Method, medium and system for estimating filter screen replacement time in ABS (Acrylonitrile Butadiene Styrene) production process

By collecting and processing filter monitoring data during the ABS production process, and establishing a neural network model embedded in the physical model, an accurate online prediction of the degree of filter clogging is achieved, and the problem of difficult to accurately predict filter clogging in the prior art is solved. Improve production stability and resource utilization efficiency.

CN119066369AInactive Publication Date: 2024-12-03QING DAO XIN QING LU CAI SE GANG BAN YOU XIAN GONG SI
View PDF 13 Cites 0 Cited by

Patent Information

Application Number
CN202411554028.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-04
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The filter blockage in the existing ABS production process is difficult to accurately predict, resulting in problems of production stability and resource waste.

Method used

By collecting historical parameter data of multiple filter monitoring points, preprocessing and normalizing, establishing a neural network model embedded in the physical model, using machine learning to predict the degree of filter clogging, and setting a threshold to determine the replacement time.

Benefits of technology

Improve the accuracy and robustness of filter clogging prediction, reduce resource waste, and ensure production stability and product quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066369A_ABST
    Figure CN119066369A_ABST
Patent Text Reader

Abstract

The invention provides an ABS production process filter screen replacement time estimation method, medium and system, and belongs to the technical field of ABS production, and the method comprises the steps: collecting historical parameter data of a plurality of filter screen monitoring points on a plurality of ABS production lines; preprocessing the collected historical parameter data to obtain standardized monitoring data; for each monitoring point, counting the change trend of historical parameter data of the monitoring point, and marking a known filter screen replacement moment as historical mark data; establishing and training a filter screen blockage machine learning model for each monitoring point based on the historical mark data; performing online prediction on the real-time data of the current filter screen by using the filter screen blockage machine learning model, and estimating the current filter screen blockage degree; and setting a threshold value of the blockage degree of the filter screen, and when the prediction result of any monitoring point exceeds the threshold value, judging that the filter screen corresponding to the monitoring point needs to be replaced. The technical problem that limitation exists when filter screen replacement time is judged according to experience of workers in the prior art is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of ABS production, and in particular, relates to a method, medium and system for estimating the time of filter replacement in an ABS production process. Background Art

[0002] ABS (Acrylonitrile Butadiene Styrene) resin is an engineering plastic widely used in the fields of automobiles, electronic appliances, daily necessities, etc. It is famous for its excellent mechanical properties, chemical resistance and processing properties. In the ABS production process, extrusion, molding and other process links need to filter the melt through the filter to remove impurities and improve product quality. However, the filter will gradually become blocked during long-term use, resulting in increased pressure and decreased flow, which seriously affects the stability of production and product performance. Therefore, how to accurately predict the blockage of the filter and replace the filter in time has always been one of the key and difficult problems in ABS production management. At present, the replacement of the filter on the ABS production line mainly relies on experience judgment. The production management personnel roughly estimate the blockage degree of the filter by observing the changes in process parameters such as pressure and temperature, and arrange the filter replacement time according to the production plan. This method has certain limitations: on the one hand, it is difficult for a single parameter to accurately reflect the overall blockage of the filter, which is prone to misjudgment; on the other hand, premature replacement will cause resource waste, and delayed replacement may lead to production interruption. Summary of the invention

[0003] In view of this, the present invention provides a method, medium and system for estimating the filter replacement time in an ABS production process, which can solve the technical problem that most of the existing technologies estimate the filter blockage degree based on the experience of production personnel and arrange the filter replacement time according to the production plan, which has limitations.

[0004] The present invention is achieved in that: A first aspect of the present invention provides a method for estimating the time of filter replacement in an ABS production process, which comprises the following steps: S10, collecting historical parameter data of multiple filter monitoring points on multiple ABS production lines, wherein the parameter data includes filter pressure, filter deformation, material fluid temperature, and material fluid flow rate; S20, preprocessing the collected historical parameter data, including removing outliers, interpolating missing values, and normalizing to obtain standardized monitoring data; S30, for each monitoring point, counting the change trend of its historical parameter data, and marking the known filter replacement time as historical marking data; S40, based on historical marking data, establish and train a filter clogging machine learning model for each monitoring point; S50, using the filter clogging machine learning model to perform online prediction on the real-time data of the current filter to estimate the current filter clogging degree; S60, setting a threshold value for the filter clogging degree, and when the prediction result of any monitoring point exceeds the threshold value, determining that the filter corresponding to the monitoring point needs to be replaced.

[0005] Among them, the specific implementation method of step S10 for collecting historical parameter data of multiple filter monitoring points on multiple ABS production lines includes: collecting historical filter pressure, filter deformation, material fluid temperature, material fluid flow rate and other parameter data of each monitoring point from multiple monitoring sensors installed on the ABS production line. These data cover the key indicators of filter operation in the ABS production process and provide a basis for subsequent predictive modeling.

