Strip shape quality abnormality marking method and system based on extreme value theory

Abstract

This invention relates to the field of quality inspection technology, specifically a method for marking anomalies in strip and sheet shape quality based on extreme value theory. The method first collects time-series data of strip and sheet shape IU values ​​and divides it into initialization and streaming data sequences, configuring control parameters. Then, based on the initialization data, it calculates an initial threshold, constructs a baseline for normal behavior, and fits a generalized Pareto distribution to obtain an initial anomaly judgment threshold. Subsequently, it calculates the residuals of the streaming data point by point with the local mean, compares them with the threshold to mark anomaly points, and uses the residuals exceeding the initial threshold as new peak values ​​to update the peak set and anomaly judgment threshold. Finally, it dynamically judges and marks anomaly points in strip and sheet shape, adaptively updates the threshold model, and outputs anomaly point information and a dynamic threshold sequence. This method achieves accurate and real-time marking of strip and sheet shape quality anomalies, adapts to the streaming data processing requirements of cold continuous rolling production lines for strip and sheet, and improves the accuracy and recall rate of anomaly marking.

Description

Technical Field

[0001] This invention belongs to the field of cold-rolled strip quality inspection technology, specifically involving a method and system for marking abnormal strip shape quality based on extreme value theory. Background Technology

[0002] As a core quality evaluation indicator for cold-rolled strip, its quality directly determines the performance of the strip product, its suitability for subsequent processing, and its final pass rate. Accurate and real-time marking of strip shape quality abnormalities is a key technical support for timely adjustment of process parameters, control of product quality, and improvement of production efficiency in the large-scale industrial production of cold-rolled strip.

[0003] Cold-rolled strip production is characterized by continuous production line operation, high rolling speed, large coil output, and high frequency of time-series data collection for strip shape IU values. This places stringent demands on the automation, real-time performance, and accuracy of strip shape anomaly marking. In the early stages of cold-rolled production technology development, the identification of strip shape anomalies mainly relied on on-site technical experts to manually judge and identify the anomalies based on their production experience and the appearance of the strip. This method has significant limitations: firstly, manual judgment is inefficient, cannot cover all the collected strip shape data, and is prone to omissions leading to continuous strip shape anomalies, resulting in a large number of defective products and increased production costs; secondly, manual judgment is highly subjective, with different experts having subjective differences in their judgment standards, easily leading to inconsistent judgment results, and it is difficult to match the real-time monitoring needs of the production line, making it unsuitable for the high-quality control requirements of modern cold-rolled strip production.

[0004] To address the drawbacks of manual judgment, the industry has gradually introduced unsupervised labeling techniques based on statistics to automate the identification of sheet and strip shape quality anomalies. Among these, the box plot method proposed by statistician Tukey has become the most widely used method for marking shape anomalies in cold-rolled sheet and strip due to its simple calculation logic and low engineering implementation difficulty. This method uses the interquartile range (IQR) of the shape detection data as the core to define the anomaly judgment standard. When a certain shape data point is lower than the first quartile (Q1) minus 1.5 times the interquartile range, or higher than the third quartile (Q3) plus 1.5 times the interquartile range, it is judged as an anomaly. In combination with the actual production requirement that the normal shape value of sheet and strip should be as small as possible, the third quartile (Q3) is usually used as the core anomaly judgment threshold in practical applications.

[0005] However, in practical applications of marking abnormal shape quality in cold-rolled strip, the box plot method reveals several insurmountable technical flaws, leading to a significant decrease in the accuracy and applicability of the marking: First, the box plot method relies on fixed static thresholds for anomaly detection, which cannot adapt to the persistent characteristics of shape anomalies in cold-rolled production. Shape anomalies in cold-rolled strip often manifest as continuously deviating values ​​from the normal range within a certain time period. When such continuous deviations do not exceed the static threshold set by the box plot method, this persistent shape anomaly cannot be effectively identified, resulting in a large number of missed detections; Second, the theoretical foundation of the box plot method... Based on the strong assumption that the data follows a normal distribution, the actual distribution of the IU value data of the strip shape is significantly different from the normal distribution due to various factors such as roll wear, rolling force fluctuations, raw material composition differences, and production line speed adjustments during the production of cold continuous rolling strip. The threshold setting based on this assumption is prone to high false alarm and false negative rates of anomaly marking, which seriously affects the accuracy of anomaly marking. Thirdly, the box plot method performs statistical analysis on the overall distribution of data, but does not specifically capture the extreme value characteristics of the strip shape data. It is not sensitive to sudden and small-scale abnormal fluctuations in strip shape and is difficult to adapt to the complex and ever-changing production conditions of cold continuous rolling production lines.

