Cigarette making machine HID detection cap abnormity diagnosis and early warning method based on center offset, difference and dimension reduction fusion

Abstract

The invention belongs to the field of cigarette manufacturing, rolling and packaging production intelligentization, and particularly relates to a cigarette making machine HID detection cap abnormity diagnosis and early warning method based on center offset, difference and dimension reduction fusion, which comprises the following steps: (1) calculating the center point positions of 36 detection curves in a set time window, carrying out preliminary screening on the curves with larger center offset, and carrying out secondary screening on the curves with larger center offset; and locking the potential anomaly detection cap. And (2) extracting a standard curve in the window through dimensionality reduction, carrying out differential comparison on the standard curve and other curves one by one, quantifying the deviation degree of each curve relative to the standard curve, and revealing trend outlier characteristics. And (3) fusing the center offset score and the tendency outlier score according to the weight to form a comprehensive scoring model which is used for accurately judging and identifying a real anomaly detection cap. According to the method, the accuracy and real-time performance of abnormal recognition of the detection cap are effectively improved by fusing two types of analysis of center offset and difference-dimension reduction and combining weighted scoring, and intelligent support is provided for equipment maintenance and quality control of the cigarette making machine.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent cigarette manufacturing and packaging production, and relates to online quality control of the packaging process and intelligent diagnosis of equipment. Specifically, it relates to a method for diagnosing and warning abnormalities of the HID detection cap of a cigarette machine based on the fusion of center offset, difference and dimensionality reduction. Background Art

[0002] During cigarette production, ventilation is a key quality indicator. Ventilation testing relies primarily on a detection cap within a ventilation test device. This cap is typically attached to the device's air pressure probe. It forms a seal or partial contact with the cigarette, measuring changes in air pressure as air enters and leaves the cigarette.

[0003] Because the detection cap frequently contacts cigarettes and mechanical components, it is prone to wear and tear. This can lead to errors in air pressure measurement and affect the accuracy of ventilation testing. Therefore, real-time monitoring of the working status and wear of the detection cap is crucial to ensuring the stability of the detection system and product quality.

[0004] Traditional single-data analysis methods (such as principal component analysis (PCA)) often struggle to balance sensitivity and stability when faced with complex fluctuations in the cap's state and high levels of data noise, leading to the risk of misjudgment or missed detection. To effectively capture performance changes caused by aging, wear, or damage to the cap, an effective anomaly diagnosis and early warning method must be developed.

[0005] In order to solve the above problems, the present invention is proposed. Summary of the Invention

[0006] The present invention achieves online intelligent diagnosis of abnormalities in the test cap through in-depth analysis of test curve data from the vent slot area of the test cap. Simultaneously, combined with real-time ventilation rejection data, early warning and alarms are issued for spare parts, effectively reducing quality misjudgments caused by equipment aging or damage, further improving production efficiency, ensuring stable product quality, and reducing economic losses caused by abnormalities.

[0007] The specific technical solutions mainly include:

[0008] Center shift analysis algorithm: Calculates the center point positions of all detection curves within a set time window, identifies curves with larger center shifts among the current 36 detection curves, and thus preliminarily screens for potential abnormal detection caps;

[0009] Difference-dimensionality reduction fusion analysis algorithm: By performing dimensionality reduction processing on the data within the time window, the current standard curve is extracted and differentially compared with other curves one by one to quantify the deviation between each curve and the standard curve;

[0010] Weighted Merge Scoring Algorithm: The center shift score and the trend outlier score are weighted and fused to form a comprehensive scoring model, which is used to accurately determine the outlier curve and identify the anomaly detection hat.

[0011] Through the above-mentioned multi-dimensional analysis and fusion judgment method, the accuracy and real-time performance of detection cap anomaly recognition are effectively improved, providing intelligent support for equipment maintenance and quality control.

[0012] In order to achieve the above purpose of the invention, the present invention provides the following technical solutions:

[0013] A method for diagnosing and warning abnormalities of HID detection caps of cigarette making machines based on the fusion of center shift, difference and dimensionality reduction is provided, which includes the following steps:

[0014] Step (1) collects the operating data of the detection cap, and extracts data points related to ventilation through the Pearson correlation coefficient and manual experience, which includes 36 real-time variable parameters of the detection cap-related equipment, the number of ventilation rejections on duty, the rejection rate, the ventilation setting value, etc., totaling 46 dimensions of valid data, and performs sliding window segmentation and normalization processing on the valid data to construct a time-feature matrix, in which rows represent time points and columns represent feature numbers;

[0015] Step (2) generates a detection cap data curve according to the time window for the data after sliding window segmentation and normalization processing in step (1), takes the median of each row of data as the center point, calculates the difference between each detection cap data curve and the center point, defines the curve with the largest difference as an outlier curve, and calculates the center point offset score based on it;

[0016] Step (3), performing first-order difference processing on the detection cap data curve generated in step (2), and then applying principal component analysis to the differential data to fuse the 36 detection cap data curves to reduce the dimension to 1 dimension, and using this curve as the 1-dimensional standard curve, calculating the difference between all detection cap data curves and the 1-dimensional standard curve, and calculating the trend outlier score based on this;

[0017] Step (4): The center shift outlier score obtained in step (2) and the trend outlier score obtained in step (3) are weighted and fused according to the preset weights to generate the final comprehensive score. Then, the threshold is dynamically calculated based on the distribution characteristics of the comprehensive score.

