A deep learning-based automobile stamping process parameter configuration method
By using an improved ConvMixer network and an improved quartile spacing method, the adaptability and accuracy problems of traditional stamping process parameter configuration methods were solved, achieving high-precision and highly adaptable optimization of stamping process parameters, thereby improving production efficiency and quality stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YUNCHENG DAHENG AUTO PARTS MFG CO LTD
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-26
AI Technical Summary
Existing methods for configuring stamping process parameters rely on traditional empirical rules, which are difficult to adapt to complex production environments and rapidly changing process requirements, resulting in low production efficiency and unstable quality. They also lack the ability to conduct in-depth analysis and comprehensive optimization of multimodal data.
An improved ConvMixer network and an improved interquartile range method are adopted to optimize the configuration of stamping process parameters through spatiotemporal feature fusion, physical feature constraints, dynamic convolutional gating and anomaly detection, and real-time adjustment is carried out in combination with deep learning technology.
It improves the accuracy and adaptability of stamping process parameter configuration, reduces quality fluctuations and equipment failures, and enhances production efficiency and overall production quality.
Smart Images
Figure CN122286940A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent manufacturing and deep learning technology, and in particular to a method for configuring automotive stamping process parameters based on deep learning. Background Technology
[0002] With the development of intelligent manufacturing and industrial automation, stamping processes, as a key link in automobile manufacturing, have received increasing attention. Existing methods for configuring stamping process parameters largely rely on traditional empirical rules and manual adjustments, resulting in low accuracy in parameter configuration during production and difficulty in adapting to complex production environments and rapidly changing process requirements. Existing methods mainly suffer from the following problems: The parameters involved in the stamping process, such as pressure, temperature, and speed, are affected by a variety of factors, and these parameters are interrelated, making it difficult to comprehensively and effectively model and optimize them using traditional methods. The process changes during stamping are highly nonlinear and time-varying. Traditional control systems typically use static models for parameter adjustment, which are difficult to respond to sudden changes in production in real time, resulting in low production efficiency and unstable quality. Traditional parameter configuration methods generally lack the ability to deeply analyze and comprehensively optimize the multimodal data of the stamping process, failing to fully utilize the temporal data and spatial characteristics of the stamping process, resulting in poor accuracy of parameter prediction and optimization, and susceptibility to noise and external disturbances.
[0003] Therefore, how to provide a method for configuring automotive stamping process parameters based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0004] One objective of this invention is to propose a deep learning-based method for configuring automotive stamping process parameters. This invention fully utilizes an improved ConvMixer network and an improved interquartile range method, and describes in detail how to optimize the configuration of stamping process parameters through spatiotemporal feature fusion, physical feature constraints, dynamic convolutional gating, anomaly detection, and weighted correction. It has the advantages of high precision, high adaptability, and real-time performance.
[0005] A method for configuring automotive stamping process parameters based on deep learning according to an embodiment of the present invention includes the following steps: Step 1: Collect raw stamping data for automotive stamping processes; Step 2: Obtain standard stamping data by preprocessing the raw stamping data; Step 3: Perform spatiotemporal separation and feature transformation on the standard stamping data to obtain the spatial stamping feature tensor and the temporal stamping feature tensor; Step 4: Input the spatial stamping feature tensor and the temporal stamping feature tensor into the improved ConvMixer network, perform spatiotemporal feature fusion and stamping parameter prediction configuration, and output the stamping parameter prediction vector; the improved ConvMixer network includes a dual-channel convolution module, a spatial-temporal co-modeling module, a physical feature constraint module, a dynamic convolution gating module, and a process parameter mapping module; Step 5: Anomaly detection and correction are performed on the stamping parameter prediction vector using an improved interquartile range method and weighted correction to obtain the optimized stamping parameter vector; the improved interquartile range method introduces the standard deviation of the prediction error sequence to adaptively adjust the anomaly factor; Step Six: Configure the automotive stamping process production equipment based on the stamping parameter optimization vector; Step 7: During the production process, continuously collect stamping feedback data and incrementally learn the improved ConvMixer network.
[0006] Optionally, the raw stamping data includes pressure, temperature, displacement, speed, stamping equipment operating status data, material thickness, material type, production cycle, ambient temperature, ambient humidity, and process variation data generated during the stamping process.
[0007] Optionally, step two specifically includes: The raw stamping data is time-aligned according to the set time steps; The time-aligned raw stamping data is smoothed by using a moving average filtering method to obtain denoised stamping data. The z-score method is used to identify and remove outliers in the denoised stamping data, and the median filling method is used to fill in the removed outliers to obtain the de-outlier stamping data. The missing data in the abnormal stamping data is filled by linear interpolation to obtain complete stamping data; Standard stamping data is obtained by standardizing the completed stamping data using the min-max normalization method.