[0006] In addition, existing methods for marking anomalies in sheet shape, such as principal component analysis (PCA) and independent component analysis (ICA), also have significant shortcomings in marking the sheet shape of cold-rolled strip: these methods have high requirements for the distribution characteristics of sheet shape data, and the model training and calculation process is complex, with poor real-time performance, making it difficult to match the streaming data processing needs of cold-rolled production lines; at the same time, their overall performance in marking anomalies is low, and they cannot balance precision and recall, making it difficult to promote and apply them on a large scale in engineering practice.

[0007] In summary, existing methods for marking anomalies in sheet and strip shape suffer from poor adaptability, low accuracy, and an inability to meet the real-time monitoring needs of production lines, making them unsuitable for the complex production conditions of cold-rolled sheet and strip. Therefore, there is an urgent need to develop a method for marking anomalies in sheet and strip shape that can overcome the limitations of the normal distribution assumption, dynamically adjust the judgment threshold, and balance precision and recall. This would enable accurate and real-time identification of shape anomalies, improve the quality control level of cold-rolled sheet and strip production, and meet the needs of large-scale industrial production. Summary of the Invention

[0008] To address the shortcomings of existing methods for marking shape quality anomalies in cold-rolled strip and sheet, such as the inability of static thresholds to handle persistent anomalies, reliance on normal distribution assumptions leading to false alarms and missed detections, weak anomaly detection capabilities, and poor real-time performance, this invention provides a method and system for marking shape quality anomalies in strip and sheet based on extreme value theory. This invention constructs a dynamic threshold mechanism based on extreme value theory, uses a generalized Pareto distribution to fit the extreme value characteristics of shape data, and combines an exponentially weighted moving average method to enhance sensitivity to data trend changes. This achieves accurate and real-time marking of shape anomalies, while also adapting to the streaming data processing requirements of cold-rolled production lines, significantly improving the accuracy and recall rate of anomaly marking, and meeting the quality control requirements of large-scale, high-precision production of cold-rolled strip and sheet.

[0009] The technical solution adopted by this invention to solve its technical problem is:

[0010] A method for marking abnormalities in the shape quality of strip and plate materials based on extreme value theory, the specific steps of which are as follows:

[0011] S1, Data preparation and parameter setting: Collect time-series data of the shape IU value of cold rolled strip; set the initial data collected as the initial data sequence, and set the subsequent continuously arriving data as the streaming data sequence; at the same time, configure a series of control parameters corresponding to the anomaly marking algorithm;

[0012] S2, Initial Threshold and Normal Behavior Baseline Construction: Based on the initial data sequence, its specified quantile is calculated as the initial threshold; at the same time, the sliding window technique is used to calculate the local mean sequence of the initial data sequence, and the local mean sequence is used to characterize the plate shape quality baseline behavior;

[0013] S3, Extreme fluctuation modeling: Fit a generalized Pareto distribution based on the initial set of peaks, and calculate the initial anomaly detection threshold; For each data point in the streaming data sequence, execute the following process iteratively;

[0014] S301, Calculate the residual between this data point and the current local mean;

[0015] S302, compare the residual with the current anomaly detection threshold, and mark it as an anomaly if it exceeds the threshold.

[0016] S303, if the residual exceeds the initial threshold, it is included as a new peak in the peak set, and the generalized Pareto distribution is refitted to update the anomaly detection threshold.

[0017] S4, Dynamic Anomaly Labeling in Streaming Data: For each data point in the streaming data sequence, the process is repeated cyclically to dynamically calculate the deviation between the data point and the current local normal behavior baseline; the deviation is compared with the anomaly judgment threshold dynamically updated based on extreme value theory and peak set to determine and label plate-shaped anomaly points; and the anomaly judgment threshold model is adaptively updated based on the comparison results.

[0018] S5, Output: Output the information of the marked plate shape anomalies and the corresponding dynamic threshold sequence.

[0019] As a preferred embodiment, a further technical solution of the present invention is:

[0020] Preferably, the series of control parameters in step S1 include risk levels used to control the probability of anomalies occurring. Initial threshold probability used to determine the initial anomaly detection boundary Length of the sliding window used to calculate local trends Exponentially weighted moving average smoothing factor is used to smooth data and sensitively capture trend changes. And an alarm switch for controlling whether to output alarm information in real time. .

