SMT steel mesh residual life prediction system based on dynamic data coupling and self-calibration method

Abstract

The invention belongs to the technical field of detection of steel meshes for manufacturing PCBs (printed circuit boards), and particularly relates to an SMT (surface mount technology) steel mesh residual life prediction system based on dynamic data coupling and a self-calibration method, the service life of a steel mesh is predicted by monitoring working state parameters of the steel mesh in real time and combining with a deep learning algorithm, and the system takes a tension value as a core monitoring parameter, so that the service life of the steel mesh is predicted. Image recognition is applied to microscopic damage detection, multi-source data fusion is realized, single index prediction limitation is broken through, a prediction model training sample is constructed through steel mesh full life cycle data, printing machine printing parameters and process environment parameters, then a prediction model is trained, online learning is performed on the prediction model in combination with newly added data, and a prediction result is obtained. The residual life of the steel mesh can be predicted in real time in combination with a prediction model output result, the system can well solve the problems of time and labor consumption and production interruption in the prior art, and the prediction precision is improved in combination with a dynamic learning mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of steel mesh detection for manufacturing PCB boards, and particularly relates to an SMT steel mesh remaining life prediction system and a self-calibration method based on dynamic data coupling. Background Art

[0002] With the rapid development of intelligent manufacturing, surface mount technology (SMT) has been widely used in the electronics manufacturing industry. An SMT production line typically consists of multiple high-precision devices, such as printers, placement machines, reflow ovens, and AOI (automated optical inspection). The printer in an SMT production line is primarily used to evenly print solder paste or red glue onto designated locations on the PCB, preparing for the subsequent placement process. A dedicated stencil is used during the printing process. The stencil's precisely designed holes ensure that solder paste is deposited only where it is needed on the PCB, preventing overflow or loss of solder paste and ensuring soldering quality.

[0003] SMT stencils typically have a limited lifespan. After a certain number of uses, the stencil's precision and performance degrade, leading to unstable printing quality. These can include clogged stencil holes, tension variations, wear, and deformation. Therefore, stencil replacement is inevitable during production. This degradation in stencil performance can lead to quality issues with printed PCBs. To ensure product quality, quality inspections are typically performed after each production run of a certain number of PCBs. SPI inspection equipment (solder paste inspection equipment) measures solder paste height, volume, area, shorts, and offset after printing to ensure print quality is not compromised. The stencil is replaced only when it no longer meets printing standards. This method requires significant manpower for inspection and maintenance, interrupts the SMT production line, and results in data fragmentation, making it inefficient.

[0004] In summary, current SMT stencil life assessment relies on quality inspection, manual experience, or fixed-cycle replacement strategies, which have the following defects: judging stencil performance by inspecting the quality of produced PCB boards can cause production line downtime, with high labor costs and low efficiency; fixed-cycle replacement can cause the stencil to be discarded before reaching its lifespan, resulting in serious material waste and the inability to predict deterioration in printing quality due to tension attenuation and microcrack expansion; judging by manual experience may result in errors, and different people have different experiences, making it difficult to obtain a unified and reliable judgment result. Summary of the Invention

[0005] The present invention aims to provide an SMT steel mesh remaining life prediction system and a self-calibration method based on dynamic data coupling, so as to realize the systematic correlation analysis of key indicators such as hole wear, steel mesh tension, and environmental parameters of the SMT steel mesh through the SMT steel mesh remaining life prediction system based on dynamic data coupling, capture the attenuation law of the steel mesh life through historical data, and make real-time prediction of the steel mesh life through real-time data printed by the printing press.

[0006] To achieve the above objectives, the present invention adopts the following technical solutions:

[0007] Provided is an SMT steel mesh remaining life prediction system based on dynamic data coupling, including:

[0008] A data acquisition module, which includes a real-time monitoring data acquisition unit and a historical data acquisition unit. The real-time monitoring data acquisition unit is used to monitor and collect data related to the stencil on the production line in real time. The data includes the tension value of the stencil collected based on time series and images that can show changes in microscopic morphology. The historical data acquisition unit is used to collect historical life data of the stencil for reference comparison, as well as to collect a fatigue parameter library of stencils of different materials.

[0009] a data processing module, the data processing module including a time series data standardization unit and an image data enhancement unit, the time series data standardization unit being used to perform standardization processing on the time series numerical data collected by the data acquisition module, and the image data enhancement unit being used to perform enhancement processing on the image data collected by the data acquisition module;

[0010] The remaining life prediction module is used to predict the remaining life of the steel mesh based on the data collected by the data collection module and the data processed by the data processing module.

