Fabric classification evaluation method based on multi-model integration

Abstract

The invention discloses a fabric classification evaluation method based on multi-model integration. The method comprises the following steps: S1, data acquisition: obtaining a fabric characteristic data set; s2, data preprocessing: screening out a fabric feature set through data cleaning, standardization and feature engineering; and S3, multi-model training: training to form multiple models. And S4, fusion of multiple models: combining prediction results of the multiple models to obtain an objective model score. And S5, model correction: giving subjective data through manual detection, integrating prediction results, and further improving the prediction accuracy. According to the method, various advanced machine learning models such as the random forest, the gradient boosting tree and the multi-layer perceptron are combined, the classification accuracy and efficiency are improved through the integrated learning technology, and intelligent classification and prediction of different fabrics are achieved. The system is particularly suitable for industries needing high-precision fabric classification, such as textile, material science and manufacturing industry, and has important practical application value and popularization prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of fabric classification, and in particular relates to a fabric classification and evaluation method based on multi-model integration. Background Art

[0002] In the field of fabric classification, traditional methods mainly rely on training classification models by designing artificial features, and using the trained models for classification operations. However, this method of artificial feature extraction relies on expert experience, the process is complex and time-consuming, and it is easily affected by human factors. The subjectivity and uncertainty of feature selection may limit the stability and accuracy of the classification model. In addition, many traditional methods rely on a single classification model, which makes it difficult to fully utilize the diversity and complexity of data. This single model performs poorly when processing data with complex nonlinear relationships, has limited generalization ability, and is difficult to maintain consistent classification performance in different fabric categories. In addition, when processing large-scale data, traditional methods have low training and prediction efficiency, which makes it difficult to meet the requirements of real-time and high efficiency in industrial production. Especially in application scenarios where classification models need to be frequently updated and adjusted, the efficiency problem of traditional methods is particularly prominent. Different fabrics have different characteristics and properties. Traditional feature extraction and classification methods are difficult to adapt to the changing application requirements. They often require special feature design and algorithm optimization for different fabric categories, which increases the complexity and cost of the application. Summary of the invention

[0003] In order to overcome the difficulty of traditional feature extraction and classification methods in adapting to the changing application requirements of fabric classification, the present invention provides the following technical solutions:

[0004] A fabric classification and evaluation method based on multi-model integration includes the following steps:

[0005] S1-Data Collection:

[0006] The relevant physical and mechanical indicators of the fabric are collected by a fabric touch measuring instrument, thereby obtaining a fabric property data set;

[0007] S2-Data Preprocessing:

[0008] Ensure the integrity and consistency of the data in the fabric property dataset through data cleaning and standardization;

[0009] Through feature engineering, the data types that have a greater impact on fabric classification in the fabric characteristic data set are screened out to form a fabric feature set;

[0010] S3-Multi-model training:

[0011] Data is extracted from the fabric feature set to form a training set and a validation set, and training is used to form multiple models based on different machine learning algorithms.

[0012] S4- Fusion of multiple models:

[0013] The prediction results of the multiple models are combined by weighted voting to reduce the prediction error and obtain an objective model score.

[0014] S5-Model Modifications:

[0015] Manual inspection provides subjective data, and the model is modified through weighted voting to obtain comprehensive prediction results, further improving the accuracy of the prediction.

[0016] Furthermore, in the step S1-data collection, the ALTAS fabric touch tester is used to measure the physical and mechanical indicators of different fabric types, thereby obtaining fabric property data of different fabric types.

[0017] Furthermore, the step S2-data preprocessing includes the following steps:

[0018] S2.1-Data Cleansing:

[0019] removing outliers in the fabric property data by using the interquartile range method (IQR);

[0020] S2.2-Data filling:

[0021] Using KNN missing value filling method to fill the missing values ​​in the fabric characteristic data;

[0022] S2.3-Data Standardization:

[0023] The fabric characteristic data is normalized and scaled to the interval [0,1] by Min-Max normalization;

[0024] S2.4-Feature Engineering:

[0025] Recursive feature elimination (RFE) and principal component analysis (PCA) are used to perform feature selection and dimension reduction, thereby screening out several fabric characteristic data that are more important for fabric classification and forming a fabric feature set.

[0026] Furthermore, in the S3-multi-model training, the multiple models include: a random forest model, a gradient boosting tree model and a multi-layer perceptron neural network model.

[0027] Furthermore, the random forest model in step S3 is composed of multiple decision trees, each tree is trained by different feature subsets and data subsets, and a majority vote is taken on the prediction results of each tree.

[0028] Furthermore, the gradient boosting tree model in step S3 reduces the prediction error by gradually adding new decision trees, and uses the negative gradient of the loss function as the target of the new tree.