[0021] Preferably, the initial threshold probability The value range is greater than 0.5 and less than 1.

[0022] Preferably, the specific formula for calculating the local mean sequence using the sliding window technique is as follows:

[0023] ;

[0024] in, The length of the sliding window. For initialization sequence The index is Data points, For the calculated first Local mean.

[0025] Preferably, in step S301, the current local mean is calculated. The method used is the exponentially weighted moving average method, and its calculation formula is as follows:

[0026] ;

[0027] in, As a smoothing factor, For the current streaming data point, This is the local mean of the previous time period. This is the local mean for the current time period.

[0028] Preferably, in step S3, the formula for calculating the initial anomaly detection threshold based on fitting a generalized Pareto distribution to the initial peak set is as follows:

[0029] ;

[0030] ;

[0031] in, As the initial threshold, For risk level parameters, is the scale parameter of the generalized Pareto distribution. The shape parameter of the generalized Pareto distribution. The upper threshold, The lower threshold is [value].

[0032] Preferred scaling parameters for the generalized Pareto distribution and shape parameters The peak set is obtained by fitting the peak set using the maximum likelihood estimation method and solved using the Grimshaw numerical optimization algorithm.

[0033] Preferably, the preprocessing of the initial data sequence in step S1 includes denoising and normalization. The denoising is used to remove impulse noise and acquisition error data in the time series data of plate shape IU value. The normalization uses minimum-maximum normalization to map the plate shape IU value to the [0,1] interval, eliminating the influence of the dimension on the abnormal marking.

[0034] Preferably, the initial peak set selection method in step S2 is as follows: calculate the residual sequence of the initial data sequence, and take the residual data that exceeds the preset high quantile and low quantile range of the residual sequence as the initial peaks and include them in the initial peak set.

[0035] This invention also discloses a strip shape quality anomaly marking system based on extreme value theory, used to implement a strip shape quality anomaly marking method based on extreme value theory, comprising: a data acquisition module for acquiring time-series data of the shape IU value of cold-rolled strip; a threshold calculation module for calculating and updating a dynamic threshold based on extreme value theory and a generalized Pareto distribution model; a trend smoothing module for smoothing the data using an exponentially weighted moving average algorithm and calculating the residual; and an anomaly judgment and marking module for comparing the residual with the dynamic threshold to complete the judgment, marking, and output of anomaly points.

[0036] The present invention, which adopts the above technical solution, has the following prominent features compared with the prior art:

[0037] 1. The extreme value theory is adopted to replace the traditional normal distribution statistical assumption. The extreme value characteristics of the plate shape IU value are modeled in a targeted manner through the generalized Pareto distribution. This achieves accurate fitting of the actual distribution law of cold rolled strip plate shape data, fundamentally solving the problem of false alarms and missed judgments caused by the distribution assumptions of traditional methods, and greatly improving the accuracy of anomaly marking.

[0038] 2. A dynamic threshold update mechanism is adopted to replace the fixed static threshold judgment method. By capturing the peak characteristics of the plate shape data in real time and updating the peak set and refitting the distribution parameters, the adaptive adjustment of the anomaly judgment threshold is realized. It can accurately identify sudden extreme value anomalies and effectively capture continuous plate shape deviation anomalies in production, adapting to the dynamic production conditions of the cold continuous rolling production line.

[0039] 3. The processing method adopts a combination of sliding window and exponentially weighted moving average. The sliding window is used to construct the baseline of normal behavior of plate shape quality. The exponentially weighted moving average method is used to smooth the flow data and give higher weight to recent data. This enables accurate capture of short-term and medium-term trend changes in plate shape data, making residual calculation more consistent with the real-time production status of the production line and improving the real-time performance and accuracy of anomaly marking.

[0040] 4. By adopting the judgment logic of upper and lower dynamic thresholds, and constructing judgment thresholds for high-value and low-value anomalies respectively, the bidirectional accurate identification of plate shape quality anomalies is realized. It effectively marks the low-value anomalies of persistently low plate shape values, which are more hidden in production, and fills the gap of the traditional method's insufficient ability to identify persistent low-value anomalies.