[0018] Step (5) combines the dynamic threshold and real-time ventilation degree to perform data diagnosis and early warning detection cap status, which is specifically divided into the following three situations:

[0019] (1) The algorithm does not return the curve ID for determining outliers, that is, there is no outlier curve. Regardless of whether the average rejection rate in the last 10 minutes of the shift is greater than the rejection rate threshold corresponding to the machine, the result returned is: the detection unit is normal;

[0020] (2) The algorithm returns the curve ID that determines the outlier, that is, there is an outlier curve, but the average rejection rate in the last 10 minutes of the shift is less than or equal to the rejection rate threshold corresponding to the machine. The result is returned as: Detection cap XX (ID corresponds to the detection cap) may be aging, please check;

[0021] (3) The algorithm returns the id of the curve that determines the outlier, that is, if there is an outlier curve, the average rejection rate in the last 10 minutes of the shift is greater than the rejection rate threshold corresponding to the machine, and the result is returned as: detection cap XX (id corresponding to the detection cap) is abnormal, please replace it in time;

[0022] Preferably, the specific process of step (1) is as follows:

[0023] Step 1.1: Collect data from the past two months through the data acquisition system. Using the Pearson correlation coefficient and manual experience, extract 46 dimensions of valid data related to ventilation, including 36 test drum slots, ventilation rejection counts, rejection rates, and ventilation setpoints. The reduced data is then constructed into a data matrix X = m*n with n columns of features and m rows of time points, where m represents the time point and n represents the number of features.

[0024] Step 1.2: Segment the data in the data matrix obtained in step 1.1 using a sliding window with a window length of 600 seconds, including the data of 36 drum slots and the number of rejections during the shift, the rejection rate, the ventilation setting value, etc. Then, adjust the processed data to the range of [0, 1] using the minimum-maximum normalization method. The minimum-maximum normalization equation is:

[0025]

[0026] Where X represents the data matrix in step 1.1, x ij is the original data value of X in the jth column and i row, min(X j ) is the minimum value of X in the jth column; max(X j ) is the maximum value of X in the jth column, x′ ij is the normalized data value, and finally x′ ij Form a new data matrix X'.

[0027] Preferably, in step (2), the 36 detection cap data obtained in step 1.1 are segmented according to a predetermined time window, and a detection cap data curve of each segmented data over time is drawn using a line graph method, and then the deviation score of each detection cap curve from the center point is calculated;

[0028] The method for selecting the center point is as follows: let the element of matrix X′ at time point i, that is, row i and column j, be d i1 , di2 , d i3 ..., d ij , then the median of row i is M i =Median(d i1 , d i2 , d i3 ..., d ij ), where d ij Represents the elements in the matrix X′, M i Indicates the median of the i-th row, which is also the center point.

[0029] Preferably, the center point offset scoring method is:

[0030] Step 1: For the element d in row i and column j ij Calculate the absolute value of the difference between it and the median, expressed as: absdiff ij =|d ij -M i |, where d ij represents the element in the i-th row and j-th column of the matrix X′, M i is the median of the i-th row obtained above, absdiff ij It represents the absolute value of the difference between the two;

[0031] Step 2: Calculate the absdiff within the current time point range in column j ij The mean of is expressed as: in, Represents the absdiff within the current time point range in column j ij of and, Indicates finding the mean of row i, avg j That is, it is represented by the absdiff within the current time point range in the jth column ij The mean of

[0032] Step 3: Calculate the judgment value p i,j , represents the curve outlier situation, and its formula is:

[0033]

[0034] Among them, absdiff ij Represents the element d in row i and column j ij The absolute value of the difference from the row median, avg j Represents the absdiff within the current time point range in column j ij The mean of , plus 1×10 -6 This is to prevent the denominator from being 0; t is the preset threshold parameter, only absdiff ij with avg jWhen the ratio is greater than the threshold parameter, p i,j Not 0, otherwise 0, where greater than 0 indicates outlier, otherwise not outlier;

[0035] The fourth step is to count outliers at each time point. j The corresponding p i,j The relationship is:

[0036] outlier_count j =outlier_count j +p i,j

[0037] Among them, outlier_count j Indicates the cumulative calculation of outliers in the j-column curve, p i,j It represents the judgment value of the j column curve in the i row, that is, the judgment of the outlier of the detection hat data curve;

[0038] If the judgment value p of a characteristic curve j at the current time point i i,j is greater than 0, then the curve j is in its outlier count outlier_count j Add x i,j , otherwise it is 0;

[0039] Step 5. Finally, after calculating the outlier counts for each detection hat data curve in the entire time interval, the center point offset score based on the outlier counts is calculated. The formula is:

[0040]

[0041] Among them, outlier_count j is the outlier count of the j-th column curve, ∑outlier_count j The sum of the outlier counts for all curves, median_scores j Score the center point offset of the jth column curve.

