Method for analyzing importance of die casting quality related variables based on multivariate standardized embedding and feature fusion

Abstract

This invention discloses a method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion. First, the die-casting process data and quality label data are cleaned, standardized, and matched. Then, a die-casting variable importance identification model is constructed and trained, including a variable standardized embedding layer, a variable selection layer, and a quality classification layer. The variable standardized embedding layer, based on the time-varying and discrete characteristics of the die-casting data, uses different embedding methods to transform different types of variables into tensors of the same shape. The variable selection layer calculates variable selection weights based on features obtained from the fusion of the tensors of the same shape for each variable or based on a learnable weight matrix. The tensors of the same shape for each variable are weighted and summed to generate a fused implicit expression. The quality classification layer takes the fused implicit expression as input and outputs quality labels. This invention innovatively combines multivariate standardized embedding and feature fusion to quantify the influence of various process parameters on quality.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and quality monitoring, specifically to a method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion. Background Technology

[0002] Die casting, as a highly efficient and precise metal forming process, is widely used in the automotive, aerospace, and electronics manufacturing industries. The die casting process involves multiple process parameters (such as pressure, temperature, and stroke) and equipment state variables (such as running time and cycle time), which collectively determine the quality of the final product. However, the die casting process is highly nonlinear and time-varying, and complex interactions may exist between variables, leading to quality fluctuations and defects. Traditional quality control methods mainly rely on empirical rules and the monitoring of single variables, making it difficult to comprehensively capture the complex relationships between multiple variables and accurately predict and classify product quality.

[0003] In recent years, with the development of industrial big data and artificial intelligence technologies, data-driven quality analysis methods have gradually become a research hotspot. Existing technologies have attempted to classify die-casting quality using machine learning models, but due to the multi-source heterogeneity of die-casting data (e.g., the coexistence of continuous numerical data and discrete enumerated data), direct modeling often yields poor results. Furthermore, existing methods typically neglect the importance of variable selection, leading to increased model sensitivity to redundant or irrelevant variables, thus reducing classification accuracy and interpretability. Therefore, there is an urgent need for a method that can effectively handle multivariate data, extract key features, and achieve high-quality classification. Summary of the Invention

[0004] To address the shortcomings of existing technologies and improve the intelligence and accuracy of die-casting quality monitoring, this invention proposes a method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion. By combining data preprocessing, variable standardized embedding, feature fusion, variable selection, classification model construction, variable importance analysis, and deep learning techniques, this method aims to solve the complexity and uncertainty of multivariate data in the die-casting production process, achieving accurate classification of die-casting quality and importance analysis of related variables. This invention is applicable to die-casting process optimization, quality monitoring, and the design and application of intelligent manufacturing systems, and has broad technological application prospects.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion includes the following steps:

[0007] S1: Acquire die-casting process data and quality label data within a predetermined range, and preprocess the data, including data cleaning, standardization, and matching operations;

[0008] S2: Construct and train a die-casting variable importance identification model, which includes a variable standardization embedding layer, a variable selection layer, and a quality classification layer. The variable standardization embedding layer is used to transform different types of variables into tensors of the same shape based on the time-varying and discrete characteristics of the die-casting data using different embedding methods. The variable selection layer calculates variable selection weights based on features obtained by fusing the tensors of the same shape of each variable or based on a learnable weight matrix, and then performs a weighted summation of the tensors of the same shape of each variable to generate a fused implicit expression. The quality classification layer takes the fused implicit expression as input and outputs quality labels that are close to the true values. After training, the die-casting variable importance identification model outputs predicted quality labels, while the variable selection layer outputs the importance of relevant variables.

[0009] Furthermore, the die-casting process data includes set parameters, sensor-collected parameters, and statistical data;

[0010] The setting parameters are parameters related to the die casting process that are set for the die casting machine during die casting production.

[0011] The sensor collects parameters including actual measured values ​​corresponding to the parameters set during the die casting process, as well as other actual measured values ​​that can be measured and monitored during the die casting process.

[0012] The statistical data includes the die-casting machine's running time, date, cycle count, number of products produced, and number of defective products produced;

[0013] The quality label data represents the change in the number of defective products produced.