[0008] Optionally, step three specifically includes: Standard stamping data is spatiotemporally separated according to spatial and temporal characteristics to obtain spatial stamping data and temporal stamping data. Among them, spatial stamping data includes pressure, temperature, displacement, stamping equipment operating status data, material thickness and material type; time-series stamping data includes production cycle, ambient temperature, ambient humidity and process variation data; The spatial stamping data is divided according to spatial location, and the spatial stamping data of each spatial location is transformed by 3x3 convolution and linear mapping to obtain the spatial stamping feature tensor. The time-series stamping data is sorted according to the time step order, and the time-series stamping data of each time step is transformed by one-dimensional convolution and linear mapping to obtain the time-series stamping feature tensor.
[0009] Optionally, step four specifically includes: The dual-channel convolution module includes a spatial convolution module and a multi-scale temporal convolution module, which respectively generate convolutional stamping spatial feature tensors and multi-scale stamping temporal feature tensors, specifically including: In the spatial convolution module, the spatial stamping feature tensor is sequentially passed through 3×3 convolution and 5×5 convolution to obtain the initial stamping feature tensor; The convolutional punching feature tensor is downsampled by a 2×2 pooling kernel to obtain the downsampled punching feature tensor. The spatial features of the downsampled stamping feature tensor are extracted by 3×3 convolution to obtain the convolutional stamping spatial feature tensor; In the multi-scale temporal convolution module, the temporal stamping feature tensor is subjected to one-dimensional convolution operation through temporal convolution with kernels of 1×3, 1×5 and 1×7, respectively, to obtain local stamping feature tensor, medium-scale stamping feature tensor and long-scale stamping feature tensor, respectively. The local stamping feature tensor, the medium-scale stamping feature tensor, and the long-scale stamping feature tensor are concatenated in the channel dimension, and the channel dimension is reduced by 1×1 convolution to obtain the multi-scale stamping temporal feature tensor. In the spatial-temporal collaborative modeling module, the convolutional stamping spatial feature tensor and the multi-scale stamping temporal feature tensor are spatiotemporally fused through bidirectional cross-attention operation to obtain the spatiotemporal collaborative stamping feature tensor. In the physical characteristic constraint module, the physical characteristics of the automotive stamping process are queried from the set materials science database. The physical characteristics of the materials include elastic modulus, yield strength and temperature-stress relationship. The material physical features are transformed into material physical feature tensors through a trainable physical embedding mapping matrix; Based on the physical laws of automotive stamping process, the physical feature constraint matrix is initialized, and the physical feature constraint matrix is multiplied with the material physical feature tensor to obtain the material physical constraint tensor. The spatiotemporal co-stamping feature tensor and the material physical constraint tensor are added element by element to obtain the physical constraint stamping feature tensor. In the dynamic convolution gating module, the physical constraint stamping feature tensor is channel-compressed through 1×1 convolution to obtain the compressed stamping feature tensor. The compressed stamping feature tensor is linearly mapped through a fully connected layer and normalized using the Sigmoid function to obtain a dynamic gating signal. Based on the dynamic gating signal, the element-wise multiplication of the physical constraint stamping feature tensor is performed to obtain the gated stamping feature tensor. The gated stamping feature tensor is convolved using a 3×3 dynamic convolution to obtain the convolutional gated stamping feature tensor. In the process parameter mapping module, the convolutional gated stamping feature tensor is flattened in the spatial and channel dimensions to obtain the flattened stamping feature vector; the flattened stamping feature vector is then linearly mapped through a fully connected layer to obtain the stamping parameter prediction vector.
[0010] Optionally, the dual-channel convolution module includes a spatial convolution module and a multi-scale temporal convolution module, which respectively generate a convolutional stamping spatial feature tensor and a multi-scale stamping temporal feature tensor, specifically including: In the spatial convolution module, the spatial stamping feature tensor is sequentially passed through 3×3 convolution and 5×5 convolution to obtain the initial stamping feature tensor; The convolutional punching feature tensor is downsampled by a 2×2 pooling kernel to obtain the downsampled punching feature tensor. The spatial features of the downsampled stamping feature tensor are extracted by 3×3 convolution to obtain the convolutional stamping spatial feature tensor; In the multi-scale temporal convolution module, the temporal stamping feature tensor is subjected to one-dimensional convolution operation through temporal convolution with kernels of 1×3, 1×5 and 1×7, respectively, to obtain local stamping feature tensor, medium-scale stamping feature tensor and long-scale stamping feature tensor, respectively. The local stamping feature tensor, the mesoscale stamping feature tensor, and the long-scale stamping feature tensor are concatenated in the channel dimension, and the channel dimension is reduced by 1×1 convolution to obtain the multi-scale stamping temporal feature tensor.