[0041] 5. The parameter solution method adopts the maximum likelihood estimation combined with the Grimshaw numerical optimization algorithm. By efficiently and accurately solving the scale and shape parameters of the generalized Pareto distribution, the distribution model can be quickly fitted and updated. With the simple algorithm logic, it takes into account both the detection performance of anomaly marking and engineering implementation. It can be adapted to the streaming data processing requirements of cold continuous rolling production lines and realize online real-time marking of plate shape anomalies. Attached Figure Description

[0042] Figure 1 This is a flowchart of the plate and strip shape quality anomaly marking method based on extreme value theory in an embodiment of the present invention;

[0043] Figure 2 This is a method for marking abnormalities in the shape quality of strip and plate materials based on extreme value theory in this embodiment of the invention (in the table and...). Figure 2 The image shows a visualization of the anomaly marking results of the IU value of a certain cold-rolled strip (represented by the EWMAPOT algorithm). Detailed Implementation

[0044] The present invention will be further illustrated below with reference to specific embodiments. The purpose of this illustration is solely to provide a better understanding of the invention. Therefore, the examples given do not limit the scope of protection of the present invention.

[0045] like Figure 1 As shown in the figure, this embodiment presents a method for marking abnormalities in the shape quality of strip materials based on extreme value theory. The specific steps are as follows:

[0046] S1, Data Preparation and Parameter Setting: Time-series data of the IU values ​​of cold-rolled strip are collected using a strip shape analyzer; the initial data is set as the initialization data sequence, and subsequent continuously arriving data is set as the streaming data sequence; simultaneously, a series of control parameters corresponding to the anomaly marking algorithm are configured. These control parameters include a risk level used to control the probability of anomalies occurring. Initial threshold probability used to determine the initial anomaly detection boundary Length of the sliding window used to calculate local trends Exponentially weighted moving average smoothing factor is used to smooth data and sensitively capture trend changes. And an alarm switch for controlling whether to output alarm information in real time. In this embodiment, the initial threshold probability The value range is greater than 0.5 and less than 1.

[0047] Specifically, the preprocessing of the initial data sequence includes denoising and normalization. Denoising is used to remove impulse noise and acquisition error data from the time series data of plate shape IU values. Normalization uses minimum-maximum normalization to map the plate shape IU values ​​to the [0,1] interval, eliminating the influence of the dimension on the anomaly labeling.

[0048] In this embodiment, time-series data of shape IU values ​​of cold-rolled strip and sheet from the 1780mm cold rolling production line in this plant were collected using a BSS-1200 shape analyzer. The collection period was 8 consecutive hours, and a total of 144,000 sets of shape IU value data were collected. The first 10,000 sets of data were set as the initial data sequence, and the remaining 134,000 sets of continuously arriving data were set as the streaming data sequence. At the same time, a series of control parameters corresponding to the anomaly marking algorithm were configured. The specific settings of each control parameter in this embodiment are as follows:

[0049] Risk level 1×10 -4 ;

[0050] Initial threshold probability 0.9;

[0051] Sliding window length :40;

[0052] Exponentially weighted moving average smoothing factor 0.5;

[0053] Alarm switch :True (Enable real-time alerts);

[0054] In this embodiment, the preprocessing of the initial data sequence includes denoising and normalization: the denoising process uses median filtering (window size 5) to remove impulse noise and acquisition error data in the time series data of plate shape IU value, effectively filtering out abnormal outliers caused by production line vibration and sensor interference; the normalization process uses minimum-maximum normalization to map the plate shape IU value to the [0,1] interval, eliminating the influence of the dimension on the outlier marking.

[0055] S2, Initial Threshold and Normal Behavior Baseline Construction: Based on the initial data sequence, the 0.9 quantile of the initial data sequence is calculated as the initial threshold. Simultaneously, a sliding window technique is used to calculate the local mean sequence of the initial data sequence, which is used to characterize the baseline behavior of the plate shape quality.

[0056] Specifically, the formula for calculating the local mean sequence using the sliding window technique is as follows:

[0057] ;

[0058] in, The length of the sliding window. For initialization sequence The index is Data points, For the calculated first Local mean.

[0059] In this embodiment, the initial peak set is selected as follows: the residual sequence of the initial data sequence is calculated (residual = initial data point - corresponding local mean), and the residual data that exceeds the range of the 0.98 high quantile and 0.02 low quantile of the residual sequence are taken as initial peaks and included in the initial peak set; after calculation, the 0.98 quantile of the residual sequence in this embodiment is 0.15 and the 0.02 quantile is -0.12, and a total of 217 initial peaks are selected to form the initial peak set peaks.

