Scrap statistics and material blockage abnormality alarm method based on production line status recognition

Abstract

This invention discloses a method for identifying equipment status in extrusion lines in the magnetic plastics industry, combining the relationships between equipment within the line to construct statistics on the duration and weight of waste generated in different sections of the line. Furthermore, the invention also provides an alarm for abnormal blockages in the auxiliary traction machine within the line. By combining machine learning methods with production line business relationships, the invention solves three major issues: real-time monitoring of production line status, alarming for production line anomalies, and worker performance assessment, without requiring technical modifications or changes to existing processes.

Description

Technical field

[0001] The present invention relates to the technical field of a method for counting scrap and alarming abnormal material blockage based on production line status identification, and in particular to the technical field of a method for counting scrap and alarming abnormal material blockage based on production line status identification. [Background Technology]

[0002] For magnetic plastic extrusion production lines, timely and accurate information on the line's operating status, scrap generated at each stage, fault alarms in certain stages, and the time it takes for line workers to respond is crucial for managers to formulate production plans, optimize production processes, and effectively evaluate worker performance. However, most companies still rely on a very traditional approach. Some companies evaluate workers by manually recording the weight of scrap produced on each production line, measured on a shift-by-shift basis; some companies even lack a monitoring and evaluation mechanism.

[0003] In recent years, some companies and researchers have attempted to identify production line status through intelligent control or video recognition.

[0004] The cost of adding a control unit is too high for most companies to afford. Furthermore, the control unit can only distinguish whether each piece of equipment on the production line is in a processing or non-processing state. Scrap statistics, abnormality alarms, and other functions require the development of machine learning models and new applications.

[0005] When using video recognition, due to the high noise and vibration on site, and the frequent need for manual processing of related abnormalities or failures, a slight change in the camera angle or an impact on the field of view will affect the recognition, resulting in unsatisfactory results.

[0006] Existing production line status recognition methods often only consider the status of the equipment itself, while ignoring the relationship between devices, which leads to certain limitations in the accuracy and real-time performance of status recognition. [Summary of the invention]

[0007] The purpose of the present invention is to solve the problems in the existing technology and propose a method for scrap statistics and abnormal blockage alarm based on production line status identification. By combining machine learning methods with production line business relationships, it solves the three major problems of real-time monitoring of production line status, production line abnormality alarm, and worker performance appraisal, without the need for technical transformation or change of the original process.

[0008] To achieve the above objectives, the present invention proposes a method for waste statistics and abnormal blockage alarm based on production line status recognition, which includes the following stages:

[0009] Step 1. Data collection and processing of each device in the production line includes the following steps:

[0010] S11: Data collection and processing of each device. The real-time current signal data of the extrusion equipment, main traction equipment, and auxiliary traction equipment are collected and cleaned to obtain processed sample data. Outliers less than or equal to 0 are first deleted, and then the IQR method is used to determine outliers. For missing data and outliers, according to the characteristics of the time series, they are filled with the previous and next values.

[0011] S12: Analyze and determine the status of each device. The extruder and main traction machine include three states: standby, start-up, and processing. The auxiliary traction machine includes four states: standby, start-up, processing, and blocking.

[0012] S13: If the extruder can collect speed, then this speed index is directly used. If the extruder cannot collect speed index, the extrusion speed of the extruder is obtained by using the corresponding relationship between the main traction machine frequency and speed, and the corresponding relationship between the extruder frequency and the main traction frequency, and using statistical and correlation methods;

[0013] S14: Feature extraction. In addition to using the processed original current value as a feature, a statistical method is selected to generate new features based on the model effect, and new features are added based on the association relationship between the equipment in the production line.

[0014] Step 2: State recognition modeling of the extruder, main tractor, and auxiliary tractor, including the following steps:

[0015] S21: clustering the equipment states of the extruder, main traction machine, and auxiliary traction machine using a clustering algorithm. If the number of clusters needs to be set, the data of different working conditions of the equipment described in S12 is the set value.

