Roll core wrinkling prediction method and device
By constructing and integrating the full receptive field convolution model, the problem of the existing core wrinkle detection method being limited in application between different production lines is solved, and efficient and accurate core wrinkle prediction and detection is achieved.
Patent Information
- Application Number
- CN202411516006.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-28
- Publication Date
- 2025-02-25
AI Technical Summary
The existing core wrinkle detection methods are limited by the differences between equipment and production lines, and cannot be effectively applied to different production lines, affecting the detection efficiency.
By collecting core production data and product data, a full receptive field convolution model is constructed, and a core wrinkle prediction model is generated using the model integration algorithm, iterative training and data updates are carried out to achieve the prediction of core wrinkle.
It realizes efficient core wrinkle inspection between different production lines, improves detection efficiency and accuracy, and is suitable for a variety of production lines.
Smart Images
Figure CN119513758A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of lithium battery manufacturing, and in particular to a method and device for predicting core wrinkling. Background Art
[0002] Winding is one of the processes in lithium battery production. In this process, the winding equipment uses a winding needle to wind the laser-cut positive and negative electrode sheets and separators in a certain arrangement order and make a winding core. In the actual winding process, the electrode sheets or separators may be wrinkled due to multiple complex reasons such as human factors, equipment, materials, and operating methods. Wrinkling of the electrode sheets or separators will lead to uneven contact areas between the separators and the electrode sheets, resulting in increased local current density, increased resistance inside the battery, and uneven distribution of the battery internal resistance, which ultimately affects the consistency of battery charging and discharging.
[0003] The existing method for detecting core wrinkles uses a CCD camera to detect the deviation fluctuation between the edge position of the pole piece and the set reference position, thereby detecting whether the core is wrinkled. Although core wrinkle detection can be achieved to a certain extent, the use of CCD for detection is limited by the core installation space and the pole piece wrapping sequence, and cannot be applied between different production lines, affecting the detection efficiency. Summary of the invention
[0004] The present invention provides a method and device for predicting core wrinkles, so as to solve the problem that the existing core wrinkle detection has great limitations and cannot be applied between different production lines, thus affecting the detection efficiency.
[0005] In order to solve the above technical problems, an embodiment of the present invention provides a method for predicting core wrinkles, comprising:
[0006] Collecting sample feature data according to the core production data and the core product data, and preprocessing the sample feature data to generate training data; the preprocessing includes data classification, data screening and data enhancement;
[0007] Constructing a plurality of full receptive field convolution models, integrating the plurality of full receptive field convolution models according to a model integration algorithm, and generating a core wrinkle prediction model;
[0008] Preset a training cycle and a number of cycles, iteratively train the core wrinkle prediction model according to the training cycle, the number of cycles and the training data, update the training data according to the prediction result and prediction weight of the core wrinkle prediction model in each training cycle, and train the core wrinkle prediction model according to the updated training data in the next training cycle;
[0009] Until the preset conditions are met, the training is stopped, the core wrinkle prediction models obtained from each round of training are integrated to obtain the final core wrinkle prediction model, and the core wrinkle prediction is performed based on the integrated core wrinkle prediction model.
[0010] Furthermore, the sample feature data is collected according to the core production data and the core product data, and the sample feature data is preprocessed to generate training data, specifically:
[0011] Collecting sample feature data according to the core wrinkle correlation, wherein the sample feature data includes label data and non-label data;
[0012] Obtain the variance, mean and data distribution of the sample feature data; remove abnormal samples in the sample feature data according to the normal distribution law, a preset deviation threshold and the data distribution; and perform consistency processing on the data distribution of the label data and the non-label data;
[0013] Dividing the labeled data into a training set and a validation set according to a preset ratio, and setting the unlabeled data as a test set;
[0014] The label data is enhanced, and the training set is updated to generate training data.
[0015] Furthermore, the label data is enhanced and the training set is updated to generate training data, specifically:
[0016] The label data is equally divided into a plurality of groups of sub-data, and full receptive field convolution models are trained respectively according to the plurality of groups of sub-data to obtain a plurality of trained full receptive field convolution models;
[0017] Predicting the test set data according to the several trained full receptive field convolution models to obtain prediction results of each full receptive field convolution model;
[0018] Generate pseudo-label data for the unwrinkled data according to the prediction results of each full receptive field convolution model, and update the training set according to the pseudo-label data.
