Method for predicting abnormity of flattening disc of cigarette making and plug assembling machine

Abstract

The invention discloses a cigarette making and plug assembling machine leveling disc abnormity prediction method. The prediction method comprises the following steps: S1, determining the correlation between a leveling disc and a process parameter index; s2, acquiring and cleaning real-time data; and S3, constructing a component anomaly prediction model. The cigarette making and plug assembling machine ZJ116 is taken as an example, according to data generated by equipment such as a sensor and a counting device of the equipment, historical fault records are combined, and clustering and classification methods are utilized, so that the abnormality of the flattening disc part is predicted in advance, and the labor cost and the loss caused by the abnormality of the flattening disc are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of cigarette manufacturing and computer data analysis, and more specifically, relates to a method for predicting abnormality of a flattening plate of a cigarette making machine. Background Art

[0002] The leveling plate is one of the key components of cigarette making equipment in the tobacco industry. Its function is to trim the tobacco strips on the smoking belt to meet the process requirements of subsequent cigarette production. The leveling plate is a wear-prone part. If the leveling plate is abnormal, it will cause problems such as empty cigarettes and cigarette weight not meeting regulations. Therefore, it is necessary to arrange maintenance personnel to regularly inspect and adjust the leveling plate.

[0003] The patent "Online Equipment Abnormality Monitoring System and Method in Silk Making" describes a method for predicting equipment abnormality in the tobacco industry. The process section targeted by this method is the silk making section, which is quite different from the cigarette making section, and adopts the threshold method. Wang Ji et al. proposed a method for detecting abnormal states of key components of aircraft engines. This detection method requires obtaining parameters directly related to the engine, such as torque, oil temperature, and speed. Xu Bin et al. proposed analyzing subway components by monitoring abnormal vibrations of tracks. All of the above analysis methods require obtaining data directly related to the target components, or adding sensors to the original equipment to enable the operation of the system. As a purely mechanical component, the leveling plate has a compact structure, and it is impossible to obtain data directly related to it, nor is there space for installing additional sensors.

[0004] In summary, it is urgent to design a flat disk prediction method based on correlation parameters to realize the abnormal prediction of the flat disk without installing additional sensors. Summary of the invention

[0005] In order to solve the technical problem that the existing flattening plate requires a lot of manual inspection and there is not enough space to install sensors, the present invention provides a flattening plate abnormality prediction method for a winder. Taking the winder ZJ116 as an example, the method of the present invention uses the data generated by the sensors and counting devices of the equipment, combined with the historical fault records, and uses clustering and classification methods to achieve early prediction of flattening plate component abnormalities, thereby reducing labor costs and losses caused by flattening plate abnormalities.

[0006] The technical solution adopted by the present invention is:

[0007] A method for predicting abnormality of a flat plate of a winder, characterized in that the prediction method comprises the following steps:

[0008] S1. Clarify the correlation between the flattening plate and process parameter indicators;

[0009] S2, obtain and clean real-time data;

[0010] S3. Build a component abnormality prediction model.

[0011] Furthermore, in step S1, the specific steps of clarifying the correlation between the flattening plate and the process parameter index are as follows:

[0012] S11. First, obtain all process parameters and part replacement records of the current machine. Take the flat plate replacement as the time node, compare the historical real-time data in each time period before and after the replacement, perform data cleaning according to the data cleaning rules, and obtain the steady-state valid data;

[0013] S12. After obtaining the effective data of steady state, the correlation between the abnormality of flat plate and process parameters, the correlation between process parameters and the historical trend of process parameters are studied respectively;

[0014] S13. The entropy method is used to calculate the correlation of abnormal process parameters of flattening plate parts.

[0015] Furthermore, in step S3, the specific process of building a component abnormality warning model is as follows:

[0016] S31, early warning model initialization;

[0017] S32. Early warning model optimization.