[0016] S22: Modeling and tuning: Through multiple rounds of adjustments to features and model parameters, the best model is selected as the state recognition model for each device. Based on the HMM model parameter adjustment and parameter selection, tuning is stopped when the accuracy of the duration of each working condition for n consecutive shifts meets the on-site customer requirements and there is little room for further model optimization.

[0017] Step 3: Calculate the weight of the extruded waste in the extrusion process, the weight of the cut waste in the main traction process, the weight of the waste generated by the blockage in the auxiliary traction process, and the weight of the waste that failed to be successfully magnetized in the magnetization process; alarm for abnormal blockage in the auxiliary traction process, including the following steps:

[0018] S31: The extrusion waste extrusion time consists of two parts: First, when the extruder is shut down and started up, waste will be extruded for a period of time, and then pulled to the main traction machine after it stabilizes; second, when the extruder is shut down for a short period of time due to mold or material change during operation, waste will also be generated. The waste generation time of the extruder is the main traction start time minus the extruder start time;

[0019] S32: waste cutting in the main traction link. This is mainly done when the product quality is unstable at the beginning of production. At this time, waste cutting is performed in the main traction link. The duration of waste cutting in the main traction link is the auxiliary traction start time minus the main traction start time.

[0020] S33: The duration of waste caused by blockage in the auxiliary traction link. This is mainly caused by wear of the auxiliary traction machine itself or the cutter head of the connected cutting machine. During a production shift, multiple blockages may occur. The auxiliary traction machine blockage duration is calculated by subtracting the blockage start time from the blockage end time.

[0021] S34: Auxiliary tractor blockage abnormal alarm. When the model identifies that the auxiliary tractor is in a blockage state, the model will push the result to the business end ERP for real-time alarm;

[0022] S35: Statistics of the weight of waste products in the magnetizing process are obtained by summing up the lengths of the plastic strips corresponding to the collected non-magnetizing counts in sections.

[0023] Preferably, if the speed of the extruder cannot be collected in step S13, the collected data is first analyzed to obtain an approximate linear relationship between the main traction frequency and the speed, and a linear regression model can be constructed for fitting. When the main traction frequency and the extruder frequency are processing the same product and both are in the processing state, the frequencies are in a proportional relationship. However, when the processed products are different, the proportions will be different. After the processing is completed during the shift, the highest frequency corresponding values of the main traction machine and the extruder appearing simultaneously in each section where the machine is stopped and the product is not changed are counted as the two frequency ratios, and then the linear relationship between the main traction machine speed and the main traction frequency is used to obtain the frequency of the extruder.

[0024] Preferably, the statistical method mentioned in step S14 is used for feature extraction, including differential features, moving average features, moving average variance, decomposition features of STL time series, etc. The size of the moving average feature window is set according to the duration of the startup state, and generally the maximum is 1 / 2 of the startup duration.

[0025] Preferably, in step S14, not only the collection characteristics and statistical characteristics of the current device are considered, but also, because there is an association between the devices in the production line, when each device is modeled, the device characteristics of the previous and next links in the production line are also added to the selectable feature set.

[0026] Preferably, the clustering algorithm mentioned in step S21 includes KMeans, GMM, HMM, DBSCAN and other clustering algorithms.

[0027] Preferably, when the model is tuned in step S22, the model features do not just use the collected single features, but incorporate the statistical features characterizing the distribution of collected data mentioned in S14. The evaluation method is through drawing display combined with on-site personnel cooperation and verification.

[0028] Preferably, in step S22, according to the HMM model parameter adjustment and parameter selection, when the accuracy of the duration of each working condition for n consecutive shifts reaches more than 90% of the on-site customer requirement and there is little room for further optimization of the model, the tuning is stopped.

[0029] As a preferred method, in step S3, the waste time of each link is counted, which not only uses the machine learning model to identify the status of a single device, but also combines the relationship between the equipment in the production line to accurately and real-time describe the working status of different links in the production line.

