Casting defect prediction method based on multi-scale and multi-mode machine learning and prediction system thereof

Abstract

The invention discloses a casting defect prediction method based on multi-scale and multi-mode machine learning and a prediction system thereof, and belongs to the field of casting. In order to solve the defects that an existing defect detection and prediction method needs a large amount of data resources and data models, the simulation result is greatly deviated from the reality, complex and rare defects cannot be detected, the simulation complexity is high, and the calculation cost is high, the method comprises the steps that multi-scale and multi-mode casting data are used, and a database is constructed; carrying out machine learning on the multi-scale and multi-modal data through a deep feedforward neural network, obtaining target casting data corresponding to nodes of each neural network, and obtaining weights of parameters of an input layer and a hidden layer of the neural network and a correlation degree corresponding to the target casting data; and outputting an accurate prediction result of the defect type of the casting by inputting casting data. The method is mainly used for detecting and predicting the defects of the casting.

Description

Technical Field

[0001] The present invention relates to a field, and in particular to a casting defect prediction method and a prediction system based on multi-scale and multi-modal machine learning. Background Art

[0002] Casting technology is the cornerstone of modern industry, possessing crucial significance in product manufacturing, industrial development, economic, and social value. However, casting defects can negatively impact the performance and quality of castings. For example, pores, shrinkage cavities, and inclusions can weaken the material's strength and toughness, reducing the fatigue life of castings. Cold shuts, insufficient pouring, and metal penetration can lead to surface irregularities or dimensions that don't meet design requirements, impacting not only the product's appearance but also machining difficulties or the inability to meet assembly requirements. Casting defects are a major cause of scrapped castings, increasing production costs and wasting resources while potentially causing serious safety concerns. Furthermore, repairing these defects can incur additional time and costs.

[0003] Predicting casting defects can help adjust casting process parameters such as pouring temperature, pouring speed, and mold design in advance, thereby reducing the occurrence of defects in castings. Defect prediction can reduce scrap caused by defects, lower production costs, and reduce subsequent inspection and repair work. Defect prediction can shorten trial-and-error cycles and reduce production interruptions caused by defects, thereby improving production efficiency. Therefore, predicting casting defects not only helps improve casting quality and production efficiency, but also significantly reduces production costs and scrap rates, which is of great significance to the sustainable development of the foundry industry.

[0004] Different casting compositions, production process parameters, and microstructures have their own impacts on casting defects. For example, excessively low pouring temperatures can lead to insufficient fluidity of the molten metal, resulting in under-casting. In cast iron, excessively high carbon content can lead to the formation of white cast iron, resulting in negative effects such as reduced strength and fluidity. Mn, as an austenitizing element, both deoxidizes and increases carbon activity, but excessive Mn can easily coarsen grains and reduce toughness. By establishing a dependency relationship between "casting composition-production process parameters-microstructure-casting defects" based on casting composition, production process parameters, and microstructure, accurate predictions of casting defects can be achieved, significantly reducing product defects while simultaneously increasing production efficiency and yield, and reducing production costs ("two increases and one decrease").

[0005] Defect prediction using trial and error methods has several problems: 1. It relies on a large number of actual production tests to discover and analyze defects, requiring repeated adjustments to process parameters and multiple tests. This is time-consuming and labor-intensive, making it difficult to meet the urgent demand for efficient production in modern industry. 2. Due to the need for multiple tests and adjustments, trial and error methods waste raw materials and energy, while increasing equipment wear and labor costs. This cost issue is particularly prominent in the production of complex castings. 3. It relies on the experience and intuition of technicians and lacks systematic theoretical support and scalability. 4. When faced with new materials, processes, or products, a large number of new tests are required, making it difficult to quickly adapt. 4. Trial and error methods are mainly based on empirical summaries of known defects, making it difficult to accurately predict some new or rare defect types through limited tests. 5. Trial and error methods usually analyze and adjust defects only after they have occurred, failing to provide early warning of defects and thus failing to effectively prevent them. 6. Trial and error methods often rely on local experience when optimizing process parameters, making it difficult to find the global optimal solution. Therefore, the experimental trial and error method has problems in casting defect prediction, such as low efficiency, high cost, data imbalance, and difficulty in fully covering defect types. It is difficult to meet the modern casting industry's demand for efficient and accurate defect prediction.