[0018] Furthermore, in step S31, the specific process of initializing the early warning model is as follows:

[0019] 1) Read reliable training data and assemble it into a directly usable DataFrame;

[0020] 2) Use the t-SNE algorithm to reduce the dimension and calculate the two-dimensional coordinates of the data;

[0021] 3) Assign each piece of data independently to a data cluster;

[0022] 4) Calculate the distance between two data in two-dimensional space;

[0023] 5) Sort by the distance between the data to obtain a list of relationships from near to far;

[0024] 6) Circulate the relationship list and merge the two closest data clusters into one data cluster;

[0025] 7) Repeat the merging step until the data clusters in the list reach a discrete state;

[0026] 8) Print discrete data clusters on a two-dimensional chart and export the data to a file;

[0027] 9) Display two-dimensional data to assist manual decision making and determine the correctness of data clusters;

[0028] 10) Read the clustered data and add labels to the data based on manual and machine-assisted decisions;

[0029] 11) Use the decision tree algorithm to randomly divide the data into two parts: training data and verification data;

[0030] 12) Use the SVC algorithm to load the training data and train the Gaussian model;

[0031] 13) Use the verification data to test the trained Gaussian model and score it;

[0032] 14) Save the verified Gaussian model to the document database for later use.

[0033] Furthermore, in step S32, the specific process of early warning model optimization is as follows:

[0034] 1) Develop an equipment data collection plan and add valid data to the data set of known abnormal data and known normal data according to the workshop operation status;

[0035] 2) Convert the data into training data according to data preprocessing rules;

[0036] 3) Divide the newly generated reliable data into two categories: training data and verification data in proportion;

[0037] 4) Add the newly generated training data to the original training set and retrain the prediction model;

[0038] 5) Use the newly generated verification data to verify the accuracy of the early warning model, and adjust the scale of training data and data cleaning preprocessing rules according to the accuracy;

[0039] 6) Adjust the training data in the training set to adapt to the new data preprocessing rules;

[0040] 7) Retrain the preprocessed model again.

[0041] Furthermore, the specific steps of using the t-SNE algorithm to reduce the dimension and calculate the two-dimensional coordinates of the data are as follows:

[0042] Step 1: Find the pairwise similarity between neighboring points in high-dimensional space;

[0043] The high-dimensional Euclidean distance between data points x and x is converted into conditional probability P(j|i), as shown in the following formula (1):

[0044]

[0045] x will select x as its neighbor according to the probability density ratio under the Gaussian distribution centered on point x;

[0046] σi is the Gaussian variance centered at data point Xi, and the probability density of a pair of points is proportional to their similarity; for nearby data points, p(j|i) will be relatively high, while for points far away, p(j|i) will be small;

[0047] The conditional probability in the high-dimensional space is symmetrized to obtain the final similarity in the high-dimensional space.

[0048] The conditional probability is symmetric by averaging the two probabilities, as shown in equation (2):

[0049]

[0050] Step 2: According to the pairwise similarity of points in the high-dimensional space, each point in the high-dimensional space is mapped to a low-dimensional mapping, as shown in the following formula (3):

[0051]

[0052] where y and y are the low-dimensional counterparts of the high-dimensional data points x and x;

[0053] Calculate the conditional probability q(j|i) similar to P(j]i) under a Gaussian distribution centered at point y, and then symmetrize the probability;

[0054] Step 3: Use gradient descent based on Kullback-Leibler divergence (KL divergence) to find a low-dimensional data representation that minimizes the mismatch between P and q, as shown in equation (4):

[0055]

[0056] Where P represents the conditional probability distribution of point x on all other data points; Q represents the conditional probability distribution of all other map points given a map point y, using gradient descent to optimize the points in the low-dimensional space;

[0057] Step 4: Use Student-t distribution to calculate the similarity between two points in low-dimensional space.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] 1. The present invention takes the ZJ116 winder as an example, based on the data generated by the sensors and counting devices of the equipment, combined with the historical fault records, and using clustering and classification methods, the abnormality of the flattening disc components can be predicted in advance.

[0060] 2. The t-SNE in the present invention uses a heavy-tailed Student-t distribution with one degree of freedom to calculate the similarity between two points in a low-dimensional space, instead of a Gaussian distribution. The T-distribution creates a probability distribution of points in a low-dimensional space, which helps to reduce crowding problems.

[0061] 3. The method of the present invention can be extended to other key components of cigarette making equipment. Based on the real-time data of each process parameter index of the machine component, the model is used to predict the accuracy of the normal and abnormal components, and the overall abnormal distribution probability of each component is obtained; according to the abnormal push rules (such as: abnormal probability size, consecutive times, etc.), the abnormal probability information of the component is pushed BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a schematic diagram of the t-SNE algorithm example.

