used to iteratively optimize the first seal quality prediction model to be trained based on the first training seal features, the first training seal detection features and the first training label, so as to obtain the trained first seal quality prediction model. The label determination module is used to take the second predicted seal pass probability output by the first trained seal quality prediction model for the first trained seal features and the first trained seal detection features as the second training label. The second model training module is used to iteratively optimize the second sealing quality prediction model to be trained based on the first training sealing features, the first training label, and the second training label, so as to obtain the target sealing quality prediction model after training.

[0026] Fifthly, an electronic device is also provided, including a memory and a processor, wherein computer instructions are stored in the memory, and when the computer instructions are executed by the processor, the method of any one of the above embodiments is implemented.

[0027] In a sixth aspect, a computer-readable storage medium is also provided, including computer instructions that, when executed on a device, cause the device to perform the method as described in any of the foregoing aspects and any of the embodiments thereof.

[0028] In a seventh aspect, embodiments of this application provide a computer program product, including a computer program or instructions, which are executed by a processor to implement the steps of any of the methods described above.

[0029] Beneficial effects: The technical solution of this application provides a sealing detection method. By inputting the sealing parameters of the workpiece to be tested into the target sealing quality prediction model, the sealing qualification probability of the workpiece to be tested is predicted, realizing the pre-detection of the sealing quality of the workpiece to be tested. Based on the sealing qualification probability obtained from the pre-detection, the corresponding sealing detection of the workpiece to be tested is performed, realizing the accurate sealing detection of the workpiece to be tested, thereby improving the sealing detection efficiency of the workpiece.

[0030] This application also provides a method for training a sealing quality prediction model. By using first training sealing features of a sample workpiece in the sealing process, first training sealing detection features in the sealing detection process, and first training labels, a first sealing quality detection model to be trained is obtained, resulting in a trained first sealing quality prediction model. This first sealing quality prediction model can accurately predict the sealing pass probability of the sample workpiece, providing an accurate and effective second training label for a second sealing quality prediction model to be trained. Then, using the first training sealing features, the first training label, and the second training label, the second sealing quality prediction model to be trained is trained, resulting in a trained target sealing quality prediction model, ensuring the training accuracy of the target sealing quality prediction model. Furthermore, by inputting the sealing parameters of the workpiece to be tested into the trained target sealing quality prediction model, without needing to additionally input the sealing detection features of the workpiece, the sealing pass probability of the workpiece can be accurately predicted, achieving both high prediction efficiency and accuracy in predicting the sealing pass probability of the workpiece. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic flowchart of a workpiece sealing detection method provided in some embodiments of this application; Figure 2 This is a flowchart of the online prediction and grading decision logic for workpiece sealing quality provided in some embodiments of this application; Figure 3 This is a schematic diagram of model training for working condition stratification provided in some embodiments of this application; Figure 4 This is a schematic diagram illustrating the principle of helium detection curve feature extraction and equivalent leak rate conversion provided in some embodiments of this application; Figure 5 This is a schematic diagram of the spatiotemporal alignment model in the data association and verification process provided in some embodiments of this application; Figure 6 This is a flowchart illustrating the training method for a sealing quality prediction model provided in some embodiments of this application; Figure 7 This is a schematic diagram of a workpiece sealing detection device provided in some embodiments of this application; Figure 8 This is a schematic diagram of a sealing quality prediction model training device provided in some embodiments of this application; Figure 9 These are internal structural diagrams of electronic devices provided in some embodiments of this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0035] This application provides a sealing detection method, a model training method, an electronic device, and a storage medium. The electronic device can be a server or a terminal, etc. In an exemplary embodiment, the terminal acquires the sealing parameters involved in the sealing process of the workpiece under test; the terminal inputs the sealing parameters into a target sealing quality prediction model to predict the sealing pass probability corresponding to the workpiece under test; the terminal performs sealing detection on the workpiece under test based on the sealing pass probability, and sends the sealing detection result corresponding to the workpiece under test to the server or other terminal for further processing. In another exemplary example, the terminal acquires the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label; based on the first training sealing features, the first training sealing detection features, and the first training label, the terminal iteratively optimizes the first sealing quality prediction model to be trained, obtaining the trained first sealing quality prediction model; the terminal uses the second predicted sealing pass probability output by the trained first sealing quality prediction model for the first training sealing features and the first training sealing detection features as the second training label; based on the first training sealing features, the first training label, and the second training label, the terminal iteratively optimizes the second sealing quality prediction model to be trained, obtaining the trained target sealing quality prediction model. The terminal can store the target sealing quality prediction model locally or send it to a server for storage. The terminal may include, but is not limited to, computers, laptops, etc. The server may be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, big data, and artificial intelligence platforms. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0036] On the one hand, this embodiment provides a method for detecting workpiece seals, such as... Figure 1 As shown, it includes the following steps: S101, Obtain the sealing parameters involved in the sealing process of the workpiece under test; S102, Input the sealing parameters into the target sealing quality prediction model to predict the sealing qualification probability of the workpiece to be tested; S103, based on the sealing qualification probability, perform sealing detection on the workpiece to be tested, and obtain the sealing detection result corresponding to the workpiece to be tested.

[0037] The workpiece to be tested refers to a workpiece that has undergone a sealing process and is about to be inspected for its seal. Sealing parameters refer to the operating parameters of the sealing equipment during the sealing process on the workpiece to be tested. The target seal quality prediction model is a pre-trained prediction model used to predict the seal quality of the workpiece to be tested based on the sealing parameters.

[0038] The probability of a successful seal refers to the probability that the seal quality of the workpiece under test, predicted by the target seal quality prediction model, meets the quality standard. The seal quality can be determined based on the seal index of the workpiece under test, such as the leakage rate. The leakage rate represents the gas leakage flow rate caused by tiny gaps in the workpiece under test during the seal inspection process. The quality standard can be a preset leakage rate threshold. Therefore, the probability of a successful seal can be the probability that the predicted leakage rate of the workpiece under test is less than the preset leakage rate threshold. The seal inspection result refers to the inspection result obtained by actually performing a seal inspection on the workpiece under test.

[0039] For example, in response to the sealing quality prediction command of the workpiece to be tested, the terminal obtains the sealing parameters involved in the sealing process of the workpiece to be tested. The sealing process is, for example, a laser welding process, and the sealing parameters can be parameters such as laser power, welding speed, and defocusing amount.

[0040] Then, the trained target seal quality prediction model is invoked, and the sealing parameters are input into the target seal quality prediction model to predict the seal pass probability of the workpiece under test. Specifically, the terminal can invoke the corresponding target seal quality prediction model based on the working conditions of the workpiece under test, such as the material type and structural type of the workpiece, to obtain the seal pass probability of the workpiece under the corresponding working conditions.

[0041] After obtaining the sealing pass probability for the workpiece under test, the terminal determines the workpiece quality type based on the sealing pass probability, such as qualified, risky, or unqualified. Risky workpieces are those with a risk of failing the sealing process. Then, according to the workpiece quality type, the terminal determines the corresponding sealing inspection type, such as standard sealing inspection, non-standard sealing inspection, or terminated sealing inspection. Following the sealing inspection procedure corresponding to this type, the terminal performs the sealing inspection on the workpiece to obtain the sealing inspection result.

[0042] In this embodiment, by inputting the sealing parameters of the workpiece to be tested into the target sealing quality prediction model, the sealing qualification probability of the workpiece to be tested is predicted, thus realizing the pre-detection of the sealing quality of the workpiece to be tested. Based on the sealing qualification probability obtained from the pre-detection, the corresponding sealing detection of the workpiece to be tested is performed, thus realizing the accurate sealing detection of the workpiece to be tested and improving the sealing detection efficiency of the workpiece.

[0043] In one exemplary embodiment, a seal test is performed on the workpiece to be tested based on the seal pass probability to obtain the seal test result corresponding to the workpiece to be tested, including: Based on the probability of successful sealing, determine the type of sealing test corresponding to the workpiece to be tested; According to the type of sealing test, the sealing test is performed on the workpiece to be tested, and the corresponding sealing test result of the workpiece to be tested is obtained.

[0044] For example, the terminal obtains the sealing pass probability corresponding to the workpiece to be tested, compares the sealing pass probability with a preset pass probability threshold, and determines the sealing detection type of the workpiece quality type corresponding to the workpiece to be tested based on the comparison result.

