Method for predicting mechanical property of steel pipe after production based on machine learning

Abstract

The invention discloses a method for predicting the mechanical property of a steel pipe after production based on machine learning, and relates to the field of artificial intelligence prediction. Aiming at the limitation of an existing prediction method in processing a complex production process and influence of multiple factors, multiple parameters and microstructure information of steel pipe production are collected, a hierarchical weighted entropy weight method is introduced to screen strong correlation features, an improved multi-modal deep learning model DeepFM is used for fusing multi-type data to perform unified modeling, and the prediction accuracy is improved. And enhancing the interpretability of the model and generating a confidence interval by integrating a decision tree XGBoost, and finally deploying the model to edge equipment to realize real-time prediction. The method can comprehensively consider influence factors, improves prediction precision, effectively fuses multi-modal data, assists a production process to adjust process parameters according to prediction, improves production efficiency and product quality, and is suitable for the field of steel pipe production.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence prediction, and in particular relates to a method for predicting the mechanical properties of steel pipes after production based on machine learning. Background Art

[0002] In the process of steel pipe production, mechanical properties are the key indicators to measure the quality of steel pipes, which directly affect the safety and reliability of steel pipes in practical applications. However, existing methods for predicting the mechanical properties of steel pipes often rely on empirical formulas and physical models, which have certain limitations when dealing with complex production processes and multi-factor influences. For example, in actual production, the mechanical properties of steel pipes are affected by many factors such as the chemical composition of raw materials, production process parameters, and microstructure information. Moreover, there are complex nonlinear relationships between these factors, which are difficult to accurately capture and model using traditional methods. In addition, the time series changes of dynamic parameters in the production process and the fusion processing of multimodal information such as microstructure image data are also problems that traditional methods are difficult to effectively solve. With the continuous improvement of industrial production requirements for steel pipe quality and the widespread application of digital and intelligent technologies in the manufacturing industry, there is an urgent need for a more accurate, efficient and comprehensive method for predicting the mechanical properties of steel pipes after production that can process multi-source data, so as to improve the production quality and efficiency of steel pipes, reduce production costs, and meet the growing market demand. Summary of the invention

[0003] In view of the technical problems existing in the prediction of steel pipe performance, the present invention proposes a method for predicting the mechanical properties of steel pipes after production based on machine learning.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is: comprising the following steps:

[0005] S1. First, collect various parameters in steel pipe production including chemical composition of raw materials, quenching temperature, tempering temperature, rolling temperature, rolling speed, cooling method, cooling time, and microstructure information of steel pipe including grain size and grain boundary distribution;

[0006] S2. Secondly, a method based on hierarchical weighted entropy weights is introduced to screen out features that are strongly related to mechanical properties and remove redundant and invalid features;

[0007] S3. For the dynamic parameters in the production process, the long short-term memory network LSTM is introduced to capture the influence of the time series changes of these parameters on the mechanical properties. For the image data of the microstructure of the steel pipe, a deep learning model based on CNN is constructed to extract image features, integrate image features, time series data with other static process parameters, and use the improved multimodal deep learning model DeepFM to achieve unified modeling of multi-dimensional data and improve prediction accuracy;

[0008] S4. Then, by integrating the decision tree XGBoost, the interpretability of the deep learning model is enhanced, the influence of different production parameters on the mechanical properties of the steel pipe is analyzed, and a confidence interval is generated for the prediction results, so that the process parameters can be adjusted according to the uncertainty of the prediction during the production process to ensure that the actual product performance meets the requirements;

[0009] S5. Finally, the model is deployed on the edge devices of the production line to achieve real-time mechanical property prediction, avoid delays, and improve production efficiency;

[0010] In step S3, the improved multimodal deep learning model DeepFM is used to implement unified modeling of multi-dimensional data in the following steps:

[0011] S31. First, the input data is divided into time series dynamic data, image data and static process data. The feature set of these multimodal data is X = {x1, x2...x n}, where n is the total number of features;

[0012] S32, then perform linear processing on the input features, the formula is: Where w0 is the bias term, x i is the i-th feature in the multimodal data, g(x i ) is the function used for feature evaluation. p k Represents feature x i The distribution probability of