[0019] Furthermore, the construction of several full receptive field convolution models, the integration of the several receptive field convolution models according to the model integration algorithm, and the generation of a core wrinkle prediction model are specifically as follows:
[0020] Constructing several full receptive field convolution models, wherein the full receptive field convolution models include three convolution layers, a pooling layer and a fully connected layer;
[0021] The multiple full receptive field convolution models are integrated according to a self-paced ensemble learning algorithm to generate a core wrinkle prediction model.
[0022] Furthermore, the core wrinkle prediction model is iteratively trained according to the training cycle, the number of cycles and the training data, the training data is updated according to the prediction result and prediction weight of the core wrinkle prediction model in each training cycle, and the core wrinkle prediction model is trained according to the updated training data in the next training cycle, specifically:
[0023] In a first training cycle, training data is extracted from the training set according to the label content, and the core wrinkle prediction model is trained according to the training data;
[0024] Predicting the data in the training set according to the trained core wrinkle prediction model, and classifying and setting weights for the sample data in the training set according to the prediction results and prediction errors;
[0025] The training set is updated according to the classified sample data and the corresponding weights, and the next training cycle is entered to train the roll core wrinkle prediction model.
[0026] Furthermore, the extracting of training data from the training set according to the label content and training of the roll core wrinkle prediction model according to the training data is specifically as follows:
[0027] The sample data in the training set is divided into wrinkled samples and unwrinkled samples according to the label content,
[0028] Randomly extracting an equal amount of wrinkled samples and unwrinkled samples from the training set as training data for this training cycle;
[0029] The core wrinkle prediction model is trained according to the training data.
[0030] Furthermore, the data in the training set are predicted according to the trained core wrinkle prediction model, and the sample data in the training set are classified and weighted according to the prediction results and prediction errors, specifically:
[0031] Predicting all the unwrinkled samples in the training set according to the trained coil core wrinkle prediction model;
[0032] The unwrinkled samples are divided into a plurality of groups of unwrinkled samples according to the size of the prediction error, and the weight of each group of unwrinkled samples is set according to the size of the prediction error.
[0033] Furthermore, the training set is updated according to the classified sample data and the corresponding weights, and the next training cycle is entered to train the roll core wrinkle prediction model, specifically:
[0034] randomly extracting sample data from each group of unwrinkled samples according to the weights, and generating training data for the current training cycle according to the sample data and wrinkled samples in the training set;
[0035] The core wrinkle prediction model is trained according to the training data of the current training cycle, and the non-wrinkled samples are regrouped according to the prediction results, and the weights of each group are updated.
[0036] Furthermore, after integrating the core wrinkle prediction models obtained from each round of training to obtain the final core wrinkle prediction model, the method further includes:
[0037] Calculating the index score of the core wrinkle prediction model according to the prediction precision and the prediction recall;
[0038] The performance of the core wrinkle prediction model is evaluated based on the indicator score.
