Float glass defect predictive monitoring method and system based on joint training

Abstract

The invention provides a float glass defect predictive monitoring method and system based on joint training, and the method comprises the steps: obtaining historical multi-dimensional process data and historical product defect data in a float glass production process, the first collection frequency of the multi-dimensional process data being higher than the second collection frequency of the product defect data; based on historical multi-dimensional process data and historical product defect data, carrying out alternate iteration type joint training on the initial time sequence prediction model and the initial defect soft measurement model to obtain a final time sequence prediction model and a final defect soft measurement model; inputting multi-dimensional process data collected in real time to the final time sequence prediction model, and outputting a process data prediction sequence; and inputting the process data prediction sequence to the final defect soft measurement model, and outputting a continuous defect prediction sequence which corresponds to the process data prediction sequence point by point in time and of which the frequency is the first acquisition frequency, so that the accuracy of defect prediction is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of defect prediction technology, and in particular to a method and system for predictive monitoring of float glass defects based on joint training. Background Technology

[0002] Float glass is widely used in industries such as construction, automotive, and electronics due to its excellent surface smoothness and optical properties. The continuous production process of float glass involves multiple complex and interdependent physicochemical processes, including melting, forming, and annealing. Even minor fluctuations in process parameters at each stage, such as furnace temperature, tin bath pressure, and atmosphere flow rate, can lead to quality problems in the finished glass product through complex nonlinear transmission mechanisms, ultimately resulting in defects such as bubbles, knots, stones, or point defects. Therefore, real-time monitoring and proactive early warning of defects in the production process are crucial for improving product yield and reducing production costs.

[0003] Currently, defect detection technology for float glass mainly relies on manual visual inspection or automated optical inspection (AOI) systems at the end of the production line. However, these methods are essentially "post-production inspection" or "passive quality control," meaning that problems can only be detected when defects have already formed and the product is about to enter the cutting and packaging stage. The fundamental limitation of this model lies in its lag; by the time batch defects are detected, upstream processes may have already been operating abnormally for some time, resulting in significant resource waste and the production of substandard products.

[0004] Therefore, model prediction has emerged. The key technical challenge in building such predictive models lies in the inherent data characteristics of industrial production sites: a significant sampling frequency mismatch exists between process data and quality data. Specifically, various sensors installed on equipment can continuously collect process parameter data from the process control system at high frequencies, ranging from minutes to even seconds, forming dense time series. However, the final defect data, whether recorded manually or statistically analyzed by an AOI system, is typically hourly in nature, representing sparse, low-frequency quality labels. This difference in data frequency leads to the "data silo" phenomenon, making it difficult for traditional supervised learning models to directly establish an accurate mapping relationship between high-frequency process fluctuations and low-frequency defect results. Conventional processing methods, such as downsampling high-frequency data (e.g., hourly averages), smooth out a large amount of crucial, transient process fluctuation information, which may be important contributing factors to defect formation. Conversely, upsampling low-frequency quality labels (e.g., linear interpolation or pre-value padding) artificially creates a large amount of unreal, false supervisory signals, severely misleading the model's training process and preventing it from learning true causal relationships, resulting in poor accuracy in defect prediction.

[0005] Therefore, how to improve the accuracy of the model in predicting defects has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This invention provides a method and system for predictive monitoring of defects in float glass based on joint training, in order to solve the problem of poor accuracy in defect prediction by existing models.

[0007] In a first aspect, the present invention provides a method for predictive monitoring of defects in float glass based on joint training, comprising: Historical multidimensional process data and historical product defect data in the float glass production process are acquired, wherein the first acquisition frequency of the multidimensional process data is higher than the second acquisition frequency of the product defect data; Based on the historical multidimensional process data and historical product defect data, the initial time series prediction model and the initial defect soft measurement model are jointly trained alternately and iteratively to obtain the final time series prediction model and defect soft measurement model; wherein, the alternate iterative joint training includes alternately executing the first stage of initial defect soft measurement model optimization and high-quality pseudo-label generation and the second stage of time series prediction model optimization under the condition that the parameters of the initial defect soft measurement model are locked. Input real-time acquired multidimensional process data into the final time series prediction model, and output the predicted process data sequence; The process data prediction sequence is input into the final defect soft measurement model, and a continuous defect prediction sequence is output, which corresponds to the process data prediction sequence point by point in time and has a frequency of the first acquisition frequency.