[0045] Specifically, the preset pass probability threshold includes a lower pass probability limit. When the pass probability is greater than or equal to the lower pass probability limit, it indicates that the pass probability of the workpiece under test has reached the minimum requirement represented by the lower pass probability limit, and the workpiece can proceed to the next sealing inspection process, i.e., perform sealing inspection on the workpiece under test. The sealing inspection type for the workpiece under test when the pass probability is greater than or equal to the lower pass probability limit can be either standard sealing inspection or non-standard sealing inspection. The terminal performs sealing inspection on the workpiece under test according to the sealing inspection process corresponding to the sealing inspection type, and obtains the corresponding sealing inspection result for the workpiece under test.

[0046] When the sealing pass probability is less than the lower limit of the pass probability, it means that the sealing pass probability of the workpiece under test has not reached the minimum requirement represented by the lower limit of the pass probability, indicating that the sealing quality of the workpiece under test is abnormal. The workpiece under test is then intercepted. The sealing test type when the sealing pass probability of the workpiece under test is less than the lower limit of the pass probability can be to terminate the sealing test, and then the workpiece under test is intercepted and processed in other ways, such as scrapping.

[0047] In this embodiment, by determining the sealing test type based on the sealing qualification probability of the workpiece to be tested, and performing the corresponding sealing test processing according to the sealing test type, it is possible to achieve graded processing of the workpiece to be tested, thereby achieving accurate sealing test of the workpiece and improving the sealing test efficiency and accuracy of the workpiece.

[0048] In one exemplary embodiment, the seal inspection type corresponding to the workpiece under test is determined based on the seal pass probability, including: When the probability of a seal passing inspection exceeds the preset passing range, the seal inspection type corresponding to the workpiece to be tested is determined to be a non-standard seal inspection. When the probability of a successful seal is within the predicted success range, the seal inspection type for the workpiece to be tested is determined to be standard seal inspection. If the sealing pass probability does not reach the preset pass range, the sealing test type for the workpiece to be tested is determined to be terminated.

[0049] Non-standard seal testing refers to a simplified seal testing procedure that deviates from the standard seal testing process. Standard seal testing refers to a seal testing procedure that must be strictly followed. Termination seal testing refers to a seal testing type that does not involve performing the seal testing.

[0050] For example, the preset pass probability threshold also includes a pass standard probability threshold, which is a judgment value used to identify the quality of the workpiece to be tested as a qualified product. If the pass standard probability threshold is greater than the lower limit of the pass probability threshold, the preset pass range can be obtained based on the pass standard probability threshold and the lower limit of the pass probability threshold.

[0051] The terminal compares the sealing pass probability of the workpiece under test with the preset pass range. When the sealing pass probability is within the preset pass range, it means that the sealing quality of the workpiece under test meets the minimum requirements but does not meet the standard requirements represented by the pass standard probability threshold. In this case, the workpiece quality type corresponding to the workpiece under test is a risky product, and the sealing test type corresponding to the risky product is the standard sealing test.

[0052] When the sealing pass probability exceeds the preset pass range, that is, when the sealing pass probability is greater than the pass standard probability threshold, it means that the sealing quality of the workpiece under test meets the standard requirements represented by the pass standard probability threshold. Then, the workpiece quality type corresponding to the workpiece under test is qualified product, and the sealing test type corresponding to the qualified product is non-standard sealing test.

[0053] When the sealing pass probability does not reach the preset pass range, i.e., the sealing pass probability is less than the lower limit of the pass probability threshold, it indicates that the sealing quality of the workpiece under test does not meet the standard requirements represented by the pass standard probability threshold. Therefore, the workpiece quality type corresponding to the workpiece under test is a non-conforming product, and the sealing test type corresponding to the non-conforming product is terminated. Furthermore, the terminal can calculate the non-conforming percentage of the workpieces under test. When the non-conforming percentage exceeds a preset percentage threshold, an early warning message is generated.

[0054] In this embodiment, by determining the sealing test type based on the sealing qualification probability of the workpiece to be tested, and performing the corresponding sealing test processing according to the sealing test type, it is possible to achieve graded processing of the workpiece to be tested, thereby achieving accurate sealing test of the workpiece and improving the sealing test efficiency and accuracy of the workpiece.

[0055] In one exemplary embodiment, the sealing process is a welding process, and the sealing test result is obtained by performing an airtightness test on the workpiece under test in the airtightness test process.

[0056] For example, the sealing process is a welding process. The sealing inspection process can be an airtightness inspection, and the sealing inspection result can be obtained by performing an airtightness inspection on the workpiece under test, and the sealing pass probability can be an airtightness pass probability. Here, the airtightness inspection is, for example, injecting helium gas into the workpiece under test to perform an airtightness inspection, which can be simply referred to as helium inspection.

[0057] For example, such as Figure 2The diagram shows a flowchart of an online prediction and grading decision-making logic for workpiece sealing quality. After welding, the sealing parameters of the new workpiece are extracted as welding data. This welding data is then input into the prediction model, i.e., the target sealing quality prediction model, to obtain the airtightness qualification probability P of the new workpiece. p For example, if a type A cover plate (aluminum alloy - 1.2mm) is welded, the welding data of the type A cover plate, including laser power P, welding speed v, and defocusing amount Δf, is collected. The welding data is then input into the target sealing quality prediction model, and the output is the probability of the type A cover plate passing the airtightness test. .

[0058] Then, based on the airtightness compliance probability, the workpieces to be tested are classified and processed according to a three-level sorting decision, as follows: Set the pass / fail probability threshold T high and the lower limit of the pass probability threshold T low The preset qualified range is [T] high T low ].

[0059] If T low ≤P pass ≤T high The workpiece was identified as a high-risk item, and the sealing inspection type was standard sealing inspection, i.e., standard helium inspection procedure; If P pass <T low The workpiece was determined to be defective, and the sealing test type was terminated. It was then sorted to a special processing channel or a scrap channel. If P pass >T high The workpiece was determined to be a qualified product, and the sealing test type was non-standard sealing test, so it entered the rapid helium test process.

[0060] The rapid helium detection process includes, for example: pre-sampling and background checking: retained; bombardment / pressurization: retained; stabilization / steady-state time t. stab Shorten the timeframe, e.g., 6-10 seconds; measurement window t meas Shorten the time window, such as to 6-10 seconds; keep the threshold at the same standard (without relaxing it), only optimize the time window to speed up the process; random inspection and verification: go through the "standard process" once for every N items.

[0061] Standard helium testing procedures include: full-process, long steady-state operation (e.g., 15-30s); long measurement window (e.g., 15-30s); strict implementation of background helium compensation and zero-point drift correction; and confirmation of risky / statistical testing threshold benchmarks.

[0062] Special processing channels include: visual / CCD weld defect verification (connectivity, voids, lack of fusion); partial re-welding / re-welding (followed by standard helium testing); extended testing: extended tstab / t meas 1. Detection of out-of-threshold locations (sniffer gun / tracer); 2. Perform offline "water inspection / bubble method" or scrap determination when necessary; 3. Simultaneously trigger equipment self-inspection (such as lens contamination, spot energy drift) and process backtracking.

[0063] In one exemplary embodiment, the preset acceptable range can be dynamically adjusted according to actual conditions. For example, during helium detection, the noise impact increases after the background helium concentration changes. To prevent missed detections, dynamic adjustments are made, for example, T... low =0.15, T high =0.95, when the background helium concentration rises to the preset alarm value, change to T. low =0.2, T high ==0.9, perform helium testing on hazardous and qualified products approaching the boundary. This can also be adjusted based on quality objectives; for example, increasing quality standards (such as a preset leak rate threshold) would increase T. low and reduce T high To prevent missed detections; conversely, to reduce T low And improve T high This can improve the production efficiency of helium detection.

[0064] In an exemplary embodiment, the target sealing quality prediction model also outputs a confidence interval corresponding to the sealing pass probability. The terminal jointly determines the workpiece quality type corresponding to the workpiece to be tested based on the sealing pass probability and the confidence interval. For example, if the sealing pass probability is 0.85, the confidence interval is 0.4-0.9, and the pass standard probability threshold is 0.8, although the sealing pass probability of 0.85 is greater than the pass standard probability threshold of 0.8, the confidence interval corresponding to the sealing pass probability of 0.85 is 0.4-0.9, indicating that the confidence level of the sealing pass probability of 0.85 is low, and the workpiece quality type corresponding to the workpiece to be tested is determined to be a risky product.