[0042] Preferably, in step (3), the method for calculating the trend outlier score is:

[0043] Step 1: Use principal component analysis to reduce the dimension and reduce the number of detection cap data curves to 1. The process of calculating the 1D standard curve is as follows:

[0044] (1) Perform differential processing on the detection cap data curve and extract the trend of each curve, which is expressed as: in is the result after the first-order difference, X′ i,j is the value of the jth feature at the i-th time point in the normalized data, X′i-1,j is the value of the jth feature at the i-1th time point in the normalized data, that is, the value of the previous time point;

[0045] (2) Find the mean of the difference in each column, expressed as: Where x′ i represents the difference operation result at the i-th time point, represents the sum of all difference results, represents the mean of the difference results of each column;

[0046] (3) Obtain the covariance Cov(X,Y) between each column of the detection hat data curve at the i-th time point, and calculate all columns in turn to form the covariance matrix C;

[0047] (4) Use the eigenvalue decomposition method to find the eigenvalues and eigenvectors of the covariance matrix, expressed as: Cv i =λ i v i , where C represents the covariance matrix, λ i represents the eigenvalue, v i represents the eigenvector;

[0048] For the eigenvalue λ i The acquisition is expressed as: det(C-λI)=0, where det represents the determinant, C represents the covariance matrix, I represents the identity matrix, and λ is the eigenvalue;

[0049] (5) For each eigenvalue λ i , the eigenvector v i It can be obtained through a linear equation, expressed as: (C-λ i I)v i =0, where C represents the covariance matrix, λ i Represents each eigenvalue, I represents the unit matrix; normalizing the eigenvector is also to normalize the eigenvector, which is expressed as: Among them, v i is the original eigenvector, indicating the trend of data change in a certain direction, ||v i || is the eigenvector v i The norm of the vector, which represents the length of the vector, u i Represents the normalized eigenvector, which has a length of 1 and the same direction as the original vector;

[0050] (6) Generation of 1-dimensional standard curve: y = V1 T X, where y represents the 1-dimensional data after dimensionality reduction, which is the projection value of the original data on the principal component direction; V1 represents the first principal component vector, indicating that the variance of the data in this direction is the largest; X represents the original data point of the detection hat curve, which is a 36-dimensional vector.

[0051] Step 2: Calculate the trend outlier score. The specific process is as follows:

[0052] (1) Based on the 1D standard curve generated above, calculate the difference between each 1D standard curve and the 1D standard curve value, expressed as: fitdata i,j =|D col (X′) i,j -pca(X′) i |, where fitdata i,j For each

[0053] The absolute value of the difference result at each time point of the curve, D col (X′) i,j is the result of the first-order difference, pca(X′) i is the value of the standard curve at time point i;

[0054] (2) Calculate the trend outlier score diff_scores based on the difference between the above detection cap data curve and the 1D standard curve value j , the formula is:

[0055]

[0056] in, Expressed as the mean of the current curve gap values, Expressed as the mean of all curve difference values, diff_scores j The larger the value, the higher the degree of outlier the curve is.

[0057] Preferably, in step (4), the calculation process of the final comprehensive score is:

[0058] The first step is to perform a weighted combination of the center point offset score and the trend outlier score. The formula is:

[0059] scores j =diff_scores j ×w1+median_scores j ×w2

[0060] Among them, scores j is the final score of the curve, diff_scores j Scoring trend outliers, median_scores j The center point offset score is calculated. w1 and w2 are the preset weight parameters of the two scores, and their values are w1+w2=1. w1 and w2 are the scale differences for comparison, ensuring that the scales of the two algorithms are the same after taking the weights.

[0061] Step 2: Talk about the scores obtained above j Sort from largest to smallest and calculate scores j The standard deviation scores_std and mean scores_means of the array are used as the basis to calculate the threshold T. The formula is:

[0062] T=scores_std×sensitivity+scores_means

[0063] Among them, sensitivity is the preset sensitivity parameter. The larger the value, the larger the T value and the fewer outlier curves.

[0064] The present invention has the following beneficial effects:

[0065] 1. The present invention utilizes an outlier detection method based on center shift and differential-dimensionality reduction fusion. Through real-time analysis of detection drum groove data, it can quickly determine online whether there are any abnormalities in the detection cap. Combined with real-time ventilation rejection data, early warning and alarm are performed, which helps to promptly discover abnormalities caused by equipment aging or damage, so as to take measures before the problem evolves.

[0066] 2. Since abnormalities in the detection cap often lead to false rejection or quality misjudgment, this method can significantly reduce false rejections caused by equipment wear or damage by accurately determining outlier data, thereby ensuring the accuracy of ventilation detection, stabilizing product quality and improving production efficiency.

[0067] 3. A real-time and effective early warning mechanism enables early replacement or maintenance of spare parts, avoiding larger-scale production downtime or quality problems caused by abnormal accumulation, thereby reducing economic losses caused by abnormalities.

[0068] 4. By calculating the center point of all detection curves within a given time window, and then judging whether the current detection cap data curve deviates greatly from the center point, potential abnormal curves can be preliminarily screened out, reflecting the fluctuation of the detection cap working status in real time.

[0069] 5. Perform first-order difference processing on the original data within the time window, and obtain the current standard curve through dimensionality reduction technology. Then, by step-by-step difference comparison, the difference value between the detection curve and the standard curve is formed to form a trend outlier score, further capture the data trend changes, and accurately identify abnormal data that deviates from the standard trend.