[0008] According to the present invention, a method for predictive monitoring of float glass defects based on joint training is provided, wherein the method comprises alternating iterative joint training of an initial time series prediction model and an initial defect soft measurement model, including: The parameters of the initial time series prediction model are locked. The initial time series prediction model is used to generate a future process data prediction sequence. The future process data prediction sequence is input into the initial defect soft measurement model to generate a high-frequency defect prediction sequence. Based on the high-frequency defect prediction sequence, the historical product defect data, and the preset smoothness loss, the parameters of the initial defect soft measurement model are updated. The high-frequency defect prediction sequence output by the updated initial defect soft measurement model is used as a high-quality pseudo-label to complete the first stage of optimization. The parameters of the updated initial defect soft measurement model are locked, and the parameters of the initial time series prediction model are updated using the high-quality pseudo-labels as supervision signals in a multi-task learning mode to complete the second stage of optimization. The first stage optimization and the second stage optimization are performed alternately until the final time series prediction model and defect soft measurement model are obtained.

[0009] According to the present invention, a method for predictive monitoring of float glass defects based on joint training is provided, wherein updating the parameters of the initial defect soft measurement model based on the high-frequency defect prediction sequence, the historical product defect data, and a preset smoothness loss includes: Based on the high-frequency defect prediction sequence, the historical product defect data, and a preset smoothness loss, a first composite loss is calculated. The first composite loss is composed of a weighted sum of the prediction accuracy loss and the smoothness loss. The prediction accuracy loss is used to measure the difference between the predicted value and the actual value of the high-frequency defect prediction sequence at the time point corresponding to the historical product defect data. The parameters of the initial defect soft measurement model are updated based on the gradient of the first composite loss.

[0010] According to the present invention, a method for predictive monitoring of float glass defects based on joint training is provided, wherein the method updates the parameters of the initial time series prediction model using a multi-task learning mode and high-quality pseudo-labels as supervision signals includes: Configure the initial time series prediction model, and output predictions for future process data and predictions for high-quality pseudo-labels; A second composite loss is calculated by combining the loss from the prediction of the future process data and the loss from the prediction of high-quality pseudo-labels. The parameters of the initial time series prediction model are updated based on the gradient of the second multi-composite loss.

[0011] According to the present invention, a method for predictive monitoring of float glass defects based on joint training is provided, wherein the second multi-composite loss is an adaptive weighted loss based on uncertainty, which is used to dynamically adjust the weight of each task loss according to the uncertainty of each task during the training process.

[0012] According to the present invention, a method for predictive monitoring of float glass defects based on joint training is provided, wherein acquiring historical multidimensional process data and historical product defect data in the float glass production process includes: Historical, multi-dimensional process data are collected from the distributed control system at the highest acquisition frequency. Historical product defect data are collected from the automated optical inspection system at a second acquisition frequency.

[0013] According to the present invention, a method for predictive monitoring of float glass defects based on joint training is provided, wherein the multidimensional process data includes at least one of the following: kiln pressure, liquid level, arch temperature, natural gas flow rate, natural gas pressure, combustion air flow rate, combustion air pressure, flow channel temperature, protective atmosphere flow rate, and tank pressure. The product defect data includes at least one of the following: bubble defects, stone defects, sand particle defects, and point defects.

[0014] According to the present invention, a method for predictive monitoring of float glass defects based on joint training is provided. The initial time series prediction model adopts a composite architecture that includes a state-space model and hybrid experts. The composite architecture utilizes the ability of the state-space model to handle long-sequence dependencies and dynamically activates the expert network through a gating network to flexibly adapt to the dynamic changes of the process under different working conditions. The defect soft measurement model adopts an architecture that integrates multi-scale attention and KAN network. The architecture captures both long-term process trends and short-term drastic fluctuations that affect defect formation through the multi-scale attention mechanism.

[0015] The method for predictive monitoring of float glass defects based on joint training provided by the present invention further includes: When the number of defects represented by the continuous defect prediction sequence exceeds the safety threshold, an alarm is issued and a warning message is pushed to the target terminal.

[0016] Secondly, the present invention provides a predictive monitoring system for float glass defects based on joint training, comprising: The acquisition module is used to acquire historical multidimensional process data and historical product defect data in the float glass production process, wherein the first acquisition frequency of the multidimensional process data is higher than the second acquisition frequency of the product defect data; The training module is used to perform alternating iterative joint training on the initial time series prediction model and the initial defect soft measurement model based on the historical multidimensional process data and historical product defect data, so as to obtain the final time series prediction model and defect soft measurement model; wherein, the alternating iterative joint training includes alternating execution of the first stage of initial defect soft measurement model optimization and high-quality pseudo-label generation and the second stage of time series prediction model optimization under the condition that the parameters of the initial defect soft measurement model are locked. The prediction module is used to input real-time acquired multidimensional process data into the final time series prediction model and output a process data prediction sequence; input the process data prediction sequence into the final defect soft measurement model and output a continuous defect prediction sequence that corresponds to the process data prediction sequence point by point in time and has a frequency of the first acquisition frequency.