[0029] Furthermore, the multilayer perceptron neural network in step S3 uses a three-part structure of an input layer, a hidden layer, and an output layer, and uses an activation function to convert the linear combination of input features into a nonlinear output, and minimizes the loss function through a back-propagation algorithm to optimize weights and biases, thereby achieving modeling and classification prediction of complex data patterns.

[0030] Furthermore, the step S5-modification of the model includes the following steps:

[0031] S5.1-User Subjective Prediction:

[0032] Users give subjective ratings to different types of fabrics based on their actual usage experience;

[0033] S5.2-Weighted distribution of subjective and objective prediction results:

[0034] A weighted score is calculated based on the subjective score and the predicted probability of the model. The weighted score comprehensively considers the evaluation of the probability predicted by the model and the subjective score. The weighted score formula is as follows:

[0035]

[0036] S5.3-Threshold Setting:

[0037] A threshold for the total score will be set. If it is greater than the threshold, the final prediction result will be output. If it is less than the threshold, the training will be resumed.

[0038] The present invention combines a variety of advanced machine learning models such as random forest, gradient boosting tree and multi-layer perceptron, improves classification accuracy and efficiency through integrated learning technology, and realizes intelligent classification and prediction of different fabrics. The system is particularly suitable for industries such as textiles, materials science, and manufacturing that require high-precision fabric classification, and has important practical application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The figure is a schematic diagram of the overall process of the fabric classification and evaluation method based on multi-model integration of the present invention.

[0040] Figure 2 The data is cleaned and standardized in the application example.

[0041] Figure 3 It is the data processed by feature engineering in the application example.

[0042] Figure 4Confusion matrix for objective model scores in application examples.

[0043] Figure 5 It is the weighted scoring result combining the subjective scoring and the objective model scoring in the application example. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is described clearly and completely with specific embodiments in conjunction with the accompanying drawings. It should be understood that the described embodiments are only used to illustrate the present invention, but not to limit the scope of the present invention.

[0045] Embodiment 1:

[0046] The present invention provides a fabric classification and evaluation method based on multi-model integration, comprising the following steps:

[0047] S1-Data Collection:

[0048] The relevant physical and mechanical indicators of the fabric are collected by a fabric touch measuring instrument, thereby obtaining a fabric property data set;

[0049] S2-Data Preprocessing:

[0050] Ensure the integrity and consistency of the data in the fabric property dataset through data cleaning and standardization;

[0051] Through feature engineering, the data types that have a greater impact on fabric classification in the fabric characteristic data set are screened out to form a fabric feature set;

[0052] S3-Multi-model training:

[0053] Data is extracted from the fabric feature set to form a training set and a validation set, and training is used to form multiple models based on different machine learning algorithms.

[0054] S4- Fusion of multiple models:

[0055] The prediction results of the multiple models are combined by weighted voting to reduce the prediction error and obtain an objective model score.

[0056] S5-Model Modifications:

[0057] Manual inspection provides subjective data, and the model is modified through weighted voting to obtain comprehensive prediction results, further improving the accuracy of the prediction.

[0058] Furthermore, in the step S1-data collection, the ALTAS fabric touch tester is used to measure the physical and mechanical indicators of different fabric types, thereby obtaining fabric property data of different fabric types; the ALTAS fabric touch tester is a conventional technical means for collecting fabric property data in the prior art;

[0059] Exemplary:

[0060] The types of fabrics include: cotton, linen, silk, and wool;

[0061] The physical and mechanical indicators include: surface friction characteristics, roughness characteristics, heat transfer characteristics, compression characteristics and bending characteristics;

[0062] The fabric characteristic data include: softness M, smoothness S, coldness and warmth W and comprehensive index T;

[0063] Furthermore, the step S2-data preprocessing includes the following steps:

[0064] S2.1-Data Cleansing:

[0065] removing outliers in the fabric property data by using the interquartile range method (IQR);

[0066] S2.2-Data filling:

[0067] Using KNN missing value filling method to fill the missing values ​​in the fabric characteristic data;

[0068] S2.3-Data Standardization:

[0069] The fabric characteristic data is normalized and scaled to the interval [0,1] by Min-Max normalization;

[0070] S2.4-Feature Engineering:

[0071] Recursive feature elimination (RFE) and principal component analysis (PCA) are used to perform feature selection and dimension reduction, thereby screening out several fabric characteristic data that are more important for fabric classification and forming a fabric feature set.

[0072] Furthermore, in the S3-multi-model training, the multiple models include: a random forest model, a gradient boosting tree model and a multi-layer perceptron neural network model.