[0006] Casting defect prediction through simulation presents several challenges: The accuracy of simulations is highly dependent on the accuracy of the model. The casting process involves complex physical phenomena, such as fluid flow, solidification heat transfer, and stress and deformation. Simulating these processes requires precise mathematical models and extensive computing resources. However, the actual casting process contains many factors that are difficult to accurately model, such as uncertainties in material properties and heat exchange between the mold and the metal. This can lead to deviations between simulation results and actual conditions. While simulation technology can predict common casting defects (such as shrinkage, porosity, and pores), current simulation methods are limited for complex or rare defects (such as cold shuts, oxide inclusions, and cracks). Simulations require a large amount of input data, including material properties, geometric models, and boundary conditions. Preparing and processing this data requires significant time and effort, and the accuracy of the data directly impacts the reliability of the simulation results. Furthermore, the complexity of the geometric model can make meshing difficult, further increasing the complexity and computational cost of the simulation. In summary, while simulation has important application value in casting defect prediction, it still faces numerous challenges in terms of model accuracy, data processing, defect prediction capabilities, and computing resources.

[0007] Therefore, there is a need for a casting defect prediction method and prediction system based on multi-scale and multi-modal machine learning with high precision, strong data processing capability and strong defect prediction capability. Summary of the Invention

[0008] In order to solve the problems of existing defect detection and prediction methods, such as the need for a large amount of data resources and data models, large deviation between simulation results and actual results, inability to detect complex and rare defects, high simulation complexity and high computational cost, the present invention provides a casting defect prediction method and prediction system based on multi-scale and multi-modal machine learning.

[0009] The present invention provides a method for predicting casting defects based on multi-scale and multi-modal machine learning, comprising the following steps:

[0010] S1. Acquire multi-scale and multi-modal casting data, and perform data processing on the casting data;

[0011] S2. performing correlation analysis on the pre-processed casting data to obtain correlation between the casting data;

[0012] S3. Select casting data with a correlation degree higher than a preset threshold, and construct an initial data set and a casting defect prediction model based on the selected data;

[0013] S4. Training the casting defect prediction model using the initial data set to obtain a deep feedforward neural network;

[0014] S5. Test the deep feedforward neural network based on the test set to obtain target casting data corresponding to each network node, and calculate parameters corresponding to casting defects based on the correlation degree corresponding to the target casting data, and predict the types of casting defects that will occur in the current batch and their probability and frequency of occurrence.

[0015] Furthermore: in S1, the casting data includes casting composition, production process parameters, tissue morphology images and casting defect images; the data processing includes data cleaning, data preprocessing and normalization processing.

[0016] Further: in S1, the data processing includes:

[0017] Screening the casting data to remove invalid data that does not conform to a preset data format, and performing data cleaning to remove outliers;

[0018] The initial valid data are divided into k clusters using a clustering model, the distance between each initial valid data and the cluster to which it belongs is calculated, and the initial valid data with the distance less than or equal to a distance threshold are normalized.

[0019] Further: in S2, the correlation analysis is performed on the pre-processed multiple casting defects and the casting data respectively to obtain the correlation between the casting defects and the casting data, including:

[0020] Performing correlation analysis on pre-processed casting data of different historical batches and casting defects to obtain the correlation between each casting data and the casting defect parameters;

[0021]

[0022] Among them, i is the serial number of each casting data, is the correlation between the i-th casting data and the L-th casting defect parameter, k is the serial number corresponding to each casting data, L is the serial number of the casting defect parameter, is the casting defect parameter, is the Lth casting defect parameter corresponding to the kth data corresponding to the i-th casting composition, production process parameters and microstructure, is the minimum difference between the k data corresponding to the casting composition, production process parameters and microstructure and the L-th casting defect parameter, is the maximum value of the Lth casting defect parameter among the k data corresponding to the material composition, production process parameters or organizational morphology, and ρ is the adjustable coefficient.

[0023] Further: in S4, the specific process of training the casting defect prediction model is: randomly sampling a preset number of subsamples from the initial data set as a training set, and the remaining samples as a test set, the input layer is composed of casting composition, production process parameters and tissue morphology, and the output layer is composed of casting defects after image recognition, and after training the casting defect prediction model according to the training set, a suitable hidden layer is established to obtain a deep feedforward neural network.