[0011] Optionally, in the spatial-temporal co-modeling module, the convolutional stamping spatial feature tensor and the multi-scale stamping temporal feature tensor are spatiotemporally fused through bidirectional cross-attention operation to obtain a spatiotemporally co-modeled stamping feature tensor, specifically including: Calculate the dot product between the convolutional stamping spatial feature tensor and the multi-scale stamping temporal feature tensor, and perform softmax normalization in the temporal dimension to obtain the spatial-temporal attention tensor. Calculate the dot product between the multi-scale stamping temporal feature tensor and the convolutional stamping spatial feature tensor, and perform softmax normalization in the spatial dimension to obtain the temporal-spatial attention tensor. The spatial-temporal attention tensor is element-wise multiplied with the multi-scale stamping temporal feature tensor to obtain the weighted temporal stamping feature tensor. The temporal-spatial attention tensor is then element-wise multiplied with the convolutional stamping spatial feature tensor to obtain the weighted spatial stamping feature tensor. The weighted temporal stamping feature tensor is added to the weighted spatial stamping feature tensor to obtain the weighted spatiotemporal stamping feature tensor; The spatiotemporal co-stamping feature tensor is obtained by performing a 3×3 dynamic convolution operation on the weighted spatiotemporal stamping feature tensor.
[0012] Optionally, step five specifically includes: Define the historical time step interval, and obtain the historical stamping parameter optimization vector for each historical time step based on the historical time step interval; The absolute error between the predicted stamping parameter vector and the historical optimized stamping parameter vector at each historical time step is calculated to obtain the prediction error vector sequence; wherein, the prediction error vector sequence is composed of the prediction error sequence of each predicted value of stamping parameter; Anomaly detection is performed on the prediction error sequence of each stamping parameter prediction value using the improved interquartile range method. The predicted stamping parameter values in the stamping parameter prediction vector are determined, and the prediction errors with anomalies in the prediction error sequence are taken as historical prediction errors. The predicted stamping parameters that exhibit abnormalities are then weighted and corrected, specifically as follows: Set a regularization constant, add the standard deviation of the prediction error sequence to the regularization constant, and take the reciprocal to obtain the weighted correction coefficient; Calculate the product of the weighted correction coefficient and the historical prediction error to obtain the optimization correction term; add the predicted stamping parameter value to the corresponding optimization correction term to obtain the optimized stamping parameter value; The stamping parameter optimization vector is obtained by replacing the corresponding stamping parameter prediction value in the stamping parameter prediction vector with the optimized stamping parameter value.
[0013] Optionally, the anomaly detection of the prediction error sequence of each stamping parameter prediction value using the improved interquartile range method specifically includes: Calculate the 25th percentile and 75th percentile of the prediction error sequence for each stamping parameter, and obtain the 1st quartile and 3rd quartile. Calculate the difference between the 3rd quartile and the 1st quartile to obtain the interquartile range; Calculate the standard deviation of the prediction error sequence and use the ratio of the standard deviation to the interquartile range as the outlier factor. Multiply the anomaly factor by the interquartile range to obtain the anomaly correction term; Calculate the difference between the first quartile and the anomaly correction term to obtain the lower limit of anomalies; calculate the sum of the third quartile and the anomaly correction term to obtain the upper limit of anomalies. If the current prediction error in the prediction error sequence is less than the lower limit of abnormality or greater than the upper limit of abnormality, it is determined that the corresponding predicted value of the stamping parameter in the stamping parameter prediction vector of the prediction error sequence is abnormal, and the prediction error with abnormality in the prediction error sequence is taken as the historical prediction error.
[0014] Optionally, step seven specifically includes: The stamping feedback data includes stamping equipment operating status data, material thickness, material type, ambient temperature, ambient humidity, and process variation data; The stamping feedback data is merged with the original stamping data, and after data preprocessing, spatiotemporal separation and feature transformation, it is input into the improved ConvMixer network. The model parameters of the improved ConvMixer network are incrementally updated by gradient descent. The model parameters include: spatial convolution kernel weights and spatial feature channel weights of the spatial convolution module; temporal convolution kernel weights and temporal feature channel weights of the multi-scale temporal convolution module; spatial-temporal attention tensors and temporal-spatial attention tensors in the spatial-temporal co-modeling module; trainable physical embedding mapping matrix in the physical feature constraint module; fully connected layer weights and dynamic convolution kernel weights of the dynamic convolution gating module; and fully connected layer weights of the process parameter mapping module.