[0063] Figure 2 It is a schematic diagram of data in high-dimensional space.

[0064] Figure 3 Is the low dimensional map will be a 2D or 3D map.

[0065] Figure 4 This is a schematic diagram of an SVM algorithm example.

[0066] Figure 5 It is a diagram of the SVM principle.

[0067] Figure 6 It is the calculation formula of the upper and lower hyperplanes of the maximum margin hyperplane.

[0068] Figure 7 It is a schematic diagram of the splitting based on the histogram.

[0069] Figure 8 This is a schematic diagram of leaf node splitting.

[0070] Fig. 9 It is a schematic diagram of the feature parallel process.

[0071] Fig.10 It is a diagram of data parallel process.

[0072] Fig.11 This is an example of an abnormal warning of the leveling plate components.

[0073] Fig.12 It is a schematic diagram of abnormal warning of machine components.

[0074] Fig.13a Correlation analysis of various process parameters that cause abnormalities in flattening plate parts Figure 1 .

[0075] Fig.13bCorrelation analysis of various process parameters that cause abnormalities in flattening plate parts Figure 2 .

[0076] Fig.14 It is a relationship matrix diagram of various process parameters that cause abnormalities in flat plate parts.

[0077] Fig.15 It is a fluctuation trend analysis chart of key process parameters.

[0078] Fig.16 It is a diagram of the component abnormality prediction model. DETAILED DESCRIPTION

[0079] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present invention, and is not used to limit the embodiment of the present invention.

[0080] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0081] The present invention will be described in detail below in conjunction with exemplary embodiments with reference to the accompanying drawings.

[0082] refer to Figures 1 to 12 , a method for predicting abnormality of a flat plate of a winder according to the present invention, the prediction method comprises the following steps:

[0083] S1. Clarify the correlation between the flattening plate and process parameter indicators;

[0084] Specifically, in step S1, the specific steps of clarifying the correlation between the flattening plate and the process parameter index are as follows:

[0085] S11. First, obtain all process parameters and part replacement records of the current machine. Take the flat plate replacement as the time node, compare the historical real-time data in each time period before and after the replacement, perform data cleaning according to the data cleaning rules, and obtain the steady-state valid data;

[0086] S12. After obtaining the effective data of steady state, the correlation between the abnormality of flat plate and process parameters, the correlation between process parameters and the historical trend of process parameters are studied respectively;

[0087] S13. The entropy method is used to calculate the correlation of abnormal process parameters of flattening plate parts.

[0088] S2, obtain and clean real-time data;

[0089] S3. Build a component abnormality prediction model.

[0090] Specifically, in step S3, the specific process of constructing the component abnormality warning model is as follows: S31, initializing the warning model;

[0091] S32. Early warning model optimization.

[0092] Specifically, in step S31, the specific process of initializing the early warning model is as follows:

[0093] 10) Read reliable training data and assemble it into a directly usable DataFrame;

[0094] 11) Use the t-SNE algorithm to reduce the dimension and calculate the two-dimensional coordinates of the data;

[0095] 12) Assign each piece of data independently to a data cluster;

[0096] 13) Calculate the distance between two data in two-dimensional space;

[0097] 14) Sort by the distance between the data to obtain a list of relationships from near to far;

[0098] 15) Circulate the relationship list and merge the two closest data clusters into one data cluster;

[0099] 16) Repeat the merging step until the data clusters in the list reach a discrete state;

[0100] 17) Print discrete data clusters on a two-dimensional chart and export the data to a file;

[0101] 18) Display two-dimensional data to assist manual decision making and determine the correctness of data clusters;

[0102] 10) Read the clustered data and add labels to the data based on manual and machine-assisted decisions;

[0103] 11) Use the decision tree algorithm to randomly divide the data into two parts: training data and verification data;

[0104] 12) Use the SVC algorithm to load the training data and train the Gaussian model;

[0105] 13) Use the verification data to test the trained Gaussian model and score it;

[0106] 14) Save the verified Gaussian model to the document database for later use.