[0006] Among them, the specific implementation method of step S20 for preprocessing the collected historical parameter data includes: first, detecting and removing outliers on the data to remove abnormal data points that deviate from the normal range; then filling in the missing data points using interpolation and other methods to ensure the integrity of the data; finally, normalizing the data of each indicator and converting it into standardized monitoring data to meet the needs of subsequent modeling and analysis.

[0007] Among them, the specific implementation method of step S30 of counting the trend of changes in historical parameter data and marking the known filter replacement time includes: performing statistical analysis on the standardized monitoring data collected at each monitoring point to find the changing rules of indicators such as pressure and deformation of the filter during use; at the same time, combining with production practice, marking the known filter replacement times in history to provide a reference basis for the subsequent training of the machine learning model.

[0008] Among them, the specific implementation method of step S40 for establishing and training a filter clogging machine learning model based on historical marking data includes: selecting a neural network embedded with a physical model as a prediction model, the model includes key components such as a physical model module, a timing module, an attention module, a feature extraction module, and a weight real-time update module, and learning and modeling the law of filter clogging through a structured neural network to lay the foundation for subsequent online predictions.

[0009] Among them, the specific implementation method of step S50 using the filter blockage machine learning model to perform online prediction on the real-time data of the current filter includes: receiving parameter data such as pressure, deformation, temperature, flow rate, etc. from each monitoring point in real time, and combining the trained physical model constraints, using the neural network model to predict the blockage degree of the current filter, providing a basis for subsequent judgment of the timing of filter replacement.

[0010] The specific implementation method of setting the filter blockage threshold and judging the filter replacement timing in step S60 includes: setting the filter blockage threshold to 70% to 80% based on production practice and experience; when the predicted blockage degree of any monitoring point exceeds the threshold, determining that the filter corresponding to the monitoring point needs to be replaced to ensure the stability of ABS production and product quality.

[0011] The filter screens include at least the extruder head filter screen, die head filter screen and filter screen of the ABS production line. These filter screens play a key filtering role in the ABS production process, and their clogging conditions need to be monitored and predicted in real time.

[0012] Wherein, the filter screen at least includes an extruder head filter screen, a die head filter screen, and a filter screen of an ABS production line.

[0013] Among them, the filter clogging machine learning model is a neural network embedded with a physical model. The physical model is specifically based on the Ergun equation between the filter pressure difference and the flow, temperature, and viscosity parameters, which is used to constrain the neural network output to conform to physical laws.

[0014] Furthermore, the neural network also includes a timing module, an attention module, a feature extraction module and a real-time weight update module. The real-time weight update module is used to optimize the weight according to the new filter replacement record. The real-time weight update module has a verification data screening function, which is used to automatically select appropriate real-time monitoring data for verifying and updating the weight of the filter clogging machine learning model.

[0015] Furthermore, the physical model adopts the Ergun equation, specifically constructing physical constraints through the relationship between the filter pressure difference, flow rate, fluid viscosity and filter blockage rate.

[0016] Furthermore, the timing module is used to capture the timing variation characteristics of each parameter, and the specific structure is a convolutional LSTM network.

[0017] Furthermore, the attention module is used to capture the correlation between different features, and the specific structure is a multi-head attention mechanism.

[0018] Furthermore, the feature extraction module is used to extract features from different parameters, and its specific structure is divided into two parts: 1) Use a one-dimensional convolutional network to extract features of filter pressure, material fluid temperature, and material fluid flow; 2) For the filter deformation data, use a two-dimensional convolutional network to extract features; The specific structure of the real-time weight update module is to update the model weights online through the Reptile meta-learning algorithm.

[0019] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for estimating the time for filter replacement in an ABS production process.

[0020] A third aspect of the present invention provides a system for estimating filter replacement timing in an ABS production process, which includes the above-mentioned computer-readable storage medium.

[0021] Compared with the prior art, the beneficial effects of the filter replacement time estimation method, medium and system provided by the present invention in the ABS production process are as follows: the scheme of the present invention makes full use of the historical monitoring data collected on the production line, including multiple key indicators such as filter pressure, deformation, material temperature, flow rate, etc., and establishes an intelligent filter clogging prediction system through data preprocessing, feature learning, dynamic modeling and other steps. Compared with the prior art, the method of the present invention has the following main advantages: 1. Comprehensive modeling using multi-source heterogeneous data: The method of the present invention collects multiple monitoring indicators including filter pressure, deformation, temperature, flow, etc., which fully reflects the operating status of the filter in the ABS production process. This fusion of multi-source data helps to improve the accuracy and robustness of the prediction, and overcomes the problem that a single indicator is difficult to accurately describe the filter blockage condition.