[0070] 6. The center shift score and the trend outlier score are weighted and combined to form the final comprehensive score. Through this combined scoring method, the degree of deviation of each curve from the overall data trend can be more comprehensively measured, and the abnormal detection caps can be accurately determined, thereby achieving accurate identification of outlier curves. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Schematic diagram of the process of abnormality diagnosis and early warning method of the present invention;

[0072] Figure 2 This is a schematic diagram of the center shift outlier scoring process in the abnormality diagnosis and early warning method of the present invention;

[0073] Figure 3 Schematic diagram of abnormal air flow detection and analysis in the abnormality diagnosis and early warning method of the present invention;

[0074] Figure 4 This is a schematic diagram of the outlier curve model test in a specific implementation manner;

[0075] Figure 5 Schematic diagram of the rejection rate test corresponding to the outlier curve model in a specific implementation manner;

[0076] Figure 6 Schematic diagram of the curve of online detection cap (abnormality) in a specific implementation manner;

[0077] Figure 7 Schematic diagram of the rejection rate of the online detection cap (abnormal) corresponding to the ventilation degree in a specific implementation method;

[0078] Figure 8 In a specific embodiment Figure 8 Schematic diagram of the online detection cap (aging) curve;

[0079] Figure 9 Schematic diagram of the rejection rate of the online detection cap (aging) corresponding to the ventilation degree in the specific implementation method. DETAILED DESCRIPTION

[0080] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the accompanying drawings and embodiments are not intended to limit the technical solutions of the present invention.

[0081] Example 1

[0082] Ventilation is an important indicator of cigarette quality. During the production process of the cigarette wrapping machine, the detection of ventilation depends on the determination of the air pressure of the detection cap. However, the detection cap is a spare part that is easily worn during production. There is no means to monitor the damage of the detection cap. Once the detection cap is damaged, a large number of air pressure misjudgments will occur, and cigarettes with abnormal ventilation quality will be mistakenly rejected.

[0083] Therefore, it is necessary to conduct real-time analysis of abnormal conditions of the detection cap during the production process of the wrapping machine to solve the problem of a large number of cigarettes with abnormal ventilation being mistakenly rejected after the detection cap is damaged. This not only improves production efficiency, but also improves the stability and reliability of the entire production.

[0084] To this end, through mining relevant data, we found that the fluctuation data in the ventilation slot area of the test cap can represent the fluctuation of the test cap. At the same time, through extensive data research and comparison of 36 sets of representative data, we found that the combination of data outlier detection and trend detection can effectively capture the aging and damage of the test cap.

[0085] In actual deployment, the abnormal situations of detection caps are divided into:

[0086] (1) There is no outlier curve, indicating that the detection unit is normal;

[0087] (2) There is an outlier curve, but the 10-minute average rejection rate is less than or equal to the rejection rate threshold corresponding to the machine, which means that the detection cap XX may be aging;

[0088] (3) There is an outlier curve, and the 10-minute average rejection rate is greater than the rejection rate threshold corresponding to the machine, which means that the detection cap XX is abnormal;

[0089] Specific reference Figure 1-3 , provides a cigarette machine HID detection cap abnormality diagnosis and early warning method based on center offset, difference and dimensionality reduction fusion, including the following steps:

[0090] Step 1: Through the data acquisition system, all collected data for the past two months, totaling more than 1,300 dimensions of data, were obtained. Using the Pearson correlation coefficient and manual experience, data points related to ventilation were extracted, including 46 dimensions of valid data, including real-time variable parameters of 36 inspection cap-related equipment, the number of ventilation rejections during the shift, the rejection rate, and the ventilation setting value. The valid data was then subjected to sliding window segmentation and normalization preprocessing. Based on the time dimension, a data matrix X = m*n with n columns of features and m rows of time points was constructed, where rows represent time points and columns represent the number of features.

[0091] The data constructed above is further segmented by sliding windows, and the data is segmented into 600-second time intervals. Then the filtered data is normalized and adjusted to the range of [0, 1] to eliminate the dimensional effects between different features and make different features have the same scale. The normalization method is the min-max method. Where, X represents the aforementioned data matrix, x ij is the original data value of X in the jth column and i row, min(X j ) is the minimum value of X in the jth column. max(X j ) is the maximum value of X in the jth column, x′ ij is the normalized data value. Finally, x′ ij Form a new data matrix X'.

[0092] Step 2: Generate a detection cap data curve based on the normalized data from step 1 according to the time window. Take the median of each row of data as the center point, calculate the difference between each curve and the center point, define the curve with the largest difference as an outlier curve, and calculate the center point offset score based on this. The specific steps are as follows:

[0093] Step 2.1: Generate a data curve. Draw a curve for the data of the 36 detection caps divided by time window. The purpose of forming a curve is to analyze the entire data within the time window and avoid the sporadic nature of the judgment.

[0094] Step 2.2: Perform center point shift analysis. The core idea is to use the median of the data points as the center point. Then, compare the difference between each curve and the center point. The curve with the largest difference is defined as an outlier curve. The process is as follows:

[0095] Step 2.2.1, center point selection. For the data matrix X′ preprocessed in step 1.2, let the element of matrix X′ at time point i, i.e. row i, be d i1 , d i2 , d i3 ..., d ij , then the median of row i is M i =Median(d i1 , d i2 , d i3 ..., d ij ), where d ij Represents the elements in the matrix X′, M i Represents the median of the i-th row, and its purpose is to find the center point of the curve.