[0107] Specifically, in step S32, the specific process of early warning model optimization is as follows:

[0108] 1) Develop an equipment data collection plan and add valid data to the data set of known abnormal data and known normal data according to the workshop operation status;

[0109] 2) Convert the data into training data according to data preprocessing rules;

[0110] 3) Divide the newly generated reliable data into two categories: training data and verification data in proportion;

[0111] 4) Add the newly generated training data to the original training set and retrain the prediction model;

[0112] 5) Use the newly generated verification data to verify the accuracy of the early warning model, and adjust the scale of training data and data cleaning preprocessing rules according to the accuracy;

[0113] 6) Adjust the training data in the training set to adapt to the new data preprocessing rules;

[0114] 7) Retrain the preprocessed model again.

[0115] Specifically, the specific steps of using the t-SNE algorithm to reduce the dimension and calculate the two-dimensional coordinates of the data are as follows:

[0116] Step 1: Find the pairwise similarity between neighboring points in high-dimensional space;

[0117] The high-dimensional Euclidean distance between data points x and x is converted into conditional probability P(j|i), as shown in the following formula (1):

[0118]

[0119] x will select x as its neighbor according to the probability density ratio under the Gaussian distribution centered on point x;

[0120] σi is the Gaussian variance centered at data point Xi, and the probability density of a pair of points is proportional to their similarity; for nearby data points, p(j|i) will be relatively high, while for points far away, p(j|i) will be small;

[0121] The conditional probability in the high-dimensional space is symmetrized to obtain the final similarity in the high-dimensional space.

[0122] The conditional probability is symmetric by averaging the two probabilities, as shown in equation (2):

[0123]

[0124] Step 2: According to the pairwise similarity of points in the high-dimensional space, each point in the high-dimensional space is mapped to a low-dimensional mapping, as shown in the following formula (3):

[0125]

[0126] where y and y are the low-dimensional counterparts of the high-dimensional data points x and x;

[0127] Calculate the conditional probability q(j|i) similar to P(j]i) under a Gaussian distribution centered at point y, and then symmetrize the probability;

[0128] Step 3: Use gradient descent based on Kullback-Leibler divergence (KL divergence) to find a low-dimensional data representation that minimizes the mismatch between P and q, as shown in equation (4):

[0129]

[0130] Where P represents the conditional probability distribution of point x on all other data points; Q represents the conditional probability distribution of all other map points given a map point y, using gradient descent to optimize the points in the low-dimensional space;

[0131] Step 4: Use Student-t distribution to calculate the similarity between two points in low-dimensional space.

[0132] Among them, t-SNE (TSNE) converts the similarity of data points into probabilities. The similarity in the original space is represented as a Gaussian joint probability (converted into probability based on the similarity between data points), and the t distribution of the similarity in the embedding space. This allows t-SNE to be particularly sensitive to local structure and has some other advantages over existing technologies.

[0133] t-SNE is currently the most effective data dimensionality reduction and visualization method, but it has disadvantages, such as large memory usage and long running time. However, when we want to classify high-dimensional data and we don’t know whether the data set is well separable (i.e., the intervals between the same type are small, and the intervals between different types are large), we can use t-SNE to project it into a 2D or 3D space to observe. If it is separable in low-dimensional space, the data is separable; if it is not separable in high-dimensional space, it may be that the data is not separable, or it may simply be because it cannot be projected into low-dimensional space.

[0134] The Kullback-Leibler (KL) divergence of the joint probability of the original space and the embedded space is used to evaluate the quality of the visualization. That is, the function related to the KL divergence is used as the loss function, and then the loss function is minimized by gradient descent to finally obtain the convergence result. In practice, multiple parameters can be adjusted to optimize the t-SNE algorithm in order to obtain the best visualization quality effect.

[0135] The main idea of ​​the Support Vector Machine (SVM) is to establish an optimal decision hyperplane that maximizes the distance between the two types of samples closest to the plane on both sides of the plane, thereby providing good generalization capabilities for classification problems. For a multidimensional sample set, the system randomly generates a hyperplane and keeps moving it to classify the samples until the sample points belonging to different categories in the training samples are exactly on both sides of the hyperplane. There may be many hyperplanes that meet this condition. SVM formally finds such a hyperplane while ensuring classification accuracy, maximizing the blank area on both sides of the hyperplane, thereby achieving optimal classification of linearly separable samples.

[0136] Advantages of SVM:

[0137] 1. It does not require many samples. It does not mean that the absolute number of training samples is small, but it means that compared with other training classification algorithms, SVM requires relatively fewer samples under the same problem complexity. And because SVM introduces kernel functions, it can easily handle high-dimensional samples.