[0024] Further: in S4, the training process of the casting defect prediction model according to the training set includes:

[0025] Calculate the weights of the input layer parameters in the neural network, calculate the hidden layer parameters based on the input layer parameters and their weights, and calculate the weights of each hidden layer parameter;

[0026] The hidden units of the hidden layer of the computational neural network include:

[0027] Use the sigmoid function to calculate the hidden units of the hidden layer:

[0028]

[0029] Among them, j represents the serial number of each casting data, d is the number of all casting parameters, x j (j=0,…,d) is the input, Z h (h=1,…,H) is the hidden unit, Z0 is the bias of the hidden layer, W hj is the weight of the first layer; H is the total number of hidden units, sigmoid is the activation function, It is the inner product of the vector consisting of bias weights and the vector consisting of input data, exp is the exponential function with e as the base, h is the ordinal number of the hidden unit, i is the ordinal number of each casting data, and w0 is the weight of x0.

[0030] Further: weighted summation is performed according to the hidden units, and different defect parameters are obtained through different weights,

[0031]

[0032] Among them, y i is the output unit, v ih is the weight of the second layer, z h is the hidden unit, H is the total number of hidden units, h is the ordinal number of the hidden unit, v i0 is the bias weight.

[0033] The casting defect prediction system of the present invention for the casting defect prediction method based on multi-scale and multi-modal machine learning comprises a data processing module, a correlation analysis module, casting data, and a casting defect prediction model;

[0034] The data processing module is used to perform data cleaning, data preprocessing and normalization on casting data;

[0035] The correlation analysis module is used to obtain the correlation between casting defects and casting data;

[0036] The casting data is used to train the casting defect prediction model, thereby obtaining a deep feedforward neural network;

[0037] The deep feedforward neural network is used to predict defects on the casting data.

[0038] The beneficial effects of the present invention are:

[0039] The casting defect prediction method and prediction system based on multi-scale and multi-modal machine learning described in the present invention have significant advantages in casting defect prediction:

[0040] 1. The present invention, through the neural network, can simultaneously process phenomena and processes from microscopic to macroscopic scales and fuse different types of data (such as images, temperature fields, stress fields, etc.), thereby significantly improving the accuracy of defect prediction;

[0041] 2. Model parameters and prediction results can be continuously optimized through real-time data collection and feedback. This continuous improvement mechanism can ensure that the model maintains high accuracy under different production conditions; it can achieve high prediction efficiency and low cost through multi-scale modeling, multi-modal data fusion and intelligent optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a structural diagram of a deep feedforward neural network;

[0043] Figure 2 It is a flow chart of the defect detection method of the casting defect recognition model;

[0044] Figure 3 It is a schematic diagram of the correlation between target composition-production process parameters-microstructure-casting defects. DETAILED DESCRIPTION

[0045] The following are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be included in the scope of protection of the present invention. The embodiments described below are only used to explain the present invention and cannot be interpreted as limiting the present invention. The scope of protection of the present invention should be based on the scope of protection of the claims. The embodiments of the present invention are described in detail below. In order to facilitate the description of the present invention and simplify the description, the technical terms used in the description of the present invention should be interpreted broadly, including but not limited to conventional replacement schemes not mentioned in this application, and also including direct implementation and indirect implementation.

[0046] Example 1

[0047] Combine Figure 1 and Figure 2 This embodiment describes a casting defect prediction method based on multi-scale and multi-modal machine learning, including the following steps:

[0048] S1. Acquire multi-scale and multi-modal casting data, including casting composition, production process parameters, microstructure images, and casting defect images, and perform cleaning, preprocessing, and normalization on the casting composition, production process parameters, microstructure images, and casting defect images in sequence;

[0049] The data processing includes:

[0050] Screening the casting data to remove invalid data that does not conform to a preset data format, and performing data cleaning to remove outliers;

[0051] The initial valid data are divided into k clusters using a clustering model, the distance between each initial valid data and the cluster to which it belongs is calculated, and the initial valid data with the distance less than or equal to a distance threshold are normalized;

[0052] S2. Perform correlation analysis on the casting defects after pretreatment and the casting composition, production process parameters and microstructure, and obtain the correlation between the casting defects and the casting composition, production process parameters and microstructure;

[0053] The correlation analysis of the various casting defects after pre-processing and the casting data is performed respectively, wherein the various casting defects such as shrinkage or cold shut are obtained, and the correlation between the casting defects and the casting data is obtained, including:

[0054] Performing correlation analysis on pre-processed casting data of different historical batches and casting defects to obtain the correlation between each casting data and the casting defect parameters;

[0055]

[0056] Among them, i is the serial number of each casting data, is the correlation between the i-th casting data and the L-th casting defect parameter, k is the serial number corresponding to each casting data, and L is the serial number of the casting defect parameter (generally, the parameter with a larger absolute value has a smaller serial number). is the casting defect parameter, is the Lth casting defect parameter corresponding to the kth data corresponding to the i-th casting composition, production process parameters and microstructure, is the minimum difference between the k data corresponding to the casting composition, production process parameters and microstructure and the L-th casting defect parameter, is the maximum value of the Lth casting defect parameter among the k data corresponding to the material composition, production process parameters or organizational morphology, and ρ is the adjustable coefficient.