[0011] Preferably, the data collected by the real-time monitoring data acquisition unit specifically include: (1.1) the tension values ​​of the four corners and the center of the steel mesh collected by the embedded MEMS strain gauge; (1.2) the image of the opening on the steel mesh captured by the industrial camera, which is used to obtain the edge burr index and diameter deviation of the opening; (1.3) the scraper pressure; (1.4) the viscosity of the solder paste; (1.5) the steel mesh temperature; (1.6) the humidity; (1.7) the tin powder particle size; (1.8) the printing speed; (1.9) the scraper width; (1.10) the printing gap;

[0012] The data collected by the historical data collection unit specifically include: (2.1) the full life cycle record of the same type of steel mesh, including tension decay curve, hole deformation rate, and maintenance record; (2.2) fatigue parameter library of steel meshes of different materials.

[0013] Preferably, the algorithm architecture and network structure of the remaining life prediction module include an input layer, a feature fusion layer, a prediction model and an output layer, wherein: the input layer includes a time series data branch unit for receiving time series data, and an image data branch unit for extracting image feature data, the image feature data including the hole diameter deviation of the steel mesh;

[0014] The feature fusion layer includes a feature cross unit, a composite index generation unit and an input preprocessing unit. The feature cross unit is used to combine the feature data extracted from the image with other features.

[0015] The composite index generating unit is used to calculate the fatigue accumulation factor and the deformation entropy value based on the following formula (1) and formula (2):

[0016]

[0017] Deformation entropy = -∑p i *ln(p i ) (2)

[0018] By dividing the actual measured hole diameter deviation range into n continuous intervals, the parameter p i Represents the number of deviation samples appearing in the i-th interval / total number of samples;

[0019] The input preprocessing unit is used to receive one-dimensional time series data T1 is the time step of the one-dimensional time series data, C1 is the feature dimension of the one-dimensional time series data, and normalization and sliding window segmentation are performed to generate fragmented data suitable for model input;

[0020] The prediction model is a TimesNet model. After pre-training, the prediction model performs prediction based on the fragmented data generated by the input pre-processing unit. The fully connected layer of the prediction model is mapped to future time steps, and the prediction result retains the time dimension feature, which includes the predicted time series data of the number of printings in the future.

[0021] The output layer outputs the time dimension characteristics of the prediction model. Based on the predicted time series data of the number of printings in the future and the historical life data of similar steel meshes, the threshold value is judged and the difference between the number of printings in which the steel mesh feature data in the predicted time series data is first lower than the threshold and the current number of printings is calculated as the predicted remaining life.

[0022] Preferably, the prediction model includes a multi-cycle detection and decomposition unit, a TimesBlock module processing unit and a fully connected unit, and the multi-cycle detection and decomposition unit includes an FFT fast Fourier transform subunit and a period rearrangement subunit, wherein:

[0023] The FFT fast Fourier transform subunit is used to perform spectrum analysis on the input data sequence and extract the first k significant frequency components P={p1, p2, ..., p k}, calculate the corresponding period length;

[0024] The period rearrangement subunit is used to reconstruct the original sequence into a two-dimensional tensor according to the detected period length The rows in the two-dimensional tensor correspond to the time points within the period, the columns correspond to the number of periods, T2 is the time step of the sequence data in the two-dimensional tensor, C2 is the feature dimension of the sequence data in the two-dimensional tensor, and p j is a certain period of data;

[0025] The TimesBlock module processing unit includes a multi-cycle feature extraction subunit, a dynamic cycle aggregation subunit and a residual connection subunit, wherein:

[0026] The multi-cycle feature extraction subunit is used to apply an Inception-style 2D convolutional network to each candidate cycle two-dimensional tensor to extract spatiotemporal features at different cycle scales in parallel. n is the nth of the k significant frequency components extracted;

[0027] The dynamic period aggregation subunit is used to fuse the period features based on the following formula (3) through learnable weights:

[0028]

[0029] Among them, F agg is the aggregate feature, α n is the spatiotemporal feature F (n) The weight of

[0030] The residual connection subunit is used to learn the difference between the input spatiotemporal feature data and the output aggregated features based on the following formula (4), and adjust the weights so as to retain the original sequence information:

[0031] Output=LayerNorm(F agg +X in ) (4)

[0032] The fully connected unit is used to extract the long-term global dependencies of the aggregated features and retain the time dimension features mapped to future time steps.

[0033] Preferably, the image data branch unit uses a network based on ConvNeXt to extract the microscopic deformation features of the openings on the stencil, including the burr index and diameter deviation; the time series data branch unit receives time series signals including the tension value, squeegee pressure, solder paste viscosity, stencil temperature, humidity, tin powder particle size, printing speed, squeegee width, and printing gap.