[0015] The beneficial effects of this invention are: This invention introduces an improved ConvMixer network and an improved interquartile range (INR) method, effectively enhancing the accuracy and adaptability of automotive stamping process parameter configuration. In the stamping process, this invention employs an improved ConvMixer network, combining depthwise convolution, spatiotemporal collaborative modeling, and physical feature constraints to achieve precise fusion of spatial and temporal features. The improved ConvMixer network processes spatial and temporal features separately through dual-channel convolution modules, and further optimizes the network's feature representation capability by incorporating material physical properties in the physical feature constraint module. Simultaneously, a dynamic convolutional gating module automatically adjusts the weights of the convolution kernels based on real-time data during production, enabling the network to flexibly respond to process changes and improving the accuracy and real-time response capability of stamping parameter configuration. Furthermore, this invention uses an improved INR method to identify and correct anomalies in the stamping parameter prediction vector. This method adaptively adjusts the anomaly factor by introducing the standard deviation of the prediction error sequence, solving the problem of fixed factor judgment for outliers in the traditional INR method, making anomaly detection more accurate and dynamic. By employing weighted correction, the accuracy of the stamping parameter optimization vector is ensured, reducing prediction errors caused by outliers and thus improving the stability and quality of the stamping process. Therefore, this invention significantly reduces manual intervention while improving the accuracy, adaptability, and real-time performance of stamping process parameter configuration, thereby mitigating quality fluctuations and equipment failures caused by improper parameter configuration during production and improving overall production efficiency. Attached Figure Description
[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a deep learning-based method for configuring automotive stamping process parameters proposed in this invention. Figure 2 This is a flowchart of the improved ConvMixer network structure in a deep learning-based method for configuring automotive stamping process parameters proposed in this invention. Figure 3 This is a flowchart of the process for generating stamping parameter optimization vectors in a deep learning-based automotive stamping process parameter configuration method proposed in this invention. Detailed Implementation
[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0018] refer to Figures 1-3 A method for configuring automotive stamping process parameters based on deep learning, comprising the following steps: Step 1: Collect raw stamping data for automotive stamping processes; Step 2: Obtain standard stamping data by preprocessing the raw stamping data; Step 3: Perform spatiotemporal separation and feature transformation on the standard stamping data to obtain the spatial stamping feature tensor and the temporal stamping feature tensor; Step 4: Input the spatial stamping feature tensor and the temporal stamping feature tensor into the improved ConvMixer network, perform spatiotemporal feature fusion and stamping parameter prediction configuration, and output the stamping parameter prediction vector; wherein, the improved ConvMixer network includes a dual-channel convolution module, a spatial-temporal co-modeling module, a physical feature constraint module, a dynamic convolutional gating module and a process parameter mapping module; Step 5: Anomaly detection and correction are performed on the stamping parameter prediction vector using the improved interquartile range method and weighted correction to obtain the optimized stamping parameter vector; wherein, the improved interquartile range method introduces the standard deviation of the prediction error sequence to adaptively adjust the anomaly factor; Step Six: Configure the automotive stamping process production equipment based on the stamping parameter optimization vector; Step 7: During the production process, continuously collect stamping feedback data and incrementally learn the improved ConvMixer network.
[0019] In this embodiment, the original stamping data includes the pressure, temperature, displacement, speed, stamping equipment operating status data, material thickness, material type, production cycle, ambient temperature, ambient humidity, and process variation data generated during the stamping process. The process variation data includes mold deformation data, workpiece deformation data, equipment status change data, material property change data, operator or process adjustment data, and ambient temperature and humidity change data.
[0020] In this embodiment, step two specifically includes: The raw stamping data is time-aligned according to the set time steps; The time-aligned raw stamping data is smoothed by using a moving average filtering method to obtain denoised stamping data. The z-score method is used to identify and remove outliers in the denoised stamping data, and the median filling method is used to fill in the removed outliers to obtain the de-outlier stamping data. In this invention, the z-score method is existing technology. For example, the denoised pressure data is [10, 12, 11, 15, 500, 13, 14, 11, 13, 10]. The fifth data point, 500, clearly deviates from the other data points, possibly due to sensor malfunction or other external factors causing an outlier. Setting the conventional threshold for the z-score method to 3, the mean of the above pressure data is calculated to be 58.9, with a standard deviation of approximately 67.8. The z-score for 500 is (500−58.9) / 67.8=6.5, which is much larger than the conventional threshold of 3, and therefore is considered an outlier. Sort the denoised stress data from smallest to largest to obtain [10, 10, 11, 11, 12, 13, 13, 14, 15, 500]. The median is (12+13) / 2=12.5. Replace 500 with 12.5 to obtain the anomaly-free stress data [10, 12, 11, 15, 12.5, 13, 14, 11, 13, 10].