[0039] In a second aspect, the present invention provides a core wrinkle prediction device, comprising: a preprocessing module, a model building module, a model training module and a prediction module;
[0040] The preprocessing module is used to collect sample feature data according to the core production data and the core product data, and preprocess the sample feature data to generate training data; the preprocessing includes data classification, data screening and data enhancement;
[0041] The model building module is used to build a plurality of full receptive field convolution models, integrate the plurality of full receptive field convolution models according to a model integration algorithm, and generate a core wrinkle prediction model;
[0042] The model training module is used to preset a training cycle and a number of cycles, iteratively train the core wrinkle prediction model according to the training cycle, the number of cycles and the training data, update the training data according to the prediction results and prediction weights of the core wrinkle prediction model in each training cycle, and train the core wrinkle prediction model according to the updated training data in the next training cycle;
[0043] The prediction module is used to stop training until a preset condition is met, integrate the core wrinkle prediction models obtained in each round of training to obtain a final core wrinkle prediction model, and perform core wrinkle prediction based on the integrated core wrinkle prediction model. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 A schematic flow chart of a method for predicting core wrinkles provided by an embodiment of the present invention;
[0045] Figure 2A schematic diagram of sample characteristic data provided by an embodiment of the present invention;
[0046] Figure 3 Another schematic diagram of sample characteristic data provided by an embodiment of the present invention;
[0047] Figure 4 A structural schematic diagram of a full receptive field convolution model of a roll core wrinkle prediction method provided by an embodiment of the present invention;
[0048] Figure 5 A schematic diagram of model integration training of a core wrinkle prediction method provided by an embodiment of the present invention;
[0049] Figure 6 A schematic diagram of changes in prediction errors of a core wrinkle prediction method provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0051] Embodiment 1
[0052] Please refer to Figure 1 , Figure 1 A schematic flow chart of a core wrinkle prediction method provided by an embodiment of the present invention includes steps 101 to 104, which are specifically as follows:
[0053] Step 101: Collect sample feature data according to the core production data and the core product data, and pre-process the sample feature data to generate training data; the pre-processing includes data classification, data screening and data enhancement;
[0054] In this embodiment, the sample feature data is collected according to the core production data and the core product data, and the sample feature data is preprocessed to generate training data, specifically:
[0055] Collecting sample feature data according to the core wrinkle correlation, wherein the sample feature data includes label data and non-label data;
[0056] Obtain the variance, mean and data distribution of the sample feature data; remove abnormal samples from the sample feature data according to the normal distribution law, the preset deviation threshold and the data distribution; and perform consistency processing on the data distribution of the label data and the unlabel data;
[0058] Dividing the labeled data into a training set and a validation set according to a preset ratio, and setting the unlabeled data as a test set;
[0059] The label data is enhanced, and the training set is updated to generate training data.
[0060] Please refer to Figure 2 and Figure 3 , Figure 2 A schematic diagram of sample characteristic data provided by an embodiment of the present invention; Figure 3 Another schematic diagram of sample characteristic data provided by an embodiment of the present invention.
[0061] In this embodiment, data such as negative pole to positive pole overhang (positive pole ear side), negative pole to positive pole overhang (negative pole ear side), diaphragm to negative pole overhang (positive pole ear side), diaphragm to negative pole overhang (negative pole ear side), and diaphragm misalignment, which have a high correlation with core wrinkling, are selected as sample feature data, and the sample feature data are divided into label data and non-label data.
[0062] In this embodiment, after eliminating abnormal samples whose data obviously exceeds the normal range according to the statistical results of the labeled data and the unlabeled data, all the data are divided into a labeled training set and a validation set, and an unlabeled test set. Then, samples in the training and validation data sets that do not conform to the data distribution of the test set are eliminated, so that the data distribution of each data set is as similar as possible, thereby ensuring the consistency of the effect of the model on the validation set and the test set.
[0063] In this embodiment, the variance mean and data distribution of the sample feature data are calculated based on data statistical tools, and abnormal samples in the sample feature data are eliminated based on the normal distribution law and the deviation threshold. The abnormal samples are samples whose deviation values from the normal distribution are greater than the deviation threshold.
[0064] In this embodiment, the deviation threshold is set to 15 times the standard deviation, which is obtained by variance calculation.
[0065] In this embodiment, samples whose sample characteristic data exceed 15 times the standard deviation σ of the normal distribution are found and eliminated during screening.
[0066] In this embodiment, outliers in the samples are removed to reduce the adverse effects of abnormal and noisy data on the model.
[0067] In this embodiment, the data distribution between the labeled data and the unlabeled data is compared, and the data distribution of the labeled data and the unlabeled data is processed for consistency, and samples in the labeled data that do not conform to the distribution of the unlabeled data are eliminated, so that the data distribution of each labeled data and the unlabeled data is as similar as possible, thereby ensuring the consistency of the model effect on the validation set and the test set.
[0068] In this embodiment, the labeled data is randomly divided into a training set and a validation set in a ratio of 7:3, and the unlabeled data is used as a test set.
[0069] In this embodiment, all sample feature data need to be standardized. Specifically, all data are mapped into standard data with a mean of 0 and a variance of 1 in the feature dimension to eliminate the data size and dimension effects, improve the input stability of the model and the model convergence speed. The label data is sorted into an array according to a preset format to generate model verification data. The array includes the number of samples, the number of features, and the length of the time series.