[0034] Preferably, the time series data normalization unit is used to perform Z-score normalization on the time series numerical data collected by the data acquisition module; the image data enhancement unit is used to enhance the image data collected by the data acquisition module by randomly rotating and adjusting the contrast.

[0035] Preferably, the trained prediction model adapts to the remaining life prediction of stencils of different specifications through transfer learning.

[0036] Preferably, the output result of the remaining life prediction module is transmitted to the automatic stencil adjustment device to directly drive the automatic stencil adjustment device to adjust the stencil.

[0037] In one embodiment, a method for predicting the remaining life and self-calibration of an SMT stencil based on dynamic data coupling is provided. The method uses the system for predicting the remaining life of an SMT stencil based on dynamic data coupling described in any one of the preceding items. The method includes:

[0038] (1) Training of the prediction model

[0039] It includes two-stage training, namely the pre-training stage and the online learning stage, where:

[0040] In the pre-training stage, samples are constructed based on the printing historical data of the stencil collected, and a basic model is obtained by training the TimesNet prediction model;

[0041] In the online learning stage, incremental training is triggered every time S1 printing data is added, and the weight parameters of the prediction model are updated through incremental training for self-calibration;

[0042] (2) Practical application of the trained prediction model

[0043] Real-time data acquisition: After each printing, the printing parameters and process environment parameters related to the stencil on the printing machine are updated, and the image data is updated every S2 printings;

[0044] Dynamic inference: The prediction model outputs the predicted value and the remaining life value for future time steps based on the real-time collected data;

[0045] Decision linkage: A visualization dashboard is generated through a display device, and an alarm is triggered when the predicted remaining life < S3% of the full cycle life.

[0046] Preferably, the S2 is 450-550; the S3 is 50-150; and the S3 is 5-15.

[0047] Preferably, the trained prediction model is adapted to the life prediction of steel meshes of different specifications through transfer learning.

[0048] Preferably, the output result of the remaining life prediction module is transmitted to the automatic mesh adjustment equipment, which directly drives the automatic mesh adjustment equipment to adjust the steel mesh.

[0049] Compared with the existing technology, the beneficial effects of the present invention are: the SMT steel mesh remaining life prediction system based on dynamic data coupling realizes the prediction of the steel mesh service life by real-time monitoring of the steel mesh working status parameters and combining deep learning algorithms. The system uses tension value as the core monitoring parameter and applies image recognition to micro damage detection to realize multi-source data fusion and break through the prediction limitations of single indicator; the prediction model training samples are constructed through the steel mesh full life cycle data, printing machine printing parameters, and process environment parameters, and then the prediction model is trained, and then the prediction model is learned online in combination with the newly added data. The remaining life of the steel mesh can be predicted in real time based on the output results of the prediction model, and a visual dashboard can be generated. If the predicted life is lower than a certain threshold, the system will issue an alarm. The system can well solve the problems of time-consuming and labor-intensive production interruptions in the existing technology, and combines the dynamic learning mechanism to improve the prediction accuracy, so that the accuracy of the remaining life prediction of the steel mesh is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0051] Figure 1 This is a general structural block diagram of an embodiment of an SMT steel mesh remaining life prediction system based on dynamic data coupling of the present invention.

[0052] Figure 2 This is a structural block diagram of the remaining life prediction module in an embodiment of the SMT steel mesh remaining life prediction system based on dynamic data coupling of the present invention.

[0053] Figure 3 This is a structural block diagram of the prediction model in an embodiment of the SMT steel mesh remaining life prediction system based on dynamic data coupling of the present invention.

[0054] Figure 4 The flowchart of an embodiment of the SMT steel mesh remaining life prediction and self-calibration method based on dynamic data coupling of the present invention. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] In one embodiment, a SMT steel mesh remaining life prediction system based on dynamic data coupling is provided. Figure 1 As shown, the SMT steel mesh remaining life prediction system based on dynamic data coupling includes a data acquisition module 100, a data processing module 200 and a remaining life prediction module 300, wherein:

[0057] The data acquisition module 100 includes a real-time monitoring data acquisition unit 101 and a historical data acquisition unit 102. The real-time monitoring data acquisition unit 101 is used to monitor and collect stencil-related data on the production line in real time, including time-series data on the stencil's tension and images showing micromorphological changes. The historical data acquisition unit 102 is used to collect historical stencil life data for reference comparison and to compile a fatigue parameter library for stencils of different materials. The real-time monitoring data acquisition unit 101 of this dynamic data-coupled SMT stencil remaining life prediction system monitors and collects stencil-related data on the production line in real time, including time-series data on the stencil's tension and images showing micromorphological changes. For the first time, tension data reflecting the tension decay curve is correlated with micromorphological changes through modeling, breaking through the limitations of single-indicator prediction.