[0021] The missing data in the abnormal stamping data is filled by linear interpolation to obtain complete stamping data; In this invention, the linear interpolation method is a prior art. Assume the anomaly-free stamping pressure data is: [10, 12, 11, 15, 12.5, 13, NaN, 14, 13, 10]. The 7th data point (NaN) is a missing value, and its preceding and following values are 13 (6th data point) and 14 (8th data point). Therefore, the linear interpolation result is: 13 + [(14−13) / (8−6)] × (7−6) = 13.5. Substituting 13.5 into the missing value position, the completed pressure data is [10, 12, 11, 15, 12.5, 13, 13.5, 14, 13, 10].
[0022] Standard stamping data is obtained by standardizing the completed stamping data using the min-max normalization method.
[0023] In this embodiment, step three specifically includes: Standard stamping data is spatiotemporally separated according to spatial and temporal characteristics to obtain spatial stamping data and temporal stamping data. Among them, spatial stamping data includes pressure, temperature, displacement, stamping equipment operating status data, material thickness and material type; time-series stamping data includes production cycle, ambient temperature, ambient humidity and process variation data; The spatial stamping data is divided according to spatial location, and the spatial stamping data of each spatial location is transformed by 3x3 convolution and linear mapping to obtain the spatial stamping feature tensor. The time-series stamping data is sorted according to the time step order, and the time-series stamping data of each time step is transformed by one-dimensional convolution and linear mapping to obtain the time-series stamping feature tensor.
[0024] In this embodiment, step four specifically includes: The dual-channel convolution module includes a spatial convolution module and a multi-scale temporal convolution module, which respectively generate convolutional stamping spatial feature tensors and multi-scale stamping temporal feature tensors; In the spatial-temporal collaborative modeling module, the convolutional stamping spatial feature tensor and the multi-scale stamping temporal feature tensor are spatiotemporally fused through bidirectional cross-attention operation to obtain the spatiotemporal collaborative stamping feature tensor. In the physical characteristic constraint module, the physical characteristics of the automotive stamping process are queried from the set materials science database. The physical characteristics of the materials include elastic modulus, yield strength and temperature-stress relationship. The material physical features are transformed into material physical feature tensors through a trainable physical embedding mapping matrix; Based on the physical laws of automotive stamping process, the physical feature constraint matrix is initialized, and the physical feature constraint matrix is multiplied with the material physical feature tensor to obtain the material physical constraint tensor. The spatiotemporal co-stamping feature tensor and the material physical constraint tensor are added element by element to obtain the physical constraint stamping feature tensor. In this invention, the physical characteristics of materials are introduced, which can effectively combine the physical laws of the stamping process. This allows the improved ConvMixer network to better understand and utilize the physical properties of materials to predict and optimize the stamping process, thereby improving the accuracy and reliability of stamping parameter configuration and achieving higher process precision and product quality in actual production.
[0025] In the dynamic convolution gating module, the physical constraint stamping feature tensor is channel-compressed through 1×1 convolution to obtain the compressed stamping feature tensor. The compressed stamping feature tensor is linearly mapped through a fully connected layer and normalized using the Sigmoid function to obtain a dynamic gating signal. Based on the dynamic gating signal, the element-wise multiplication of the physical constraint stamping feature tensor is performed to obtain the gated stamping feature tensor. The gated stamping feature tensor is convolved using a 3×3 dynamic convolution to obtain the convolutional gated stamping feature tensor. In the process parameter mapping module, the convolutional gated stamping feature tensor is flattened in the spatial dimension and the channel dimension to obtain the flattened stamping feature vector; the flattened stamping feature vector is linearly mapped through a fully connected layer to obtain the stamping parameter prediction vector. The stamping parameter prediction vector consists of stamping pressure, die temperature, stamping speed, material thickness, material type, die clearance, stamping process time, annealing temperature and time, forming force distribution, die wear condition, and stamping accuracy.