[0070] In this embodiment, since the proportion of crumpled cores is relatively small in actual production, the proportion of crumpled battery cell samples in the actual samples obtained is low, resulting in sample imbalance. Therefore, in order to balance the ratio of wrinkled samples and non-crumpled samples in the label data, the label data needs to be enhanced.
[0071] In this embodiment, the label data is enhanced according to the label content, and the training set is updated to generate training data, specifically:
[0072] The label data is equally divided into a plurality of groups of sub-data, and full receptive field convolution models are trained respectively according to the plurality of groups of sub-data to obtain a plurality of trained full receptive field convolution models;
[0073] Predicting the test set data according to the several trained full receptive field convolution models to obtain prediction results of each full receptive field convolution model;
[0074] Generate pseudo-label data for the unwrinkled data according to the prediction results of each full receptive field convolution model, and update the training set according to the pseudo-label data.
[0075] In this embodiment, the labeled data is divided into 10 groups of sub-data containing equal samples, and one group is selected as verification data each time, and the others are selected as training data. According to the training data, several full receptive field 1D-CNN (one-dimensional convolutional neural network) models, i.e., full receptive field convolution models, are trained respectively to obtain 10 fitted full receptive field convolution models. Then, the test set data is input into the full receptive field convolution model respectively and the prediction results are given. The intersection of the unwrinkled labels of the prediction results of each model is taken, and this data is added to the training set as pseudo-label data.
[0076] Step 102: constructing a plurality of full receptive field convolution models, integrating the plurality of full receptive field convolution models according to a model integration algorithm, and generating a core wrinkle prediction model;
[0077] Please refer to Figure 4 , Figure 4 A structural schematic diagram of a full receptive field convolution model of a roll core wrinkle prediction method provided in an embodiment of the present invention.
[0078] In this embodiment, several full receptive field convolution models are constructed, and the several receptive field convolution models are integrated according to the model integration algorithm to generate a core wrinkle prediction model, specifically:
[0079] Constructing several full receptive field convolution models, wherein the full receptive field convolution models include three convolution layers, a pooling layer, and a fully connected layer;
[0080] The multiple full receptive field convolution models are integrated according to a self-paced ensemble learning algorithm to generate a core wrinkle prediction model.
[0081] In this embodiment, each of the first three layers of the full receptive field convolution model network consists of a convolution layer and a corresponding BN layer, and its activation function is set to ReLU. The fourth and fifth layers of the network are the global mean pooling layer and the fully connected layer, respectively, and the fully connected layer uses the SoftMax activation function to output the classification results.
[0082] In this embodiment, the plurality of full receptive field convolution models are integrated according to a self-paced ensemble learning algorithm (SPE) to generate a roll core wrinkle prediction model.
[0083] Step 103: Preset a training cycle and the number of cycles, iteratively train the core wrinkle prediction model according to the training cycle, the number of cycles and the training data, update the training data according to the prediction result and prediction weight of the core wrinkle prediction model in each training cycle, and train the core wrinkle prediction model according to the updated training data in the next training cycle;
[0084] Please refer to Figure 5 , Figure 5 A schematic diagram of model integration training for a method for predicting roll core wrinkles provided in an embodiment of the present invention.
[0085] In this embodiment, machine learning frameworks such as scikit-learn and Pytorch are used for training. The hyperparameters of the model are used as variables, and are input into each full receptive field 1D-CNN sub-model according to the constructed training data and SPE model integration method to obtain the corresponding integrated model. The corresponding indicators of the integrated model under different hyperparameters are compared to obtain the corresponding hyperparameters of the integrated model with the best prediction effect, and the F1 score and other indicators of the model on the validation set are calculated. Using the optimal hyperparameters, the validation set data is added to the training set at the same time, and the final integrated model is trained, and finally the integrated model is used to give the test set prediction.
[0086] In this embodiment, the NG sample is a wrinkled sample, and the OK sample is a non-wrinkled sample.
[0087] In this embodiment, the core wrinkle prediction model is iteratively trained according to the training cycle, the number of cycles and the training data, the training data is updated according to the prediction result and prediction weight of the core wrinkle prediction model in each training cycle, and the core wrinkle prediction model is trained according to the updated training data in the next training cycle, specifically:
[0088] In a first training cycle, training data is extracted from the training set according to the label content, and the core wrinkle prediction model is trained according to the training data;
[0089] Predict the data in the training set according to the trained core wrinkle prediction model, and classify and set weights for the sample data in the training set according to the prediction results and prediction errors;
[0090] The training set is updated according to the classified sample data and the corresponding weights, and the next training cycle is entered to train the roll core wrinkle prediction model.