[0073] Furthermore, the step S2.1-data cleaning includes the following steps:

[0074] S2.1.1-Calculate quartiles:

[0075] The first quartile Q1: the 25th percentile of the data set;

[0076] The third quartile Q3: the 75th percentile of the data set;

[0077] S2.1.2-Calculate the interquartile range (IQR):

[0078] IQR=Q3-Q1

[0079] S2.1.3 - Determine the range of outliers:

[0080] Lower Bound: Q1-1.5*IQR

[0081] Upper Bound: Q3+1.5*IQR

[0082] S2.1.4-Detecting outliers:

[0083] If a data point is smaller than the lower limit or larger than the upper limit, it is considered an outlier and is deleted to complete data cleaning.

[0084] Furthermore, the step S2.2-data filling includes the following steps:

[0085] S2.2.1-Select K value:

[0086] Select K nearest neighbor data points;

[0087] S2.2.2-Calculate distance:

[0088] Use Euclidean distance to calculate the distance between the data point to be filled and other data points;

[0089] S2.2.3-Select nearest neighbor:

[0090] Select the K nearest neighbors;

[0091] S2.2.4—Calculate the padding value:

[0092] Use the most common category among the K neighbors as fill-in value.

[0093] Furthermore, in the step S2.3-data standardization, the formula for Min-Max standardization is:

[0094]

[0095] in:

[0096] x is the original data point, x min is the minimum value of feature x, x max is the maximum value of feature x, x normmalized are the normalized data points.

[0097] Furthermore, the recursive feature elimination (RFE) and principal component analysis (PCA) in step S2.4-feature engineering are only commonly used technical means for extracting feature data in the prior art and will not be described in detail.

[0098] Furthermore, the random forest model in step S3 is composed of multiple decision trees, each tree is trained by different feature subsets and data subsets, and the prediction results of each tree are voted by majority, and its training formula is as follows:

[0099]

[0100] in, Represents the final prediction result, which is the predicted category of the input sample x;

[0101] mode: represents the mode; T1, T2, …, Tn: represents the base learners in the ensemble learning model;

[0102] x: represents the input sample, which is a point in the feature space, and we want to predict which category it belongs to;

[0103] Ti(x): represents the prediction result of the i-th tree for the input sample X.

[0104] Furthermore, the gradient boosting tree model in step S3 reduces the prediction error by gradually adding new decision trees, and uses the negative gradient of the loss function as the target of the new tree. The training formula is as follows:

[0105]

[0106] in:

[0107] is the predicted value of the model for sample i after the m+1th iteration;

[0108] is the model's predicted value for sample i after the mth iteration;

[0109] H is the learning rate, which is a parameter between 0 and 1 that controls the contribution or influence of each tree on the final prediction result.

[0110] h m (x i ) is the prediction value of the mth tree for sample i, which is the prediction of the newly trained tree for the sample;

[0111] x i is the feature vector of the i-th sample.

[0112] Furthermore, the multilayer perceptron neural network in step S3 is structured as an input layer, a hidden layer, and an output layer. It uses an activation function (such as ReLU, Sigmoid) to convert the linear combination of input features into a nonlinear output, and minimizes the loss function through a back-propagation algorithm to optimize weights and biases, thereby achieving modeling and classification prediction of complex data patterns. The specific settings are as follows:

[0113] Hidden layer output:

[0114] Output layer probability:

[0115] Among them, w and b are weights and biases, φ is the activation function, and Softmax is used for the probability distribution of multi-classification problems.

[0116] Furthermore, in the step S4-multi-model fusion, the prediction results of multiple models are fused by soft voting, and the formula of soft voting is as follows:

[0117]

[0118] in:

[0119] is the objective model score;

[0120] M is the total number of models;

[0121] P m (x) is the probability distribution of the input sample x belonging to each category for the mth model;

[0122] argmax: is a function that finds the class index that maximizes the sum of probabilities.

[0123] Furthermore, the step S5-modification of the model includes the following steps:

[0124] S5.1-User Subjective Prediction:

[0125] Users give subjective ratings to different types of fabrics based on their actual usage experience;

[0126] S5.2-Weighted distribution of subjective and objective prediction results:

[0127] A weighted score is calculated based on the subjective score and the predicted probability of the model. The weighted score comprehensively considers the evaluation of the probability predicted by the model and the subjective score. The weighted score formula is as follows:

[0128]

[0129] S5.3-Threshold Setting:

[0130] A threshold for the total score will be set. If it is greater than the threshold, the final prediction result will be output. If it is less than the threshold, the training will be resumed.