[0026] In this embodiment, the dual-channel convolution module includes a spatial convolution module and a multi-scale temporal convolution module, which respectively generate a convolutional stamping spatial feature tensor and a multi-scale stamping temporal feature tensor, specifically including: In the spatial convolution module, the spatial stamping feature tensor is sequentially passed through 3×3 convolution and 5×5 convolution to obtain the initial stamping feature tensor; The convolutional punching feature tensor is downsampled by a 2×2 pooling kernel to obtain the downsampled punching feature tensor. The spatial features of the downsampled stamping feature tensor are extracted by 3×3 convolution to obtain the convolutional stamping spatial feature tensor; In the multi-scale temporal convolution module, the temporal stamping feature tensor is subjected to one-dimensional convolution operation through temporal convolution with kernels of 1×3, 1×5 and 1×7, respectively, to obtain local stamping feature tensor, medium-scale stamping feature tensor and long-scale stamping feature tensor, respectively. The local stamping feature tensor, the mesoscale stamping feature tensor, and the long-scale stamping feature tensor are concatenated in the channel dimension, and the channel dimension is reduced by 1×1 convolution to obtain the multi-scale stamping temporal feature tensor.
[0027] In this invention, a dual-channel convolutional module processes spatial and temporal stamping features separately, enabling accurate extraction and fusion of data features from different dimensions. The spatial convolutional module effectively extracts spatial features during the stamping process through multi-layer convolution and pooling operations, enhancing the expressive power of spatial information. The multi-scale temporal convolutional module processes temporal data using convolutional kernels of different sizes, capturing dynamic changes at different time scales and ensuring the acquisition of effective feature information across various time scales. By concatenating and performing dimensionality reduction operations along the channel dimension, the multi-dimensional information of spatial and temporal features is fused to generate a stamping feature tensor. Through multi-dimensional and comprehensive feature extraction and fusion, the improved ConvMixer network can better adapt to complex stamping process parameter configurations, thereby improving prediction accuracy and real-time performance.
[0028] In this embodiment, in the spatial-temporal co-modeling module, the convolutional stamping spatial feature tensor and the multi-scale stamping temporal feature tensor are spatiotemporally fused through bidirectional cross-attention operation to obtain a spatiotemporally co-modeled stamping feature tensor, specifically including: Calculate the dot product between the convolutional stamping spatial feature tensor and the multi-scale stamping temporal feature tensor, and perform softmax normalization in the temporal dimension to obtain the spatial-temporal attention tensor. Calculate the dot product between the multi-scale stamping temporal feature tensor and the convolutional stamping spatial feature tensor, and perform softmax normalization in the spatial dimension to obtain the temporal-spatial attention tensor. The spatial-temporal attention tensor is element-wise multiplied with the multi-scale stamping temporal feature tensor to obtain the weighted temporal stamping feature tensor. The temporal-spatial attention tensor is then element-wise multiplied with the convolutional stamping spatial feature tensor to obtain the weighted spatial stamping feature tensor. The weighted temporal stamping feature tensor is added to the weighted spatial stamping feature tensor to obtain the weighted spatiotemporal stamping feature tensor; The spatiotemporal co-stamping feature tensor is obtained by performing a 3×3 dynamic convolution operation on the weighted spatiotemporal stamping feature tensor.
[0029] In this invention, the spatial-temporal collaborative modeling module performs spatiotemporal fusion of spatial and temporal stamping feature tensors through bidirectional cross-attention operations, significantly improving the accuracy and depth of feature fusion. By performing softmax normalization on both the temporal and spatial dimensions, the spatial-temporal collaborative modeling module effectively captures and weights the interrelationships between different dimensions, ensuring accurate interaction between multi-scale temporal data and spatial features. By fusing the weighted temporal and spatial features through element-wise multiplication, the improved ConvMixer network can adaptively focus on more critical features in time and space, further enhancing its ability to predict complex process parameters. The spatiotemporal collaborative stamping feature tensor, after undergoing 3×3 dynamic convolution, captures richer spatial-temporal information, making the configuration of stamping process parameters more accurate and efficient.
[0030] In this invention, the improved ConvMixer network inherits the convolutional structure of the standard ConvMixer network. The convolutional layers in the standard ConvMixer network extract spatial information from the input data through depthwise convolution, perform feature combination through pointwise convolution, and then perform channel fusion through a multilayer perceptron to extract higher-order features, thereby enhancing the model's expressive power.
[0031] Compared to the standard ConvMixer network, the improved ConvMixer network incorporates the following structural improvements: It introduces a dual-channel convolutional structure to handle spatial and temporal features separately. The spatial convolutional module processes spatial information related to location and geometric parameters, while the multi-scale temporal convolutional module focuses on capturing dynamically changing temporal information during production, such as process variations and production cycles, through three sets of convolutional kernel sizes. In the spatial-temporal collaborative modeling module, bidirectional cross-attention operations are used to further optimize the fusion effect of spatiotemporal features through the interaction between spatial and temporal features.