[0065] In this embodiment, real-time welding data is input into the target sealing quality prediction model during online application, outputting the airtightness qualification probability and its confidence interval, and then combining it with an adaptive threshold for risk assessment. Specifically, for workpieces that have completed welding but have not undergone helium testing, the airtightness qualification probability is predicted based on their welding data, and a three-level sorting decision is made according to a preset qualification interval. This achieves optimal allocation of helium testing resources, intercepting non-conforming products in advance and allowing qualified products to pass quickly, significantly improving sealing inspection efficiency and production efficiency while ensuring quality. Furthermore, by monitoring the proportion of non-conforming products, a real-time monitoring radar is provided for the production line equipment status and incoming material quality, enabling early detection and early warning of quality risks.

[0066] In an exemplary embodiment, the target seal quality prediction model is obtained by using the second predicted seal pass probability output by the trained first seal quality prediction model as the second training label, and training the model based on the input first training seal features and the second training label. The first seal quality prediction model is obtained by training the model based on the input first training seal features and the first training seal detection features.

[0067] The first sealing quality prediction model is obtained by training the model using sealing features and sealing detection features. The target sealing quality prediction model is obtained by training the model using sealing features. The second predicted sealing pass probability is the sealing pass probability predicted by the trained first sealing quality prediction model based on the input sealing features and sealing detection features.

[0068] For example, for a workpiece that has completed the sealing process but has not undergone sealing inspection, in order to improve the accuracy of predicting the sealing qualification probability of the workpiece, the first training sealing feature and the first training sealing inspection feature can be used in advance to train the first sealing quality prediction model to be trained, so as to obtain the trained first sealing quality prediction model.

[0069] The first, trained seal quality prediction model is used as the offline prediction model. Based on the first trained seal features and the first trained seal detection features, this offline prediction model outputs a second predicted seal pass probability, which is then used as the second training label. A target seal quality prediction model is then trained based on the first trained seal features and the second training label, and this trained target seal quality prediction model is used as the online prediction model.

[0070] For a workpiece that has completed the sealing process but has not undergone sealing inspection, the sealing characteristics corresponding to the sealing parameters of the workpiece are input into the online prediction model, and the sealing pass probability of the workpiece is output.

[0071] In this embodiment, a first sealing quality prediction model is trained using sealing features and sealing detection features, enabling the first sealing quality prediction model to output an accurate sealing pass probability based on these features. The predicted sealing pass probability output by the first sealing quality prediction model is then used to train a target sealing quality prediction model, allowing the target sealing quality prediction model to indirectly learn the relationship between sealing features, sealing detection features, and sealing pass probability. Furthermore, when the target sealing quality prediction model is applied online, it can accurately predict the sealing pass probability of the workpiece even with only sealing features as input, ensuring the accuracy of the sealing pass probability prediction.

[0072] In an exemplary embodiment, to address the problem of insufficient model generalization ability caused by the variability of product models and production conditions during mass production, the first sealing quality prediction model and the target sealing quality prediction model can be trained hierarchically according to the working conditions.

[0073] like Figure 3 The diagram illustrates a model training approach based on stratified operating conditions. In the historical production data pool, historical samples are stratified according to key operating condition variables such as material type and plate thickness, resulting in subsets A, B, ... of different operating conditions. This makes the predictive model's learning more targeted. Then, within each operating condition layer, corresponding predictive models A, B, ... are constructed to be trained. These models are trained using the sealing features corresponding to welding data and the sealing detection features corresponding to helium testing parameters as input, outputting the probability of airtightness compliance.

[0074] In this embodiment, the working condition hierarchical modeling effectively addresses changes in product and production conditions, reduces the risk of model performance degradation, and also improves the model's generalization ability and robustness.

[0075] In one exemplary embodiment, the workpiece seal detection method further includes: Acquire the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label; Based on the first training sealing features, the first training sealing detection features, and the first training label, the first sealing quality prediction model to be trained is iteratively optimized to obtain the first sealing quality prediction model after training. The second predicted seal pass probability output by the first trained seal quality prediction model based on the first trained seal features and the first trained seal detection features is used as the second training label. Based on the first training sealing features, the first training label, and the second training label, the second sealing quality prediction model to be trained is iteratively optimized to obtain the target sealing quality prediction model after training.

[0076] For example, during the training process of the first sealing quality prediction model to be trained, the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label are acquired. Specifically, based on the unique identifier of the sample workpiece, the first training sealing features can be acquired simultaneously, such as sealing features of welding process parameters (e.g., laser power P, welding speed v, defocusing amount Δf, etc.), and the first training sealing detection features can be acquired, such as acquiring the entire helium mass spectrometry leak detection process data: evacuation / leakage curve, ambient temperature, effective volume, background helium concentration, etc. The first training label includes the actual sealing detection results, such as helium gas detection result data.

[0077] The terminal uses the first training sealing features, the first training sealing detection features, and the first training label to iteratively optimize the first sealing quality prediction model to be trained, and obtain the first sealing quality prediction model after training.

[0078] Then, the first training seal features and the second seal training features are input again into the first seal quality prediction model after training to obtain the second predicted seal qualification probability, and the second predicted seal qualification probability is used as the second training label.

[0079] Using the first training seal features, the first training label, and the second training label, the second seal quality prediction model to be trained is iteratively optimized to obtain the target seal quality prediction model after training.

[0080] In an exemplary example, second training sealing features involved in the sealing process and second training sealing detection features involved in the sealing detection process of other sample workpieces can be obtained, along with corresponding training labels for other samples. The terminal inputs the second training sealing features and second training sealing detection features into the trained first sealing quality prediction model to obtain the sealing pass probability of other samples, and uses the sealing pass probability of other samples as the second training label. Using the second training sealing features, the training labels of other samples, and the second training label, the second sealing quality prediction model to be trained is iteratively optimized to obtain the trained target sealing quality prediction model.

[0081] In this embodiment, a first sealing quality prediction model is trained using sealing features and sealing detection features, enabling the first sealing quality prediction model to output an accurate sealing pass probability based on these features. The predicted sealing pass probability output by the first sealing quality prediction model is then used to train a target sealing quality prediction model, allowing the target sealing quality prediction model to indirectly learn the relationship between sealing features, sealing detection features, and sealing pass probability. Furthermore, when the target sealing quality prediction model is applied online, it can accurately predict the sealing pass probability of the workpiece even with only sealing features as input, ensuring the accuracy of the sealing pass probability prediction.

[0082] In an exemplary embodiment, the first training label includes a seal qualification classification label and an actual seal leakage rate; based on the first training seal features, the first training seal detection features, and the first training label, the first seal quality prediction model to be trained is iteratively optimized to obtain the trained first seal quality prediction model, including: The first training seal features and the first training seal detection features are input into the first main network and the second auxiliary network of the first seal quality prediction model to be trained. Obtain the first predicted seal qualification probability through the first main network, and obtain the predicted seal leakage rate through the second auxiliary network; Determine the first target model loss based on the seal qualification classification label, the first predicted seal qualification probability, the actual seal leakage rate, and the predicted seal leakage rate; Based on the first target model loss, iteratively optimize the first seal quality prediction model to be trained to obtain the trained first seal quality prediction model.

[0083] Exemplarily, the first seal quality prediction model is a multi-task learning model. The first main network is a sub-network in the first seal quality prediction model that performs the main task. The main task is the core task for evaluating the model effect, such as the classification main task for whether the seal quality is qualified. The second auxiliary network is a sub-network in the first seal quality prediction model that performs the auxiliary task. The auxiliary task is an auxiliary task for sharing knowledge and providing additional learning signals to improve the performance of the main task, such as a regression auxiliary task. The seal qualification classification label refers to the actual seal detection result label of the sample workpiece, which can be determined according to the actual seal leakage rate. The actual seal leakage rate refers to the actual leakage rate value of the sample workpiece during the actual seal detection process. Among them, the seal detection result label and the actual leakage rate value can be used as the first training label.