[0060] S3, Extreme Fluctuation Modeling: Based on fitting a generalized Pareto distribution to the initial peak set, the initial anomaly detection threshold is calculated. In this embodiment, the formula for calculating the initial anomaly detection threshold based on fitting a generalized Pareto distribution to the initial peak set is as follows:

[0061] ;

[0062] ;

[0063] in, As the initial threshold, For risk level parameters, is the scale parameter of the generalized Pareto distribution. The shape parameter of the generalized Pareto distribution. The upper threshold, The lower threshold is [value].

[0064] In this embodiment, the scale parameter of the generalized Pareto distribution is... and shape parameters The peak set is obtained by fitting the peak set using the maximum likelihood estimation method and solved using the Grimshaw numerical optimization algorithm.

[0065] Then, for each data point in the streaming data sequence, the following process is executed iteratively.

[0066] S301, calculate the residual between the data point and the current local mean.

[0067] In this embodiment, the current local mean is calculated. The method used is the exponentially weighted moving average method, and its calculation formula is as follows:

[0068] ;

[0069] in, As a smoothing factor, For the current streaming data point, This is the local mean of the previous time period. This is the local mean for the current time period.

[0070] S302, compare the residual with the current anomaly detection threshold, and mark it as an anomaly if it exceeds the threshold.

[0071] The residual calculated by S301 is compared with the current upper threshold. Lower threshold In comparison, if the residual > or residual If the data point is marked as an anomaly, an immediate alarm message will be sent to the production line control room through the Siemens TDC control system. The alarm message will include the timestamp of the anomaly, the plate shape IU value, the residual value, and the anomaly type (high value anomaly / low value anomaly).

[0072] S303 If the residual exceeds the initial threshold, it is included as a new peak in the peak set, and the generalized Pareto distribution is refitted to update the anomaly detection threshold.

[0073] If the absolute value of the residual exceeds the initial threshold, the residual is included in the peak set as a new peak. The generalized Pareto distribution is refitted using the maximum likelihood estimation method combined with the Grimshaw numerical optimization algorithm, the scale parameters and shape parameters are updated, and the upper and lower anomaly judgment thresholds are recalculated by substituting them into the threshold formula to achieve adaptive dynamic updating of the threshold. In this embodiment, the distribution parameters are refitted every 20 new peaks added to the peak set, taking into account both computational efficiency and the timeliness of threshold updates.

[0074] S4, Dynamic Anomaly Labeling in Streaming Data: For each data point in the streaming data sequence, S301-S303 are executed iteratively to dynamically calculate the deviation (residual) between the data point and the current local normal behavior baseline. The deviation is compared with the anomaly judgment threshold dynamically updated based on extreme value theory and peak set to determine and label plate-shaped anomalies. Based on the comparison results of the residual with the initial threshold and the anomaly judgment threshold, different threshold model update logics are triggered to achieve adaptive update of the anomaly judgment threshold model: If the residual exceeds the initial threshold but does not reach the current anomaly judgment threshold, the residual is included in the peak set as a new peak. The generalized Pareto distribution is refitted using the maximum likelihood estimation method combined with the Grimshaw numerical optimization algorithm to update the scale and shape parameters of the distribution. Then, the upper and lower anomaly judgment thresholds are re-solved using the threshold calculation formula to complete the dual update of the threshold model parameters and threshold results. If the residual does not exceed the initial threshold, the threshold model is not updated, and the current distribution parameters and anomaly judgment threshold remain unchanged.

[0075] The core of the anomaly detection threshold model in this invention is a generalized Pareto distribution (GPD) fitted to the peak set. The update of the threshold model is essentially a dynamic expansion of the peak set + refitting of the GPD distribution parameters + recalculation of the anomaly threshold. The specific update steps are completely bound to the triggering conditions, and are divided into two stages: trigger update and execution update. No update occurs without a trigger, ensuring the rationality and real-time nature of the update.

[0076] Step 1: Update the triggering conditions (the threshold model update will only be initiated if this condition is met).

[0077] When the residual of a streaming data point (the deviation of a data point from the current local mean) exceeds the initial threshold but does not reach the current anomaly detection threshold, the threshold model is updated. If the residual directly exceeds the anomaly detection threshold, it is only marked as an anomaly and is not included in the peak set, nor is the threshold model updated, to avoid extreme outliers interfering with model fitting. If the residual does not exceed the initial threshold, it is judged as normal data fluctuation, and the model is not updated.