[0058] The data processing module 200 includes a time series data normalization unit 201 and an image data enhancement unit 202. The time series data normalization unit 201 is used to normalize the time series numerical data collected by the data acquisition module, and the image data enhancement unit 202 is used to enhance the image data collected by the data acquisition module 100. In this embodiment, the time series data normalization unit 201 is used to perform Z-score normalization processing (mean = 0, standard deviation = 1) on the time series numerical data collected by the data acquisition module; the image data enhancement unit 202 is used to enhance the image data collected by the data acquisition module by randomly rotating (±5°) and adjusting the contrast (0.8-1.2 times).

[0059] The remaining life prediction module 300 is used to predict the remaining life of the stencil based on the data collected by the data acquisition module 100 and the data processed by the data processing module 200. Because the previously collected data includes the stencil's tension value and an image reflecting changes in the stencil's microscopic morphology, the remaining life prediction module 300 uses both types of data to make predictions, improving prediction accuracy.

[0060] Furthermore, the data collected by the real-time monitoring data acquisition unit 101 in the SMT stencil remaining life prediction system based on dynamic data coupling specifically include: (1.1) the tension values ​​of the four corners and the center of the stencil collected by the embedded MEMS strain gauge; (1.2) the image of the opening on the stencil captured by the industrial camera, which is used to obtain the burr index and diameter deviation of the edge of the opening; (1.3) the scraper pressure; (1.4) the viscosity of the solder paste; (1.5) the temperature of the stencil; (1.6) the humidity; (1.7) the size of the tin powder particles; (1.8) the printing speed; (1.9) the scraper width; and (1.10) the printing gap. The scraper pressure here refers to the pressure applied by the scraper on the steel mesh when scraping solder paste on it; the solder paste viscosity refers to the viscosity of the solder paste on the steel mesh; the humidity refers to the humidity in the workshop where the production line is located; the tin powder particle size refers to the particle size of the tin powder used to melt into solder paste; the printing speed refers to the number of PCB boards printed per unit time on the steel mesh during the printing process; the scraper width refers to the width of the scraper when scraping solder paste on the steel mesh; and the printing gap refers to the distance between the steel mesh and the printed PCB board.

[0061] The data collected by the historical data collection unit 102 specifically include: (2.1) the full life cycle record of the same type of steel mesh, including tension decay curve, hole deformation rate, and maintenance record; (2.2) fatigue parameter library (elastic modulus, fatigue coefficient) of steel meshes of different materials (stainless steel, nickel alloy).

[0062] Further, combined Figure 2 As shown, the algorithm architecture and network structure of the remaining life prediction module 300 in the SMT steel mesh remaining life prediction system based on dynamic data coupling include an input layer 301, a feature fusion layer 302, a prediction model 303 and an output layer 304, wherein: the input layer 301 includes a time series data branch unit 3011 for receiving time series data, and an image data branch unit 3012 for extracting image feature data, the image feature data includes the hole diameter deviation of the steel mesh, and also includes the burr index of the hole.

[0063] The image data branching unit 3012 in the SMT steel mesh remaining life prediction system based on dynamic data coupling uses a ConvNeXt-based network to extract the micro-deformation characteristics of the openings on the steel mesh, including burr index and diameter deviation; the time series data branching unit 3011 receives time series signals including tension value, scraper pressure, solder paste viscosity, steel mesh temperature, humidity, tin powder particle size, printing speed, scraper width and printing gap.

[0064] The feature fusion layer 302 includes a feature intersection unit 3021, a composite index generation unit 3022, and an input preprocessing unit 3023. The feature intersection unit 3021 is used to splice the feature data extracted from the image with other features, that is, the two features of burr index and diameter deviation calculated based on the opening image, and the tension value, scraper pressure, solder paste viscosity, steel mesh temperature, humidity, tin powder particle size, printing speed, scraper width and printing gap are put together as factors affecting the life of the steel mesh.

[0065] The composite index generating unit 3022 is used to calculate the fatigue accumulation factor and the deformation entropy value based on the following formula (1) and formula (2):

[0066]

[0067] Deformation entropy = -∑p i *ln(p i ) (2)

[0068] By dividing the actual measured hole diameter deviation range into n continuous intervals, the parameter p i Represents the number of deviation samples appearing in the i-th interval / total number of samples.