[0138] 2. Minimize structural risk. This risk refers to the cumulative error between the classifier’s approximation of the true model of the problem and the true solution to the problem.

[0139] Nonlinearity means that SVM is good at dealing with situations where sample data is linearly inseparable, which is mainly achieved through slack variables (also called penalty variables) and kernel function technology. This part is also the essence of SVM. The principle of the SVM algorithm is to find an optimal hyperplane to separate data points of different categories and maximize the interval from the hyperplane to the nearest data point. Specifically, SVM finds the optimal hyperplane by maximizing the interval to achieve data classification and regression analysis. The relevant concepts and diagrams are as follows:

[0140] 1) Hyperplane: In two-dimensional space, a hyperplane is a straight line; in three-dimensional space, it is a plane; and in higher-dimensional space, it is a hyperplane. The goal of SVM is to find an optimal hyperplane that can separate data points of different categories.

[0141] 2) Maximum Margin: SVM attempts to find a hyperplane that maximizes the distance (called margin) from the hyperplane to the nearest data point (support vector), which can increase the robustness and generalization performance of the classification.

[0142] 3) Kernel Technique: When dealing with nonlinearly separable data, SVM uses the kernel technique to map the data to a higher-dimensional space in order to find a better hyperplane in the new space to classify the data.

[0143] 4) Loss function and optimization: SVM solves the optimal hyperplane by minimizing the loss function and introducing regularization terms to obtain the best classification results.

[0144] 5) Support vector: During the training process, only some data points (support vectors) closest to the hyperplane have an impact on the final hyperplane, and these data points determine the final classification result.

[0145] Before talking about LightGBM, let's first introduce GBDT. GBDT (Gradient Boosting Decision Tree) is a long-standing model in machine learning. Its main idea is to use weak classifiers (decision trees) for iterative training to obtain the optimal model. This model has the advantages of good training effect and low overfitting. GBDT is widely used in the industry and is usually used for tasks such as click-through rate prediction and search ranking.

[0146] LightGBM is a new member of the boosting ensemble model. Like XGBoost, it is an efficient implementation of GBDT. In principle, it is similar to GBDT and XGBoost, and uses the negative gradient of the loss function as the residual approximation of the current decision tree to fit the new decision tree. LightGBM performs better than XGBoost in many aspects, and can speed up the training of the GBDT model without compromising accuracy.

[0147] LightGBM is nearly 10 times faster than XGBoost, uses about 1 / 6 of the memory of XGBoost, and has improved accuracy.

[0148] LightGBM uses a histogram-based decision tree algorithm. The specific principles are as follows:

[0149] 1. Histogram-based splitting:

[0150] LightGBM discretizes continuous eigenvalues ​​into K bins (buckets), and then constructs a histogram to speed up the calculation and selection of eigenvalues. Before enumerating the split points, the histogram is sorted according to the label mean corresponding to each category; then the optimal split points are enumerated in sequence according to the sorting results. Of course, this method is prone to overfitting, so LightGBM also adds a lot of constraints and regularization for this method.

[0151] 2. Leaf node splitting:

[0152] Unlike the "depth-first" splitting implemented by most GBDTs, LightGBM adopts a "leaf node splitting" strategy, selecting the leaf node with the largest splitting gain for splitting each time, so as to find the global optimal solution faster.

[0153] 3. Support parallel learning:

[0154] Through strategies based on feature parallelism and data parallelism, LightGBM can be trained efficiently in a distributed environment.

[0155] The main idea of ​​feature parallelism is that different machines search for the optimal split point on different feature sets, and then synchronize the optimal split point between machines. XGBoost uses this feature parallel method. This feature parallel method has a big disadvantage: it divides the data vertically, each machine contains different data, and then uses different machines to find the optimal split point for different features. The division results need to be communicated to each machine, which adds additional complexity.

[0156] LightGBM does not perform vertical data partitioning, but instead stores all training data on each machine. After obtaining the best partitioning solution, the partitioning can be performed locally, reducing unnecessary communication.