[0084] The terminal inputs the first training seal feature and the first training seal detection feature into the first main network and the second auxiliary network in the first seal quality prediction model to be trained. The first main network performs a classification task prediction on the first training seal feature and the first training seal detection feature to obtain the first predicted seal qualification probability, and the second auxiliary network performs a regression task prediction on the first training seal feature and the first training seal detection feature to obtain the predicted seal leakage rate. Then, according to the first predicted seal qualification probability and the seal qualification classification label, calculate the first model loss corresponding to the first main network. The calculation of the first model loss is shown in formula (1).

[0085] (1) Where represents the first model loss; represents the first predicted seal qualification probability, which characterizes the probability that the prediction model believes the sample is "qualified". ; represents the seal qualification classification label.

[0086] Seal qualification classification label can be a binary label, which can be determined according to the actual seal leakage rate, as shown in formula (2).

[0087] (2) Where This represents the actual sealing leakage rate converted to the equivalent standard conditions; This indicates the preset leakage rate threshold; It is an indicator function, and formula (2) can be understood as: if (If the leakage rate meets the standard), then (Qualified), if (If the leakage rate exceeds the standard), then (Unqualified). Therefore, the model predicts the probability of a successful seal. Defined as the sealing leakage rate, converted to equivalent values ​​under standard conditions and parameters under current operating conditions and parameters. Not exceeding the preset leakage rate threshold The probability of can be expressed as in formula (3).

[0088] (3) Based on the actual sealing leakage rate and the predicted sealing leakage rate, the second model loss corresponding to the second auxiliary network is calculated. The calculation of the second model loss is shown in formula (4).

[0089] (4) in, This represents the loss of the second model; This represents the logarithm of the predicted seal leakage rate under standard conditions output by the second auxiliary network; This represents the logarithmic value of the actual sealing leakage rate under standard conditions.

[0090] The first loss model and the second loss model are fused to obtain the first objective model loss. The calculation of the first objective model loss is shown in formula (5).

[0091] (5) in, This represents the loss of the first objective model; and These represent the weight parameters of the first main network and the second auxiliary network, respectively, for example, α=0.7 and β=0.3.

[0092] Then, based on the loss of the first target model, the first sealing quality prediction model to be trained is iteratively optimized to obtain the first sealing quality prediction model after training.

[0093] In this embodiment, by using a first target model loss that integrates the first model loss and the second model loss to train the first sealing quality prediction model, the first main network and the second auxiliary network can be jointly optimized, thereby improving the training accuracy of the first sealing quality prediction model. For example, since the helium detection curve features are strongly correlated with the leak rate, the characterization results can be closer to the actual physical truth, thus simultaneously improving the first main network performing the classification task; welding data are causal features, important for both classification and regression tasks, and can provide optimization gains to the second auxiliary network performing the regression task during iterative optimization.

[0094] In an exemplary embodiment, the seal detection process is, for example, a helium detection process. During the helium detection, helium detection parameters of the sample workpiece are collected, an equivalent leak rate is calculated, and the seal detection features corresponding to the helium detection parameters are extracted. The collected helium detection parameters include data from the entire helium mass spectrometry leak detection process, such as evacuation / leakage curves, ambient temperature, effective volume, and background helium concentration. The extracted seal detection features and equivalent leak rate calculation are as follows: Feature extraction is performed on the raw helium detection curve, such as the evacuation / leakage curve: Slope k of the evacuation curve pump : Calculate the average slope of the pressure drop during the last 3 seconds of the evacuation curve, i.e., k pump = (P t-3 -P t ) / 3.0.

[0095] Peak leakage rate L peak : Directly take the maximum value of the leak rate curve during the detection phase.

[0096] Leakage rate stability value L stable : Take the arithmetic mean of the leak rate in the last 2 seconds of the detection phase.

[0097] Equivalent leakage rate conversion: The actual sealing leakage rate L of the sample workpiece was collected. m The actual sealing leakage rate L m Converted to the equivalent actual sealing leakage rate L under standard conditions eq .

[0098] The standard condition is defined as follows: reference pressure P std =1.0 mbar, reference temperature T std =20℃ (293.15K), reference volume V std =2.0L, standard background helium concentration C bg_std =0.

[0099] Monthly calibration: Use the same standard leak every month, such as the nominal value L. std =5.0×10 -9mbar·L / s, for calibrating the helium detector, the correction factors include: pressure correction factor C. P Temperature correction factor C T Volume correction factor C V Background helium compensation coefficient C bg and zero-point drift compensation coefficient C zero The equivalent leakage rate conversion is shown in formula (6).

[0100] (6) In this embodiment, based on the correction coefficients obtained from the calibration experiment, an equivalent leak rate conversion model is introduced to convert the measured leak rate to the equivalent leak rate under standard reference conditions. This eliminates the fluctuations in leak rate readings caused by differences in helium detection equipment configuration and environmental conditions, establishes a cross-platform and cross-time comparable quality grading benchmark, achieves unified quality criteria, solves the fundamental problem of consistency in quality assessment, and thus improves the accuracy of model predictions.

[0101] For example, such as Figure 4 As shown, a schematic diagram of the principle of helium detection curve feature extraction and equivalent leak rate conversion is provided. Since physical correction is to eliminate errors caused by differences in detection conditions, relying solely on statically calibrated coefficients is prone to errors. Features can be extracted from the original helium detection curve, such as the slope of the evacuation curve, the slope of the leak rate curve, and the stable value of the leak rate. Based on the extracted curve features and the equivalent leak rate conversion formula (6), a physical correction model is established to convert the measured sealing leak rate L m Converted to the equivalent actual sealing leakage rate L under standard conditions eq .

[0102] In this embodiment, a physical correction model is established based on the extracted curve features and the equivalent leakage rate conversion formula (6) to convert the measured sealing leakage rate L. m Converted to the equivalent actual sealing leakage rate L under standard conditions eq This can effectively improve the accuracy of correction and the adaptability of the model.

[0103] In an exemplary embodiment, based on the first training sealing features, the first training label, and the second training label, the second sealing quality prediction model to be trained is iteratively optimized to obtain the trained target sealing quality prediction model, including: The first training seal features are input into the second seal quality prediction model to be trained to obtain the third seal qualification probability; The loss of the second target model is determined based on the sealing qualification classification label, the second training label, and the third sealing qualification probability; Based on the loss of the second target model, the second sealing quality prediction model to be trained is iteratively optimized to obtain the trained target sealing quality prediction model.

[0104] Exemplarily, the second predicted seal qualification probability output by the first seal quality prediction model that has completed training based on the first training seal features and the first training seal detection features is used as the second training label. The second seal quality prediction model to be trained is called. The second seal quality prediction model can be a classification prediction model, and its model structure can be the same as or different from the model structure of the first main network in the first seal quality prediction model.

[0105] The terminal inputs the first training seal features into the second seal quality prediction model to be trained, obtains the third seal qualification probability, then calculates the third model loss based on the seal qualification classification label and the third seal qualification probability, and calculates the fourth model loss based on the third seal qualification probability and the second training label, and fuses the third model loss and the fourth model loss to obtain the second target model loss. Among them, the third model loss can be calculated according to formula (1), and the fourth model loss is shown in formula (7).

[0106] (7) Among them, L KL represents the fourth model loss; represents the KL divergence loss function; P t represents the second training label, that is, the second seal qualification probability; P s represents the third seal qualification probability.

[0107] Then, according to the second target model loss, the second seal quality prediction model to be trained is iteratively optimized to obtain the target seal quality prediction model that has completed training.

[0108] In this embodiment, by using the first training seal features, the first training label, and the second training label to train the second seal quality prediction model, the second seal quality prediction model indirectly learns the seal detection features, such as the law contained in the helium detection curve, by imitating the second training label, thereby ensuring the training accuracy and prediction accuracy of the target seal quality prediction model.

[0109] In an exemplary embodiment, before obtaining the first training seal features, the first training seal detection features, and the first training label involved in the seal process of the sample workpiece, the method further includes: Based on the seal end time and the seal detection start time corresponding to multiple initial sample workpieces, determine the actual process time difference corresponding to each initial sample workpiece; Based on the actual process time difference of each initial sample workpiece and the preset standard process time difference, determine the abnormal time sample workpieces among the multiple initial sample workpieces; Abnormal time sample workpieces are screened out from multiple initial sample workpieces to obtain sample workpieces.