[0057] S3. Select key casting components, production process parameters, and microstructures with a correlation degree higher than a correlation threshold from among the casting components, production process parameters, and microstructures, and construct an initial data set using these corresponding multi-scale and multi-modal data and casting defects;

[0058] S4. Obtain a preset number of subsamples from the initial data set by random sampling, form an input layer consisting of casting composition, production process parameters, and microstructure, and form an output layer consisting of casting defects after image recognition, establish an appropriate hidden layer after training based on the subsamples, and obtain a deep feedforward neural network;

[0059] The specific process of training the casting defect prediction model is as follows: randomly sampling a preset number of subsamples from the initial data set as a training set, and the remaining samples as a test set. The input layer is composed of casting composition, production process parameters and tissue morphology, and the output layer is composed of casting defects after image recognition. After training the casting defect prediction model according to the training set, a suitable hidden layer is established to obtain a deep feedforward neural network.

[0060] In S4, the process of training the casting defect prediction model according to the training set includes:

[0061] Calculate the weights of the input layer parameters in the neural network, calculate the hidden layer parameters based on the input layer parameters and their weights, and calculate the weights of each hidden layer parameter;

[0062] The hidden units of the hidden layer of the computational neural network include:

[0063] Use the sigmoid function to calculate the hidden units of the hidden layer:

[0064]

[0065] Among them, j represents the serial number of each casting data, d is the number of all casting parameters, x j (j=0,…,d) is the input, Z h (h=1,…,H) is the hidden unit, Z0 is the bias of the hidden layer, W hj is the weight of the first layer; H is the total number of hidden units, sigmoid is the activation function, It is the inner product of the vector consisting of bias weights and the vector consisting of input data, exp is the exponential function with e as the base, h is the ordinal number of the hidden unit, i is the ordinal number of each casting data, and w0 is the weight of x0.

[0066] According to the weighted summation of the hidden units, different defect parameters are obtained through different weights.

[0067]

[0068] Among them, y i is the output unit, v ih is the weight of the second layer, z h is the hidden unit, H is the total number of hidden units, h is the ordinal number of the hidden unit, v i0 is the bias weight.

[0069] S5. Based on the multi-scale multimodal data corresponding to the key casting components, production process parameters and tissue morphology in the remaining data and the preset number of neural networks, obtain the target key casting components, production process parameters and tissue morphology corresponding to each network node, and calculate the parameters corresponding to the casting defects based on the correlation between the target key casting components, production process parameters and tissue morphology (a large correlation means a greater weight in the calculation process), and accurately predict the types of casting defects that will occur in the current batch and their probability and frequency of occurrence.

[0070] like Figure 3 As shown in the figure, the relationship between "target composition-production process parameters-microstructure-casting defects" is constructed based on multi-scale and multi-modal machine learning to obtain accurate prediction results of casting defects.

[0071] Example 2

[0072] This embodiment is described in conjunction with Example 1. This embodiment discloses a casting defect prediction system based on a casting defect prediction method using multi-scale multimodal machine learning, comprising a data processing module, a correlation analysis module, casting data, and a casting defect prediction model.

[0073] The data processing module is used to perform data cleaning, data preprocessing and normalization on casting data;

[0074] The correlation analysis module is used to obtain the correlation between casting defects and casting data;

[0075] The casting data is used to train the casting defect prediction model, thereby obtaining a deep feedforward neural network;

[0076] The deep feedforward neural network is used to predict defects on the casting data.