[0157] The traditional data parallel strategy is to divide the data horizontally, let different machines construct histograms locally first, then merge them globally, and finally find the optimal split point on the merged histogram. LightGBM uses Reduce Scatter in data parallelism to distribute the task of histogram merging to different machines, reducing communication and computing, and using histogram difference to further reduce the communication volume by half.

[0158] The present invention takes the flattening plate of the ZJ116 type winder as an example to illustrate how to determine the correlation between the flattening plate and the process parameter index. The steps are as follows:

[0159] First, obtain all the process parameters and part replacement records of the current machine. Take the flat plate replacement as the time node, compare the historical real-time data in each time period before and after the replacement, and perform data cleaning according to the data cleaning rules to obtain the steady-state valid data.

[0160] After obtaining effective steady-state data, the correlation between the leveling plate anomaly and process parameters, the correlation between process parameters, and the historical trend of process parameters are studied respectively.

[0161] The entropy method is used to calculate the correlation of process parameters of abnormal flat plate parts.

[0162] Fig.13a and Fig.13b The figure shows the correlation analysis of various process parameters that cause abnormal flattening plate parts. The covariance method is used to analyze the relationship matrix of various process parameters, and the results are as follows: Fig.14The relationship matrix diagram of each process parameter that causes the abnormality of the flat plate parts. When analyzing the data of the process parameters of the abnormal flat plate parts, it is also necessary to analyze the fluctuation trend and parameter characterization status of the historical real-time process parameters. If the parameter trend is abnormal, an abnormal prompt will be triggered to confirm whether the status of each process parameter is really the process parameter that affects the abnormality of the flat plate parts.

[0163] The key process parameter fluctuation trend analysis chart is as follows Fig.15 shown.

[0164] After the above steps, the first 30 preliminary indicators of the current machine that have an important impact on the flattening plate parts are obtained. The process parameters that cause the abnormality of the flattening plate parts are initially confirmed, and then the manual experience of the equipment experts is reviewed and confirmed. In actual situations, continuous verification and adjustment are made to ensure the effectiveness of the parameters. The process parameter indicators related to the flattener parts are shown in Table 1 below:

[0165] Table 1 Process parameter indicators related to ZJ116 leveler parts

[0166]

[0167] According to the selected process parameter indicators related to each component, according to certain data cleaning rules (for example: the output is greater than 20,000 pieces, the vehicle speed is greater than the steady-state speed of 2,000 pieces / minute, the continuous operation time is greater than 3 minutes, and the unit time frequency of data acquisition is 1 minute), the data of the process parameter indicators is cleaned and preprocessed, invalid data and abnormal data are filtered or marked, and valid data (real-time value or incremental value) for a period of time is obtained, and the data is saved in the time series database. The effective real-time data of the process parameter indicators of the components obtained are shown in Table 2 below:

[0168] Table 2 Examples of real-time data of process parameter indicators of parts

[0169]

[0170] By confirming the good process parameter indicators, valid data for a long period of time (including data before and after maintenance and replacement) is obtained from the real-time database. According to the current valid data, the entropy method is used to calculate the different weight values ​​of all process parameter indicators corresponding to the component. At the same time, the current valid data is also used as the training data sample of the valid data of the flat plate component (including the verification data set); for the valid data, the t-SNE clustering algorithm is used to perform multi-dimensional indicator dimensionality reduction clustering to achieve dimensionality reduction and vector quantization (Vector Quantization), and compress the high-order features into two columns; then the SVM classification algorithm is used to discretize the data and group and label it (divided into two categories: normal and abnormal); the component abnormality prediction model is trained according to the LightGBM decision tree algorithm, and the model parameters are continuously optimized in actual tests to obtain the best component abnormality prediction model. The flat plate abnormality prediction model is constructed as shown in the figure. Fig.16 As shown in the figure, the X-axis and the Y-axis represent the component feature vector after dimensionality reduction. The green area represents the normal TAG point, and the red area represents the abnormal TAG point.

Claims

1. A method for predicting abnormality of a flat plate of a winder, characterized in that: The prediction method comprises the following steps: S1. Clarify the correlation between the flattening plate and process parameter indicators; S2, obtain and clean real-time data; S3. Build a component abnormality prediction model.