[0131] Application Example: Application example of the fabric classification and evaluation method based on Example 1:

[0132] S1-Data Collection:

[0133] The relevant physical and mechanical indicators of the fabric are collected by a fabric touch measuring instrument, thereby obtaining a fabric property data set;

[0134] S2-Data Preprocessing:

[0135] The integrity and consistency of the data in the fabric characteristic data set are ensured by data cleaning and standardization; the data after data cleaning and standardization are as follows: Figure 2 As shown;

[0136] Through feature engineering, the data types that have a greater impact on fabric classification in the fabric characteristic data set are screened out to form a fabric feature set. The processing results of feature engineering are as follows: Figure 3 As shown;

[0137] S3-Multi-model training:

[0138] Data is extracted from the fabric feature set to form a training set and a validation set, and training is used to form multiple models based on different machine learning algorithms.

[0139] S4- Fusion of multiple models:

[0140] The prediction results of the multiple models are combined by weighted voting to reduce the prediction error and obtain an objective model score. The confusion matrix of the objective model score is as follows: Figure 4 As shown;

[0141] S5-Model Modifications:

[0142] Manual testing gives subjective data, and the model is modified by weighted voting to obtain a comprehensive prediction result. The weighted scoring result of the combination of subjective scoring and objective model scoring is as follows: Figure 5 shown.

[0143] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are within the scope of the technical solution of the present invention.

Claims

1. A fabric classification and evaluation method based on multi-model integration, characterized in that: The steps include: S1-Data Collection: The relevant physical and mechanical indicators of the fabric are collected by a fabric touch measuring instrument, thereby obtaining a fabric property data set; S2-Data Preprocessing: Ensure the integrity and consistency of the data in the fabric property dataset through data cleaning and standardization; Through feature engineering, the data types that have a greater impact on fabric classification in the fabric characteristic data set are screened out to form a fabric feature set; S3-Multi-model training: Data is extracted from the fabric feature set to form a training set and a validation set, and training is used to form multiple models based on different machine learning algorithms. S4- Fusion of multiple models: The prediction results of the multiple models are combined by weighted voting to reduce the prediction error and obtain an objective model score. S5-Model Modifications: Manual inspection provides subjective data, and the model is modified through weighted voting to obtain comprehensive prediction results, further improving the accuracy of the prediction.

2. The fabric classification and evaluation method based on multi-model integration according to claim 1 is characterized in that: In the step S1-data collection, the ALTAS fabric touch tester is used to measure the physical and mechanical indicators of different fabric types, thereby obtaining fabric property data of different fabric types.

3. The fabric classification and evaluation method based on multi-model integration according to claim 1 is characterized in that: The step S2-data preprocessing includes the following steps: S2.1-Data Cleansing: removing outliers in the fabric property data by using the interquartile range method (IQR); S2.2-Data filling: Using KNN missing value filling method to fill the missing values ​​in the fabric characteristic data; S2.3-Data Standardization: The fabric characteristic data is normalized and scaled to the interval [0,1] by Min-Max normalization; S2.4-Feature Engineering: Recursive feature elimination (RFE) and principal component analysis (PCA) are used to perform feature selection and dimension reduction, thereby screening out several fabric characteristic data that are more important for fabric classification and forming a fabric feature set.

4. The fabric classification and evaluation method based on multi-model integration according to claim 1 is characterized in that: In the S3-multi-model training, the multiple models include: a random forest model, a gradient boosting tree model, and a multi-layer perceptron neural network model.

5. The fabric classification and evaluation method based on multi-model integration according to claim 4 is characterized in that: The random forest model in step S3 is composed of multiple decision trees, each tree is trained by different feature subsets and data subsets, and the prediction results of each tree are taken by majority vote.

6. The fabric classification and evaluation method based on multi-model integration according to claim 4 is characterized in that: The gradient boosting tree model in step S3 reduces the prediction error by gradually adding new decision trees, and uses the negative gradient of the loss function as the target of the new tree.

7. The fabric classification and evaluation method based on multi-model integration according to claim 4 is characterized in that: The multilayer perceptron neural network in step S3 is composed of three parts: input layer, hidden layer and output layer. It uses activation function to convert linear combination of input features into nonlinear output, and minimizes loss function through back propagation algorithm to optimize weights and bias, thereby achieving modeling and classification prediction of complex data patterns.

8. The fabric classification and evaluation method based on multi-model integration according to claim 1 is characterized in that: The step S5-model modification includes the following steps: S5.1-User Subjective Prediction: Users give subjective ratings to different types of fabrics based on their actual usage experience; S5.2-Weighted distribution of subjective and objective prediction results: A weighted score is calculated based on the subjective score and the predicted probability of the model. The weighted score comprehensively considers the evaluation of the probability predicted by the model and the subjective score. The weighted score formula is as follows: S5.3-Threshold Setting: A threshold for the total score will be set. If it is greater than the threshold, the final prediction result will be output. If it is less than the threshold, the training will be resumed.

Images