[0091] In this embodiment, N training cycles are preset.
[0092] In this embodiment, the extracting of training data from the training set according to the label content and training of the core wrinkle prediction model according to the training data are specifically as follows:
[0093] The sample data in the training set is divided into wrinkled samples and unwrinkled samples according to the label content,
[0094] Randomly extracting an equal amount of wrinkled samples and unwrinkled samples from the training set as training data for this training cycle;
[0095] The core wrinkle prediction model is trained according to the training data.
[0096] In this embodiment, in the first training cycle, the data in the training set is divided into wrinkled samples and unwrinkled samples, and samples are randomly selected from the unwrinkled samples to generate training data, so that the ratio of wrinkled samples to unwrinkled samples in the training data used to train the model is 1:1.
[0097] In this embodiment, the data in the training set are predicted according to the trained core wrinkle prediction model, and the sample data in the training set are classified and weighted according to the prediction results and prediction errors, specifically:
[0098] According to the trained coil core wrinkle prediction model, all the unwrinkled samples in the training set are predicted;
[0099] The unwrinkled samples are divided into a plurality of groups of unwrinkled samples according to the size of the prediction error, and the weight of each group of unwrinkled samples is set according to the size of the prediction error.
[0100] In this embodiment, the generated training data is used to train the integrated core wrinkle prediction model to obtain the fitted core wrinkle prediction model, the non-wrinkled samples are predicted according to the core wrinkle prediction model, and the prediction error is calculated, and all non-wrinkled samples are divided into M groups according to the prediction error Loss, and each group is assigned a weight that is positively correlated with the prediction error.
[0101] In this embodiment, the training set is updated according to the classified sample data and the corresponding weights, and the next training cycle is entered to train the core wrinkle prediction model, specifically:
[0102] randomly extracting sample data from each group of unwrinkled samples according to the weights, and generating training data for the current training cycle according to the sample data and wrinkled samples in the training set;
[0103] The core wrinkle prediction model is trained according to the training data of the current training cycle, and the non-wrinkled samples are regrouped according to the prediction results, and the weights of each group are updated.
[0104] In this embodiment, except for the first training cycle, that is, in the subsequent N-1 training cycles, a certain number of samples are randomly selected from the divided groups of unwrinkled samples according to the weights, and the extracted unwrinkled samples and wrinkled data are used to generate the training data of the current cycle, and the core wrinkle prediction model is trained according to the training data of the current training cycle. After the current training cycle is completed, all unwrinkled samples are predicted according to the trained core wrinkle prediction model, and all unwrinkled samples are re-divided into several groups according to the prediction error Loss, and the weights of each group are updated.
[0105] In this embodiment, the prediction error size is represented according to the Loss function.
[0106] Step 104: until the preset conditions are met, stop training, output the core wrinkle prediction model, and perform core wrinkle prediction based on the core wrinkle prediction model.
[0107] In this embodiment, the weights are updated with each round of training and decay to a constant in the last round of training, that is, each group of weights is 1 / M, and 1 / M*number of wrinkled samples are selected from each group. Finally, these N models are combined to obtain an integrated core wrinkle prediction model.
[0108] In this embodiment, the integrated core wrinkle prediction model is specifically:
[0109]
[0110] Among them, F i (x) is the output of each sub-model.
[0111] In this embodiment, after integrating the core wrinkle prediction models obtained in each round of training to obtain the final core wrinkle prediction model, the method further includes:
[0112] Calculating the index score of the core wrinkle prediction model according to the prediction precision and the prediction recall;
[0113] The performance of the core wrinkle prediction model is evaluated based on the indicator score.
[0114] In this embodiment, the integrated model indicators are set to prediction accuracy, recall rate and indicator score F1, and the indicator score F1 is specifically:
[0115]
[0116] Among them, F Score Represents the F1 score, Precision is the prediction accuracy, and Recall is the prediction recall.
[0117] Please refer to Figure 6 , Figure 6 A schematic diagram of changes in prediction errors of a core wrinkle prediction method provided in an embodiment of the present invention.