[0013] S33. Then, in order to enhance the understanding of complex multimodal features, the third-order feature interaction is added. The formula is as follows: where v i ,v j ,v k is feature x i , x j , x k The embedding vector of i ,v j ,v k > represents the triple inner product of the embedding vector of the third-order interaction feature, x i x j x k Represents the third-order interaction of features;

[0014] S34, combined with the deep feature representation of multimodal data, high-dimensional feature interactions are learned through deep neural networks, and high-order relations of features are further extracted through multi-layer nonlinear networks;

[0015] S35, combine the linear part, the third-order feature interaction and the deep part, and finally output the prediction result

[0016] Preferably, the steps for implementing feature selection by introducing a method based on hierarchical weighted entropy weights in step 2 are as follows:

[0017] S21. First, according to the different data sources and attributes in the steel pipe production process, the features are divided into multiple categories, and the data of each category is preprocessed to ensure the normalization of the values ​​and avoid calculation deviations caused by different dimensions between features;

[0018] S22. For each feature category, there are m samples in the data set, and each sample has n features. For each feature x i The proportion of eigenvalues ​​obtained through normalization Calculate the entropy value of the feature based on the normalized eigenvalue in Used to ensure the normalization of entropy values ​​and calculate the weight of each feature based on the entropy value Get the feature weight distribution within the category;

[0019] S23, calculate the importance evaluation between categories, use the correlation analysis method to evaluate the contribution of each feature category to the mechanical properties, assign an initial category weight coefficient to each category, and normalize the weights of each category to ensure that the sum of the importance of different categories is 1;

[0020] S24. Multiply the feature weight in each category by the category weight to calculate the final weight W of each feature final (x i )=W k ×W i , where W final (x i ) is the feature x i The final weight, W k is the weight coefficient of the category to which the feature belongs, W i is the feature weight within the category;

[0021] S25. According to the finally calculated feature weight distribution, a threshold of 0.32 is set, and features with weights higher than the threshold are retained.

[0022] Preferably, the implementation process of integrating decision tree XGBoost in step S4 is:

[0023] S41, first prepare multimodal data processed by the deep learning model DeepFM, and obtain the prediction output result of the model;

[0024] S42. Use the XGBoost model to analyze feature importance. By training the XGBoost model to regress the features, we can get the importance score of each feature: Where T iis a feature x i All decision trees of |T i | is the number of decision trees, Δy j is due to the feature x i changes in forecasts due to the introduction of

[0025] S43. The output of the deep learning model is trained through the XGBoost model to generate more accurate prediction results. The objective function of the model training is: Where m is the number of samples, y i is the actual label, is the prediction result, λ is the regularization parameter, Ω(f k ) is the penalty term for model complexity;

[0026] S44. Then, in order to generate a confidence interval for the prediction result, the output distribution of XGBoost is used to calculate the confidence interval, and the formula for setting the confidence interval is: in is the predicted value, z is the critical value of the standard normal distribution, and σ is the standard deviation of the predicted result;

[0027] S45. Finally, make adjustments based on the prediction results and confidence intervals generated by the model. When the confidence interval is large, consider increasing the control of related parameters to reduce uncertainty. When the importance of features is high, give priority to monitoring and optimizing these features to ensure that the final product performance meets the requirements.

[0028] Compared with the prior art, the advantages and positive effects of the present invention are that it comprehensively collects various parameters and microstructure information of steel pipe production, can comprehensively consider the influence of multiple factors, and can make more accurate predictions than existing methods that rely on empirical formulas and physical models. It is based on hierarchical weighted entropy weight screening features to overcome the shortcomings of traditional entropy weight methods and improve accuracy. The improved multimodal deep learning model DeepFM effectively integrates multiple types of data, improves the comprehensiveness and accuracy of predictions, and makes up for the defects of traditional models in integrating multimodal data. XGBoost is integrated to enhance the interpretability of the model, generate confidence intervals to assist production adjustments, and can also make real-time predictions to improve production efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 The stress-strain curve diagram simulated according to the predicted results of mechanical properties;

[0031] Figure 2 Schematic diagram of the microstructure of steel pipe material; DETAILED DESCRIPTION

[0032] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0033] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments of the following disclosure.