[0096] Step 2.2.2 Outlier Counting. If the judgment value x of a certain characteristic curve j at the current time point i i,j is greater than 0, then the curve j is in its outlier count outlier_count j Accumulate x i,j , otherwise it is 0. The following are the details of the steps:

[0097] The first step is to find the element d in row i and column j. ij Calculate the absolute value of the difference between it and the row median calculated in step 2.1.1, expressed as: absdiff ij =|d ij -M i |. Among them, d ij represents the element in the i-th row and j-th column of the matrix X′, M i is the median of the i-th row obtained in the previous step, absdiff ij It represents the absolute value of the difference between the two.

[0098] The second step is to calculate the absdiff within the current time point range in column j. ij The mean of is expressed as: in, Represents the absdiff within the current time point range in column j ij of and, Indicates finding the mean of row i, avg j That is, it is represented by the absdiff within the current time point range in the jth column ij The mean of .

[0099] The third step is to calculate the judgment value p i,j , represents the outlier condition of the curve. If it is greater than 0, it means outlier, otherwise it is not outlier. The formula is:

[0100]

[0101] Among them, absdiff ij Represents the element d in row i and column j ij The absolute value of the difference from the row median, avg j Represents the absdiff within the current time point range in column j ij The mean of , plus 1×10 -6 This is to prevent the denominator from being 0. t is the preset threshold parameter, only absdiff ij with avg j When the ratio is greater than the threshold, p i,j Not 0, otherwise 0.

[0102] The fourth step is outlier_count at each time point j p corresponding to the outlier count i,j , the formula is:

[0103] outlier_count j =outlier_count j +p i,j

[0104] Among them, outlier_count j Indicates the cumulative calculation of outliers in the j-column curve, p i,j It represents the judgment value of the j-column curve in the i-row, that is, the outlier judgment of the 36 curves.

[0105] Finally, after calculating the outlier counts for each characteristic curve in the entire time interval, the results based on the outlier counts are scored using the following formula:

[0106]

[0107] Among them, outlier_count j is the outlier count of the j-th column curve, ∑outlier_count j The sum of the outlier counts for all curves, median_scores j Score the center point offset of the jth column curve.

[0108] Step 3: Use difference-dimensionality reduction fusion analysis to fuse the 36 curves and reduce their dimension to 1. Use this curve as the standard curve, analyze and compare the gaps between all curves and the standard curve, and form a gap score. The steps are as follows:

[0109] Step 3.1 First, perform differential processing. For the curves drawn in step 2.1, perform differential processing on each curve. This method can extract the trend of each curve, which can be expressed as: in, is the result after the first-order difference, X′ i,j is the value of the jth feature at the i-th time point in the original data, X′ i-1,j It is the value of the jth feature at the i-1th time point (the value of the previous time point) in the original data.

[0110] Step 3.2: Based on the difference calculation results obtained in step 3.1, use principal component analysis (PCA) to reduce the dimension and reduce the number of characteristic curves to 1, which is the standard curve. The process is as follows:

[0111] The first step is to find the mean of the difference in each column, which is expressed as: Among them, x i represents the difference operation result at the i-th time point, represents the sum of all difference results, Represents the mean of the difference results of each column.

[0112] The second step is to find the covariance Cov(X,Y) between each column of the curve at the i-th time point. All columns are calculated in sequence to form the covariance matrix C.

[0113] The third step is to use the eigenvalue decomposition method to find the eigenvalues and eigenvectors of the covariance matrix, expressed as: Cv i =λ i v i . Where C represents the covariance matrix, λ i represents the eigenvalue, v i represents the feature vector.

[0114] The acquisition of the eigenvalue λ is expressed as: det(C-λI) = 0. Here, det represents the determinant, C represents the covariance matrix, I represents the identity matrix, and λ is the eigenvalue.

[0115] For each eigenvalue λ i , the eigenvector v i It can be obtained through a linear equation, expressed as: (C-λ i I)v i = 0. Where C represents the covariance matrix, λ i represents each eigenvalue and I represents the identity matrix.

[0116] The fourth step is to normalize the feature vector, which is expressed as: Among them, v i is the original eigenvector, indicating the trend of data change in a certain direction, ||v i || is the eigenvector v i The norm of the vector, which represents the length of the vector, u i Represents the normalized eigenvector, which has a length of 1 and the same direction as the original vector.

[0117] Step 5: Generate a 1D standard curve: y = V1 T X, where y represents the 1-dimensional data after dimensionality reduction, which is the projection value of the original data on the principal component direction; V1 represents the first principal component vector, indicating that the variance of the data in this direction is the largest; X represents the original data point of the detection hat curve, which is a 36-dimensional vector.

[0118] Step 3.3: Based on the standard curve generated in step 3.2, calculate the difference between each curve and the standard curve value, expressed as: fitdata i,j =|D col (X′) i,j -pca(X′) i |. Among them, fitdata i,j is the absolute value of the difference result at each time point of each curve, D col (X′) i,j is the result of the first-order difference in step 3.1, pca(X′) i is the value of the standard curve at time point i.

[0119] Step 3.4 calculates the trend outlier score based on the difference between each curve and the standard curve calculated in step 3.3. The formula is:

[0120]

[0121] in, Expressed as the mean of the current curve gap values, Expressed as the mean of all curve difference values, diff_scores j The larger the value, the higher the degree of outlier the curve is.