[0069] In the above formula (1), the single scraper pressure and steel mesh temperature are from the data collected by the real-time monitoring data acquisition unit 101, and the material fatigue coefficient is from the fatigue parameter library of steel meshes of different materials collected by the historical data acquisition unit 102. In the above formula (2), the parameter p i Represents the probability that the hole diameter deviation falls into the i-th discrete interval. The fatigue accumulation factor and deformation entropy value generated by the compliance index generation unit 3022 are also important factors affecting the life of the steel mesh. They are important indicators that reflect the fatigue and deformation of the steel mesh as the number of printings increases. The deformation entropy value and other characteristic data collected previously are the current characteristic data of the steel mesh. They are indicators that reflect the deformation unevenness of the current steel mesh. The fatigue accumulation factor is the cumulative parameter of the steel mesh in all historical printing operations, which is used to reflect the fatigue index of the steel mesh. Therefore, the fatigue accumulation factor and the deformation entropy value are used as two important factors and are put together with the other characteristic data above as prediction data.

[0070] The input preprocessing unit 3023 is used to receive one-dimensional time series data T1 is the time step of the one-dimensional time series data, and C1 is the feature dimension of the one-dimensional time series data. Normalization and sliding window segmentation are performed to generate fragmented data suitable for model input. For example, normalization converts different feature data into a distribution with a mean of 0 and a standard deviation of 1 to eliminate the dimensional effects between different variables and ensure that all features are on the same scale. Sliding window segmentation divides long series data into multiple shorter segments. Specifically, a fixed-size window is set and the window is gradually moved to generate multiple segments.

[0071] parameter It represents one-dimensional time series data, which is used for training prediction models. It is a T×C matrix, where T represents time (number of prints) and C represents the number of features, the number of features and the number of different types of data. In one-dimensional time series data, each row represents all the feature data of each print. The C features include all the previous feature data: burr index, diameter deviation, tension value, scraper pressure, solder paste viscosity, steel mesh temperature, humidity, tin powder particle size, printing speed, scraper width, printing gap, fatigue accumulation factor, and deformation entropy value.

[0072] The prediction model 303 is a TimesNet model. After pre-training, the prediction model 3024 performs prediction based on the fragmented data generated by the input preprocessing unit 3023. The fully connected layer of the prediction model 3024 is mapped to the future time step, and the prediction result retains the time dimension feature, which includes the predicted time series data of the number of printings in the future.

[0073] The output layer 304 outputs the time dimension characteristics of the prediction model. Based on the predicted time series data of the number of printings in the future and the historical life data of similar steel meshes, the threshold is determined and the difference between the number of printings in which the steel mesh feature data in the predicted time series data is first lower than the threshold and the current number of printings is calculated as the predicted remaining life.

[0074] The input layer 301, feature fusion layer 302, prediction model 303, output layer 304 and related modules, units, and sub-units in this article are all functional modules that can be executed by computers. The terms "module, unit, layer, sub-unit, model" are only used to distinguish different functional models.

[0075] Further, combined Figure 3 As shown, the prediction model 303 in the SMT steel mesh remaining life prediction system based on dynamic data coupling includes a multi-cycle detection and decomposition unit 3031, a TimesBlock module processing unit 3032 and a fully connected unit 3033. The multi-cycle detection and decomposition unit 3031 includes an FFT fast Fourier transform subunit 3031-a and a period rearrangement subunit 3031-b, wherein:

[0076] The FFT subunit 3031-a is used to perform spectrum analysis on the input data sequence and extract the first k significant frequency components P={p1, p2, ..., p k}, calculate the corresponding period length.

[0077] The FFT fast Fourier transform subunit 3031-a performs Fourier transform (FFT), which can identify the significant main periodic components in the time series (such as daily cycle, weekly cycle, etc.) through spectrum analysis, convert the one-dimensional time series data into frequency domain signals, and screen out the period lengths corresponding to the top k frequencies with the largest amplitude. In the prediction of steel mesh life, Fourier transform is used to identify the periodicity of data and separate long-term fatigue accumulation from single printing impact.

[0078] The period rearrangement subunit 3031-b is used to reconstruct the original sequence into a two-dimensional tensor according to the detected period length. The rows in the two-dimensional tensor correspond to the time points within the period, the columns correspond to the number of periods, T2 is the time step of the sequence data in the two-dimensional tensor, C2 is the feature dimension of the sequence data in the two-dimensional tensor, and p j For example, if the original data is T×C (e.g., 100×13), and the Fourier transform calculates a period of 5, then the data is reconstructed into 5×20×13. The purpose of this is to treat the time series data as image data, and to subsequently use convolution to extract features.

[0079] The TimesBlock module processing unit 3032 includes a multi-cycle feature extraction subunit 3032-a, a dynamic cycle aggregation subunit 3032-b, and a residual connection subunit 3033-c, wherein:

[0080] The multi-cycle feature extraction subunit 3032-a is used to apply an Inception-style 2D convolutional network to each candidate cycle two-dimensional tensor to extract spatiotemporal features at different cycle scales in parallel. n is the nth significant frequency component among the k extracted ones.