[0118] In this embodiment, during training, the SPE method parameters are set as follows: training rounds are 10, and the number of groups is 20. The hyperparameters of a typical full receptive field 1D-CNN sub-model are set as follows: initial learning rate is 0.001, batch size is 48, epoch is 500, optimizer is Adam, and Loss is the cross entropy loss function. The calculation formula of Loss is as follows:
[0119]
[0120] Among them, y i Represents the label of the sample, 1 for wrinkled and 0 for unwrinkled, p i represents the probability of wrinkling of sample i predicted by the model, and N is the number of predicted samples.
[0121] In this embodiment, by using the data of the misalignment process between the positive electrode sheet, the negative electrode sheet and the separator as samples, after data preprocessing of the samples, multiple full receptive field 1D-CNN models are integrated and trained through the SPE (Self-paced Ensemble) model integration method, and finally a core wrinkle prediction model is obtained. The SPE model integration algorithm reduces the impact of the imbalance and noise of the mold core wrinkle samples on the model effect. At the same time, the full receptive field 1D-CNN can achieve better prediction results without adjusting the size of the 1D-CNN convolution kernel, reducing the computing resources required for the adjustment of the model hyperparameters, thereby improving the prediction accuracy and prediction efficiency of the model, and can be applied to various core production lines.
[0122] The embodiment of the present invention also provides a core wrinkle prediction device, comprising: a preprocessing module, a model building module, a model training module and a prediction module;
[0123] The preprocessing module is used to collect sample feature data according to the core production data and the core product data, and preprocess the sample feature data to generate training data; the preprocessing includes data classification, data screening and data enhancement;
[0124] The model building module is used to build a plurality of full receptive field convolution models, integrate the plurality of full receptive field convolution models according to a model integration algorithm, and generate a core wrinkle prediction model;
[0125] The model training module is used to preset a training cycle and the number of cycles, iteratively train the core wrinkle prediction model according to the training cycle, the number of cycles and the training data, update the training data according to the prediction results and prediction weights of the core wrinkle prediction model in each training cycle, and train the core wrinkle prediction model according to the updated training data in the next training cycle;
[0126] The prediction module is used to stop training until a preset condition is met, integrate the core wrinkle prediction models obtained in each round of training to obtain a final core wrinkle prediction model, and perform core wrinkle prediction based on the integrated core wrinkle prediction model.
[0127] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for predicting core wrinkles, characterized in that: include: Collecting sample feature data according to the core production data and the core product data, and preprocessing the sample feature data to generate training data; The preprocessing includes data classification, data screening and data enhancement; Constructing a plurality of full receptive field convolution models, integrating the plurality of full receptive field convolution models according to a model integration algorithm, and generating a core wrinkle prediction model; Preset a training cycle and a number of cycles, iteratively train the core wrinkle prediction model according to the training cycle, the number of cycles and the training data, update the training data according to the prediction result and prediction weight of the core wrinkle prediction model in each training cycle, and train the core wrinkle prediction model according to the updated training data in the next training cycle; Until the preset conditions are met, the training is stopped, the core wrinkle prediction models obtained from each round of training are integrated to obtain the final core wrinkle prediction model, and the core wrinkle prediction is performed based on the integrated core wrinkle prediction model.
2. A method for predicting core wrinkles according to claim 1, characterized in that: The collecting of sample feature data based on the core production data and the core product data, and preprocessing of the sample feature data to generate training data, specifically: Collecting sample feature data according to the core wrinkle correlation, wherein the sample feature data includes label data and non-label data; Obtain the variance, mean and data distribution of the sample feature data; remove abnormal samples in the sample feature data according to the normal distribution law, a preset deviation threshold and the data distribution; and perform consistency processing on the data distribution of the label data and the non-label data; Dividing the labeled data into a training set and a validation set according to a preset ratio, and setting the unlabeled data as a test set; The label data is enhanced, and the training set is updated to generate training data.
3. A method for predicting core wrinkles according to claim 2, characterized in that: The step of performing data enhancement on the label data and updating the training set to generate training data is specifically as follows: The label data is equally divided into a plurality of groups of sub-data, and full receptive field convolution models are trained respectively according to the plurality of groups of sub-data to obtain a plurality of trained full receptive field convolution models; Predicting the test set data according to the several trained full receptive field convolution models to obtain prediction results of each full receptive field convolution model; Generate pseudo-label data for the unwrinkled data according to the prediction results of each full receptive field convolution model, and update the training set according to the pseudo-label data.