[0030] Preferably, step S33 assists in the identification of abnormal traction blockage, which includes two parts. Due to the limitations of acquisition capabilities and cost control requirements, blockages that generally last for more than 1 minute can be directly identified by the model. The blockage duration is obtained by subtracting the blockage start time from the auxiliary traction blockage end time identified by the model. If the blockage duration is less than 1 minute and the model cannot accurately identify it, it is judged by the changes in the magnetized and non-magnetized counts in the magnetization link.

[0031] The beneficial effects of the present invention are as follows: the present invention adopts a machine learning algorithm to identify the status of the extrusion equipment, main traction equipment, and auxiliary traction equipment in the magnetic plastic extrusion production line, and combines the magnetic PLC recording information to identify the duration of the status of different links in the production line through the business relationship between the status of each device, and then combines the collected speed and statistical density (meter length) data to calculate the weight of waste in each link and push it to the manager in real time. At the same time, the auxiliary traction equipment often causes blockage problems due to failures. The model can identify its status and issue a real-time blockage alarm to remind the workers in charge of the production line to deal with it in time, thereby improving the efficiency of the production line. This model only needs to collect the equipment spindle current signal by adding a gateway device, and does not require the existing production line to be modified, so the collection cost is low. By adopting a machine learning model combined with the correlation between different equipment in the production line, the status of different links of the production line can be monitored in real time, waste statistics can be counted, and abnormal alarms can be issued, which can help managers to optimize production, evaluate employee performance, and deal with production line abnormalities in a timely manner, thereby improving work efficiency.

[0032] The features and advantages of the present invention will be described in detail through embodiments with reference to the accompanying drawings.

Brief Description of the Drawings

[0033] Figure 1 This is an overall flow chart of the method for waste statistics and abnormal blockage alarm based on production line status identification of the present invention. [Specific implementation method]

[0034] See Figure 1 The present invention provides a method for waste statistics and abnormal blockage alarm based on production line status recognition, which includes the following stages:

[0035] Step 1. Data collection and processing of each device in the production line includes the following steps:

[0036] S11: Data collection and processing of each device. The real-time current signal data of the extrusion equipment, main traction equipment, and auxiliary traction equipment are collected and cleaned to obtain processed sample data. Outliers less than or equal to 0 are first deleted, and then the IQR method is used to determine outliers. For missing data and outliers, according to the characteristics of the time series, they are filled with the previous and next values.

[0037] S12: Analyze and determine the status of each device. The extruder and main traction machine include three states: standby, start-up, and processing. The auxiliary traction machine includes four states: standby, start-up, processing, and blocking.

[0038] S13: If the extruder can collect speed, then this speed index is directly used. If the extruder cannot collect speed index, the extrusion speed of the extruder is obtained by using the corresponding relationship between the main traction machine frequency and speed, and the corresponding relationship between the extruder frequency and the main traction frequency, and using statistical and correlation methods;

[0039] S14: Feature extraction. In addition to using the processed original current value as a feature, a statistical method is selected to generate new features based on the model effect, and new features are added based on the association relationship between the equipment in the production line.

[0040] Step 2: State recognition modeling of the extruder, main tractor, and auxiliary tractor, including the following steps:

[0041] S21: clustering the equipment states of the extruder, main traction machine, and auxiliary traction machine using a clustering algorithm. If the number of clusters needs to be set, the data of different working conditions of the equipment described in S12 is the set value.

[0042] S22: Modeling and tuning: Through multiple rounds of adjustments to features and model parameters, the best model is selected as the state recognition model for each device. Based on the HMM model parameter adjustment and parameter selection, tuning is stopped when the accuracy of the duration of each working condition for n consecutive shifts meets the on-site customer requirements and there is little room for further model optimization.

[0043] Step 3: Calculate the weight of the extruded waste in the extrusion process, the weight of the cut waste in the main traction process, the weight of the waste generated by the blockage in the auxiliary traction process, and the weight of the waste that failed to be successfully magnetized in the magnetization process; alarm for abnormal blockage in the auxiliary traction process, including the following steps:

[0044] S31: The extrusion waste extrusion time consists of two parts: First, when the extruder is shut down and started up, waste will be extruded for a period of time, and then pulled to the main traction machine after it stabilizes; second, when the extruder is shut down for a short period of time due to mold or material change during operation, waste will also be generated. The waste generation time of the extruder is the main traction start time minus the extruder start time;

[0045] S32: waste cutting in the main traction link. This is mainly done when the product quality is unstable at the beginning of production. At this time, waste cutting is performed in the main traction link. The duration of waste cutting in the main traction link is the auxiliary traction start time minus the main traction start time.