[0034] Embodiment: In a steel pipe production enterprise, in order to accurately predict the mechanical properties of steel pipes after production, thereby improving product quality, optimizing production processes and improving production efficiency, and to solve the limitations of existing prediction methods when dealing with complex production processes and multiple factors due to reliance on empirical formulas and physical models, a method for predicting the mechanical properties of steel pipes after production based on machine learning is adopted to achieve this.

[0035] First, it is necessary to collect a variety of parameters in steel pipe production, including the chemical composition of raw materials, quenching temperature, tempering temperature, rolling temperature, rolling speed, cooling method, cooling time, and microstructure information of steel pipes, including grain size and grain boundary distribution; the accurate collection of these data provides rich original information for subsequent accurate prediction of mechanical properties. By collecting various parameters, the factors affecting the mechanical properties of steel pipes can be fully considered, laying a solid foundation for accurate prediction.

[0036] Considering that the defect of the entropy weight method is that it ignores the difference in feature categories and does not consider the importance between categories, it leads to unreasonable weight distribution and inaccurate feature screening. Based on the hierarchical weighted entropy weight method, these problems are overcome, the category characteristics are considered, the category importance is evaluated, the features are accurately screened, and the prediction accuracy is improved. First, according to the different data sources and attributes in the steel pipe production process, the features are divided into multiple categories, and the data of each category is preprocessed to ensure numerical normalization and avoid calculation deviations due to different dimensions between features; for each feature category, there are m samples in the data set, and each sample has n features. For each feature x i The proportion of eigenvalues ​​obtained through normalization Calculate the entropy value of the feature based on the normalized eigenvalue in Used to ensure the normalization of entropy values ​​and calculate the weight of each feature based on the entropy value Obtain the feature weight distribution within the category; calculate the importance evaluation between categories, use the correlation analysis method to evaluate the contribution of each feature category to the mechanical properties, assign an initial category weight coefficient to each category, normalize the weights of each category to ensure that the sum of the importance of different categories is 1; multiply the feature weight in each category by the category weight to calculate the final weight W of each feature final (x i )=W k ×W i , where W final (x i ) is the feature x i The final weight, W k is the weight coefficient of the category to which the feature belongs, W i is the feature weight within the category; finally, according to the final calculated feature weight distribution, a threshold of 0.32 is set to retain features with weights higher than the threshold.

[0037] In order to target different parameters in the production process, the present invention adopts an improved multimodal deep learning model DeepFM. Traditional methods are difficult to effectively integrate time-series dynamic data, image data, and static process parameters in steel pipe production. These data types have different structures and feature representations. Traditional models cannot fully explore the complex correlations between them, resulting in information loss or insufficient utilization. The DeepFM model can effectively integrate multimodal data such as time-series dynamic data, image data, and static process parameters. Through reasonable structural design, different types of data are processed separately, and their features are organically combined within the model, which makes full use of the information in various types of data and improves the comprehensiveness and accuracy of the prediction. First, the input data is divided into time-series dynamic data, image data, and static process data. The feature set of these multimodal data is X={x1,x2...x n}, where n is the total number of features; secondly, the input features are linearly processed, and the formula is: Where w0 is the bias term, x i is the i-th feature in the multimodal data, g(x i ) is the function used for feature evaluation. p k Represents feature x i The distribution probability of ; Then, in order to enhance the understanding of complex multimodal features, the third-order feature interaction is added, and the formula is as follows: where v i ,v j ,v k is feature x i , x j , x k The embedding vector of i ,v j,v k > represents the triple inner product of the embedding vector of the third-order interaction feature, x i x j x k Represents the third-order interaction of features; combines the deep feature representation of multimodal data, learns high-dimensional feature interactions through deep neural networks, and further extracts high-order relationships of features through multi-layer nonlinear networks; combines the linear part, the third-order feature interaction and the deep part, and finally outputs the prediction results