[0017] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the predictive monitoring method for float glass defects based on joint training as described above.

[0018] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the float glass defect predictive monitoring method based on joint training as described above.

[0019] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the predictive monitoring method for float glass defects based on joint training as described above.

[0020] Beneficial effects: It has realized the transformation from "post-event detection" to "pre-event early warning" mode, and can predict the possible quality defects, their types and quantities, tens of minutes or even an hour in advance. This provides process engineers with valuable lead time to diagnose problems and adjust operating parameters, and realizes a leap from passive quality control to proactive quality assurance.

[0021] It has overcome the technical challenge of multi-rate data fusion in the industrial field. By creating high-quality pseudo-labels as an intermediate bridge and adopting an iterative optimization training mechanism, it has cleverly solved the industry problem of difficulty in effectively modeling minute-level process data and hour-level quality data due to frequency mismatch, and fully released the potential value contained in high-frequency process data that was previously ignored or discarded.

[0022] This significantly improves the accuracy and interpretability of the prediction model. The joint training framework forces the model to learn the deep causal relationship between process dynamics and defect formation, rather than simple surface correlations. Co-training enables the time series prediction model to focus more on extracting key process features, and the defect soft measurement model can map based on more accurate future states. As a result, the overall prediction accuracy and robustness of the system far exceed those of simple cascade models. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1This is a flowchart illustrating the predictive monitoring method for float glass defects based on joint training provided in this embodiment. Figure 2 This is a schematic diagram illustrating the principle of the joint training method provided in this embodiment; Figure 3 This is a schematic diagram of the structure of the float glass defect prediction monitoring system based on joint training provided in this embodiment; Figure 4 This is a schematic diagram of the structure of the electronic device provided in this embodiment. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] Figure 1 This is a flowchart illustrating the predictive monitoring method for float glass defects based on joint training provided in this embodiment.

[0027] like Figure 1 As shown in the figure, the present invention provides a method for predictive monitoring of float glass defects based on joint training, which mainly includes the following steps: 101. Obtain historical multidimensional process data and historical product defect data during the float glass production process, wherein the first collection frequency of multidimensional process data is higher than the second collection frequency of product defect data.

[0028] Specifically, multi-dimensional process data, i.e. historical multi-dimensional process data, is collected from the distributed control system (DCS) of the float glass production line. Since the DCS is the core system for unified monitoring of process parameters in the industrial field, the acquisition frequency is set to the first acquisition frequency (high frequency, such as 1-5 seconds / time). The parameters to be collected include at least: furnace pressure (reflecting the pressure stability inside the furnace), liquid level (controlling the glass melt volume), and arch temperature (affecting the glass melting quality) in the melting process; natural gas flow rate / pressure in the combustion system (controlling heating intensity) and combustion air flow rate / pressure (ensuring complete combustion); and flow channel temperature (controlling the glass melt flow state), protective atmosphere flow rate (preventing tin bath oxidation), and bath pressure (stabilizing the forming environment) in the forming process.

[0029] Product defect data, i.e., historical product defect data, is collected from the automated optical inspection (AOI) system or manual inspection and recording system at the end of the production line. The AOI system can automatically scan for defects on the glass surface, while manual recording is suitable for supplementing minor defects that the AOI may miss. The collection frequency is set to the second collection frequency (low frequency, such as 1-2 hours / time), and the data format is "defect type + statistical value", such as "bubble defect: 5 per square meter (level 2)" and "stone defect: 2 per square meter (level 1)".

[0030] Among them, historical multi-dimensional process data and historical product defect data refer to multi-dimensional process data and product defect data within a preset time period in the past.

[0031] Preprocess historical multi-dimensional process data and historical product defect data. For example, if the DCS collects process data every 2 minutes and the AOI collects defect data every 1 hour, the timestamp of the AOI is used as the basis to bind the 30 process data points in the corresponding hour with the defect data point to ensure that the "process parameters at a certain time" and the "defect results at that time" can correspond.

[0032] The units of different parameters differ too much. For example, the unit for kiln pressure is Pa, with a value range of 5-6; the unit for natural gas flow rate is m³ / h, with a value range of 100-500. Directly inputting these into the model will cause the model to be inaccurate. By normalizing all process data to the [0,1] interval, the influence of units can be eliminated, allowing the model to learn the role of each parameter fairly.