[0046] S33: The duration of waste caused by blockage in the auxiliary traction link. This is mainly caused by wear of the auxiliary traction machine itself or the cutter head of the connected cutting machine. During a production shift, multiple blockages may occur. The auxiliary traction machine blockage duration is calculated by subtracting the blockage start time from the blockage end time.

[0047] S34: Auxiliary tractor blockage abnormal alarm. When the model identifies that the auxiliary tractor is in a blockage state, the model will push the result to the business end ERP for real-time alarm;

[0048] S35: Statistics of the weight of waste products in the magnetizing process are obtained by summing up the lengths of the plastic strips corresponding to the collected non-magnetizing counts in sections.

[0049] Among them, if the speed of the extruder cannot be collected in step S13, the collected data is first analyzed to obtain an approximate linear relationship between the main traction frequency and the speed, and a linear regression model can be constructed for fitting. The main traction frequency and the extruder frequency are in a directly proportional relationship when processing the same product and both are in the processing state. However, when the processed products are different, the ratio will be different. After the processing is completed during the shift, the corresponding value with the highest frequency of the main traction machine and the extruder appearing at the same time in each shutdown period without changing the product is counted as the two frequency ratios, and then the linear relationship between the main traction machine speed and the main traction frequency is used to obtain the frequency of the extruder.

[0050] Among them, the statistical method mentioned in step S14 is used for feature extraction, including differential features, moving average features, moving average variance, decomposition features of STL time series, etc. The size of the moving average feature window is set according to the duration of the startup state, and generally the maximum is 1 / 2 of the startup duration.

[0051] Among them, in step S14, not only the collection characteristics and statistical characteristics of the current device are considered, but also because there is a correlation between the devices in the production line, when each device is modeled, the device characteristics of the previous link and the next link in the production line are also added to the selectable feature set.

[0052] The clustering algorithms mentioned in step S21 include KMeans, GMM, HMM, DBSCAN and other clustering algorithms.

[0053] Among them, when tuning the model in step S22, the model features do not only use the single features collected, but also incorporate the statistical features mentioned in S14 that characterize the distribution of collected data. The evaluation method is to combine drawing display with verification by on-site personnel.

[0054] Among them, in step S22, according to the HMM model parameter adjustment and parameter selection, the tuning is stopped when the accuracy of the duration of each working condition for n consecutive shifts reaches more than 90% of the on-site customer requirement and there is little room for further optimization of the model.

[0055] Among them, in step S3, the waste time of each link is counted, which not only uses the machine learning model to identify the status of a single device, but also combines the relationship between the equipment in the production line to accurately and real-time describe the working status of different links in the production line.

[0056] Among them, step S33 assists in the identification of traction blockage anomalies, which includes two parts. Due to the limitations of acquisition capabilities and cost control requirements, blockages that generally exceed 1 minute can be directly identified by the model. The blockage duration is obtained by subtracting the blockage start time from the auxiliary traction blockage end time identified by the model. If the blockage duration is less than 1 minute and the model cannot accurately identify it, it is judged by the changes in the magnetized and non-magnetized counts in the magnetization link.

[0057] Working process of the present invention:

[0058] During the working process of the scrap statistics and blockage abnormal alarm method based on production line status identification of the present invention, in order to facilitate understanding of the technical solution of the present invention, the production line status identification and scrap statistics service of a magnetic plastic extrusion production line in a factory in a real working environment are described in detail below.

[0059] A specific implementation example of the state identification and waste statistics of a factory's magnetic plastic extrusion production line. The entire implementation process is divided into two stages: training stage and online use stage. Figure 1 .