[0122] Step 4: Perform weighted merging of the scoring results in Step 2 and Step 3, and finally output the algorithm merging result to accurately determine the abnormality of the detection cap.

[0123] Step 4.1 performs weighted merging, that is, weighted merging of the center point offset score and the trend outlier score. The formula is:

[0124] scores j =diff_scores j ×w1+median_scores j ×w2

[0125] Among them, scores j is the final score of the curve, diff_scores j Score the curve trend outlier obtained in step 3, median_scores j The center point offset score of the curve obtained in step 2 is w1 and w2, which are the preset weight parameters of the two scores, and their values are w1+w2=1. w1 and w2 are the scale differences of the comparison, ensuring that the scales of the two algorithms are the same after taking the weights.

[0126] Step 4.2: Sorting the 36 curve scores from largest to smallest for the weighted combined scoring results from step 4.1. Calculating the standard deviation scores_std and mean scores_means of the scores array. Based on this, calculate the threshold T and return only the curve scores greater than the threshold. The formula is:

[0127] T=scores_std×sensitivity+scores_means

[0128] Among them, sensitivity is the preset sensitivity parameter. The larger the value, the larger the T value, and the fewer curves returned.

[0129] Finally, the algorithm returns the curve ID of the outlier, which is the code defined by the data collection.

[0130] Step 5: The algorithm returns the ID of the curve that is judged as outlier, performs anomaly diagnosis based on the real-time ventilation rejection rate, and outputs anomaly detection caps and treatment suggestions. The situations are divided into the following three cases:

[0131] 4) The algorithm does not return the curve ID for determining outliers, that is, there is no outlier curve. Regardless of whether the average rejection rate in the last 10 minutes of the shift is greater than the rejection rate threshold corresponding to the machine, the result returned is: the detection unit is normal;

[0132] 5) The algorithm returns the curve ID that determines the outlier. This means that there is an outlier curve, but the average rejection rate in the last 10 minutes of the shift is less than or equal to the rejection rate threshold corresponding to the machine. The result returned is: Detection cap XX (ID corresponds to the detection cap) may be aging, please check;

[0133] 1) The algorithm returns the ID of the curve that determines the outlier, that is, if there is an outlier curve and the average rejection rate in the last 10 minutes of the shift is greater than the rejection rate threshold corresponding to the machine, the result will be returned as: Detection cap XX (ID corresponds to the detection cap) is abnormal, please replace it in time;

[0134] More specifically, the method of this embodiment is applied to the Protos 2C model in a cigarette factory, where 15 online production devices are deployed. Due to certain differences between the devices, in terms of parameter setting, we take the rejection rate data of the corresponding machine for the past year, segment it into time windows of 600 seconds, and obtain 372,813 window samples. We then calculate the mean of the samples and finally obtain the 10-minute historical rejection rate mean of each corresponding machine. The threshold is set as follows:

[0135]

[0136]

[0137] In the algorithm analysis, the analysis is based on the data point codes of the 36 drum grooves defined by the data acquisition, that is, the ID. Therefore, the point codes also need to be converted. The conversion is as follows:

[0138] Detect drum groove data encoding Corresponding detection cap Detect drum groove data encoding Corresponding detection cap 57342 Detection cap 1 57360 Detection Cap 19 57343 Detection Cap 2 57361 Detection Cap 20 57344 Detection Cap 3 57362 Detection Cap 21 57345 Detection cap 4 57363 Detection cap 22 57346 Detection Cap 5 57364 Detection cap 23 57347 Detection Cap 6 57365 Detection cap 24 57348 Detection Cap 7 57366 Detection cap 25 57349 Detection Cap 8 57367 Detection Cap 26 57350 Detection Cap 9 57368 Detection Cap 27 57351 Detection cap 10 57369 Detection cap 28 57352 Detection cap 11 57370 Detection Cap 29 57353 Detection cap 12 57371 Detection Cap 30 57354 Detection cap 13 57372 Detection cap 31 57355 Detection cap 14 57373 Detection cap 32 57356 Detection cap 15 57374 Detection cap 33 57357 Detection cap 16 57375 Detection cap 34 57358 Detection cap 17 57376 Detection cap 35 57359 Detection cap 18 57377 Detection cap 36

[0139] By analyzing the results of the algorithm and combining it with the set threshold, we can achieve timely warning and effective presentation of abnormal conditions of the detection cap during the online production process. The following is the analysis results of the detection cap conditions of 15 devices captured at a certain time, and the results are as follows:

[0140] Real-time judgment of 15 Protos 2C detection caps in a factory

[0141]

[0142] in, Figure 4 、 Figure 5 This is a schematic diagram of the drum groove curve outlier detection and the corresponding rejection rate curve in the algorithm test of the structure of the present invention.

[0143] Figure 6 、 Figure 7 This diagram shows the outlier detection and rejection rate curves for the online anomaly diagnosis and early warning method. As can be seen from the diagram, the outlier curve is present and the rejection rate is greater than the set threshold. Therefore, the algorithm's final output is: "Detection cap XX is abnormal. Please replace it immediately."

[0144] Figure 8 、 Figure 9 This diagram shows the rejection rate curve corresponding to outliers in the drum groove curve detected in the online anomaly diagnosis and early warning method. As can be seen from the figure, there is an outlier curve, but the rejection rate is less than the set threshold. Therefore, the algorithm's final output is: Detection cap XX may be aging. Please check.