[0033] Sensor malfunctions are inevitable in production, leading to missing process data for a few minutes. If the loss is small (e.g., 1-2 points), linear interpolation is used. For example, if the data for the 10th minute is 2.3 and for the 12th minute it is 2.7, then the 11th minute would be filled with 2.5. If the loss is large (e.g., more than 5 consecutive points), LSTM interpolation is used. A small LSTM model is trained using surrounding normal process data to predict the missing values, which better reflects the changing trends of process parameters.

[0034] The processed data will form a paired dataset of "historical process data sequence + corresponding defect label", which not only preserves the transient fluctuations in high-frequency process data, but also solves the problems of inconsistent data format and omissions, thereby ensuring the accuracy of model training.

[0035] 102. Based on historical multidimensional process data and historical product defect data, the initial time series prediction model and the initial defect soft measurement model are jointly trained alternately and iteratively to obtain the final time series prediction model and defect soft measurement model. The alternate iterative joint training includes alternately executing the first stage of initial defect soft measurement model optimization and high-quality pseudo-label generation, and the second stage of time series prediction model optimization under the condition that the parameters of the initial defect soft measurement model are locked.

[0036] The initial time series prediction model's task is to predict future process data based on past process data. It also needs to learn to predict defect pseudo-labels in the second stage of joint training. Therefore, the architecture must be able to handle long sequences and adapt to different operating conditions. The chosen architecture is a hybrid architecture of a state-space model (such as Mamba) and a hybrid expert (MoE), with details as follows: The state-space model was chosen because it processes long sequences much faster than traditional LSTM and can remember process changes over a longer period. For example, in the process of float glass melting, the change in the arch temperature one hour ago may affect the current quality of the molten glass. Mamba can efficiently capture this long-term dependency. The length of the input sequence of the model is set to "process data of the past 2 hours" (e.g., 1 point every 2 minutes, for a total of 60 points), and the length of the output sequence is set to "process data of the next 1 hour" (30 points), which just covers the critical time window for defect formation.

[0037] The number of expert networks is determined by the number of process variables. For example, if there are 10 process variables, 12 expert networks would be set up. Each expert network focuses on learning the changing patterns of 1-2 variables. For instance, expert 1 learns about kiln pressure and liquid level, while expert 2 learns about natural gas flow and pressure. A gating network is then used to "select experts." For example, when kiln pressure fluctuates greatly, the gating network will activate more experts 1 and 3, which are related to kiln pressure; when combustion air parameters change, it will activate more experts 4 and 5. In this way, the model can flexibly adapt to different operating conditions.

[0038] The task of the soft measurement model for defects is to "transform future process data into high-frequency defect predictions," and its core is to solve the nonlinear mapping problem from high-dimensional process data to low-frequency defect data. Architecture selection: A fusion architecture of multi-scale attention and Kolmogorov-Arnold network (KAN) is used, details of which are as follows: Defect formation is related to both short-term drastic fluctuations (such as a sudden drop of 5°C in flow channel temperature within 5 minutes) and long-term trends (such as a persistently low protective atmosphere flow rate within 1 hour). The attention scale is divided into two layers: a short-term scale (1-5 minutes) specifically for capturing transient changes in parameters; and a long-term scale (30-60 minutes) specifically for capturing slow parameter drift. During training, the defect soft-sensor model automatically weights the scales with greater impact. For example, since bubble defects are often related to short-term temperature fluctuations, the defect soft-sensor model will give higher attention weights to the short-term scale; similarly, since stone defects are often related to long-term kiln pressure instability, the long-term scale will receive higher weights.

[0039] Traditional neural networks, using a fixed ReLU activation function, struggle to fit the strong nonlinear relationship in float glass production where small parameter changes lead to a significant increase in defects. The KAN network, however, uses a learnable B-spline function as its activation function. For instance, when natural gas pressure is between 0.8 and 1.0 MPa, the number of defects changes gradually; above 1.0 MPa, the number of defects surges. KAN can accurately fit this inflection point relationship, with a prediction error more than 30% lower than traditional networks. By setting three hidden layers in KAN and twice the number of nodes per layer (e.g., 10 process variables, 20 nodes per layer), complex mappings can be effectively modeled.

[0040] After the two initial models are built, they are pre-trained independently. The initial time series prediction model is trained using historical process data, enabling it to make preliminary predictions about future process data. The initial defect soft measurement model is trained using high-frequency process data and corresponding low-frequency defect labels, enabling it to initially map process data to defects. The purpose of pre-training is to lay the foundation for subsequent joint training and avoid the model parameters being too chaotic at the beginning, which could lead to training non-convergence.

[0041] After the two initial models have been trained independently, joint training is required. Through a collaborative optimization training framework, the two deep learning models, which focus on process dynamic learning and defect mapping respectively, guide each other and evolve together. The association is completed through an alternating, iterative joint training strategy, which is divided into two stages.