[0060] The specific process of the training phase is as follows:

[0061] Step 1.1: Data collection and processing: First, the gateway collects the device's real-time current signal data (the collected real-time current signal is a low-frequency current signal with a frequency of 1 second). The current signal includes a timestamp and current value. After data cleaning, the processed sample data is obtained. The data cleaning steps include: first, deleting values less than or equal to 0, then using the IQR method to determine outliers. For missing data and outliers, based on the characteristics of time series, forward (backward) filling is used to handle them.

[0062] Step 1.2: Determine the number of equipment processing states, analyze and determine the states of interest for each equipment, and through visual analysis of collected historical data combined with equipment status research, classify the equipment states of interest as follows: extruders and main haul-off machines have three states: {standby, start-up, and processing}, and auxiliary haul-off machines have four states: {standby, start-up, processing, and blocking};

[0063] Step 1.3: If the speed indicator cannot be collected for the extruder, the frequency of the main traction machine is obtained by combining the correspondence between the extruder frequency and speed with the relationship between the extrusion frequency and the main traction frequency using statistical and correlation methods. Specifically, after the status of each device is identified in step 1.5, the collected data is first analyzed to determine that the main traction frequency and speed are approximately linearly related, and a linear regression model can be constructed for fitting. When the main traction frequency and the extruder frequency are processing the same product and are both in the processing state, the frequencies are directly proportional, but when the products being processed are different, the ratio will vary. After the processing is completed during the shift, the highest frequency of the main traction machine and the extruder appearing simultaneously during each shutdown period without changing the product (changing the product requires changing the material, and the shutdown may exceed 10 minutes) is counted as the two frequency ratios. Then, the linear relationship between the main traction machine speed and the main traction frequency is used to obtain the frequency of the extruder.

[0064] Step 1.4: Feature extraction. In addition to using the processed raw current value as a feature, statistical methods are selected to generate new features based on the model effect, and new features are added based on the relationship between the equipment in the production line. Specifically, statistical features include features that describe the current distribution characteristics and changes, including differential features, moving average features (the window size is set according to the duration of the startup state, generally taking a maximum of 1 / 2 the startup duration), moving average variance, and STL time series decomposition features. In addition, the equipment features of the previous and next links in the production line are also added to the optional feature set. Together, they constitute all the optional features of the current equipment status.

[0065] Step 1.5: Use a clustering algorithm to cluster the equipment states of the extruder, main tractor, and auxiliary tractor. If the number of clusters needs to be set, the data for the different equipment operating conditions described in S12 will be used as the set value. Clustering algorithms include KMeans, GMM, HMM, and DBSCAN. KMeans, GMM, and HMM are trained on historical data from the past month or so. DBSCAN, due to its higher computational complexity, is trained on historical data from the past 10 days.

[0066] Step 1.6: Modeling and tuning. Through multiple rounds of adjustments to features and model parameters, the model with the best effect is selected as the state recognition model for each device. The model is selected first, and then the features and parameters are comprehensively tuned. Due to the complexity of the equipment status, after the first round of initial modeling using all features, HMM performed better for all three devices; the second round of parameter and feature selection was then used for parameter tuning, and the evaluation method was visualization combined with on-site working condition feedback results. Because verification is carried out through historical training data plotting and on-site verification, when the consistency between the duration of the model for three consecutive shifts and the on-site feedback of each device in the production line reaches more than 90% and it is difficult to optimize the model, the tuning is stopped. Finally, the model with the best effect is saved as a model file for subsequent calls.

[0067] Step 1.7: Next, based on the model results trained in 1.6, write the waste statistics logic for the extrusion, main traction, auxiliary traction, and magnetic threading stages. The core of this logic is to calculate the length of time that waste is generated based on the equipment status and its relationship in the production line, and then calculate the waste weight:

[0068] Waste weight = time of waste generation * speed of the link * density of the link

[0069] The speed is calculated based on the collected speed index or the frequency relationship, and the density of each link is provided after statistics on the client side.