[0081] Multi-cycle feature extraction is to rearrange the cycle lengths corresponding to the first k frequencies with the largest amplitudes, and then extract features. Features can be extracted from multiple cycles separately, and the features are richer, which can be better used for life prediction.

[0082] The dynamic period aggregation subunit 3032-b is used to fuse the period features based on the following formula (3) through learnable weights:

[0083]

[0084] Among them, Fagg is the aggregate feature, α n is the spatiotemporal feature F (n) The weight of .

[0085] The role of dynamic periodic aggregation is to convert one-dimensional time series data into two-dimensional space-time tensor data (rows represent changes between cycles, and columns represent fluctuations within cycles) by automatically identifying multi-periodic features (such as daily / weekly / yearly cycles), and adaptively weighting key cycles based on spectral amplitude, thereby unifying modeling and solving the problems of periodic drift, multi-period superposition and noise interference in complex time series.

[0086] The residual connection subunit 3032-c is used to learn the difference between the input spatiotemporal feature data and the output aggregated features based on the following formula (4), and adjust the weights so as to retain the original sequence information:

[0087] Output=LayerNorm(F agg +X in ) (4)

[0088] The role of residual connections is to transmit the original input through skip connections, so that the neural network can directly learn the difference (residual) between the input and output, thereby alleviating the gradient vanishing / exploding problem, suppressing network performance degradation, accelerating deep network training, enhancing feature expression capabilities, realizing cross-layer information fusion, and improving the feature reuse capability and performance of the prediction model.

[0089] Based on the above, it can be seen that the TimesBlock module processing unit 3032 includes three core components: a multi-cycle feature extraction subunit 3032-a, a dynamic cycle aggregation subunit 3032-b and a residual connection subunit 3033-c, and the prediction model is composed of multiple stacked TimesBlock module processing units 3032.

[0090] The fully connected unit 3033 is used to extract the long-term global dependency of the aggregated features and retain the time dimension features mapped to future time steps. By retaining the time dimension features mapped to future time steps, data support is provided for predicting the remaining life of the steel mesh.

[0091] Based on the above embodiments, it can be seen that the SMT steel mesh remaining life prediction system based on dynamic data coupling realizes the prediction of the steel mesh service life by real-time monitoring of the steel mesh working status parameters and combining deep learning algorithms. The system uses tension value as the core monitoring parameter and applies image recognition to micro damage detection. The samples for prediction model training are constructed through the steel mesh full life cycle data, printing machine printing parameters, and process environment parameters. The prediction model (timesnet) is then trained and then combined with the newly added data for online learning of the prediction model. The remaining life of the steel mesh can be predicted in real time based on the output results of the prediction model, and a visual dashboard can be generated. If the predicted life is lower than a certain threshold, the system will issue an alarm. The system can well solve the problems of time-consuming, labor-intensive and production interruption in the existing technology, and combines the dynamic learning mechanism to improve the prediction accuracy, so that the accuracy of the remaining life prediction of the steel mesh is further improved.

[0092] In one embodiment, a method for predicting the remaining life of an SMT steel mesh and self-calibration based on dynamic data coupling is provided. The method adopts the SMT steel mesh remaining life prediction system based on dynamic data coupling in the previous embodiment, combined with Figure 4 As shown, the method includes:

[0093] Step S100: Training the prediction model

[0094] The training process includes two stages: pre-training stage S101 and online learning stage S102, where:

[0095] The pre-training stage S101 is to construct samples based on the collected printing history data of the steel mesh, and obtain a basic model by training the TimesNet prediction model.

[0096] In the online learning stage S102, incremental training is triggered every time 500 new printing data are added, and the weight parameters of the prediction model are updated through incremental training to perform self-calibration.

[0097] Since every 500 new printing data will trigger incremental training and update the weight parameters of the prediction model, a self-calibration effect can be achieved.

[0098] Step S200: Practical application of the trained prediction model

[0099] S201, real-time data collection: after each printing, the collected printing parameters and process environment parameters related to the stencil on the printing press are updated, and the image data is updated every 100 printings;

[0100] S202, dynamic reasoning: the prediction model outputs the predicted value and remaining life value of the future time step based on the real-time collected data;

[0101] S203, decision linkage: Generate a visual dashboard through a display device, and trigger an alarm when the predicted remaining life is less than 10% of the full cycle life.

[0102] The trained prediction model is adapted to the life prediction of steel meshes of different specifications (such as 0.12mm / 0.15mm thickness steel mesh) through transfer learning. During transfer learning, self-calibration is used to make the prediction model adapt to the prediction of steel meshes of corresponding specifications.