[0014] Furthermore, in the data preprocessing stage, the sample data is first divided according to the abrupt difference in the number of products produced, resulting in sample V = {V1, V2, ..., V...}. m}∈R m×T×1 Then, samples with abnormal durations, abnormal changes in the number of produced samples, and abnormal changes in the number of produced defective samples are removed. Subsequently, the maximum sequence length is determined based on the sample duration, and a masking matrix is ​​generated. For non-enumerated data, Z-score normalization is used for standardization. Finally, data matching is performed, that is, the timestamp of the quality label data is matched with the end timestamp of the die-casting process data sample, and samples with a difference between the two timestamps exceeding a preset threshold are removed to ensure the accuracy and consistency of the data.

[0015] Furthermore, the implementation process of the variable standardization embedding layer is as follows:

[0016] Given a sample of die-casting process data V = {V1, V2, ..., V...} m}∈R m×T×1 and implicit expression dimension f h Where m is the number of variables in the sample, and V i ={V i,1 V i,2 ,…,V i,T}∈R T×1 V represents the sampled value of the i-th variable. i,t ∈R 1 This represents the sampled value of the i-th variable at time t;

[0017] Variable standardization embedding layer M vf Represented as That is, each variable V i The feature dimensions are mapped to a uniform dimension size, specifically:

[0018] For continuous numerical variables, a fully connected linear layer is used to expand their feature dimension to f. h ;

[0019] For discrete enumerated variables, after passing through the embedding layer and the fully connected linear layer, the feature dimension is expanded to f. h ;

[0020] Then, all the expanded variables are arranged according to their data sample V = {V1, V2, ..., V} in the die-casting process. m The inputs in the order specified in the table are concatenated to obtain a tensor of the same shape.

[0021] Furthermore, the variable selection layer M vs Represented as That is, for tensors V of the same shape h Feature extraction and feature selection are performed to generate fused implicit variables. Specifically:

[0022] (1) For V h The sampled value of the i-th variable at time t Feature extraction is performed using a gated residual network (GRN) to obtain...

[0023] (2) Obtain variable selection weights through the following two methods

[0024] a. V h By swapping and flattening the last two dimensions, we obtain... Then, it is generated through a gated residual network (GRN), a linear fully connected layer, and a softmax function. Right now

[0025] b. Provide a learnable matrix w vs ∈R T×m Generated by the softmax function Right now Variable selection weights It represents the importance of variables and is used in variable importance analysis;

[0026] (3) Generate the implicit representation at time t using a weighted summation method. Ultimately, the implicit expression of fusion is obtained.

[0027] Furthermore, the quality classification layer M c Expressed as: That is, the implicit expression of fusion As input, H is transformed into a quality label probability vector. Where N c Given the number of quality label categories, select the label with the highest probability as the output predicted quality label.

[0028] Furthermore, during the training of the die-casting variable importance identification model, the input is a sample V = {V1, V2, ..., V...} of the die-casting process data. m}∈R m×T×1 and the corresponding real quality label Where V is a real matrix containing T time steps, m variables, and each variable having a feature dimension of 1; It is a discrete enumeration value that can be transformed into a true quality label probability vector through one-hot encoding. The output is the predicted quality label probability vector.

[0029] By minimizing the difference between the predicted value P and the actual value The cross-entropy loss function, which calculates the difference between the predicted and actual values, makes the predicted value P as close as possible to the true value.

[0030] An apparatus for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion includes one or more processors for implementing a method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion.

[0031] An electronic device, comprising:

[0032] One or more processors;

[0033] A storage device for storing one or more programs, which, when executed by the electronic device, enable the electronic device to implement a method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion.

[0034] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion.