[0022] 2. Hybrid modeling combining physical mechanism and machine learning: The method of the present invention combines physical models with neural networks. On the one hand, physical laws such as the Ergun equation are used to constrain the output of the neural network to ensure that the prediction results conform to physical laws; on the other hand, through the powerful learning ability of the neural network, the complex nonlinear relationships hidden in the historical data can be mined. This modeling strategy combining physical constraints with data-driven makes the prediction model both well interpretable and fully utilizes the prediction ability of the data.

[0023] 3. Considering the time series characteristics of the dynamic changes of filter blockage: The method of the present invention uses a convolutional LSTM network to capture the characteristics of the changes of each monitoring parameter over time, which can effectively model the dynamic evolution of the filter blockage condition. Compared with the existing method that only focuses on the prediction of a single moment, this modeling method based on time series data can better reflect the continuity and dynamics of the actual operation of the filter.

[0024] 4. Capable of online adaptive optimization: Based on the trained prediction model, the method of the present invention uses the Reptile meta-learning algorithm to achieve real-time update of weights. When a new filter replacement record appears, the model can automatically adjust parameters according to the latest sample and continuously optimize the prediction performance, thereby adapting to changes in the production environment and providing guarantee for long-term stable application.

[0025] In summary, the present invention solves the technical problem that most of the prior arts estimate the clogging degree of the filter based on the experience of production personnel and arrange the filter replacement time according to the production plan, which has limitations. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0027] Figure 1 A flow chart of the method provided by the present invention; Figure 2 It is the graph of filter pressure changing with time, and the horizontal axis is the date; Figure 3 This is a graph showing the change of filter deformation over time, with the horizontal axis representing the date; Figure 4 It is a graph of material temperature changing with time, with the horizontal axis being the date; Figure 5 It is a graph of material flow rate changing with time, with the horizontal axis being the date; Figure 6 This is the filter blockage prediction result diagram, and the horizontal axis is the date; Figure 7 This is a comparison chart of the congestion trends at different monitoring points, with the horizontal axis representing the date. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0029] like Figure 1 As shown, it is a flow chart of a method for estimating the time of filter replacement in an ABS production process provided by the present invention, and the method comprises the following steps: S10, collecting historical parameter data of multiple filter monitoring points on multiple ABS production lines, the parameter data including filter pressure, filter deformation, material fluid temperature, and material fluid flow rate; S20, preprocessing the collected historical parameter data, including removing outliers, interpolating missing values, and normalizing to obtain standardized monitoring data; S30, for each monitoring point, counting the change trend of its historical parameter data, and marking the known filter replacement time as historical marking data; S40, based on historical marking data, establish and train a filter clogging machine learning model for each monitoring point; S50, using the filter clogging machine learning model to perform online prediction on the real-time data of the current filter to estimate the current filter clogging degree; S60, setting a threshold value for the filter clogging degree, and when the prediction result of any monitoring point exceeds the threshold value, determining that the filter corresponding to the monitoring point needs to be replaced.

[0030] The specific implementation methods of the above steps are described in detail below: Step S10 collects historical parameter data of multiple filter monitoring points on multiple ABS production lines: The main purpose of this step is to collect and organize the historical parameter data related to the operation of the filter screen during the ABS production process. Specifically, it is necessary to collect historical monitoring data including multiple indicators such as filter screen pressure, filter screen deformation, material fluid temperature, and material fluid flow. These data come from multiple monitoring points on the ABS production line. The purpose is to obtain sufficient sample data to lay the foundation for the subsequent prediction model establishment.

[0031] Step S20 pre-processes the collected historical parameter data: The collected raw monitoring data may contain some outliers or missing values, so the data needs to be preprocessed. First, the data needs to be detected and removed for outliers, and those abnormal data points that are obviously deviated from the normal range need to be removed. Then, for missing data points, interpolation and other methods can be used to fill them in to ensure the integrity of the data. Finally, the data of each indicator is normalized and converted into standardized monitoring data to prepare for subsequent modeling and analysis.

[0032] Step S30 counts the changing trend of historical parameter data and marks the known filter replacement time: The standardized monitoring data collected at each monitoring point needs to be carefully analyzed for its historical change trends. Through statistical analysis of these data, the changing patterns of pressure, deformation and other indicators of the filter during use can be found. At the same time, combined with actual production, it is necessary to mark the known filter replacement times in history to provide a reference for subsequent machine learning model training.