[0103] In addition, by transmitting the output results of the remaining life prediction module to the automatic mesh adjustment equipment, the automatic mesh adjustment equipment is directly driven to adjust the steel mesh, and a visual dashboard is generated through the display device on the automatic mesh adjustment equipment for operators to view the remaining life information of the steel mesh.

[0104] The advantages of this dynamic data-coupling-based SMT stencil remaining life prediction and self-calibration method are as follows: 1. Multi-source data fusion: Correlating tension decay curves with micromorphological changes, overcoming the limitations of single-metric prediction. 2. Adaptive learning: Using transfer learning technology, it can adapt to stencils of varying specifications (e.g., 0.12mm / 0.15mm thickness). 3. Prediction-maintenance closed loop: Output results can directly drive automated stencil adjustment equipment, reducing manual intervention.

[0105] Therefore, this dynamic data-coupled SMT stencil remaining life prediction and self-calibration method is used to predict the service life of SMT printer stencils. By collecting printing parameters, stencil parameters, environmental parameters, and historical attenuation data during the printing process, and training it using a time series model (TimesNet), the trained model can be used to predict the stencil's service life, solving the problem of constant manual visual inspection in existing technologies. This method significantly improves the overall production efficiency of SMT production lines and is suitable for SMT production environments of all sizes, with broad application prospects.

[0106] It should be noted that, in this article, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements that are inherent to such process, method, article or apparatus.

[0107] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A SMT steel mesh remaining life prediction system based on dynamic data coupling, characterized in that: include: A data acquisition module, which includes a real-time monitoring data acquisition unit and a historical data acquisition unit. The real-time monitoring data acquisition unit is used to monitor and collect data related to the stencil on the production line in real time. The data includes the tension value of the stencil collected based on time series and images that can show changes in microscopic morphology. The historical data acquisition unit is used to collect historical life data of the stencil for reference comparison, as well as to collect a fatigue parameter library of stencils of different materials. a data processing module, the data processing module including a time series data standardization unit and an image data enhancement unit, the time series data standardization unit being used to perform standardization processing on the time series numerical data collected by the data acquisition module, and the image data enhancement unit being used to perform enhancement processing on the image data collected by the data acquisition module; The remaining life prediction module is used to predict the remaining life of the steel mesh based on the data collected by the data collection module and the data processed by the data processing module.

2. The SMT steel mesh remaining life prediction system based on dynamic data coupling according to claim 1, characterized in that: The data collected by the real-time monitoring data acquisition unit specifically include: (1.1) the tension values ​​of the four corners and the center of the steel mesh collected by the embedded MEMS strain gauge; (1.2) the image of the opening on the steel mesh captured by the industrial camera, which is used to obtain the edge burr index and diameter deviation of the opening; (1.3) the scraper pressure; (1.4) the viscosity of the solder paste; (1.5) the steel mesh temperature; (1.6) the humidity; (1.7) the tin powder particle size; (1.8) the printing speed; (1.9) the scraper width; (1.10) the printing gap; The data collected by the historical data collection unit specifically include: (2.1) the full life cycle record of the same type of steel mesh, including tension decay curve, hole deformation rate, and maintenance record; (2.2) fatigue parameter library of steel meshes of different materials.

3. The SMT steel mesh remaining life prediction system based on dynamic data coupling according to claim 2 is characterized in that: The algorithm architecture and network structure of the remaining life prediction module include an input layer, a feature fusion layer, a prediction model and an output layer, wherein: the input layer includes a time series data branch unit for receiving time series data, and an image data branch unit for extracting image feature data, the image feature data including the opening diameter deviation of the steel mesh; The feature fusion layer includes a feature cross unit, a composite index generation unit and an input preprocessing unit. The feature cross unit is used to combine the feature data extracted from the image with other features. The composite index generating unit is used to calculate the fatigue accumulation factor and the deformation entropy value based on the following formula (1) and formula (2): Deformation entropy = -∑p i *ln(p i )(2) By dividing the actual measured hole diameter deviation range into n continuous intervals, the parameter p i Represents the number of deviation samples appearing in the i-th interval / total number of samples; The input preprocessing unit is used to receive one-dimensional time series data T1 is the time step of the one-dimensional time series data, C1 is the feature dimension of the one-dimensional time series data, and normalization and sliding window segmentation are performed to generate fragmented data suitable for model input; The prediction model is a TimesNet model. After pre-training, the prediction model performs prediction based on the fragmented data generated by the input pre-processing unit. The fully connected layer of the prediction model is mapped to future time steps, and the prediction result retains the time dimension feature, which includes the predicted time series data of the number of printings in the future. The output layer outputs the time - dimension features of the prediction model. Based on the predicted time - series data in a future period of printing times, combined with the historical life data of the same type of stencil, and according to the threshold judgment, the difference between the printing times when the stencil feature data in the predicted time - series data first drops below the threshold and the current printing times is calculated as the predicted remaining life.