[0110] For example, the terminal obtains sealing data and sealing detection data associated with the workpiece identifiers of multiple initial sample workpieces. Based on the sealing data and sealing detection data, it obtains the sealing end time and sealing detection start time corresponding to each of the multiple initial sample workpieces, calculates the time difference between the sealing detection start time and sealing end time of each initial sample, and obtains the actual process time difference corresponding to each initial sample.

[0111] Then, the actual process time difference is compared with the preset standard process time difference. Based on the comparison results, abnormal time sample workpieces are identified among multiple initial sample workpieces. The terminal then filters out the abnormal time sample workpieces from the multiple initial sample workpieces to obtain the sample workpieces.

[0112] For example, to ensure production line efficiency, the preset standard time difference between the welding process and the helium inspection process is 130 seconds, and an error baseline of 15 seconds is set. Based on the preset standard time difference and the error baseline, the tolerance time window is 115 seconds to 145 seconds. If the actual time difference exceeds the tolerance time window (i.e., the actual time difference is greater than 145 seconds), it may be due to an abnormal stoppage in the next or subsequent process when the workpiece needs to be transferred from the welding process, causing the workpiece's process time to exceed the time limit. In this case, the workpiece is identified as an abnormal time sample workpiece. If the actual time difference does not reach the tolerance time window (i.e., the actual time difference is less than 115 seconds), it may be due to a scheduling error causing the workpiece to not complete the entire transmission path (e.g., bypassing the buffer area) and be directly sent to the helium inspection process. In this case, the workpiece is identified as an abnormal time sample workpiece.

[0113] In an exemplary embodiment, the standard process time difference can be output by establishing a transmission model. First, based on the workpiece flow path: welding station → roller conveyor → buffer area → robot gripper → helium inspection equipment, the transmission time of each path is calculated, and a transmission model is established as shown in formula (8).

[0114] (8) Where Δt represents the standard process time difference; d1 represents the roller conveyor length; v1 represents the roller conveyor speed; d2 represents the robot gripping distance; v2 represents the robot's average speed; τ buffer Indicates the average dwell time in the buffer; τ queue This indicates the average queuing time in front of the helium detection equipment.

[0115] In this embodiment, by determining and filtering out abnormal time sample workpieces based on the actual process time difference and the preset standard process time difference, the accuracy of the training data corresponding to the sample workpieces can be guaranteed, thereby ensuring the training accuracy of the target sealing quality prediction model.

[0116] In an exemplary embodiment, before acquiring the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label, the method further includes: Obtain the reference sealing end time corresponding to the pre-selected workpiece. The pre-selected workpiece is the workpiece whose workpiece identifier is successfully associated with the sealing data but fails to be associated with the sealing detection data. The sealing data is used to extract sealing features, and the sealing detection data is used to extract sealing detection features. The target time search range is determined based on the difference between the reference sealing end time and the preset standard process time. Acquire multiple sealing detection data points whose start time falls within the target time search range, and determine the target sealing detection data point from among the multiple sealing detection data points; Pre-selected workpieces that are successfully associated with the target sealing test data are used as sample workpieces.

[0117] For example, the terminal successfully associates sealing data with workpiece identifiers in the historical production database, while workpieces whose sealing detection data fails to be associated with workpiece identifiers are selected as pre-selected workpieces. That is, workpieces with sealing data but missing sealing detection data are selected as pre-selected workpieces.

[0118] The sealing end time in the sealing data corresponding to the pre-selected workpiece is used as the reference sealing end time for the pre-selected workpiece. Then, based on the difference between the reference sealing end time and the preset standard process time, the target time search range is determined. The target time search range can be expressed as (T 参考 +Δt-T 误 T 参考 +Δt+ T 误 ), T 参考 T represents the reference sealing end time, Δt represents the preset standard process time difference, and T represents the reference sealing end time. 误 This represents the error reference. For example, Δt = 130s, T 误 =15s, then the target time search range is [T] after the timestamp of the reference sealing end time. 参考 +115, T 参考 A time window of +145 seconds.

[0119] Then, in the sealing inspection dataset, search for multiple sealing inspection data whose start time is within the target time search range. Based on the preset matching dimensions, filter the target sealing inspection data that matches the pre-selected workpiece from the multiple sealing inspection data. The preset matching dimensions are, for example, time proximity, logistics sequence consistency, product model, working condition matching degree, etc.

[0120] Specifically, this can involve statistically analyzing the matching degree between multiple sealing test data and the pre-selected workpiece in each preset matching dimension, weighting and fusing the multiple matching degrees to obtain the comprehensive matching degree of the multiple sealing test data, and determining the sealing test data with the highest comprehensive matching degree as the target sealing test data.

[0121] The target sealing test data is associated with the workpiece identifier of the pre-selected workpiece, and the pre-selected workpiece that is successfully associated is used as the sample workpiece.

[0122] In this embodiment, by supplementing the missing sealing test data for pre-selected workpieces and using them as sample workpieces, the amount of training data corresponding to the sample workpieces can be increased, thereby ensuring the training accuracy of the target sealing quality prediction model.

[0123] In one exemplary embodiment, such as Figure 5 The diagram illustrates a spatiotemporal alignment model for data association and verification. By scanning workpiece identifiers (such as QR codes), welding data, helium inspection data, and helium inspection results are automatically collected from the controller of the welding equipment and the data interface of the helium inspection equipment, and stored in a time-series database bound to the workpiece identifiers.

[0124] In the time-series database, enable a data association engine such as SQL statements to associate the welding data table and the helium inspection result table using the workpiece identifier as the key. The welding data in the welding data table carries a timestamp T. w The helium test result table contains helium test data and helium test result data. The helium test data carries a timestamp T. h .

[0125] The workpiece flow path is determined based on production line logistics information such as path length, transmission speed, and buffer time. A transmission model (also described as a spatiotemporal alignment model) is established based on the workpiece flow path. The theoretical time delay Δt is calculated using the spatiotemporal alignment model, which is the standard process time difference. Then, the welding end time and helium inspection start time of the initial sample workpiece are verified based on the theoretical time delay. The initial sample that passes the verification is used as the sample workpiece, and the sample workpieces that fail the verification and are judged as abnormal data are manually reviewed.

[0126] For multiple helium inspection data lacking workpiece identification, a time window (i.e., target time search range) is determined based on the reference welding end time, theoretical time delay, and error benchmark of the pre-selected workpiece. Data gaps are filled according to the time window, that is, target helium inspection data matching the pre-selected workpiece is searched from the multiple helium inspection data lacking workpiece identification according to the time window, and the target helium inspection data is associated with the pre-selected workpiece.

[0127] In this embodiment, it is not only used to verify the correctness of data directly associated with the identifier code, but also to fill in the gaps in records with missing identifiers. This solves the problem of time asynchrony caused by data transmission and caching across processes, thereby ensuring the integrity and reliability of the modeling data foundation.

[0128] In one exemplary embodiment, the intelligent sorting and risk warning process for laser welding quality based on helium calibration and working condition stratification includes an offline modeling stage and an online sorting stage.

[0129] The offline modeling phase includes the following steps: S1: Data acquisition and binding steps, using the workpiece identifier as the key index, synchronously acquire and bind welding process parameters, helium mass spectrometry leak detection process parameters and leak detection result data; S2: Data association and verification steps, using the workpiece identifier as the primary key for data association, and establishing a spatiotemporal alignment model based on the workpiece transmission path to verify and fill in any gaps in the association results; S3: Equivalent leak rate conversion step, extracting time-series features from the helium detection process curve, and converting the measured leak rate to the equivalent leak rate under standard reference conditions based on predefined physical correction coefficients; S4: Supervised learning modeling steps: The samples are stratified according to key operating condition variables, and a sealing quality prediction model is built in each stratum with welding process parameters and helium inspection process characteristics as inputs and airtightness qualification as the output target for training.

[0130] The online sorting stage includes the following steps: S5: Online prediction step. For new workpieces that have been welded but have not yet undergone helium testing, the welding process parameters are acquired in real time and input into the model trained in step S4 to obtain the probability of airtightness. S6: Three-level sorting decision steps, with preset high and low thresholds: If the probability of passing the airtightness test is higher than the high threshold, it is judged as a qualified product and is allowed to proceed directly to the regular helium testing process. If the airtightness compliance rate is lower than the low threshold, it is judged as a non-conforming product and diverted to a special processing channel. If the probability of passing the airtightness test is between the high and low thresholds, it is judged as a risky product and enters the standard helium testing process.