[0032] The improved ConvMixer network significantly enhances its ability to capture multi-scale information of temporal features by introducing a dual-channel convolutional structure and a bidirectional cross-attention mechanism. The dual-channel convolutional structure better handles features at different time scales in time-series data, improving the model's response to long-term and short-term changes in the stamping process. The bidirectional cross-attention mechanism optimizes the fusion of spatiotemporal features, making the interaction between spatial and temporal features more accurate. Through these improvements, the improved ConvMixer network can more accurately predict parameter configurations, improving the stability and reliability of predictions and enhancing the model's adaptability to different production environments.
[0033] In this embodiment, step five specifically includes: Define the historical time step interval, and obtain the historical stamping parameter optimization vector for each historical time step based on the historical time step interval; The absolute error between the predicted stamping parameter vector and the historical optimized stamping parameter vector at each historical time step is calculated to obtain the prediction error vector sequence; wherein, the prediction error vector sequence is composed of the prediction error sequence of each predicted value of stamping parameter; Anomaly detection is performed on the prediction error sequence of each stamping parameter prediction value using the improved interquartile range method. The predicted stamping parameter values in the stamping parameter prediction vector are determined, and the prediction errors with anomalies in the prediction error sequence are taken as historical prediction errors. The predicted stamping parameters that exhibit abnormalities are then weighted and corrected, specifically as follows: Set a regularization constant, add the standard deviation of the prediction error sequence to the regularization constant, and take the reciprocal to obtain the weighted correction coefficient; Calculate the product of the weighted correction coefficient and the historical prediction error to obtain the optimization correction term; add the predicted stamping parameter value to the corresponding optimization correction term to obtain the optimized stamping parameter value; The stamping parameter optimization vector is obtained by replacing the corresponding stamping parameter prediction value in the stamping parameter prediction vector with the optimized stamping parameter value.
[0034] For example, in a specific application, suppose the predicted stamping pressure in the automotive stamping process is 1200 Pa. A historical time step interval is defined, and the absolute error between the current predicted stamping parameter value and the historical optimized stamping parameter vector for each historical time step is calculated, resulting in a prediction error sequence. The standard deviation of the prediction error sequence is calculated to be 10 Pa. For example, if the historical optimized stamping parameter value for a certain historical time step is 1150 Pa, then the prediction error is 1200 - 1150 = 50 Pa. Anomaly detection is performed using the improved interquartile range method. If the current prediction error is found to be abnormal, then the corresponding predicted stamping parameter value of 1200 Pa is abnormal, and 50 Pa is recorded as the historical prediction error. In the weighted correction process, a regularization constant of 0.5 is set. The standard deviation of the prediction error sequence, 10 Pa, is added to the regularization constant 0.5 to obtain 10.5 Pa, and the reciprocal is taken to obtain the weighted correction coefficient of 0.0952. By multiplying the weighted correction coefficient by the historical prediction error of 50 Pa, the optimization correction term is obtained as 4.76 Pa. Finally, the optimized correction term is added to the predicted stamping parameter value to obtain the optimized stamping parameter value 1200Pa+4.76Pa=1204.76Pa, which replaces the original 1200Pa, thus obtaining the optimized stamping parameter vector.
[0035] In this embodiment, the prediction error sequence of each stamping parameter prediction value is used for anomaly detection through an improved interquartile range method, specifically including: Calculate the 25th percentile and 75th percentile of the prediction error sequence for each stamping parameter, and obtain the 1st quartile and 3rd quartile. Calculate the difference between the 3rd quartile and the 1st quartile to obtain the interquartile range; Calculate the standard deviation of the prediction error sequence and use the ratio of the standard deviation to the interquartile range as the outlier factor. Multiply the anomaly factor by the interquartile range to obtain the anomaly correction term; Calculate the difference between the first quartile and the anomaly correction term to obtain the lower limit of anomalies; calculate the sum of the third quartile and the anomaly correction term to obtain the upper limit of anomalies. If the current prediction error in the prediction error sequence is less than the lower limit of abnormality or greater than the upper limit of abnormality, it is determined that the corresponding predicted value of the stamping parameter in the stamping parameter prediction vector of the prediction error sequence is abnormal, and the prediction error with abnormality in the prediction error sequence is taken as the historical prediction error.