4. The SMT steel mesh remaining life prediction system based on dynamic data coupling according to claim 3 is characterized by: The prediction model includes a multi - cycle detection and decomposition unit, a TimesBlock module processing unit, and a fully - connected unit. The multi - cycle detection and decomposition unit includes an FFT (Fast Fourier Transform) sub - unit and a period rearrangement sub - unit, where: The FFT fast Fourier transform subunit is used to perform spectrum analysis on the input data sequence and extract the first k significant frequency components P={p1, p2, ..., p k }, calculate the corresponding period length; The period rearrangement subunit is used to reconstruct the original sequence into a two-dimensional tensor according to the detected period length The rows in the two-dimensional tensor correspond to the time points within the period, the columns correspond to the number of periods, T2 is the time step of the sequence data in the two-dimensional tensor, C2 is the feature dimension of the sequence data in the two-dimensional tensor, and p j is a certain period of data; The TimesBlock module processing unit includes a multi - cycle feature extraction sub - unit, a dynamic period aggregation sub - unit, and a residual connection sub - unit, where: The multi-cycle feature extraction subunit is used to apply an Inception-style 2D convolutional network to each candidate cycle two-dimensional tensor to extract spatiotemporal features at different cycle scales in parallel. n is the nth of the k significant frequency components extracted; The dynamic period aggregation sub - unit is used to fuse each period feature based on the following formula (3) through learnable weights: Among them, F agg is the aggregate feature, α n is the spatiotemporal feature F (n) The weight of The residual connection sub - unit is used for the model to learn the difference between the input spatio - temporal feature data and the output aggregated feature based on the following formula (4), and adjust the weights to retain the original sequence information: Output=LayerNorm(F agg +X in ) (4) The fully - connected unit is used to extract the long - term global dependencies of the aggregated features and retain the time - dimension features mapped to future time steps.

5. The SMT steel mesh remaining life prediction system based on dynamic data coupling according to claim 3, characterized in that: The image data branch unit uses a network based on ConvNeXt to extract the microscopic deformation features of the openings on the stencil, including the burr index and diameter deviation; the time - series data branch unit receives time - series signals including the tension value, squeegee pressure, solder paste viscosity, stencil temperature, humidity, tin powder particle size, printing speed, squeegee width, and printing gap.

6. The SMT steel mesh remaining life prediction system based on dynamic data coupling according to claim 1, characterized in that: The time - series data normalization unit is used to perform Z - score normalization on the time - series numerical data collected by the data acquisition module; the image data enhancement unit is used to enhance the image data collected by the data acquisition module by randomly rotating and adjusting the contrast.

7. A method for predicting the remaining life of an SMT steel mesh and self-calibration based on dynamic data coupling, characterized in that: Using the SMT stencil remaining life prediction system based on dynamic data coupling according to any one of claims 1 - 5, the method includes: (1) Training of the prediction model It includes two - stage training, namely the pre - training stage and the online learning stage, where: The pre - training stage is to construct samples based on the collected printing historical data of the stencil, and obtain a basic model by training the TimesNet prediction model; The online learning stage is that every time S1 new printing data is added, it triggers incremental training, and updates the weight parameters of the prediction model through incremental training for self - calibration; (2) Practical application of the trained prediction model Real - time data acquisition: After each printing, update the printing parameters and process environment parameters related to the stencil on the printing machine, and update the image data every S2 prints; Dynamic inference: The prediction model outputs the predicted values and remaining life values for future time steps according to the real - time collected data; Decision linkage: Generate a visual dashboard through a display device, and trigger an alarm when the predicted remaining life < S3% of the full - cycle life.

8. The SMT steel mesh remaining life prediction and self-calibration method based on dynamic data coupling according to claim 7, characterized in that: S2 is 450 - 550; S3 is 50 - 150; S3 is 5 - 15.

9. The SMT steel mesh remaining life prediction and self-calibration method based on dynamic data coupling according to claim 7, characterized in that: The trained prediction model adapts to the life prediction of different - specification stencils through transfer learning.

10. The SMT steel mesh remaining life prediction and self-calibration method based on dynamic data coupling according to claim 7, characterized in that: The output result of the remaining life prediction module is transmitted to the automatic mesh adjustment equipment, which directly drives the automatic mesh adjustment equipment to adjust the steel mesh.

Images