[0145] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for diagnosing and warning abnormalities of HID detection caps of cigarette making machines based on the fusion of center shift, difference and dimensionality reduction, characterized in that: The following steps are involved: Step (1) collects the operating data of the detection cap, and extracts data points related to ventilation through the Pearson correlation coefficient and manual experience, which includes 36 real-time variable parameters of the detection cap-related equipment, the number of ventilation rejections on duty, the rejection rate, the ventilation setting value, etc., totaling 46 dimensions of valid data, and performs sliding window segmentation and normalization processing on the valid data to construct a time-feature matrix, in which rows represent time points and columns represent feature numbers; Step (2) generates a detection cap data curve according to the time window for the data after sliding window segmentation and normalization processing in step (1), takes the median of each row of data as the center point, calculates the difference between each detection cap data curve and the center point, defines the curve with the largest difference as an outlier curve, and calculates the center point offset score based on it; Step (3), performing first-order difference processing on the detection cap data curve generated in step (2), and then applying principal component analysis to the differential data to fuse the 36 detection cap data curves to reduce the dimension to 1 dimension, and using this curve as the 1-dimensional standard curve, calculating the difference between all detection cap data curves and the 1-dimensional standard curve, and calculating the trend outlier score based on this; Step (4) performs weighted fusion of the center offset outlier score obtained in step (2) and the trend outlier score obtained in step (3) according to the preset weights to generate the final comprehensive score, and then dynamically calculates the threshold based on the distribution characteristics of the comprehensive score. Step (5) combines the dynamic threshold and the real-time ventilation rejection rate to perform data diagnosis and early warning detection cap status, which is divided into the following three situations: 1) The algorithm does not return the curve ID for determining outliers, that is, there is no outlier curve. Regardless of whether the average rejection rate in the last 10 minutes of the shift is greater than the rejection rate threshold corresponding to the machine, the result returned is: the detection unit is normal; 2) The algorithm returns the curve ID that determines the outlier. This means that there is an outlier curve, but the average rejection rate in the last 10 minutes of the shift is less than or equal to the rejection rate threshold corresponding to the machine. The result is returned as: Detection cap XX may be aging, please check, where the ID corresponds to the detection cap. 3) The algorithm returns the curve ID that determines the outlier, that is, if there is an outlier curve and the average rejection rate in the last 10 minutes of the shift is greater than the rejection rate threshold corresponding to the machine, the result is returned as: Detection cap XX is abnormal, please replace it in time, where the id corresponds to the detection cap.

2. The method for abnormal diagnosis and early warning of HID detection cap of cigarette making machine based on center shift, difference and dimensionality reduction fusion according to claim 1 is characterized in that: The specific process of step (1) is as follows: Step 1.1: Collect data from the past two months through the data acquisition system. Using the Pearson correlation coefficient and manual experience, extract 46 dimensions of valid data related to ventilation, including 36 test drum slots, ventilation rejection counts, rejection rates, and ventilation setpoints. The reduced data is then constructed into a data matrix X = m*n based on the time dimension, with n columns of features and m rows of time points, where m represents the time point and n represents the number of features. Step 1.2: Segment the data in the data matrix obtained in step 1.1 using a sliding window with a window length of 600 seconds. This includes the data of 36 drum slots, the number of rejects during the shift, the reject rate, the ventilation setting value, etc. Then, adjust the processed data to the range of [0, 1] using the minimum-maximum normalization method. The minimum-maximum normalization equation is: Where X represents the data matrix in step 1.1, x ij is the original data value of X in the jth column and i row, min(X j ) is the minimum value of X in the jth column; max(X j ) is the maximum value of X in the jth column, x′ ij is the normalized data value, and finally x′ ij Form a new data matrix X'.

3. The method for diagnosing and warning abnormalities of HID detection caps of cigarette making machines based on the fusion of center shift, difference and dimensionality reduction according to claim 2 is characterized in that: Specifically, step (2) involves segmenting the 36 detection cap data obtained in step 1.1 according to a predetermined time window, plotting a detection cap data curve of each segmented data over time using a line graph method, and then calculating the offset score of each detection cap curve from the center point; The method for selecting the center point is as follows: let the element of matrix X′ at time point i, that is, row i and column j, be d i1 , d i2 , d i3 ..., d ij , then the median of row i is M i =Median(d i1 , d i2 , d i3 ..., d ij ), where d ij Represents the elements in the matrix X′, M i Indicates the median of the i-th row, which is also the center point.