[0033] Step S40 builds and trains a filter clogging machine learning model based on historical tag data: With the historical data and labeling information accumulated in the previous steps, we can start to build and train the machine learning prediction model for filter clogging. Here we choose to use a neural network embedded with a physical model as the prediction model. Specifically, this neural network model includes the following key modules: 1) Physical model module: Based on the Ergun equation, the relationship between filter pressure difference, flow rate, fluid viscosity and blockage rate is established to constrain the output of the neural network to conform to physical laws.

[0034] 2) Timing module: Use convolutional LSTM network to capture the changing characteristics of each parameter over time.

[0035] 3) Attention module: Use multi-head attention mechanism to learn the correlation between different features.

[0036] 4) Feature extraction module: Different convolutional networks are used to extract features for different types of input data (pressure, temperature, flow, deformation).

[0037] 5) Real-time weight update module: Using the Reptile meta-learning algorithm, the weight parameters of the neural network are optimized in real time according to the new filter replacement records.

[0038] Through such a structure-rich neural network model, combined with supervised learning of historical labeled data, the law of filter clogging can be well learned and modeled, laying the foundation for subsequent online prediction.

[0039] Step S50 uses the filter blockage machine learning model to perform online prediction on the real-time data of the current filter: With the trained filter clogging prediction model, we can use it to make online predictions on the real-time monitoring data of the filter on the current production line. Specifically, the model will receive real-time pressure, deformation, temperature, flow and other parameter data from each monitoring point, and combine the physical model constraints to predict the current filter clogging degree. This prediction result will provide an important basis for the subsequent judgment of the timing of filter replacement.

[0040] Step S60 sets the filter blockage threshold to determine when to replace the filter: After obtaining the predicted results of the filter blockage degree, a reasonable blockage degree threshold needs to be set. When the predicted blockage degree of any monitoring point exceeds the threshold, it can be determined that the filter corresponding to the monitoring point needs to be replaced. The setting of this threshold needs to be combined with production practice and experience, and can generally be set to a blockage degree of 70%~80%. When this threshold is reached, the filter should be replaced immediately to ensure the stability of ABS production and product quality.

[0041] In general, this ABS production process filter replacement time estimation method makes full use of the historical monitoring data on the production line, combines machine learning and physical models, and establishes a powerful online prediction system. The system can monitor the filter blockage status in real time and predict the optimal time for filter replacement in advance, greatly improving the automation level and operational efficiency of ABS production.

[0042] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for estimating the time for filter replacement in an ABS production process.

[0043] A third aspect of the present invention provides a system for estimating filter replacement timing in an ABS production process, which includes the above-mentioned computer-readable storage medium.

[0044] In order to better understand and implement the present invention, the following provides an embodiment 1 of the specific steps of a computer-readable program running in a computer-readable medium or a computer-like device of the present invention: In this embodiment 1, for step S10 to collect historical parameter data of multiple filter monitoring points on multiple ABS production lines, the following variables can be defined: Indicates the number of ABS production lines, Indicates the number of filter monitoring points on each production line. The first production line Monitoring points, collect the following historical parameter data: :Filter pressure; : Filter deformation; : Material fluid temperature; : Material fluid flow rate; in Represents the time series. By collecting these historical parameter data, comprehensive information on the filter operation status during the ABS production process is obtained.

[0045] When preprocessing the collected historical parameter data in step S20, it is necessary to first detect and remove abnormal values. Pressure data If there is an abnormality, it can be detected by the following methods: ; in and Respectively The sample mean and standard deviation of . If , then determine outliers and remove them.

[0046] For missing data points, interpolation can be used to fill them. Assume that in the time interval [ t 1 , t 2 ] Memory The missing values ​​can be filled using cubic spline interpolation: ; in are the coefficients of the cubic polynomial, through the known and And the derivative conditions are determined.

[0047] Finally, in order to meet the needs of subsequent modeling, it is necessary to normalize the data of each indicator. are the normalized filter pressure, deformation, temperature and flow data respectively, then: ; ; ; ; In the formula, and are the maximum and minimum values ​​of the corresponding items respectively.

[0048] Through the above preprocessing steps, standardized monitoring data were obtained, laying the foundation for subsequent analysis and modeling.

[0049] When calculating the trend of historical parameter data and marking the known filter replacement time in step S30, the following method can be used: For each monitoring point , No. For a production line, the rate of change of each parameter over time can be calculated: ; By analyzing these change rates, we can find the changing rules of various indicators during the use of the filter. At the same time, combined with actual production records, it is necessary to mark the known filter replacement times in history. , which provides a reference for subsequent machine learning model training.

[0050] In step S40, when establishing and training the filter blockage machine learning model based on historical marking data, a neural network embedded with a physical model is selected as the prediction model. Specifically, the model includes the following key components: 1. Physical model module: Using Ergun equation to establish filter pressure difference ,flow , fluid viscosity and congestion rate The relationship between: ; in, is the filter aperture, is the fluid density. This physical model is used to constrain the output of the neural network to make it conform to the laws of physics.