4. A method for predicting core wrinkles according to claim 1, characterized in that: The method of constructing a plurality of full receptive field convolution models and integrating the plurality of receptive field convolution models according to a model integration algorithm to generate a core wrinkle prediction model is as follows: Constructing several full receptive field convolution models, wherein the full receptive field convolution models include three convolution layers, a pooling layer and a fully connected layer; The multiple full receptive field convolution models are integrated according to a self-paced ensemble learning algorithm to generate a core wrinkle prediction model.
5. A method for predicting core wrinkles according to any one of claims 1 to 4, characterized in that: The core wrinkle prediction model is iteratively trained according to the training cycle, the number of cycles and the training data, the training data is updated according to the prediction result and prediction weight of the core wrinkle prediction model in each training cycle, and the core wrinkle prediction model is trained according to the updated training data in the next training cycle, specifically: In a first training cycle, the training data is extracted from the training set according to the label content, and the core wrinkle prediction model is trained according to the training data; Predicting the data in the training set according to the trained core wrinkle prediction model, and classifying and setting weights for the sample data in the training set according to the prediction results and prediction errors; The training set is updated according to the classified sample data and the corresponding weights, and the next training cycle is entered to train the roll core wrinkle prediction model.
6. A method for predicting core wrinkles according to claim 5, characterized in that: The extracting of training data from the training set according to the label content and training the core wrinkle prediction model according to the training data is specifically as follows: The sample data in the training set is divided into wrinkled samples and unwrinkled samples according to the label content, Randomly extracting an equal amount of the wrinkled samples and the unwrinkled samples from the training set as training data for this training cycle; The core wrinkle prediction model is trained according to the training data.
7. A method for predicting core wrinkles according to claim 6, characterized in that: The method predicts the data in the training set according to the trained core wrinkle prediction model, and classifies and sets weights for the sample data in the training set according to the prediction results and prediction errors, specifically: Predicting all the unwrinkled samples in the training set according to the trained coil core wrinkle prediction model; The unwrinkled samples are divided into a plurality of groups of unwrinkled samples according to the size of the prediction error, and the weight of each group of unwrinkled samples is set according to the size of the prediction error.
8. A method for predicting core wrinkles according to claim 7, characterized in that: The training set is updated according to the classified sample data and the corresponding weights, and the next training cycle is entered to train the core wrinkle prediction model, specifically: randomly extracting sample data from each group of unwrinkled samples according to the weights, and generating training data for the current training cycle according to the sample data and wrinkled samples in the training set; The roll core wrinkle prediction model is trained according to the training data of the current training cycle, and the non-wrinkled samples are regrouped according to the prediction results, and the weights of each group are updated.
9. A method for predicting core wrinkles according to claim 1, characterized in that: After the core wrinkle prediction models obtained from each round of training are integrated to obtain the final core wrinkle prediction model, the method further includes: Calculating the index score of the core wrinkle prediction model according to the prediction precision and the prediction recall; The performance of the core wrinkle prediction model is evaluated based on the indicator score.
10. A core wrinkle prediction device, characterized in that: include: Preprocessing module, model building module, model training module and prediction module; The preprocessing module is used to collect sample feature data according to the core production data and the core product data, and preprocess the sample feature data to generate training data; the preprocessing includes data classification, data screening and data enhancement; The model building module is used to build a plurality of full receptive field convolution models, integrate the plurality of full receptive field convolution models according to a model integration algorithm, and generate a core wrinkle prediction model; The model training module is used to preset a training cycle and a number of cycles, iteratively train the core wrinkle prediction model according to the training cycle, the number of cycles and the training data, update the training data according to the prediction results and prediction weights of the core wrinkle prediction model in each training cycle, and train the core wrinkle prediction model according to the updated training data in the next training cycle; The prediction module is used to stop training until a preset condition is met, integrate the core wrinkle prediction models obtained from each round of training to obtain a final core wrinkle prediction model, and perform core wrinkle prediction based on the integrated core wrinkle prediction model.