[0042] Phase 1: Optimize the defect soft measurement module and generate high-quality pseudo-labels. The goal of this phase is to make the defect soft measurement module more accurate, and at the same time generate high-frequency monitoring signals (pseudo-labels) that can guide the time series prediction module. The steps are as follows: Step 1: Lock down the parameters of the time series prediction module and use its current capabilities to predict future process data to avoid parameter changes affecting the optimization of subsequent soft measurement modules.

[0043] Step 2: Generate a future process data prediction sequence. Take the pre-trained time series prediction module, input historical multidimensional process data, and let it output a high-frequency process data prediction sequence for the next hour (e.g., 1 point every 2 minutes, for a total of 30 points).

[0044] Step 3: Generate an initial high-frequency defect prediction sequence. Input the above process data prediction sequence into the initial defect soft measurement module to obtain a high-frequency defect prediction sequence of 30 points (corresponding to the process data point by point).

[0045] Step 4: Calculate the first composite loss and optimize the soft measurement module. The first composite loss consists of two weighted parts: The prediction accuracy loss is calculated using mean squared error (MSE). It compares the difference between the predicted high-frequency defect sequence and the actual low-frequency defect labels at corresponding time points. For example, if the actual label is 10 bubbles per hour, then the average of 30 points in the high-frequency sequence should be close to 10. The sum of squared errors between each point and the actual label is the MSE. This part is weighted at 0.7 to ensure the model does not deviate from the actual defect patterns.

[0046] The smoothness loss is calculated using the sum of the squares of the differences between adjacent points. For example, if the first point predicts 5 bubbles and the second point suddenly predicts 15, the squared difference is 100. This abrupt change is unrealistic, so a penalty is applied. This part of the weight is set to 0.3 to ensure that the prediction sequence is smooth and continuous.

[0047] Using the Adam optimizer, based on the gradient of the first composite loss, only the parameters of the defect soft measurement module are updated. The learning rate is set to 1e-4 for each iteration, and the iteration is repeated m times. The stopping condition for m is that the decrease of the first composite loss is less than 1e-5 for three consecutive times, indicating that the module performance has stabilized and no further iteration is needed.

[0048] The first composite loss takes into account both the accuracy of the model's predictions at sparse real defect data time points (e.g., mean squared error loss) and the temporal continuity and smoothness of the entire high-frequency prediction sequence itself (e.g., penalty term for the rate of change of adjacent points).

[0049] Step 5: Generate high-quality pseudo-labels. Save the high-frequency defect prediction sequence output by the optimized soft measurement module; this is the high-quality pseudo-label. It has both high-frequency characteristics (1 point every 2 minutes) and closely matches the actual defect patterns, essentially building a bridge between high-frequency process data and low-frequency defect labels.

[0050] Through the first stage of optimization, the prediction error of the defect soft measurement model can be effectively reduced, and the generated pseudo-labels can provide dense supervision signals, solving the pain point of traditional methods that do not have high-frequency defect labels.

[0051] Phase Two: Optimize the time series forecasting module so that it can not only predict process data but also monitor changes in defect-related parameters. The steps are as follows: Step 1: Lock down the parameters of the defect soft measurement module and use the optimized data from the first stage to measure the performance of the time series prediction module.

[0052] Step 2: Configure the multi-task learning mode so that the time series prediction model can do two things simultaneously: Task 1: Predict future process data, as before, output a 30-point process data sequence. Task 2: Predict high-quality pseudo-labels, outputting a 30-point defect prediction sequence, corresponding to the pseudo-labels.

[0053] Step 3: Calculate the second composite loss to optimize the time series model. The second composite loss uses an adaptive weighted loss based on uncertainty, which is more flexible than fixed weights. Task 1 uses MSE to calculate the prediction error of the process data, and Task 2 uses MSE to calculate the prediction error of the pseudo-labels. Then calculate the homoscedastic uncertainty for each task: lower the weight of tasks with high uncertainty, and increase the weight of tasks with low uncertainty.

[0054] The ingenious aspect of using a second composite loss function to drive model optimization lies in the fact that it forces time series prediction models not only to learn the evolution of the process data itself, but also to focus on and learn the specific process dynamic features that have a strong causal relationship with the formation of the final defect. Only in this way can high-quality defect pseudo-labels be accurately fitted.

[0055] Step 4: Optimization operation: Using the Adam optimizer again, based on the gradient of the second composite loss, only update the parameters of the time series prediction model, set the learning rate to 1e-4, and iterate repeatedly n times. The stopping condition for n is that the pseudo-label prediction error on the validation set remains unchanged for 5 consecutive times, indicating that the model has learned to pay attention to defect-related features.