[0051] 2. Timing module: The convolutional LSTM network is used to capture the changing characteristics of each parameter over time. Production line monitoring points, the input sequence is [ p ̂ ij ( t - T +1), p ̂ ij ( t - T +2),..., p ̂ ij ( t ) , d ̂ ij ( t - T +1),..., d ̂ ij ( t ), T ̂ ij ( t - T +1),..., T ̂ ij ( t ), Q ̂ ij ( t - T +1),..., Q ̂ ij ( t )] ,in is the time window length. The convolutional LSTM network can extract these time series features to provide support for subsequent congestion prediction.

[0052] 3. Attention Module: A multi-head attention mechanism is used to learn the correlation between different features. Production line Monitoring points at time The feature representation , the attention mechanism can be expressed as: ; ; in, is a learnable weight matrix. The attention mechanism can capture the correlation between different features and improve the expressiveness of the model.

[0053] 4. Feature extraction module: For different types of input data, different convolutional networks are used for feature extraction.

[0054] For filter pressure , Material fluid temperature and flow For one-dimensional data, use a one-dimensional convolutional network to extract features; For filter deformation For two-dimensional data such as image processing, a two-dimensional convolutional network is used to extract features.

[0055] Through this differentiated feature extraction method, the characteristics of different types of data can be better utilized.

[0056] 5. Weight real-time update module: In order to enable the model to dynamically adjust parameters over time and adapt to changes in the production environment, the Reptile meta-learning algorithm is used to optimize and update the model weights in real time. Specifically, when a new filter replacement record appears, it is used as a validation set, and the Reptile algorithm is used to fine-tune the model parameters to better fit the new sample distribution. At the same time, the module also has the function of automatically screening suitable validation data to further improve the effect of online learning.

[0057] Through the design of the above neural network model, combined with the constraints of the physical model, we can make full use of historical monitoring data, learn and model the laws of filter clogging, and provide strong support for subsequent online predictions.

[0058] In step S50, when the filter blockage machine learning model is used to perform online prediction on the real-time data of the current filter, the pressure from each monitoring point is received in real time. ,deformation ,temperature and flow Combined with the neural network model trained in step S40, the current filter blockage degree can be predicted. : ; in, Represents the parameters of the model. This prediction result will provide an important basis for the subsequent judgment of the timing of filter replacement.

[0059] In step S60, when setting the filter blockage threshold and determining the filter replacement timing, the filter blockage threshold can be set based on production practice and experience. Set between 70% and 80%. No. Predicted congestion levels for production lines When the threshold is exceeded, it can be determined that the filter corresponding to the monitoring point needs to be replaced: ; Through this method based on real-time monitoring and prediction, signs of filter blockage can be discovered in a timely manner, providing decision-making support for production managers and ensuring the stability of ABS production and product quality.

[0060] In general, this ABS production process filter replacement time estimation method makes full use of the historical monitoring data on the production line, combines physical model constraints and powerful neural network prediction capabilities, and establishes an intelligent online monitoring and prediction system. The system can monitor the operating status of the filter in real time and predict the optimal time for filter replacement in advance, which greatly improves the automation level and operational efficiency of ABS production.

[0061] Specifically, the principle of the present invention is to establish an intelligent prediction system based on the combination of machine learning and physical models. The key to this system is to make full use of the historical monitoring data collected on the production line, and finally obtain a machine learning model that can accurately predict the degree of filter blockage through steps such as data preprocessing, feature learning, and dynamic modeling. This prediction model contains the following key innovations: 1. Physical model constraints: When establishing a machine learning prediction model, the method of the present invention incorporates physical models such as the Ergun equation to constrain the output of the neural network. The Ergun equation describes the relationship between the filter pressure difference, flow rate, fluid viscosity and blockage rate, which conforms to the laws of physics. Embedding this physical model into the neural network can ensure that the prediction results do not violate the basic laws of physics, and improve the interpretability and reliability of the model.

[0062] 2. Time series feature modeling: The blockage of the filter is a dynamic process, and it is necessary to consider the change characteristics of each monitoring parameter over time. The method of the present invention uses a convolutional LSTM network to capture this time series feature. The network structure can effectively extract the time-dependent information contained in the input sequence and provide support for dynamic prediction. Through this time series modeling method, the blockage trend of the filter can be predicted more accurately, rather than just being limited to the state at a certain moment.