[0131] The specific implementation details are as follows: The data acquisition and binding steps include: device configuration initialization; Data Acquisition and Cleaning: Welding data and helium testing data were acquired. Welding data included laser power P, welding speed v, swing frequency / amplitude / duty cycle, focal length offset Δf, shielding gas flow rate Q, and purity, etc.; helium testing data included helium filling pressure P. bomb / time t bomb Evacuation / leakage curve p(t) or R(t), detection temperature T, effective volume V of the measured cavity. cel l. Background Helium (He) bg Equipment zero-point / sensitivity calibration data; preprocessing welding data and helium detection data, such as cleaning preprocessing, including deduplication, time-based correction, missing / outlier labeling, recording background helium fluctuations and zero-point drift, etc.

[0132] Data association and verification steps include: Cross-process time-space alignment: The theoretical time delay Δt between stations i is calculated as shown in formula (9).

[0133] (9) in, Indicates the length of the path segment for workstation i; Indicates the segmental linear velocity of workstation i; Indicates cache retention time; Indicates queuing / switching time; This indicates the device clock skew.

[0134] Matching strategy: Two levels of tolerance (primary tolerance ±0.6~1.0Δt; secondary tolerance ±1.5~2.0Δt) + non-zero priority + phase self-correction are adopted to ensure the accuracy and fault tolerance of data correlation.

[0135] The equivalent leakage rate conversion steps include: Curve feature extraction and equivalent leakage rate conversion: Curve feature extraction includes extracting curve slopes such as the slope of the emptying curve and the slope of the leakage rate curve, as well as shape features such as the mean of the steady-state segment, the time constant τ, the integral area, and the inflection point t; Equivalent leakage rate conversion: , among which, each To correct for factors (pressure, temperature, volume, background helium, zero-point drift), the measured leak rate is estimated using standard samples / historical calibration. Unified conversion to standard reference conditions This enables cross-device / cross-configuration comparability.

[0136] Supervised learning modeling steps include: Layered modeling of working conditions: Layered variables: material type, plate thickness, structural form, fixtures, etc.; Offline model (first sealing quality prediction model) training input: welding data features + helium detection data features (including equivalent leak rate); Offline model training output: airtightness compliance probability (main task); Online model (target seal quality prediction model) training input: welding data features; Online model training output: probability of airtightness compliance; Model selection: XGBoost / LightGBM or lightweight neural networks; Validation: Time-block and out-of-batch validation; Metrics: AUC, recall, etc.

[0137] Online prediction steps include: Online prediction and three-level sorting decision-making: Deploy an online model (target seal quality prediction model). For new workpieces that have been welded but have not yet undergone helium testing, acquire welding data in real time and input it into the online model to obtain the probability of airtightness. ; The three-level sorting decision-making process includes: Preset high threshold With low threshold Perform three-level sorting: Predicted qualified products (green channel): If ≥ The workpiece can be directly connected to the conventional or rapid helium testing process.

[0138] Predicting non-conforming products (red channel): If ≤ The workpieces are diverted to special processing channels (such as being scrapped directly, downgraded to grade B, or entering the priority review channel). Risky goods (yellow channel): If < < The workpiece enters the standard helium testing process and undergoes precise standard testing; Threshold Adaptive: High Threshold With low threshold It can be dynamically adjusted according to the background helium concentration of the helium detection system, the zero-point status of the equipment, or the production quality target.

[0139] Risk warning and root cause analysis: Real-time statistical prediction of the proportion and characteristics of non-conforming products. When the proportion exceeds the preset control limit, a production line anomaly warning is issued to the production / equipment engineer. Combined with cluster analysis of welding parameters of non-conforming workpieces, it helps to locate potential root causes, such as laser power attenuation, protective lens contamination, or incoming material batch issues.

[0140] Traceability and visualization: Output data alignment details, equivalent leak rate logs, equipment status trajectories, sorting decision records, quality KPI trends and early warning logs provide full-chain data support for quality backtracking and continuous improvement.

[0141] In an exemplary embodiment, taking the application in a laser sealing welding production line for power battery covers as an example, the process steps of intelligent sorting and risk warning for laser welding quality based on helium detection calibration and working condition stratification are described, including the following: 1. Application Scenarios and Data Preparation This embodiment uses a laser sealing welding production line for aluminum cover plates of a new energy vehicle power battery as an example. This production line produces two main types of cover plates: Type A (material: 3003 aluminum alloy, plate thickness: 1.2mm) and Type B (material: 304 stainless steel, plate thickness: 1.0mm). The production line is equipped with three different models of helium mass spectrometry leak detection equipment (equipment numbers: HLJ-01, HLJ-02, HLJ-03).

[0142] To build the model, complete production data for 30 consecutive working days, totaling 85,632 workpieces, was collected from this production line. Each data record contains: Welding data: laser power (P, unit: watt-W), welding speed (v, unit: mm / s), defocusing amount (Δf, unit: mm). Raw helium detection data: complete evacuation-detection pressure curve and leak rate curve (sampling frequency 10Hz), chamber pressure at the start of detection (P act (Unit: millibar, mbar) and ambient temperature (T) act (Unit: degrees Celsius - ℃), effective volume estimated based on workpiece and fixture (V_act, unit: liters - L), background helium concentration measured before inspection (C). bg_act (Unit: millibar·L / s) Helium leak test results data: the final leak rate value determined by the equipment (L meas (Unit: mbar·L / s) and pass / fail status (pass / fail, threshold set to 1.0×10⁻⁶). -9 mbar·L / s).

[0143] Working conditions and logistics information: workpiece ID, product model (corresponding material and plate thickness), welding station number, welding completion timestamp, helium inspection station number, and helium inspection start timestamp.

[0144] 2. Specific Implementation Steps The system automatically retrieves welding data, raw helium inspection data, and helium inspection results from the PLC controller of the welding equipment and the data interface of the helium inspection equipment by scanning the workpiece identification, and stores them in a time-series database by binding them with the workpiece identification.

[0145] Use SQL statements in the database to link the welding data table and the helium inspection result table, with the workpiece identifier as the key.

[0146] Establish a transport model: The workpiece flow path is: welding station → roller conveyor with length d1 = 8 meters → average dwell time τ buffer =45-second buffer zone → Robot grasping distance d2=3 meters → Helium detection equipment.

[0147] The theoretical time delay Δt is calculated as follows: roller conveyor speed v1 = 0.4 m / s, robot average speed v2 = 0.6 m / s, and average queuing time τ in front of the helium detector. queue =60 seconds.

[0148] The theoretical delay is calculated as follows: Δt=d1 / v1+d2 / v2+τ buffer +τ queue =8 / 0.4+3 / 0.6+45+60=20+5+45+60=130 seconds.

[0149] Time verification: For data successfully associated with the ID, calculate the actual difference δt between the welding completion time and the helium detection start time. If |δt-130|>15 seconds (error baseline), mark the data as "time anomaly" for manual review. In this embodiment, the percentage of anomaly data is approximately 0.5%.

[0150] Missing records: For records with missing IDs in the helium detection data table, the most matching welding record was searched for and associated within a time window of [115, 145] seconds after the welding completion timestamp. 1,203 missing records were successfully added. In the end, the total number of valid samples was 86,835.

[0151] Curve feature extraction: slope k of the evacuation curve pump Leakage peak L peak Leakage rate stability value L stable .

[0152] Taking a calibration of the HLJ-01 equipment as an example, the equivalent leakage rate L is calculated. eq Conversion, obtain the correction factor in advance: Pressure correction factor C P =P std / P act =1.0 / 0.95≈1.0526; Temperature correction factor C T =sqrt(T act / T std = sqrt(298.15 / 293.15)≈1.0085 (assuming the ambient temperature is 25℃); Volume correction factor C V =V act / V std =2.1 / 2.0=1.05; Background helium compensation coefficient C bg =1+(C bg_act / L std =1 + (2.0e-10 / 5.0e-9) = 1.04; Zero drift compensation coefficient C zero =0.995 (read from the device calibration log) .

[0153] Pressure correction factor C P Temperature correction factor C T Volume correction factor C V Background helium compensation coefficient C bg and zero-point drift compensation coefficient C zero This applies to the calculation of leakage rate for all workpieces within this cycle of the equipment.