[0056] Through the second stage of optimization, the time series prediction model can not only predict process data more accurately, but also actively capture defect-related features. Subsequent predictions will focus on changes in defect-related features, providing more valuable process data for the defect soft measurement model.

[0057] The process is iterated until the model converges. The first and second stages are treated as one iteration cycle, and this cycle is repeated. For example, after the first cycle, the defect soft measurement model becomes more accurate, and the pseudo-label quality is higher. In the second cycle, the time series model is optimized with higher-quality pseudo-labels to output more accurate process data. In the third cycle, the soft measurement model is optimized again with more accurate process data, until the mean absolute error (MAE) of defect prediction on the validation set is less than 2 per hour for three consecutive cycles. This indicates that the performance of the entire model combination has stabilized, at which point the iteration stops, and the final time series prediction model and defect soft measurement model are obtained.

[0058] By using a collaborative training approach that guides each other, the overall defect prediction error of the model can be effectively reduced compared to traditional cascaded models, thus solving the core challenge of multi-rate data fusion.

[0059] 103. Input the real-time collected multidimensional process data into the final time series prediction model, and output the predicted sequence of process data.

[0060] Once the final model is obtained, real-time online prediction can be performed. The current multi-dimensional process data (kiln pressure, liquid level, temperature, etc.) is collected every 2 minutes. After preprocessing (normalization, missing data filling), the data is input into the final time series prediction model in the form of a sliding window. For example, if the current time is 10:00, the input is the process data (60 points) from 8:00 to 10:00.

[0061] The time series prediction model outputs a predicted sequence of process data from 10:00 to 11:00 (30 points, one every 2 minutes), such as predicting kiln pressure of 5.25 Pa at 10:02 and liquid level of 5.8 m at 10:04.

[0062] 104. Input the process data prediction sequence into the final defect soft measurement model, and output a continuous defect prediction sequence that corresponds point-by-point in time to the process data prediction sequence and has a frequency of the first acquisition frequency.

[0063] Input the above process data prediction sequence into the final defect soft measurement module, and output a continuous defect prediction sequence from 10:00 to 11:00. For example, 3 bubbles are predicted at 10:02, 4 bubbles at 10:04, and 6 bubbles at 11:00. The average value of the whole sequence is 5 bubbles / hour, which is right at the edge of the bubble threshold (5 bubbles) for high-end glass.

[0064] After the model predicts the defect risk, it needs to accurately trigger alarms to provide useful information: Step 1: Set dynamic safety thresholds. These thresholds should be dynamically adjusted based on production conditions: for example, during daytime production of high-end glass (with high requirements), set the bubble threshold to 5 bubbles / square meter; at night, during production of ordinary glass (with lower requirements), set the threshold to 15 bubbles / square meter. The thresholds can be set by process engineers based on historical yield data, or the system can automatically calculate them based on normal production data from the past month.

[0065] Step 2: Triggering Alarms and Information Push. When the predicted value of a certain type of defect in the continuous defect prediction sequence output by the final defect soft measurement model exceeds the corresponding threshold, the system will do three things: A red warning window will pop up on the display screen in the central control room, and a buzzer will sound to alert the on-duty personnel; a push notification will be sent to the process engineer via WeChat / SMS, including the defect type (e.g., bubbles), the predicted occurrence time (e.g., in the next 30 minutes), related process parameters (e.g., the current kiln pressure is 6.3 Pa, exceeding the normal range by 0.3 Pa), and suggested adjustment measures (e.g., reducing natural gas flow by 5%). The alarm time, defect type, and processing result will be stored in the database for subsequent review and optimization of thresholds and models.

[0066] The entire process of online forecasting, from data collection to forecast output, can track production status in real time, provide early warnings of defect risks, and allow staff to shift from waiting for defects to occur before taking action to proactively adjust and prevent defects. This can improve product yield and effectively reduce defective product losses.

[0067] Figure 2 This is a schematic diagram illustrating the principle of the joint training method provided in this embodiment.

[0068] like Figure 2 As shown, the core is to achieve a coordinated upgrade of the time series prediction model and the defect soft measurement model through two-stage optimization. The specific process is as follows: Historical data input: Historical data containing multidimensional process data and real defect labels Y are used as input to provide basic samples for model training.

[0069] Step 1: Time Series Model Predictor variables .

[0070] Model (Even in architectures containing state space and hybrid experts) it receives historical process data and outputs a predicted sequence of future process data. (High frequency, multi-dimensional). It can capture the dynamic evolution trend of process parameters and provide input for defect prediction.