[0078] Step 2: Perform threshold model update (3 core operations, executed in sequence)

[0079] ① Expand the peak set: The residual that triggers the update is taken as the new peak and added to the existing peak set to realize the dynamic expansion of the peak set, so that the peak set always fits the latest extreme fluctuation characteristics of the plate shape data;

[0080] ② Refit the GPD distribution parameters: Based on the expanded new peak set, the maximum likelihood estimation method is used to fit the generalized Pareto distribution, and the scale parameter σ and shape parameter ξ of the distribution are solved again by the Grimshaw numerical optimization algorithm to complete the update of the core parameters of the threshold model;

[0081] ③ Recalculate the anomaly detection threshold: Substitute the updated scale parameter σ and shape parameter ξ into the formula for calculating the anomaly detection threshold (upper threshold). Lower threshold By combining the risk level and the initial threshold, new upper and lower anomaly judgment thresholds are obtained, thus updating the final judgment result of the threshold model.

[0082] In this embodiment, for 134,000 sets of streaming data, a sliding window with a step size of 1 is used for point-by-point processing. The processing time for a single data point is ≤0.01s, which meets the 50Hz acquisition frequency requirement and realizes real-time online processing of streaming data. During the processing, a total of 326 plate shape anomaly points were identified, including 107 high-value anomaly points and 219 low-value anomaly points, effectively capturing the hidden anomaly of continuously low plate shape values ​​in the production line.

[0083] Figure 2 This is a visualization of the actual effect of using the method of this invention to mark anomalies in the shape IU value of a certain coil of cold-rolled strip. The horizontal axis represents the sequence number of continuous sample points of the shape IU value, and the vertical axis represents the shape IU value, a core indicator characterizing the quality of the strip shape. In the figure, the dynamic threshold lines are upper and lower threshold curves updated in real time by the algorithm based on extreme value theory and generalized Pareto distribution. The area between the two lines represents the normal fluctuation range of the shape IU value. Unlike traditional static thresholds, this dynamic threshold line can adaptively adjust according to the actual fluctuation characteristics of the production line shape data, adapting to the dynamic production conditions of the cold rolling production line. Anomalies are shape IU value data points distributed outside the upper and lower dynamic threshold lines. Low-value anomalies below the lower dynamic threshold line are hidden anomalies that are easily missed by traditional methods in cold rolling production. The algorithm of this invention achieves accurate bidirectional identification of high-value and low-value anomalies through dual dynamic threshold judgment logic. The number of false positives and false negatives in the figure is controlled to an extremely low range, which demonstrates the technical advantages of the algorithm of this invention in marking board shape anomalies, which has both high precision and high recall, and effectively solves the problem of high false positive and false negative rates in traditional methods.

[0084] from Figure 2It can be clearly seen that the algorithm of this invention can accurately capture the continuous deviation anomaly and sudden extreme value anomaly of the plate shape IU value in cold continuous rolling production. In particular, it realizes the effective marking of low value anomalies with continuously low plate shape values, and verifies the practicality and accuracy of the dynamic marking step of the flow cytometry anomaly.

[0085] S5, Output Results: Output the information of the marked plate shape anomalies and the corresponding dynamic threshold sequence. The anomaly information includes the anomaly index, timestamp, original plate shape IU value, local mean, residual value, anomaly type, and alarm time. The dynamic threshold sequence includes the upper threshold, lower threshold, and update time corresponding to each data point in the streaming data processing. The output results are stored in Excel format on the industrial control computer and simultaneously uploaded to the production line MES system to provide data support for process engineers to adjust rolling parameters (such as rolling force and bending force).

[0086] This invention also discloses a strip shape quality anomaly marking system based on extreme value theory, used to implement a strip shape quality anomaly marking method based on extreme value theory, including: a data acquisition module for acquiring time-series data of the shape IU value of cold-rolled strip; in this embodiment, a BSS-1200 shape meter and Siemens Profinet industrial bus are used to acquire the time-series data of the shape IU value of cold-rolled strip; a threshold calculation module for calculating and updating dynamic thresholds based on extreme value theory and a generalized Pareto distribution model; deployed on an industrial control computer, developed using Python, for calculating and updating dynamic thresholds based on extreme value theory and a generalized Pareto distribution model, integrating the parameter solving functions of the Grimshaw numerical optimization algorithm and the maximum likelihood estimation method; and a trend smoothing module for smoothing the data using an exponentially weighted moving average algorithm and calculating residuals; linked with the threshold calculation module, for smoothing the data using an exponentially weighted moving average algorithm and calculating residuals, supporting online adjustment of the smoothing factor (adjustment range 0.1~0.9). The anomaly detection and marking module is used to compare residuals with dynamic thresholds to determine, mark, and output anomalies. In this embodiment, it is used to interface with the Siemens TDC control system and MES system of the production line to compare residuals with dynamic thresholds, determine, mark, and output anomalies, and support remote control of alarm switches and multi-terminal push of anomaly information.