[0035] The beneficial effects of this invention are as follows:

[0036] This invention presents a method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion. Through a variable standardized embedding layer, it transforms multi-source heterogeneous data from the die-casting process into a unified expression, resolving data compatibility issues. The variable selection layer, based on a learnable weight matrix or unified expression, automatically identifies key variables, reducing redundant interference and improving classification accuracy and interpretability. The classification model utilizes implicit expressions to capture complex relationships between multiple variables, achieving accurate quality classification and reducing misclassification rates. Simultaneously, the variable selection weights provide a quantitative analysis of the impact of each variable on quality, offering a scientific basis for process optimization. This method is not dependent on specific equipment or processes, is applicable to various die-casting scenarios, and its modular design facilitates expansion and integration. Through deep learning and data-driven technologies, it automates data preprocessing, feature extraction, and classification analysis, reducing reliance on human experience and significantly improving the intelligence and automation level of quality monitoring. This invention effectively enhances the quality control capabilities and process optimization efficiency of die-casting production, providing crucial technical support for the field of intelligent manufacturing. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion, according to an embodiment of the present invention.

[0038] Figure 2 This is a schematic diagram of the variable standardization embedding layer and variable selection layer in a specific embodiment of the present invention.

[0039] Figure 3 This is a schematic diagram illustrating the importance of variables obtained by applying the method of the present invention to a certain die-casting dataset. Detailed Implementation

[0040] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0041] like Figure 1As shown, the method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion provided by this invention includes the following steps:

[0042] Step 1: Obtain die-casting process data and quality label data within the predetermined range, and preprocess the data, including data cleaning, standardization, and matching operations.

[0043] The specific implementation process is as follows: Acquire actual die-casting process data and its corresponding quality label data. Die-casting process data includes set parameters, sensor-collected parameters, and statistical data. Set parameters are those related to the die-casting process set for the die-casting machine during production, such as pressure and stroke. Sensor-collected parameters are determined by the die-casting machine brand and include actual measured values ​​corresponding to the set parameters during the die-casting process, as well as other measurable and monitorable values ​​during the die-casting process. Statistical data covers the die-casting machine's running time, date, cycle time, number of products produced, and number of defective products produced. Quality label data represents the change in the number of defective products produced.

[0044] In the data preprocessing stage, the sample data is first divided according to the abrupt difference in the number of products produced, resulting in sample V = {V1, V2, ..., V...}. m}∈R m×T×1 Specifically, the production process of a product typically lasts 45 to 90 seconds. During this time, the number of products produced remains constant. However, when the previous product finishes production and the next product begins, the number of products produced changes abruptly. By capturing this abrupt change, the data can be inversely divided into independent samples according to the product production process. Then, samples with abnormal durations, abnormal changes in the number of produced samples, and abnormal changes in the number of produced defective samples are removed. Subsequently, the maximum sequence length is determined based on the sample duration, and a masking matrix is ​​generated. For non-enumerated data, Z-score normalization is used for standardization. Finally, data matching is performed, that is, the timestamp of the quality label data is matched with the end timestamp of the die-casting process data sample, and samples with a timestamp difference exceeding a preset threshold are removed to ensure data accuracy and consistency.

[0045] Step 2: Construct and train a die-casting variable importance identification model. This model includes a variable standardization embedding layer, a variable selection layer, and a quality classification layer. The variable standardization embedding layer transforms different types of variables into tensors of the same shape using different embedding methods based on the time-varying and discrete characteristics of the die-casting data. The variable selection layer calculates variable selection weights based on features obtained by fusing the tensors of the same shape for each variable or based on a learnable weight matrix, and then performs a weighted summation of the tensors of the same shape for each variable to generate a fused implicit representation. The quality classification layer takes the fused implicit representation as input and outputs quality labels that are close to the true values. After training, the die-casting variable importance identification model outputs predicted quality labels, while the variable selection layer outputs the importance of relevant variables.

[0046] The specific implementation process of the variable standardization embedding layer is as follows:

[0047] Given a sample of die-casting process data V = {V1, V2, ..., V...} m}∈R m×T×1 and implicit expression dimension f h Where m is the number of variables in the sample, and V i ={V i,1 V i,2 ,…,V i,T}∈R T×1 V represents the sampled value of the i-th variable. i,t ∈R 1 This represents the sampled value of the i-th variable at time t.

[0048] Variable standardization embedding layer Each variable V i The feature dimensions are mapped to a uniform dimension size. For example... Figure 2 As shown, for continuous numerical variables, a linear fully connected layer is used to expand its feature dimension to f. h For discrete enumerated variables, after passing through the embedding layer and the fully connected linear layer, the feature dimension is expanded to f. h All expanded variables are then processed according to their representation in the die-casting process data sample V = {V1, V2, ..., V...} m The inputs in the order specified in the table are concatenated to obtain a tensor of the same shape.