[0071] Phase 1 optimization C1: Optimizing the defect soft measurement model And generate pseudo tags.

[0072] Step 2: Model Generate pseudo tags ,Model (e.g., an architecture that integrates multi-scale attention with KAN) receives Output high-frequency defect prediction sequence (i.e., pseudo-labels).

[0073] Step 3: Update the model (m iterations) combining the real defect label Y with the pseudo label Calculate the composite loss (prediction accuracy loss + smoothness loss), and iteratively update the model based on the loss gradient. The parameters are adjusted until the model converges.

[0074] Two-stage optimization C2: Optimizing the time series model Based on model Multi-task prediction output [ , ],Model Using historical data as input, it simultaneously performs two tasks: predicting new future process data. Predicting pseudo-labels (and Alignment).

[0075] Step 5: Update the model (n iterations) Calculate the adaptive weighted loss based on uncertainty (the loss between balancing process data prediction and pseudo-label prediction), and iteratively update the model based on the loss gradient. The parameters are adjusted until the model converges.

[0076] Repeat the above iterative process until the defect prediction accuracy of the model combination converges. If the validation set error no longer decreases significantly, the best-performing joint model is finally obtained.

[0077] It is also important to note that the specific figures listed in this embodiment are merely illustrative and do not constitute any limitation on the solution. The figures in actual applications may be different values.

[0078] Figure 3 This is a schematic diagram of the structure of the float glass defect prediction monitoring system based on joint training provided in this embodiment.

[0079] like Figure 3 As shown in the figure, this embodiment provides a predictive monitoring system for float glass defects based on joint training, comprising: The acquisition module 301 is used to acquire historical multidimensional process data and historical product defect data in the float glass production process, wherein the first acquisition frequency of the multidimensional process data is higher than the second acquisition frequency of the product defect data. Training module 302 is used to perform alternating iterative joint training on the initial time series prediction model and the initial defect soft measurement model based on historical multidimensional process data and historical product defect data, so as to obtain the final time series prediction model and defect soft measurement model. The alternating iterative joint training includes alternating execution of the first stage of initial defect soft measurement model optimization and high-quality pseudo-label generation, and the second stage of time series prediction model optimization under the condition that the parameters of the initial defect soft measurement model are locked. The prediction module 303 is used to input real-time acquired multidimensional process data into the final time series prediction model and output the process data prediction sequence; input the process data prediction sequence into the final defect soft measurement model and output a continuous defect prediction sequence that corresponds to the process data prediction sequence point by point in time and has a frequency of the first acquisition frequency.

[0080] Figure 4 This is a schematic diagram of the structure of the electronic device provided in this embodiment.

[0081] like Figure 4 As shown, the electronic device may include a processor 401, a communications interface 402, a memory 403, and a communication bus 404. The processor 401, communications interface 402, and memory 403 communicate with each other via the communication bus 404. The processor 401 can call logic instructions from the memory 403 to execute a jointly trained predictive monitoring method for float glass defects.

[0082] Furthermore, the logical instructions in the aforementioned memory 403 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), or optical discs.

[0083] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the float glass defect prediction monitoring method based on joint training provided by the above methods.

[0084] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the float glass defect predictive monitoring method based on joint training provided by the methods described above.

[0085] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0086] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A predictive monitoring method for defects in float glass based on joint training, characterized in that, include: Historical multidimensional process data and historical product defect data in the float glass production process are acquired, wherein the first acquisition frequency of the multidimensional process data is higher than the second acquisition frequency of the product defect data; Based on the historical multidimensional process data and historical product defect data, the initial time series prediction model and the initial defect soft measurement model are jointly trained alternately and iteratively to obtain the final time series prediction model and defect soft measurement model; wherein, the alternate iterative joint training includes alternately executing the first stage of initial defect soft measurement model optimization and high-quality pseudo-label generation and the second stage of time series prediction model optimization under the condition that the parameters of the initial defect soft measurement model are locked. Input real-time acquired multidimensional process data into the final time series prediction model, and output the predicted process data sequence; The process data prediction sequence is input into the final defect soft measurement model, and a continuous defect prediction sequence is output, which corresponds to the process data prediction sequence point by point in time and has a frequency of the first acquisition frequency.