[0038] In order to enhance the interpretability of the deep learning model, the present invention generates confidence intervals for the prediction results by integrating the decision tree XGBoost, so that the process parameters can be adjusted according to the uncertainty of the prediction during the production process to ensure that the actual product performance meets the requirements. First, prepare the multimodal data processed by the deep learning model DeepFM and obtain the prediction output results of the model; use the XGBoost model to analyze the importance of the features, and regress the features by training the XGBoost model to obtain the importance score of each feature: Where T i is a feature x i All decision trees of |T i | is the number of decision trees, Δy j is due to the feature x i The prediction changes caused by the introduction of ; the output of the deep learning model is trained through the XGBoost model to generate more accurate prediction results. The objective function of the model training is: Where m is the number of samples, y i is the actual label, is the prediction result, λ is the regularization parameter, Ω(f k ) is the penalty term for model complexity; then, in order to generate the confidence interval of the prediction result, the output distribution of XGBoost is used to calculate the confidence interval, and the formula for setting the confidence interval is: in is the predicted value, z is the critical value of the standard normal distribution, and σ is the standard deviation of the predicted result. Finally, adjustments are made based on the predicted result generated by the model and its confidence interval. When the confidence interval is large, consider increasing the control of related parameters to reduce uncertainty. When the importance of a feature is high, give priority to monitoring and optimizing these features to ensure that the final product performance meets the requirements.

[0039] Finally, after the model training and optimization is completed, it is deployed to the edge devices of the production line. These edge devices are located at the production site and can collect various data on steel pipe production in real time, such as the chemical composition of raw materials, various process parameters and microstructure information. The collected data will be immediately input into the model, and the model will quickly calculate and output real-time mechanical property prediction results. This process is closely connected, avoiding delays caused by data transmission and remote computing. The production process can be adjusted in time based on this result to ensure efficient production, stabilize product quality, effectively improve production efficiency, and ensure the smooth operation of the production line.

[0040] Figure 1 The stress-strain curve is simulated according to the prediction results of mechanical properties. The curve trends of the predicted value and the true value can be clearly seen in the figure. The predicted value curve is close to the true value curve, which shows that the method for predicting the mechanical properties of steel pipes after production based on machine learning proposed in the present invention has high accuracy. In the low strain area, both the predicted value and the true value show a small stress change. With the increase of strain, the stress gradually increases, and within a certain strain range, the growth trends of the two are basically the same. Figure 2 The detailed information of the microstructure of the steel pipe material is shown. From the figure, we can observe the characteristics of the grain shape, size and clarity of the grain boundary. This microstructural information is of great significance for understanding the mechanical properties of the steel pipe. Smaller and uniform grain size usually helps to improve the strength and toughness of the steel pipe, because smaller grains can hinder the movement of dislocations, thereby enhancing the material's ability to resist deformation. The uniformity of grain boundary distribution will also affect the mechanical properties of the material. Uniformly distributed grain boundaries can make stress more evenly transmitted in the material, reduce stress concentration, and thus improve the overall performance of the steel pipe.