[0063] 3. Multi-source feature fusion: The operating status of the filter is the result of the combined effect of multiple factors, and a single monitoring indicator is difficult to fully reflect its blockage degree. The method of the present invention collects a variety of heterogeneous data such as filter pressure, deformation, material temperature, flow, etc., and adopts differentiated feature extraction methods (one-dimensional convolutional network and two-dimensional convolutional network) to fully utilize the information of these multi-source data. This fusion of multi-source features can provide richer and more comprehensive inputs and improve the generalization ability of the prediction model.

[0064] 4. Adaptive online optimization: In actual production, due to the continuous changes in process conditions, the operating status of the filter will also change to a certain extent. In order to enable the prediction model to adapt to these changes, the method of the present invention adopts the Reptile meta-learning algorithm to achieve real-time optimization and update of model parameters. Whenever a new filter replacement record appears, the model will automatically adjust the weight according to the latest sample to ensure that the prediction results can continuously track the actual operating status of the filter and improve the accuracy and stability of the prediction.

[0065] In summary, the technical principles of the method of the present invention are mainly reflected in: 1) combining physical model constraints to improve model interpretability; 2) using time series feature modeling to capture dynamic change laws; 3) using multi-source feature fusion to enhance prediction capabilities; 4) enhancing the adaptability of the model through adaptive online optimization.

[0066] In order to better understand the present invention, a specific application scenario embodiment 2 is designed below: A certain enterprise has 3 ABS production lines, each of which is equipped with 5 filter monitoring points, which are located at key positions such as the extruder head, die head and filter. The enterprise has been seeking to improve the automation level and operational efficiency of the ABS production process, hoping to detect the signs of filter clogging in time and provide decision support for filter replacement. Based on this, the enterprise decided to adopt the ABS production process filter replacement time estimation method proposed by the present invention, and the specific implementation is as follows: 1. Data Collection First, the company collected historical operating data from five filter monitoring points on three ABS production lines in the past two years, including parameters such as filter pressure, filter deformation, material fluid temperature, and material fluid flow. Taking the third monitoring point of the first production line as an example, some of the collected historical data are shown in Table 1: Table 1 Historical data table time Filter pressure (MPa) Filter deformation(mm) Material temperature (°C) Material flow rate (L / min) 2022-01-0108:00 2.1 0.3 210 80 2022-01-0108:15 2.2 0.4 211 79 2022-01-0108:30 2.3 0.4 209 78 ... ... ... ... ... 2023-12-3123:45 3.2 0.8 215 72 Figure 2-Figure 5This is a trend chart of filter monitoring data over time, showing the change trends of four key parameters: filter pressure, deformation, material temperature, and material flow rate from January 1, 2022 to December 31, 2023. The figure contains four sub-graphs, which respectively show the changes of these parameters over time. It can be observed that the filter pressure and deformation show an upward trend, the material temperature is relatively stable, and the material flow rate shows a downward trend. These trends intuitively reflect the process of gradual clogging of the filter. By collecting these historical data, the company has accumulated a wealth of filter operation status samples, laying the foundation for subsequent predictive modeling.

[0067] 2. Data Preprocessing The collected raw monitoring data will inevitably contain some outliers and missing values, so necessary preprocessing is required. For the data of the third monitoring point of the first production line, the specific preprocessing steps are as follows: (1) Outlier detection and removal: Take the filter pressure data as an example to calculate its standard score: ; in, is the sample mean, is the sample standard deviation. If , then the data point is determined to be an outlier and is removed. In this way, a total of 18 abnormal pressure data points are removed.

[0068] (2) Missing value filling: For missing data points, cubic spline interpolation is used to fill them. Taking the missing flow data in a certain period as an example, let the missing interval be [ t 1 , t 2 ] , we can construct the following interpolation function: ; in is the unknown coefficient, which can be determined by the known boundary conditions. Using this method, a total of 27 missing flow data points were filled.

[0069] (3) Data normalization: In order to eliminate the impact of dimension on subsequent modeling, the minimum-maximum normalization processing is performed on the monitoring index data. Taking the filter pressure as an example, the normalized pressure data The calculation is as follows: ; The normalization of other indicators is similar. Through the above preprocessing steps, the enterprise obtains standardized monitoring data, laying the foundation for subsequent analysis and modeling.

[0070] 3. Historical data analysis After completing data preprocessing, the company began to analyze the changing trend of its historical monitoring data for each monitoring point. Taking the third monitoring point of the first production line as an example, the analysis results are as follows: (1) The filter pressure showed a trend of gradual increase, from the initial 2.1MPa to 3.2MPa, an increase of 52%.

[0071] (2) The deformation of the filter screen also gradually increased from 0.3 mm to 0.8 mm, an increase of 167%.

[0072] (3) The material temperature remains basically stable, fluctuating between 209 and 215°C.

[0073] (4) The material flow rate showed a gradual downward trend, from the initial 80 L / min to 72 L / min, a decrease of 10%.