2. The method for predictive monitoring of float glass defects based on joint training according to claim 1, characterized in that, The alternating iterative joint training of the initial time series prediction model and the initial defect soft measurement model includes: The parameters of the initial time series prediction model are locked. The initial time series prediction model is used to generate a future process data prediction sequence. The future process data prediction sequence is input into the initial defect soft measurement model to generate a high-frequency defect prediction sequence. Based on the high-frequency defect prediction sequence, the historical product defect data, and the preset smoothness loss, the parameters of the initial defect soft measurement model are updated. The high-frequency defect prediction sequence output by the updated initial defect soft measurement model is used as a high-quality pseudo-label to complete the first stage of optimization. The parameters of the updated initial defect soft measurement model are locked, and the parameters of the initial time series prediction model are updated using the high-quality pseudo-labels as supervision signals in a multi-task learning mode to complete the second stage of optimization. The first stage optimization and the second stage optimization are performed alternately until the final time series prediction model and defect soft measurement model are obtained.

3. The method for predictive monitoring of float glass defects based on joint training according to claim 2, characterized in that, The step of updating the parameters of the initial defect soft measurement model based on the high-frequency defect prediction sequence, the historical product defect data, and a preset smoothness loss includes: Based on the high-frequency defect prediction sequence, the historical product defect data, and a preset smoothness loss, a first composite loss is calculated. The first composite loss is composed of a weighted sum of the prediction accuracy loss and the smoothness loss. The prediction accuracy loss is used to measure the difference between the predicted value and the actual value of the high-frequency defect prediction sequence at the time point corresponding to the historical product defect data. The parameters of the initial defect soft measurement model are updated based on the gradient of the first composite loss.

4. The method for predictive monitoring of float glass defects based on joint training according to claim 2, characterized in that, The step of updating the parameters of the initial time series prediction model using a multi-task learning mode and the high-quality pseudo-labels as supervision signals includes: Configure the initial time series prediction model, and output predictions for future process data and predictions for high-quality pseudo-labels; A second composite loss is calculated by combining the loss from the prediction of the future process data and the loss from the prediction of high-quality pseudo-labels. The parameters of the initial time series prediction model are updated based on the gradient of the second multi-composite loss.

5. The method for predictive monitoring of float glass defects based on joint training according to claim 4, characterized in that, The second multi-composite loss is an adaptive weighted loss based on uncertainty, which is used to dynamically adjust the weights of the loss for each task according to the uncertainty of each task during the training process.

6. The method for predictive monitoring of float glass defects based on joint training according to claim 1, characterized in that, The acquisition of historical multidimensional process data and historical product defect data in the float glass production process includes: Historical, multi-dimensional process data are collected from the distributed control system at the highest acquisition frequency. Historical product defect data are collected from the automated optical inspection system at a second acquisition frequency.

7. The method for predictive monitoring of float glass defects based on joint training according to any one of claims 1-6, characterized in that, The multidimensional process data includes at least one of the following: kiln pressure, liquid level, arch temperature, natural gas flow rate, natural gas pressure, combustion air flow rate, combustion air pressure, flow channel temperature, protective atmosphere flow rate, and tank pressure. The product defect data includes at least one of the following: bubble defects, stone defects, sand particle defects, and point defects.

8. The method for predictive monitoring of float glass defects based on joint training according to any one of claims 1-6, characterized in that, The initial time series prediction model adopts a composite architecture that includes a state-space model and hybrid experts. The composite architecture utilizes the ability of the state-space model to handle long-sequence dependencies and dynamically activates the expert network through a gating network to flexibly adapt to the dynamic changes of the process under different operating conditions. The defect soft measurement model adopts an architecture that integrates multi-scale attention and KAN network. The architecture captures both long-term process trends and short-term drastic fluctuations that affect defect formation through the multi-scale attention mechanism.

9. The method for predictive monitoring of float glass defects based on joint training according to any one of claims 1-6, characterized in that, Also includes: When the number of defects represented by the continuous defect prediction sequence exceeds the safety threshold, an alarm is issued and a warning message is pushed to the target terminal.

10. A predictive monitoring system for float glass defects based on joint training, characterized in that, include: The acquisition module is used to acquire historical multidimensional process data and historical product defect data in the float glass production process, wherein the first acquisition frequency of the multidimensional process data is higher than the second acquisition frequency of the product defect data; The training module is used to perform alternating iterative joint training on the initial time series prediction model and the initial defect soft measurement model based on the historical multidimensional process data and historical product defect data, so as to obtain the final time series prediction model and defect soft measurement model; wherein, the alternating iterative joint training includes alternating execution of the first stage of initial defect soft measurement model optimization and high-quality pseudo-label generation and the second stage of time series prediction model optimization under the condition that the parameters of the initial defect soft measurement model are locked. The prediction module is used to input real-time acquired multidimensional process data into the final time series prediction model and output a process data prediction sequence; input the process data prediction sequence into the final defect soft measurement model and output a continuous defect prediction sequence that corresponds to the process data prediction sequence point by point in time and has a frequency of the first acquisition frequency.

Images