[0049] Variable selection layer For tensors of the same shape V h Feature extraction and feature selection are performed to generate fused implicit variables. The specific process is as follows:

[0050] For V h The sampled value of the i-th variable at time t Feature extraction is performed using a gated residual network (GRN) to obtain... Where GRN(x) = norm9x + GELU(x), norm is the normalization operation, GELU(x) = x·Φ(x), and Φ(x) is the cumulative distribution function of the standard normal distribution.

[0051] Variable selection weights It can be obtained in the following two ways:

[0052] a. V h By swapping and flattening the last two dimensions, we obtain... Then, it is generated through a gated residual network (GRN), a linear fully connected layer, and a softmax function. Right now

[0053] b. Provide a learnable matrix w vs ∈R T×m Generated by the softmax function Right now Variable selection weights It can represent the importance of variables and be used in variable importance analysis.

[0054] For time t, the implicit representation at time t is generated using a weighted summation method. Ultimately, the implicit expression of fusion is obtained.

[0055] The quality classification layer is expressed as follows: That is, the implicit expression of fusion As input, H is transformed into a quality label probability vector. Where N c Given the number of quality label categories, the label with the highest probability is selected as the output predicted quality label. Simultaneously, the variable selection layer M... vs You will get the variable selection weights. This weight will be used as the output for variable importance.

[0056] During model training, the embedding layer M containing variables is standardized. vf Variable selection layer M vs And quality classification model M c The overall importance identification model of die casting variables (represented as) Joint training is performed, with the input being: sample data of the die-casting process, V = {V1, V2, ..., V...}. m}∈R m×T×1 and the corresponding real quality label V is a real matrix containing T time steps, m variables, and each variable having a feature dimension of 1. It is a discrete enumeration value that can be transformed into a true quality label probability vector through one-hot encoding. The output is the predicted quality label probability vector.

[0057] By minimizing the difference between the predicted value P and the actual value The cross-entropy loss function, which calculates the difference between the predicted and actual values, makes the predicted value P as close as possible to the true value.

[0058] This invention also provides a device for analyzing the importance of die-casting quality correlation variables based on multivariate embedding and variable selection, comprising one or more processors for implementing the method for analyzing the importance of die-casting quality correlation variables based on multivariate standardized embedding and feature fusion in the above embodiments. Embodiments of the device for analyzing the importance of die-casting quality correlation variables based on multivariate embedding and variable selection can be applied to any device with data processing capabilities, such as a computer. The device can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of the device with data processing capabilities reading the corresponding computer program instructions from non-volatile memory into memory and running them. From a hardware perspective, a hardware structure diagram of the device for analyzing the importance of die-casting quality correlation variables based on multivariate embedding and variable selection, in addition to the processor, memory, network interface, and non-volatile memory, may also include other hardware depending on the actual function of the device, which will not be elaborated further.

[0059] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion as described in the above embodiments. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computational program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0060] The effectiveness of this invention is verified using a specific die-casting dataset. This dataset was collected via an IoT device at a frequency of 1 second. The dataset contains feature values ​​representing the production quantity of die-cast products, which are used to partition the samples. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The length of each sample is limited to a maximum of 60 characters; samples exceeding this length are truncated by 60 steps from the end. After multiple rounds of training, the variable selection weight distribution is as follows: Figure 3 As shown in the figure, the results indicate that "BoostPrs" is the variable most relevant to die casting quality in the die casting process data.