[0087] To verify the present invention's method for marking abnormalities in strip and sheet shape quality based on extreme value theory (in the table and...) Figure 2The anomaly labeling effect of the EWMAPOT algorithm is shown in Table 1. In this embodiment, the shape IU value data of 50 coils of strip steel from the cold rolling production line (each coil contains 2880 sets of data, totaling 144000 sets) is selected as the test set. The test set has been manually anomaly labeled by senior process experts of the production line and used as the real labels. The EWMAPOT algorithm of the present invention is compared with the box plot labeling method, POT algorithm and classic machine learning methods PCA and ICA commonly used in the industry. The test indicators are precision (%), recall (%) and F1 score (%), where F1 score is the comprehensive evaluation index. The test results are shown in Table 1.

[0088] Table 1

[0089] Algorithm Precision (%) Recall (%) F1 Score (%) Boxplot notation 100.00 68.06 80.99 PCA 61.87 20.06 30.30 ICA 58.58 12.32 20.36 POT 97.46 84.21 95.35 EWMAPOT algorithm 96.50 95.35 95.92

[0090] The comparison results above show that the EWMAPOT algorithm of this invention has significant performance advantages over other methods:

[0091] 1. Compared with the box plot labeling method, the algorithm of this invention improves the recall rate by 27.29 percentage points and the F1 score by 14.93 percentage points, solving the problem of missed detection of persistent anomalies and low-value anomalies by the traditional static threshold method;

[0092] 2. Compared with the POT algorithm, the precision of the algorithm of this invention is only slightly reduced by 0.96 percentage points, the recall rate is significantly improved by 11.14 percentage points, and the F1 score is improved by 5.57 percentage points. By introducing the exponential weighted moving average method, it achieves accurate capture of the trend changes of board shape data and greatly improves the comprehensiveness of anomaly identification.

[0093] 3. Compared with machine learning methods such as PCA and ICA, the algorithm of this invention achieves several times the improvement in precision, recall and F1 score. It does not require complex model training, has simple calculation logic, and is more suitable for the streaming data processing needs of cold rolling production lines.

[0094] This invention presents a method for marking abnormal shape in strip and sheet materials based on extreme value theory. Compared with traditional techniques, it achieves significant performance improvements in marking abnormal shape in cold-rolled strip and sheet materials. The specific technical effects are as follows:

[0095] 1. Significantly improved anomaly marking accuracy: For the 50 steel coils marked by experts, the EWMAPOT algorithm achieved a precision of 96.50%, a recall of 95.35%, and an F1 score of 95.92%, which is the best overall performance compared to the box plot method, POT algorithm, PCA, and ICA algorithm, and greatly reduces false alarms and false negatives.

[0096] 2. Adapts to the true distribution characteristics of plate shape data: Without relying on the assumption of normal distribution, it can accurately model the extreme value distribution of plate shape data by fitting the GPD, and can still accurately model the non-normally distributed plate shape IU value.

[0097] 3. Meets the real-time requirements of industry: The POT algorithm has low computational complexity, while the EWMAPOT algorithm optimizes the Grimshaw parameter search process and adopts a sliding window with a step size of 1, balancing computational efficiency and recall rate. It can process the flowing plate shape data of continuous cold rolling production, realize online dynamic threshold updates and anomaly marking, and is suitable for industrial large-scale production monitoring scenarios.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention. All equivalent changes made based on the description and drawings of the present invention are included within the scope of the present invention.