Claims

1. A casting defect prediction method based on multi-scale and multi-modal machine learning, characterized in that: The steps include: S1. Acquire multi-scale and multi-modal casting data, and perform data processing on the casting data; S2. performing correlation analysis on the pre-processed casting data to obtain correlation between the casting data; S3. Select casting data with a correlation higher than a preset threshold, and construct an initial data set and a casting defect prediction model based on the selected data; S4. Training the casting defect prediction model using the initial data set to obtain a deep feedforward neural network; S5. Test the deep feedforward neural network based on the test set to obtain target casting data corresponding to each network node, and calculate parameters corresponding to casting defects based on the correlation degree corresponding to the target casting data, and predict the types of casting defects that will occur in the current batch and their probability and frequency of occurrence.

2. The casting defect prediction method based on multi-scale and multi-modal machine learning according to claim 1, characterized in that: In S1, the casting data includes casting composition, production process parameters, tissue morphology images and casting defect images; the data processing includes data cleaning, data preprocessing and normalization processing.

3. The casting defect prediction method based on multi-scale and multi-modal machine learning according to claim 1, characterized in that: In S1, the data processing includes: Screening the casting data to remove invalid data that does not conform to a preset data format, and performing data cleaning to remove outliers; The initial valid data are divided into k clusters using a clustering model, the distance between each initial valid data and the cluster to which it belongs is calculated, and the initial valid data with the distance less than or equal to a distance threshold are normalized.

4. The casting defect prediction method based on multi-scale and multi-modal machine learning according to claim 1, characterized in that: In S2, the correlation analysis is performed on the pre-processed multiple casting defects and the casting data respectively to obtain the correlation between the casting defects and the casting data, including: Performing correlation analysis on pre-processed casting data of different historical batches and casting defects to obtain the correlation between each casting data and the casting defect parameters; Among them, i is the serial number of each casting data, is the correlation between the i-th casting data and the L-th casting defect parameter, k is the serial number corresponding to each casting data, L is the serial number of the casting defect parameter, is the casting defect parameter, is the Lth casting defect parameter corresponding to the kth data corresponding to the i-th casting composition, production process parameters and microstructure, is the minimum difference between the k data corresponding to the casting composition, production process parameters and microstructure and the L-th casting defect parameter, is the maximum value of the Lth casting defect parameter among the k data corresponding to the material composition, production process parameters or organizational morphology, and ρ is the adjustable coefficient.

5. The casting defect prediction method based on multi-scale and multi-modal machine learning according to claim 1, characterized in that: In S4, the specific process of training the casting defect prediction model is as follows: randomly sampling a preset number of subsamples from the initial data set as a training set, and the remaining samples as a test set, the input layer is composed of casting composition, production process parameters and tissue morphology, and the output layer is composed of casting defects after image recognition, and after training the casting defect prediction model according to the training set, a suitable hidden layer is established to obtain a deep feedforward neural network.

6. The casting defect prediction method based on multi-scale and multi-modal machine learning according to claim 5, characterized in that: In S4, the process of training the casting defect prediction model according to the training set includes: Calculate the weights of the input layer parameters in the neural network, calculate the hidden layer parameters based on the input layer parameters and their weights, and calculate the weights of each hidden layer parameter; The hidden units of the hidden layer of the computational neural network include: Use the sigmoid function to calculate the hidden units of the hidden layer: Among them, j represents the serial number of each casting data, d is the number of all casting parameters, x j (j=0,…,d) is the input, Z h (h=1,…,H) is the hidden unit, Z0 is the bias of the hidden layer, W hj is the weight of the first layer; H is the total number of hidden units, sigmoid is the activation function, It is the inner product of the vector consisting of bias weights and the vector consisting of input data, exp is the exponential function with e as the base, h is the ordinal number of the hidden unit, i is the ordinal number of each casting data, and w0 is the weight of x0.

7. The casting defect prediction method based on multi-scale and multi-modal machine learning according to claim 6, characterized in that: According to the weighted summation of the hidden units, different defect parameters are obtained through different weights. Among them, y i is the output unit, v ih is the weight of the second layer, z h is the hidden unit, H is the total number of hidden units, h is the ordinal number of the hidden unit, v i0 is the bias weight.

8. A casting defect prediction system for the casting defect prediction method based on multi-scale and multi-modal machine learning according to any one of claims 1 to 7, characterized in that: Including data processing module, correlation analysis module, casting data and casting defect prediction model, The data processing module is used to perform data cleaning, data preprocessing and normalization on casting data; The correlation analysis module is used to obtain the correlation between casting defects and casting data; The casting data is used to train the casting defect prediction model, thereby obtaining a deep feedforward neural network; The deep feedforward neural network is used to predict defects on the casting data.

Images