4. The method for diagnosing and warning abnormalities of HID detection caps of cigarette making machines based on center shift, difference and dimensionality reduction fusion according to claim 3 is characterized in that: The center point offset scoring method is: Step 1: For the element d in row i and column j ij Calculate the absolute value of the difference between it and the median, expressed as: absdiff ij =|d ij -M i |, where d ij represents the element in the i-th row and j-th column of the matrix X′, M i is the median of the i-th row obtained above, absdiff ij It represents the absolute value of the difference between the two; Step 2: Calculate the absdiff within the current time point range in column j ij The mean of is expressed as: in, Represents the absdiff within the current time point range in column j ij of and, Indicates finding the mean of row i, avg j That is, it is represented by the absdiff within the current time point range in the jth column ij The mean of Step 3: Calculate the judgment value p i,j , represents the curve outlier situation, and its formula is: Among them, absdiff ij Represents the element d in row i and column j ij The absolute value of the difference from the row median, avg j Represents the absdiff within the current time point range in column j ij The mean of , plus 1×10 -6 This is to prevent the denominator from being 0; t is the preset threshold parameter, only absdiff ij with avg j When the ratio is greater than the threshold parameter, p i,j Not 0, otherwise 0, where greater than 0 indicates outlier, otherwise not outlier; The fourth step is to count outliers at each time point. j The corresponding p i,j The relationship is: outlier_count j =outlier_count j +p i,j Among them, outlier_count j Indicates the cumulative calculation of outliers in the j-column curve, p i,j It represents the judgment value of the j column curve in the i row, that is, the judgment of the outliers of the 36 detection hat data curves; If the judgment value p of a characteristic curve j at the current time point i i,j is greater than 0, then the curve j is in its outlier count outlier_count j Add x i,j , otherwise it is 0; Step 5. Finally, after calculating the outlier counts for each detection hat data curve in the entire time interval, the center point offset score based on the outlier counts is calculated. The formula is: Among them, outlier_count j is the outlier count of the j-th column curve, ∑outlier_count j The sum of the outlier counts for all curves, median_scores j Score the center point offset of the jth column curve.

5. The method for diagnosing and warning abnormalities of HID detection caps of cigarette making machines based on the fusion of center shift, difference and dimensionality reduction according to claim 1 is characterized in that: In step (3), the method for calculating the trend outlier score is: Step 1: Use principal component analysis to reduce the dimensionality of the 36 test cap data curves to 1. The process of calculating the 1D standard curve is as follows: (1) Perform differential processing on the detection cap data curve and extract the trend of each curve, which is expressed as: in is the result after the first-order difference, X′ i,j is the value of the jth feature at the i-th time point in the normalized data, X′ i-1,j is the value of the jth feature at the i-1th time point in the normalized data, that is, the value of the previous time point; (2) Find the mean of the difference in each column, expressed as: Among them, x i ′ represents the difference operation result at the i-th time point, represents the sum of all difference results, Represents the mean of the difference results of each column; (3) Obtain the covariance Cov(X,Y) between each column of the detection hat data curve at the i-th time point, and calculate all columns in turn to form the covariance matrix C; (4) Use the eigenvalue decomposition method to find the eigenvalues and eigenvectors of the covariance matrix, expressed as: Cv i =λ i v i , where C represents the covariance matrix, λ i represents the eigenvalue, v i represents the eigenvector; For the eigenvalue λ i The acquisition is expressed as: det(C-λI)=0, where det represents the determinant, C represents the covariance matrix, I represents the identity matrix, and λ is the eigenvalue; (5) For each eigenvalue λ i , the eigenvector v i It can be obtained through a linear equation, expressed as: (C-λ i I)v i =0, where C represents the covariance matrix, λ i Represents each eigenvalue, I represents the unit matrix; normalizing the eigenvector is also to normalize the eigenvector, which is expressed as: Among them, v i is the original eigenvector, indicating the trend of data change in a certain direction, ||v i || is the eigenvector v i The norm of the vector, which represents the length of the vector, u i Represents the normalized eigenvector, which has a length of 1 and the same direction as the original vector; (6) Generation of 1-dimensional standard curve: y = V1 T X, where y represents the 1-dimensional data after dimensionality reduction, which is the projection value of the original data on the principal component direction; V1 represents the first principal component vector, indicating that the variance of the data in this direction is the largest; X represents the original data point of the detection hat curve, which is a 36-dimensional vector. Step 2: Calculate the trend outlier score. The specific process is as follows: (1) Based on the 1D standard curve generated above, calculate the difference between each 1D standard curve and the 1D standard curve value, expressed as: fitdata i,j =|D col (X′) i,j -pca(X′) i |, where fitdata i,j is the absolute value of the difference result at each time point of each curve, D col (X′) i,j is the result of the first-order difference, pca(X′) i is the value of the standard curve at time point i; (2) Calculate the trend outlier score diff_scores based on the difference between the above detection cap data curve and the 1D standard curve value j , the formula is: in, Expressed as the mean of the current curve gap values, Expressed as the mean of all curve difference values, diff_scores j The larger the value, the higher the degree of outlier the curve is.

6. The method for diagnosing and warning abnormalities of HID detection caps of cigarette making machines based on center shift, difference and dimensionality reduction fusion according to claim 1 is characterized in that: In step (4), the calculation process of the final comprehensive score is: The first step is to perform a weighted combination of the center point offset score and the trend outlier score. The formula is: scores j =diff_scores j ×w1+median_scores j ×w2 Among them, scores j is the final score of the curve, diff_scores j Scoring trend outliers, median_scores j The center point offset score is calculated. w1 and w2 are the preset weight parameters of the two scores, and their values are w1+w2=1. w1 and w2 are the scale differences for comparison, ensuring that the scales of the two algorithms are the same after taking the weights. The second step is to convert the scores obtained above into j Sort from largest to smallest and calculate scores j The standard deviation scores_std and mean scores_means of the array are used as the basis to calculate the threshold T. The formula is: T=scores_std×sensitivity+scores_means Among them, sensitivity is the preset sensitivity parameter. The larger the value, the larger the T value and the fewer outlier curves.

Images