[0070] Specific process of online use stage:

[0071] Step 2.1 Real-time acquisition and processing of new data: Every 30 seconds, the acquired data is processed according to the processing method of the training process and features are constructed;

[0072] Step 2.2 Call model file clustering: Call the saved model file to predict the different processing states of each device in the production line based on the feature set;

[0073] Step 2.3: Count the duration of each state of each device every 30 seconds and feed it back to the front-end display; count the weight of waste generated by each device every hour and feed it back to the front-end display; every 30 seconds, determine whether it is currently in a material blockage state based on the auxiliary traction machine status identification result, and if so, feed it back to the front-end for an alarm prompt.

[0074] The present invention adopts a machine learning algorithm to identify the status of extrusion equipment, main traction equipment, and auxiliary traction equipment in the magnetic plastic extrusion production line, and combines the magnetic PLC recording information to identify the duration of the status of different links in the production line through the business relationship between the status of each device, and then combines the collected speed and statistical density (meter length) data to calculate the weight of waste in each link and push it to the manager in real time. At the same time, the auxiliary traction equipment often causes blockage problems due to failures. The model can identify its status and issue a real-time blockage alarm to remind the workers in charge of the production line to deal with it in time, thereby improving the efficiency of the production line. This model only needs to collect the equipment spindle current signal by adding a gateway device, and does not need to modify the existing production line, so the collection cost is low. By adopting a machine learning model and combining the correlation between different equipment in the production line, the status of different links of the production line is monitored in real time, waste statistics, and abnormal alarms are performed, which helps managers to optimize production, evaluate employee performance, and deal with production line abnormalities in a timely manner, thereby improving work efficiency.

[0075] The above embodiments are intended to illustrate the present invention, not to limit the present invention. Any solution that is a simple transformation of the present invention falls within the protection scope of the present invention.

Claims

1. A method for waste statistics and abnormal blockage alarm based on production line status recognition, characterized by: The following stages are included: Step 1. Data collection and processing for each device in the production line includes the following steps: S11: Data collection and processing of each device. The real-time current signal data of the extrusion equipment, main traction equipment, and auxiliary traction equipment are collected and cleaned to obtain processed sample data. Outliers less than or equal to 0 are first deleted, and then the IQR method is used to determine outliers. For missing data and outliers, according to the characteristics of the time series, they are filled with the previous and next values. S12: Analyze and determine the status of each device. The extruder and main traction machine include three states: standby, start-up, and processing. The auxiliary traction machine includes four states: standby, start-up, processing, and blocking. S13: If the extruder can collect speed, then this speed index is directly used. If the extruder cannot collect speed index, the extrusion speed of the extruder is obtained by using the corresponding relationship between the main traction machine frequency and speed, and the corresponding relationship between the extruder frequency and the main traction frequency, and using statistical and correlation methods; S14: Feature extraction: In addition to using the processed original current value as a feature, a statistical method is selected based on the model effect to generate new features. New features are also added based on the relationship between the equipment in the production line. Step 2: State recognition modeling of the extruder, main tractor, and auxiliary tractor, including the following steps: S21: clustering the equipment states of the extruder, the main tractor, and the auxiliary tractor using a clustering algorithm. If the number of clusters needs to be set, the data of the different working conditions of the equipment described in S12 is the set value; S22: Modeling and tuning: Through multiple rounds of adjustments to features and model parameters, the best model is selected as the state recognition model for each device. Based on the HMM model parameter adjustment and parameter selection, tuning is stopped when the accuracy of the duration of each working condition for n consecutive shifts meets the on-site customer requirements and there is little room for further model optimization. Step 3: Calculate the weight of the extruded waste in the extrusion process, the weight of the cut waste in the main traction process, the weight of the waste generated by the blockage in the auxiliary traction process, and the weight of the waste that failed to be successfully magnetized in the magnetization process; alarm for abnormal blockage in the auxiliary traction process, including the following steps: S31: The extrusion waste extrusion time consists of two parts: First, when the extruder is shut down and started up, waste will be extruded for a period of time, and then pulled to the main traction machine after it stabilizes; second, when the extruder is shut down for a short period of time due to mold or material change during operation, waste will also be generated. The waste generation time of the extruder is the main traction start time minus the extruder start time; S32: waste cutting in the main traction link. At the beginning of production, product quality is unstable, so cutting will be performed in the main traction link. The duration of waste cutting in the main traction link is the auxiliary traction start start time minus the main traction start start time; S33: The duration of waste caused by blockage in the auxiliary traction link. This is caused by wear of the auxiliary traction machine itself or the cutter head of the connected cutting machine. During a production shift, multiple blockages may occur. The auxiliary traction machine blockage duration is calculated by subtracting the blockage start time from the blockage end time. S34: Auxiliary tractor blockage abnormal alarm. When the model identifies that the auxiliary tractor is in a blockage state, the model will push the result to the business end ERP for real-time alarm; S35: Statistics of the weight of waste products in the magnetizing process are obtained by summing up the lengths of the plastic strips corresponding to the collected non-magnetizing counts in sections.