[0154] Working condition stratification: The data is divided into two layers according to "material-plate thickness": Layer 1 (aluminum alloy-1.2mm, sample size 52,101) and Layer 2 (stainless steel-1.0mm, sample size 34,734).

[0155] Model training (taking Layer 1 as an example): Input features (6 dimensions): [laser power P, welding speed v, defocusing amount Δf, evacuation slope k] pump Leakage peak L peak Leakage rate stability value L stable Z-score normalization is applied to all features; Model selection and training: The LightGBM framework is used to build a multi-task learning model, namely the first sealing quality prediction model (offline model) and the second sealing quality prediction model (online model). Offline model: Main task (classification): Air tightness compliance (compliant = 1, non-compliant = 0), using the binary cross-entropy loss function. ; Auxiliary task (regression): Natural logarithm of equivalent leak rate log(L) eq Using the mean squared error loss function ; Total loss function: ; Hyperparameters: learning rate set to 0.05, maximum tree depth to 6, number of iterations to 500, training set and test set divided in an 8:2 ratio; After the offline and online models were trained, the AUC area of ​​the model on the test set for predicting the airtightness compliance probability reached 0.963.

[0156] Model deployment: Deploy the trained online model on the production line edge computing server.

[0157] Online prediction: When a type A cover plate (aluminum alloy -1.2mm) is welded, its welding parameters [P=1550W, v=65mm / s, Δf=+1mm] are collected in real time, input into the online model of Layer 1, and its airtightness compliance probability P is output. pass =0.89.

[0158] Three-level sorting decision: Preset high threshold T high =0.95, low threshold T low =0.15; Since 0.15 < 0.89 < 0.95, the workpiece was identified as a high-risk item and entered into the standard helium testing process. If the probability of a certain workpiece is 0.07 (<0.15), it is judged as a defective product and sorted to the scrap channel; If the probability of a certain workpiece is 0.97 (>0.95), it is judged as a qualified product and enters the fast helium inspection channel.

[0159] Risk warning: The system detected that the proportion of defective products rose from an average of 1% to 4.5% and continued to rise. It immediately issued a warning to the equipment engineer. After investigation, it was found that the laser protective lens was contaminated, which caused the energy to be unstable. After timely treatment, the proportion returned to normal.

[0160] On the other hand, this embodiment provides a method for training a sealing quality prediction model, such as... Figure 6 As shown, it includes the following steps: S601, acquire the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label; S602, Based on the first training sealing features, the first training sealing detection features and the first training label, the first sealing quality prediction model to be trained is iteratively optimized to obtain the first sealing quality prediction model after training. S603, the second predicted seal pass probability output by the first trained seal quality prediction model for the first trained seal features and the first trained seal detection features is used as the second training label; S604. Based on the first training sealing features, the first training label, and the second training label, the second sealing quality prediction model to be trained is iteratively optimized to obtain the target sealing quality prediction model after training.

[0161] In this embodiment, a first sealing quality detection model is trained using the first training sealing features of the sample workpiece in the sealing process, the first training sealing detection features in the sealing detection process, and the first training label. This results in a trained first sealing quality prediction model, which can accurately predict the sealing pass probability of the sample workpiece, providing an accurate and effective second training label for the second sealing quality prediction model. Then, the first training sealing features, the first training label, and the second training label are used to train the second sealing quality prediction model, resulting in a trained target sealing quality prediction model, ensuring the training accuracy of the target sealing quality prediction model. Furthermore, by inputting the sealing parameters of the workpiece to be tested into the trained target sealing quality prediction model, without needing to additionally input the sealing detection features of the workpiece, the sealing pass probability of the workpiece can be accurately predicted, achieving both high prediction efficiency and accuracy in predicting the sealing pass probability of the workpiece.

[0162] Thirdly, Figure 7 This is a schematic diagram of a workpiece sealing detection device according to an embodiment of this application, as shown below. Figure 7 As shown, the above-mentioned workpiece sealing detection device 700 includes: a data acquisition module 701, a sealing quality prediction module 702, and a sealing detection module 703. The device will be described below.

[0163] The data acquisition module 701 is used to acquire the sealing parameters involved in the sealing process of the workpiece under test; The sealing quality prediction module 702 is used to input sealing parameters into the target sealing quality prediction model and predict the sealing qualification probability of the workpiece to be tested. The sealing detection module 703 is used to perform sealing detection on the workpiece under test based on the sealing pass probability, and obtain the sealing detection result corresponding to the workpiece under test.

[0164] In an exemplary embodiment, the sealing detection module 703 is further configured to determine the sealing detection type corresponding to the workpiece to be tested based on the sealing pass probability; and perform sealing detection on the workpiece to be tested according to the sealing detection type to obtain the sealing detection result corresponding to the workpiece to be tested.

[0165] In an exemplary embodiment, the sealing detection module 703 is further configured to determine that the sealing detection type corresponding to the workpiece under test is non-standard sealing detection when the sealing pass probability exceeds a preset pass range; determine that the sealing detection type corresponding to the workpiece under test is standard sealing detection when the sealing pass probability is within the predicted pass range; and determine that the sealing detection type corresponding to the workpiece under test is terminated sealing detection when the sealing pass probability does not reach the preset pass range.

[0166] In one exemplary embodiment, the sealing process is a welding process, and the sealing test result is obtained by performing an airtightness test on the workpiece under test in the airtightness test process.

[0167] In an exemplary embodiment, the target seal quality prediction model is obtained by using the second predicted seal pass probability output by the trained first seal quality prediction model as the second training label, and training the model based on the input first training seal features and the second training label. The first seal quality prediction model is obtained by training the model based on the input first training seal features and the first training seal detection features.

[0168] In an exemplary embodiment, the workpiece sealing detection device 700 is further configured to acquire a first training sealing feature involved in the sealing process of the sample workpiece, a first training sealing detection feature involved in the sealing detection process, and a first training label; based on the first training sealing feature, the first training sealing detection feature, and the first training label, iteratively optimize the first sealing quality prediction model to be trained to obtain a first sealing quality prediction model that has been trained; use the second predicted sealing pass probability output by the first training sealing feature and the first training sealing detection feature as the second training label; and based on the first training sealing feature, the first training label, and the second training label, iteratively optimize the second sealing quality prediction model to be trained to obtain a target sealing quality prediction model that has been trained.

[0169] In an exemplary embodiment, the first training label includes a seal qualification classification label and an actual seal leakage rate; the workpiece seal detection device 700 is further configured to input the first training seal features and the first training seal detection features into the first main network and the second auxiliary network of the first seal quality prediction model to be trained; obtain the first predicted seal qualification probability through the first main network and the predicted seal leakage rate through the second auxiliary network; determine the first target model loss based on the seal qualification classification label, the first predicted seal qualification probability, the actual seal leakage rate and the predicted seal leakage rate; and iteratively optimize the first seal quality prediction model to be trained based on the first target model loss to obtain the trained first seal quality prediction model.

[0170] In an exemplary embodiment, the workpiece sealing detection device 700 is further configured to input the first training sealing features into the second sealing quality prediction model to be trained to obtain the third sealing qualification probability; determine the second target model loss based on the sealing qualification classification label, the second training label and the third sealing qualification probability; and iteratively optimize the second sealing quality prediction model to be trained based on the second target model loss to obtain the trained target sealing quality prediction model.

[0171] In an exemplary embodiment, the workpiece sealing detection device 700 is further configured to determine the actual process time difference corresponding to each initial sample workpiece based on the sealing end time and sealing detection start time corresponding to the multiple initial sample workpieces respectively; determine the abnormal time sample workpiece among the multiple initial sample workpieces based on the actual process time difference of each initial sample workpiece and the preset standard process time difference; and remove the abnormal time sample workpiece from the multiple initial sample workpieces to obtain the sample workpiece.

[0172] In an exemplary embodiment, the workpiece sealing detection device 700 is further configured to acquire a reference sealing end time corresponding to a pre-selected workpiece, wherein the pre-selected workpiece is a workpiece whose workpiece identifier is successfully associated with the sealing data but fails to be associated with the sealing detection data, the sealing data is used to extract sealing features, and the sealing detection data is used to extract sealing detection features; based on the reference sealing end time and the time difference of a preset standard process, a target time search range is determined; multiple sealing detection data whose sealing detection start time is within the target time search range are acquired, and a target sealing detection data is determined from the multiple sealing detection data; and the pre-selected workpiece whose workpiece identifier is successfully associated with the target sealing detection data is used as a sample workpiece.