2. A method for predicting abnormality of a flat plate of a winder as claimed in claim 1, characterized in that: In step S1, the specific steps of clarifying the correlation between the flattening plate and the process parameter index are as follows: S11. First, obtain all process parameters and part replacement records of the current machine. Take the flat plate replacement as the time node, compare the historical real-time data in each time period before and after the replacement, perform data cleaning according to the data cleaning rules, and obtain the steady-state valid data; S12. After obtaining the effective data of steady state, the correlation between the abnormality of flat plate and process parameters, the correlation between process parameters and the historical trend of process parameters are studied respectively; S13. The entropy method is used to calculate the correlation of abnormal process parameters of flattening plate parts.

3. A method for predicting abnormality of a flat plate of a winder as claimed in claim 1, characterized in that: In step S3, the specific process of building a component abnormality warning model is as follows: S31, early warning model initialization; S32. Early warning model optimization.

4. A method for predicting abnormality of a flat plate of a winder as claimed in claim 3, characterized in that: In step S31, the specific process of initializing the early warning model is as follows: 1) Read reliable training data and assemble it into a directly usable DataFrame; 2) Use the t-SNE algorithm to reduce the dimension and calculate the two-dimensional coordinates of the data; 3) Assign each piece of data independently to a data cluster; 4) Calculate the distance between two data in two-dimensional space; 5) Sort by the distance between the data to obtain a list of relationships from near to far; 6) Circulate the relationship list and merge the two closest data clusters into one data cluster; 7) Repeat the merging step until the data clusters in the list reach a discrete state; 8) Print discrete data clusters on a two-dimensional chart and export the data to a file; 9) Display two-dimensional data to assist manual decision making and determine the correctness of data clusters; 10) Read the clustered data and add labels to the data based on manual and machine-assisted decisions; 11) Use the decision tree algorithm to randomly divide the data into two parts: training data and verification data; 12) Use the SVC algorithm to load the training data and train the Gaussian model; 13) Use the verification data to test the trained Gaussian model and score it; 14) Save the verified Gaussian model to the document database for later use.

5. The method for predicting abnormality of a flat plate of a winder as claimed in claim 3, characterized in that: In step S32, the specific process of early warning model optimization is as follows: 1) Develop an equipment data collection plan and add valid data to the data set of known abnormal data and known normal data according to the workshop operation status; 2) Convert the data into training data according to data preprocessing rules; 3) Divide the newly generated reliable data into two categories: training data and verification data in proportion; 4) Add the newly generated training data to the original training set and retrain the prediction model; 5) Use the newly generated verification data to verify the accuracy of the early warning model, and adjust the scale of training data and data cleaning preprocessing rules according to the accuracy; 6) Adjust the training data in the training set to adapt to the new data preprocessing rules; 7) Retrain the preprocessed model again.

6. A method for predicting abnormality of a flat plate of a winder as claimed in claim 4, characterized in that: The specific steps for using the t-SNE algorithm to reduce dimension and calculate the two-dimensional coordinates of the data are as follows: Step 1: Find the pairwise similarity between neighboring points in high-dimensional space; The high-dimensional Euclidean distance between data points x and x is converted into conditional probability P(j|i), as shown in the following formula (1): x will select x as its neighbor according to the probability density ratio under the Gaussian distribution centered on point x; σi is the Gaussian variance centered at data point Xi, and the probability density of a pair of points is proportional to their similarity; for nearby data points, p(j|i) will be relatively high, while for points far away, p(j|i) will be small; The conditional probability in the high-dimensional space is symmetrized to obtain the final similarity in the high-dimensional space. The conditional probability is symmetric by averaging the two probabilities, as shown in equation (2): Step 2: According to the pairwise similarity of points in the high-dimensional space, each point in the high-dimensional space is mapped to a low-dimensional mapping, as shown in the following formula (3): where y and y are the low-dimensional counterparts of the high-dimensional data points x and x; Calculate the conditional probability q(j|i) similar to P(j]i) under a Gaussian distribution centered at point y, and then symmetrize the probability; Step 3: Use gradient descent based on Kullback-Leibler divergence (KL divergence) to find a low-dimensional data representation that minimizes the mismatch between P and q, as shown in equation (4): Where P represents the conditional probability distribution of point x on all other data points; Q represents the conditional probability distribution of all other map points given a map point y, using gradient descent to optimize the points in the low-dimensional space; Step 4: Use Student-t distribution to calculate the similarity between two points in low-dimensional space.

Images