[0074] At the same time, the company marked the three filter replacement times in the monitoring point in the production records, which occurred on June 15, 2022, January 10, 2023, and July 5, 2023. Through the analysis of these historical data, the company has mastered the basic change rules of the filter operation status of the monitoring point, which provides an important basis for the establishment of subsequent prediction models.

[0075] 4. Prediction model training Based on the above historical data analysis results, the company began to establish a prediction model for filter clogging. Considering that filter clogging is a complex process involving multiple physical factors, the company decided to use a neural network embedded with a physical model as the prediction model. The model specifically includes the following key modules: (1) Physical model module: The company uses the Ergun equation to describe the filter pressure difference ,flow , fluid viscosity and congestion rate The relationship between them is used as a constraint: ;in, is the filter aperture, is the fluid density. This physical model is embedded in the neural network to ensure that the output results conform to the laws of physics.

[0076] (2) Time series module: In order to capture the changing characteristics of each monitoring parameter over time, the company uses a convolutional LSTM network as a time series feature extraction module. The input of this module is the past The monitoring data sequence within a time step, such as [ p ̂ ( t - T +1), p ̂ ( t - T +2),..., p ̂ ( t ), d ̂ ( t - T +1),..., d ̂ ( t ), T ̂ ( t - T +1),..., T ̂ ( t ), Q ̂ ( t - T +1),..., Q ̂ ( t )] ,in They are normalized pressure, deformation, temperature and flow data respectively. The convolutional LSTM network can effectively extract these time series features and provide support for subsequent congestion prediction.

[0077] (3) Attention module: In order to capture the correlation between different monitoring indicators, the company integrates a multi-head attention mechanism into the model. The feature representation of ,The attention module is calculated as follows: ; ; in, is a learnable weight matrix. This attention mechanism can adaptively assign different weights to different features and enhance the model’s ability to capture key features.

[0078] (4) Feature extraction module: For different types of monitoring data, the company adopts differentiated feature extraction methods. For one-dimensional pressure, temperature, and flow data, a one-dimensional convolutional network is used to extract features; for two-dimensional deformation data, a two-dimensional convolutional network is used for feature extraction. Through this differentiated feature extraction, the characteristics of various types of data can be better utilized to provide high-quality input features for subsequent congestion prediction.

[0079] (5) Real-time weight update module: In order to enable the model to dynamically adjust parameters over time and adapt to changes in the production environment, the company integrated the Reptile meta-learning algorithm into the prediction model to achieve real-time optimization of weights. Specifically, whenever a new filter replacement record appears, the company will use it as a validation set and fine-tune the model parameters through the Reptile algorithm to better fit the new sample distribution. At the same time, this module also has the function of automatically screening suitable validation data to further improve the effect of online learning.

[0080] With the above model design, the company used the historical data of the third monitoring point of the first production line to train and verify the model. During the training process, the company adopted a stratified sampling strategy to ensure that the sample ratio of each type of blockage state was balanced, which improved the generalization of the model. After repeated iterative optimization, a machine learning model that can accurately predict the degree of filter blockage at the monitoring point was finally obtained.

[0081] 5. Online prediction and decision-making: With the trained prediction model, the company began to apply it in real time on the production line. Taking the third monitoring point of the first production line as an example, the specific process is as follows: (1) Real-time data collection: The sensors installed on the production line collect the filter pressure, deformation, material temperature and material flow data of the monitoring point in real time and transmit them to the prediction model.

[0082] (2) Blockage degree prediction: The prediction model receives these monitoring data in real time and predicts the current filter blockage degree of the monitoring point based on the physical constraints, time series characteristics, multi-source feature fusion and other mechanisms introduced above. .

[0083] (3) Decision support: Based on production practice experience, the enterprise sets the filter blockage threshold Set at 80%. When the predicted congestion level at any monitoring point When the threshold is exceeded, the decision to replace the filter is triggered.

[0084] like Figure 6As shown in the figure, the filter blockage prediction results from November 1, 2023 to December 31, 2023 are shown. The blue solid line in the figure represents the actual blockage degree, the red dotted line represents the predicted blockage degree, the green dotted line represents the replacement threshold (0.8), and the yellow shaded area represents the warning area. As can be seen from the figure, the prediction result is very close to the actual blockage degree. When the predicted blockage degree exceeds the threshold, it enters the warning area, prompting the need to replace the filter. In actual application, the company found that the predicted blockage degree of the third monitoring point of the first production line reached 82% after 2 hours, exceeding the threshold of 80%. After receiving this warning signal, the production management personnel immediately arranged to replace the filter corresponding to the monitoring point, avoiding production interruptions and product quality problems caused by filter blockage.