[0061] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion, characterized in that, Includes the following steps: S1: Acquire die-casting process data and quality label data within a predetermined range, and preprocess the data, including data cleaning, standardization, and matching operations; S2: Construct and train a die-casting variable importance identification model, which includes a variable standardization embedding layer, a variable selection layer, and a quality classification layer. The variable standardization embedding layer is used to transform different types of variables into tensors of the same shape based on the time-varying and discrete characteristics of the die-casting data using different embedding methods. The variable selection layer calculates variable selection weights based on features obtained by fusing the tensors of the same shape of each variable or based on a learnable weight matrix, and then performs a weighted summation of the tensors of the same shape of each variable to generate a fused implicit expression. The quality classification layer takes the fused implicit expression as input and outputs quality labels that are close to the true values. After training, the die-casting variable importance identification model outputs predicted quality labels, while the variable selection layer outputs the importance of relevant variables. The variable selection layer Represented as That is, for tensors of the same shape Feature extraction and feature selection are performed to generate fused implicit variables. Specifically: (1) To The Middle Variables in Sample value at time Feature extraction was performed using a gated residual network (GRN) to obtain... ; (2) Obtain variable selection weights through the following two methods : a. will By swapping and flattening the last two dimensions, we obtain... Then, through a gated residual network GRN, a linear fully connected layer, and Function generation ,Right now ; b. Provide a learnable matrix ,pass Function generation ,Right now Variable selection weights It represents the importance of variables and is used in variable importance analysis; (3) Generate the implicit representation at time t using a weighted summation method. Ultimately, a fused implicit expression is obtained. .

2. The method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion according to claim 1, characterized in that, The die-casting process data includes set parameters, sensor-collected parameters, and statistical data. The setting parameters are parameters related to the die casting process that are set for the die casting machine during die casting production. The sensor collects parameters including actual measured values ​​corresponding to the parameters set during the die casting process, as well as other actual measured values ​​that can be measured and monitored during the die casting process. The statistical data includes the die-casting machine's running time, date, cycle count, number of products produced, and number of defective products produced; The quality label data represents the change in the number of defective products produced.

3. The method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion according to claim 1, characterized in that, In the data preprocessing stage, the sample data is first divided according to the abrupt difference in the number of products already produced, resulting in a dataset containing... Each time step A real sample matrix of n variables, each with a feature dimension of 1. ; Then, samples with abnormal duration, abnormal changes in the number of produced samples, and abnormal changes in the number of produced non-conforming samples are removed. Subsequently, the maximum sequence length is determined based on the duration of the samples, and a masking matrix is ​​generated. For non-enumerated data, Z-score normalization is used for standardization. Finally, data matching is performed, which involves matching the timestamp of the quality label data with the end timestamp of the die-casting process data samples to the nearest value, and removing samples whose timestamp difference exceeds a preset threshold to ensure the accuracy and consistency of the data.

4. The method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion according to claim 1, characterized in that, The implementation process of the variable standardization embedding layer is as follows: Given a die casting process data Each time step A real sample matrix of n variables, each with a feature dimension of 1. and implicit expression dimensions ,in, The number of variables in the sample. Indicates the first Sampled values ​​of each variable, Indicates the first Variables in The sampled value at time; Variable standardization embedding layer Represented as That is, each variable The feature dimensions are mapped to a uniform dimension size, specifically: For continuous numerical variables, a linear fully connected layer is used to expand their feature dimensions. ; For discrete enumerated variables, after passing through the embedding layer and the fully connected linear layer, the feature dimension is expanded to... ; Then, all the expanded variables are arranged according to their data samples in the die-casting process. The inputs are concatenated in sequence to obtain tensors of the same shape. .

5. The method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion according to claim 1, characterized in that, The quality classification layer Expressed as That is, the implicit expression of fusion. As input, Transform into a quality label probability vector ,in Given the number of quality label categories, select the label with the highest probability as the output predicted quality label.

6. The method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion according to claim 5, characterized in that, During the training of the die-casting variable importance identification model, the input is a sample of die-casting process data. and the corresponding real quality label ;in, It is a collection Each time step A real matrix of n variables, each with a feature dimension of 1; It is a discrete enumeration value that can be transformed into a true quality label probability vector through one-hot encoding. The output is the predicted quality label probability vector. ; By minimizing the measured predicted value Compared with the true value The cross-entropy loss function of the difference between the predicted values ​​makes the predicted values... As close as possible to the true value .

7. A device for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion, characterized in that, It includes one or more processors for implementing the method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by the electronic device, cause the electronic device to implement the method for importance analysis of die-casting quality-related variables based on multivariate standardized embedding and feature fusion as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the method for analyzing the importance of die-casting quality-related variables based on multivariate standardized embedding and feature fusion as described in any one of claims 1 to 6.

Images