Claims

1. A method for marking abnormalities in the shape quality of strip and plate materials based on extreme value theory, characterized in that, The specific steps are as follows: S1, Data preparation and parameter setting: Collect time-series data of the shape IU value of cold rolled strip; set the initial data collected as the initial data sequence, and set the subsequent continuously arriving data as the streaming data sequence; at the same time, configure a series of control parameters corresponding to the anomaly marking algorithm; S2, Initial Threshold and Normal Behavior Baseline Construction: Based on the initial data sequence, its specified quantile is calculated as the initial threshold; at the same time, the sliding window technique is used to calculate the local mean sequence of the initial data sequence, and the local mean sequence is used to characterize the plate shape quality baseline behavior; S3, Extreme fluctuation modeling: Fitting a generalized Pareto distribution based on the initial peak set and calculating the initial anomaly detection threshold; For each data point in the streaming data sequence, the following process is executed repeatedly; S301, Calculate the residual between this data point and the current local mean; S302, compare the residual with the current anomaly detection threshold, and mark it as an anomaly if it exceeds the threshold. S303, if the residual exceeds the initial threshold, it is included as a new peak in the peak set, and the generalized Pareto distribution is refitted to update the anomaly detection threshold. S4, Dynamic Anomaly Labeling in Streaming Data: For each data point in the streaming data sequence, execute it cyclically and dynamically calculate the deviation between the data point and the current local normal behavior baseline. The deviation is compared with the anomaly detection threshold based on extreme value theory and peak set dynamic update to identify and mark plate shape anomalies; and the anomaly detection threshold model is adaptively updated according to the comparison results. S5, Output: Output the information of the marked plate shape anomalies and the corresponding dynamic threshold sequence.

2. The method for marking abnormal plate and strip shape quality based on extreme value theory according to claim 1, characterized in that: The series of control parameters in step S1 include risk levels used to control the probability of anomalies occurring. Initial threshold probability used to determine the initial anomaly detection boundary Length of the sliding window used to calculate local trends Exponentially weighted moving average smoothing factor is used to smooth data and sensitively capture trend changes. And an alarm switch for controlling whether to output alarm information in real time. .

3. The method for marking abnormal plate and strip shape quality based on extreme value theory according to claim 2, characterized in that: Initial threshold probability The value range is greater than 0.5 and less than 1.

4. The plate shape quality anomaly marking method based on extreme value theory according to claim 1, characterized in that: In step S2, the specific formula for calculating the local mean sequence using the sliding window technique is as follows: ； in, The length of the sliding window. For initialization sequence The index is Data points, For the calculated first Local mean.

5. The plate shape quality anomaly marking method based on extreme value theory according to claim 1, characterized in that: In step S301, the current local mean is calculated. The method used is the exponentially weighted moving average method, and its calculation formula is as follows: ； in, As a smoothing factor, For the current streaming data point, This is the local mean of the previous time period. This is the local mean for the current time period.

6. The plate shape quality anomaly marking method based on extreme value theory according to claim 1, characterized in that: In step S3, the formula for calculating the initial anomaly detection threshold based on fitting a generalized Pareto distribution to the initial peak set is as follows: ； ； in, As the initial threshold, For risk level parameters, is the scale parameter of the generalized Pareto distribution. The shape parameter of the generalized Pareto distribution. The upper threshold, The lower threshold is [value].

7. The plate shape quality anomaly marking method based on extreme value theory according to claim 6, characterized in that: Scale parameter of generalized Pareto distribution and shape parameters The peak set is obtained by fitting the peak set using the maximum likelihood estimation method and solved using the Grimshaw numerical optimization algorithm.

8. The method for marking abnormal plate and strip shape quality based on extreme value theory according to claim 1, characterized in that: The preprocessing of the initial data sequence in step S1 includes denoising and normalization. Denoising is used to remove impulse noise and acquisition error data from the time series data of plate shape IU values. Normalization uses minimum-maximum normalization to map the plate shape IU values ​​to the [0,1] interval, eliminating the influence of the dimension on the anomaly label.

9. The method for marking abnormal shape quality of strip and plate materials based on extreme value theory according to claim 1, characterized in that, The initial peak set selection method in step S2 is as follows: calculate the residual sequence of the initial data sequence, and take the residual data that exceeds the preset high quantile and low quantile range of the residual sequence as the initial peaks and include them in the initial peak set.

10. A plate and strip shape quality anomaly marking system based on extreme value theory, used to implement the plate and strip shape quality anomaly marking method based on extreme value theory as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire time-series data of the shape IU value of cold rolled strip; The threshold calculation module is used to calculate and update dynamic thresholds based on extreme value theory and the generalized Pareto distribution model. The trend smoothing module is used to smooth the data and calculate the residuals using an exponentially weighted moving average algorithm. The anomaly detection and marking module is used to compare residuals with dynamic thresholds to complete the detection, marking, and output of anomalies.

Images