[0085] Through continuous real-time monitoring and prediction, the company found that there were certain differences in the filter blockage conditions at each monitoring point. For example, due to the adjustment of process parameters, the filter pressure and deformation at the fourth monitoring point of the second production line changed slowly, and the predicted blockage degree never exceeded the threshold. However, the first monitoring point of the third production line showed a faster blockage trend, and the company also replaced the filter earlier accordingly. Figure 7 The blockage trends of three different monitoring points (line 1-point 3, line 2-point 4, line 3-point 1) throughout 2023 are compared. As can be seen from the figure, there are obvious differences in the blockage speed of different monitoring points. The blockage speed of line 3-point 1 is the fastest, while the blockage speed of line 2-point 4 is the slowest. The black dotted line represents the replacement threshold (0.8). This comparison chart helps production managers understand the blockage characteristics of different monitoring points and formulate more targeted maintenance strategies.

[0086] In general, by adopting the method of the present invention, enterprises can realize intelligent monitoring and prediction of the filter status during the ABS production process, providing timely and accurate decision support for filter replacement. This not only improves production stability and product quality, but also greatly saves manpower and resource consumption for filter replacement. At the same time, the large amount of monitoring and prediction data accumulated by this method also provides valuable data support for further optimizing ABS production process parameters.

[0087] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for estimating the time of filter replacement in an ABS production process, characterized in that: The following steps are involved: S10, collecting historical parameter data of multiple filter monitoring points on multiple ABS production lines, wherein the parameter data includes filter pressure, filter deformation, material fluid temperature, and material fluid flow rate; S20, preprocessing the collected historical parameter data, including removing outliers, interpolating missing values, and normalizing to obtain standardized monitoring data; S30, for each monitoring point, counting the change trend of its historical parameter data, and marking the known filter replacement time as historical marking data; S40, based on historical marking data, establish and train a filter clogging machine learning model for each monitoring point; S50, using the filter clogging machine learning model to perform online prediction on the real-time data of the current filter to estimate the current filter clogging degree; S60, setting a threshold value for the filter clogging degree, and when the prediction result of any monitoring point exceeds the threshold value, determining that the filter corresponding to the monitoring point needs to be replaced.

2. The method for estimating the time of filter replacement in the ABS production process according to claim 1, characterized in that: The filter screen at least includes an extruder head filter screen, a die head filter screen, and a filter filter screen of an ABS production line.

3. The method for estimating the time of filter replacement in the ABS production process according to claim 1, characterized in that: The filter clogging machine learning model is a neural network embedded with a physical model. The physical model is specifically based on the Ergun equation between the filter pressure difference and the flow, temperature, and viscosity parameters, which is used to constrain the neural network output to conform to physical laws.

4. The method for estimating the time of filter replacement in the ABS production process according to claim 3, characterized in that: The neural network also includes a timing module, an attention module, a feature extraction module and a real-time weight update module. The real-time weight update module is used to optimize the weight according to the new filter replacement record. The real-time weight update module has a verification data screening function, which is used to automatically select appropriate real-time monitoring data for verifying and updating the weight of the filter clogging machine learning model.

5. The method for estimating the time of filter replacement in the ABS production process according to claim 4, characterized in that: The physical model adopts the Ergun equation, specifically constructing physical constraints through the relationship between filter pressure difference, flow rate, fluid viscosity and filter blockage rate.

6. The method for estimating the time of filter replacement in the ABS production process according to claim 4, characterized in that: The timing module is used to capture the timing change characteristics of each parameter, and the specific structure is a convolutional LSTM network.

7. The method for estimating the time of filter replacement in the ABS production process according to claim 4, characterized in that: The attention module is used to capture the correlation between different features, and the specific structure is a multi-head attention mechanism.

8. The method for estimating the time of filter replacement in the ABS production process according to claim 4, characterized in that: The feature extraction module is used to extract features from different parameters, and its specific structure is divided into two parts: 1) Use a one-dimensional convolutional network to extract features of filter pressure, material fluid temperature, and material fluid flow; 2) For the filter deformation data, use a two-dimensional convolutional network to extract features; The specific structure of the real-time weight update module is to update the model weights online through the Reptile meta-learning algorithm.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the method for estimating the time for filter replacement in an ABS production process as described in any one of claims 1-8.

10. A filter replacement time estimation system for ABS production process, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.

Citation Information

Patent Citations

  • Operation index abnormity monitoring method

    CN109615226A

  • Residual oil distribution prediction method and device based on historical and predicted oil reservoir knowledge

    CN114638401A

  • Drying temperature determination method for automobile coating and application

    CN116432529A

  • Filter screen adaptive adjustment method and system based on big data

    CN117390428A

  • Cable joint fault detection method and system based on long short-term memory (LSTM)

    CN117743961A