[0173] Fourthly, Figure 8 This is a schematic diagram of a sealing quality prediction model training device according to an embodiment of this application, as shown below. Figure 8 As shown, the above-mentioned sealing quality prediction model training device 800 includes: a data acquisition module 801, a first model training module 802, a label determination module 803, and a second model training module 804. The device will be described below.

[0174] The data acquisition module 801 is used to acquire the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label; The first model training module 802 is used to iteratively optimize the first seal quality prediction model to be trained based on the first training seal features, the first training seal detection features and the first training label, so as to obtain the trained first seal quality prediction model. The label determination module 803 is used to take the second predicted seal pass probability output by the trained first seal quality prediction model for the first trained seal features and the first trained seal detection features as the second training label. The second model training module 804 is used to iteratively optimize the second sealing quality prediction model to be trained based on the first training sealing features, the first training label, and the second training label, so as to obtain the target sealing quality prediction model after training.

[0175] Each module in the aforementioned sealing detection device and sealing quality prediction model training device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0176] Fifthly, this embodiment provides an electronic device, including a memory and a processor. The memory stores computer instructions, and when the computer instructions are executed by the processor, they implement the method of any of the above embodiments.

[0177] In one embodiment, this embodiment also provides an electronic device, which may be a server, and its internal structure diagram may be as follows. Figure 9 As shown, this electronic device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer instructions, and a database. The internal memory provides the environment for the operation of the operating system and computer instructions stored in the non-volatile storage media. The database stores data involved in the business data processing methods. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer instructions are executed by the processor, they implement a workpiece sealing detection method and a model training method.

[0178] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0179] Sixthly, embodiments of this application provide a computer-readable storage medium storing computer instructions thereon, which are loaded by a processor to execute the arrangements in any of the methods described above. In embodiments of this application, the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0180] In a seventh aspect, embodiments of this application provide a computer program product, including a computer program or instructions, which are executed by a processor to implement the steps of any of the methods described above.

[0181] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0182] The above provides a detailed description of the workpiece sealing detection method, model training method, electronic device, and computer-readable storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for detecting the seal of a workpiece, characterized in that, include: Obtain the sealing parameters involved in the sealing process of the workpiece under test; The sealing parameters are input into the target sealing quality prediction model to predict the sealing qualification probability of the workpiece under test. Based on the sealing pass probability, the workpiece to be tested is subjected to sealing test, and the sealing test result corresponding to the workpiece to be tested is obtained.

2. The method according to claim 1, characterized in that, The process of performing a seal test on the workpiece under test based on the seal pass probability to obtain the seal test result corresponding to the workpiece under test includes: Based on the sealing pass probability, determine the sealing test type corresponding to the workpiece to be tested; According to the described sealing test type, the workpiece to be tested is subjected to sealing test, and the sealing test result corresponding to the workpiece to be tested is obtained.

3. The method according to claim 2, characterized in that, The process of determining the seal inspection type corresponding to the workpiece under test based on the seal pass probability includes: When the sealing pass probability exceeds the preset pass range, the sealing test type corresponding to the workpiece to be tested is determined to be non-standard sealing test; When the sealing pass probability is within the predicted pass range, the sealing test type corresponding to the workpiece to be tested is determined to be standard sealing test; When the sealing pass probability does not reach the preset pass range, the sealing test type corresponding to the workpiece to be tested is determined to be terminated sealing test.

4. The method according to any one of claims 1-3, characterized in that, The sealing process is a welding process, and the sealing test result is obtained by performing an airtightness test on the workpiece under test in the airtightness test process.

5. The method according to any one of claims 1-4, characterized in that, The target sealing quality prediction model is obtained by using the second predicted sealing pass probability output by the first sealing quality prediction model after training as the second training label, and training the model based on the input first training sealing features and the second training label. The first sealing quality prediction model is obtained by training the model based on the input first training sealing features and the first training sealing detection features.

6. The method according to any one of claims 1-5, characterized in that, The method further includes: Acquire the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label; Based on the first training sealing features, the first training sealing detection features and the first training label, the first sealing quality prediction model to be trained is iteratively optimized to obtain the first sealing quality prediction model after training. The second predicted seal pass probability output by the first trained seal quality prediction model based on the first trained seal features and the first trained seal detection features is used as the second training label. Based on the first training sealing features, the first training label, and the second training label, the second sealing quality prediction model to be trained is iteratively optimized to obtain the target sealing quality prediction model after training.

7. The method according to claim 6, characterized in that, The first training label includes a seal pass classification label and an actual seal leakage rate; the step of iteratively optimizing the first seal quality prediction model to be trained based on the first training seal features, the first training seal detection features, and the first training label to obtain the trained first seal quality prediction model includes: The first training sealing features and the first training sealing detection features are input into the first main network and the second auxiliary network of the first sealing quality prediction model to be trained. The first predicted seal pass probability is obtained through the first main network, and the predicted seal leakage rate is obtained through the second auxiliary network; Based on the seal pass classification label, the first predicted seal pass probability, the actual seal leakage rate, and the predicted seal leakage rate, the first target model loss is determined. Based on the loss of the first target model, the first sealing quality prediction model to be trained is iteratively optimized to obtain the first sealing quality prediction model after training.

8. The method according to claim 7, characterized in that, The step of iteratively optimizing the second seal quality prediction model to be trained based on the first training seal features, the first training label, and the second training label to obtain the trained target seal quality prediction model includes: The first training sealing features are input into the second sealing quality prediction model to be trained to obtain the third sealing qualification probability; Based on the sealing qualification classification label, the second training label, and the third sealing qualification probability, the second target model loss is determined; Based on the loss of the second target model, the second sealing quality prediction model to be trained is iteratively optimized to obtain the trained target sealing quality prediction model.

9. The method according to any one of claims 5-8, characterized in that, Before acquiring the first training sealing feature involved in the sealing process of the sample workpiece, the first training sealing detection feature involved in the sealing detection process, and the first training label, the method further includes: Based on the sealing end time and sealing detection start time corresponding to multiple initial sample workpieces, the actual process time difference corresponding to each initial sample workpiece is determined. Based on the actual process time difference and the preset standard process time difference for each initial sample workpiece, abnormal time sample workpieces are identified among the multiple initial sample workpieces. The abnormal time sample workpieces are screened out from the multiple initial sample workpieces to obtain the sample workpieces.

10. The method according to any one of claims 5-8, characterized in that, Before acquiring the first training sealing feature involved in the sealing process of the sample workpiece, the first training sealing detection feature involved in the sealing detection process, and the first training label, the method further includes: Obtain the reference sealing end time corresponding to the pre-selected workpiece. The pre-selected workpiece is the workpiece whose workpiece identifier is successfully associated with the sealing data but fails to be associated with the sealing detection data. The sealing data is used to extract sealing features. The sealing detection data is used to extract sealing detection features. Based on the reference sealing end time and the time difference of the preset standard process, the target time search range is determined; Acquire multiple sealing detection data whose start time is within the target time search range, and determine the target sealing detection data among the multiple sealing detection data; Pre-selected workpieces that are successfully associated with the target sealing test data are used as sample workpieces.

11. A method for training a sealing quality prediction model, characterized in that, include: Acquire the first training sealing features involved in the sealing process of the sample workpiece, the first training sealing detection features involved in the sealing detection process, and the first training label; Based on the first training sealing features, the first training sealing detection features and the first training label, the first sealing quality prediction model to be trained is iteratively optimized to obtain the first sealing quality prediction model after training. The second predicted seal pass probability output by the first trained seal quality prediction model based on the first trained seal features and the first trained seal detection features is used as the second training label. Based on the first training sealing features, the first training label, and the second training label, the second sealing quality prediction model to be trained is iteratively optimized to obtain the target sealing quality prediction model after training.

12. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing a computer program configured to be executed by the processor to implement the steps of the method according to any one of claims 1 to 11.

13. A computer storage medium, characterized in that, The computer storage medium stores a computer program configured to be executed by a processor to implement the method of any one of claims 1 to 11.

Images