2. The method for waste statistics and abnormal material blockage alarm based on production line status recognition according to claim 1, characterized in that: If the speed of the extruder cannot be collected in step S13, the collected data is first analyzed to obtain an approximate linear relationship between the main traction frequency and the speed, and a linear regression model can be constructed for fitting. When the main traction frequency and the extruder frequency are processing the same product and both are in the processing state, the frequencies are in a proportional relationship. However, when the processed products are different, the proportions will be different. After the processing is completed during the shift, the highest frequency corresponding value of the main traction machine and the extruder appearing simultaneously in each shutdown period without changing the product is counted as the two frequency ratios, and then the linear relationship between the main traction machine speed and the main traction frequency is used to obtain the frequency of the extruder.

3. The method for waste statistics and abnormal blockage alarm based on production line status recognition according to claim 1, characterized in that: The statistical method mentioned in step S14 is used to extract features, including differential features, moving average features, moving average variance, and decomposition features of the STL time series. The size of the moving average feature window is set according to the duration of the startup state, with a maximum of 1 / 2 of the startup duration.

4. The method for waste statistics and abnormal material blockage alarm based on production line status recognition according to claim 1 is characterized in that: In step S14, not only the collection features and statistical features of the current device are considered, but also, because there is a correlation between the devices in the production line, when each device is modeled, the device features of the previous and next links in the production line are also added to the selectable feature set.

5. The method for waste statistics and abnormal blockage alarm based on production line status recognition according to claim 2, characterized in that: The clustering algorithms mentioned in step S21 include KMeans, GMM, HMM, and DBSCAN clustering algorithms.

6. The method for waste statistics and abnormal material blockage alarm based on production line status recognition according to claim 5, characterized in that: When the model is tuned in step S22, the model features do not only use the single features collected, but also incorporate the statistical features mentioned in S14 that characterize the distribution of the collected data. The evaluation method is to combine drawing display with verification by on-site personnel.

7. The method for waste statistics and abnormal material blockage alarm based on production line status recognition according to claim 6, characterized in that: In step S22, according to the HMM model parameter adjustment and parameter selection, when the accuracy of the duration of each working condition for n consecutive shifts reaches more than 90% required by the on-site customer and there is little room for further optimization of the model, the tuning is stopped.

8. The method for waste statistics and abnormal material blockage alarm based on production line status recognition according to claim 3 is characterized in that: In step S3, the waste time of each link is counted. This not only uses the machine learning model to identify the status of a single device, but also combines the relationship between the equipment in the production line to accurately and real-time describe the working status of different links in the production line.

9. The method for waste statistics and abnormal material blockage alarm based on production line status recognition according to claim 3, characterized in that: Step S33 is the identification of auxiliary traction blockage anomalies, which includes two parts. Due to the limitations of acquisition capabilities and cost control requirements, blockages exceeding 1 minute are directly identified through the model, and the blockage duration is obtained by subtracting the blockage start time from the auxiliary traction blockage end time identified by the model. If the blockage duration is less than 1 minute and the model cannot accurately identify it, it is judged by the changes in the magnetized and non-magnetized counts in the magnetization link.

Images