[0041] The above description is only a preferred embodiment of the present invention and does not limit the present invention in other forms. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for predicting the mechanical properties of steel pipes after production based on machine learning, characterized in that: The following steps are involved: S1. First, collect various parameters in steel pipe production including chemical composition of raw materials, quenching temperature, tempering temperature, rolling temperature, rolling speed, cooling method, cooling time, and microstructure information of steel pipe including grain size and grain boundary distribution; S2. Secondly, a method based on hierarchical weighted entropy weights is introduced to screen out features that are strongly related to mechanical properties and remove redundant and invalid features; S3. For the dynamic parameters in the production process, the long short-term memory network LSTM is introduced to capture the influence of the time series changes of these parameters on the mechanical properties. For the image data of the microstructure of the steel pipe, a deep learning model based on CNN is constructed to extract image features, integrate image features, time series data with other static process parameters, and use the improved multimodal deep learning model DeepFM to achieve unified modeling of multi-dimensional data and improve prediction accuracy; S4. Then, by integrating the decision tree XGBoost, the interpretability of the deep learning model is enhanced, the influence of different production parameters on the mechanical properties of the steel pipe is analyzed, and a confidence interval is generated for the prediction results, so that the process parameters can be adjusted according to the uncertainty of the prediction during the production process to ensure that the actual product performance meets the requirements; S5. Finally, the model is deployed on the edge devices of the production line to achieve real-time mechanical property prediction, avoid delays, and improve production efficiency; In step S3, the improved multimodal deep learning model DeepFM is used to implement unified modeling of multi-dimensional data in the following steps: S31. First, the input data is divided into time series dynamic data, image data and static process data. The feature set of these multimodal data is X = {x1, x2...x n }, where n is the total number of features; S32, then perform linear processing on the input features, the formula is: Where w0 is the bias term, x i is the i-th feature in the multimodal data, g(x i ) is the function used for feature evaluation. p k Represents feature x i The distribution probability of S33. Then, in order to enhance the understanding of complex multimodal features, the third-order feature interaction is added. The formula is as follows: where v i ,v j ,v k is feature x i , x j , x k The embedding vector of i ,v j ,v k > represents the triple inner product of the embedding vector of the third-order interaction feature, x i x j x k Represents the third-order interaction of features; S34, combined with the deep feature representation of multimodal data, high-dimensional feature interactions are learned through deep neural networks, and high-order relations of features are further extracted through multi-layer nonlinear networks; S35, combine the linear part, the third-order feature interaction and the deep part, and finally output the prediction result 2. The method for predicting mechanical properties of steel pipes after production based on machine learning according to claim 1, characterized in that: The steps for implementing feature selection by introducing the method based on hierarchical weighted entropy weight in step 2 are as follows: S21. First, according to the different data sources and attributes in the steel pipe production process, the features are divided into multiple categories, and the data of each category is preprocessed to ensure the normalization of the values ​​and avoid calculation deviations caused by different dimensions between features; S22. For each feature category, there are m samples in the data set, and each sample has n features. For each feature x i The proportion of eigenvalues ​​obtained through normalization Calculate the entropy value of the feature based on the normalized eigenvalue in Used to ensure the normalization of entropy values ​​and calculate the weight of each feature based on the entropy value Get the feature weight distribution within the category; S23, calculate the importance evaluation between categories, use the correlation analysis method to evaluate the contribution of each feature category to the mechanical properties, assign an initial category weight coefficient to each category, and normalize the weights of each category to ensure that the sum of the importance of different categories is 1; S24. Multiply the feature weight in each category by the category weight to calculate the final weight W of each feature final (x i )=W k ×W i , where W final (x i ) is the feature x i The final weight, W k is the weight coefficient of the category to which the feature belongs, W i is the feature weight within the category; S25. According to the finally calculated feature weight distribution, a threshold of 0.32 is set, and features with weights higher than the threshold are retained.

3. The method for predicting mechanical properties of steel pipes after production based on machine learning according to claim 1, characterized in that: The implementation process of integrating decision tree XGBoost in step S4 is as follows: S41, first prepare multimodal data processed by the deep learning model DeepFM, and obtain the prediction output result of the model; S42. Use the XGBoost model to analyze feature importance. By training the XGBoost model to regress the features, we can get the importance score of each feature: Where T i is a feature x i All decision trees of |T i | is the number of decision trees, Δy j is due to the feature x i changes in forecasts due to the introduction of S43. The output of the deep learning model is trained through the XGBoost model to generate more accurate prediction results. The objective function of the model training is: Where m is the number of samples, y i is the actual label, is the prediction result, λ is the regularization parameter, Ω(f k ) is the penalty term for model complexity; S44. Then, in order to generate a confidence interval for the prediction result, the output distribution of XGBoost is used to calculate the confidence interval, and the formula for setting the confidence interval is: in is the predicted value, z is the critical value of the standard normal distribution, and σ is the standard deviation of the predicted result; S45. Finally, make adjustments based on the prediction results and confidence intervals generated by the model. When the confidence interval is large, consider increasing the control of related parameters to reduce uncertainty. When the importance of features is high, give priority to monitoring and optimizing these features to ensure that the final product performance meets the requirements.

Images