[0036] In this invention, the improved interquartile range (IQR) method introduces the ratio of standard deviation to IQR as an anomaly factor, and dynamically adjusts the sensitivity of anomaly detection by multiplying the anomaly factor by the IQR. Compared to the traditional IQR method, it can more adaptively handle the volatility and changes in data, making anomaly detection more accurate. Traditional methods use a fixed factor, such as 1.5 times the IQR, to determine outliers, while the improved method incorporates the standard deviation, considering the dispersion of the data, and can flexibly adjust the anomaly detection standard under different conditions. The purpose of multiplying the anomaly factor by the IQR is to adjust the anomaly threshold according to the volatility of the data, thereby more accurately identifying data points that deviate significantly from the normal fluctuation range. By calculating the ratio of standard deviation to IQR as a variable factor, the IQR factor is dynamically adjusted to obtain an anomaly correction term that better matches the data distribution characteristics. This approach can effectively improve the accuracy and adaptability of anomaly detection, avoiding the one-size-fits-all approach of traditional methods.
[0037] In this embodiment, step seven specifically includes: The stamping feedback data includes stamping equipment operating status data, material thickness, material type, ambient temperature, ambient humidity, and process variation data; The stamping feedback data is merged with the original stamping data, and after data preprocessing, spatiotemporal separation and feature transformation, it is input into the improved ConvMixer network. The model parameters of the improved ConvMixer network are incrementally updated by gradient descent. The model parameters include: spatial convolution kernel weights and spatial feature channel weights of the spatial convolution module; temporal convolution kernel weights and temporal feature channel weights of the multi-scale temporal convolution module; spatial-temporal attention tensors and temporal-spatial attention tensors in the spatial-temporal co-modeling module; trainable physical embedding mapping matrices in the physical feature constraint module; fully connected layer weights and dynamic convolution kernel weights of the dynamic convolution gating module; and fully connected layer weights of the process parameter mapping module.
[0038] Example 1: To verify the feasibility of this invention in practice, the method of this invention was applied to the stamping process parameter configuration system of an automobile manufacturing plant. This automobile manufacturing plant stamps automobile body parts through an automated production line. Traditional stamping parameter configuration methods rely on manual experience and simple rules, which leads to insufficient accuracy and real-time response capability in complex production environments. Furthermore, it is difficult to adjust parameters in real time, resulting in suboptimal production efficiency and product quality.
[0039] During implementation, the production line first collects various operational data from the stamping equipment. Through data preprocessing, spatiotemporal separation, and feature transformation, spatial stamping feature tensors and temporal stamping feature tensors are generated and input into an improved ConvMixer network. Through spatiotemporal feature fusion and stamping parameter prediction configuration, a stamping parameter prediction vector is output. Based on this vector, an improved interquartile range method and weighted correction are used to detect and correct anomalies in the stamping parameters, obtaining an optimized stamping parameter vector, which is then used to configure the stamping production equipment. During production, continuously collected stamping feedback data is used to continuously update the improved ConvMixer network through incremental learning, enabling the stamping process parameter configuration system to adapt to changes in production in real time, thus improving the stability and accuracy of the stamping process.
[0040] To verify the effectiveness of the present invention, the method of the present invention was experimentally compared with two sets of comparative schemes. Comparative scheme 1 is a stamping parameter configuration method based on the standard ConvMixer network, and comparative scheme 2 is a traditional manual experience configuration method. The comparison results are shown in Table 1.
[0041] Table 1 Performance Comparison of Different AI-Generated Reports
[0042] As shown in Table 1, the method of this invention significantly outperforms the comparative schemes in several comparative indicators. Specifically, the method of this invention achieves a stamping product qualification rate of 98.5%, significantly higher than 95.2% for comparative scheme 1 and 91.7% for comparative scheme 2, indicating that the method of this invention can better control the stamping process and ensure higher product quality. In terms of production efficiency, the method of this invention achieves 150 pieces / hour, compared to 140 pieces / hour for comparative scheme 1 and 120 pieces / hour for comparative scheme 2, representing an increase in production efficiency of approximately 7% and 25%, respectively, fully demonstrating the superiority of the method of this invention in improving production efficiency. The adjustment response time of the method of this invention is only 5 seconds, significantly better than the comparative schemes, enabling the stamping process parameter configuration system to quickly adapt to changes and adjustments in production, thereby reducing production downtime and improving production flexibility.
[0043] Furthermore, the equipment failure rate of the method of the present invention is 2%, compared to 4% for comparative scheme 1 and 7% for comparative scheme 2. The method of the present invention significantly reduces the frequency of equipment failures, contributing to improved stability and reliability of the production line. Regarding the accuracy of stamping parameter configuration, the method of the present invention achieves an accuracy of 98.2%, significantly higher than the comparative schemes, indicating that the method of the present invention can provide more precise parameter configurations, thereby improving the precision and consistency of the stamping process. The method of the present invention achieves cost savings of 18%, compared to 12% for comparative scheme 1 and 5% for comparative scheme 2. The method of the present invention can significantly reduce production costs and improve overall economic efficiency.
[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for configuring automotive stamping process parameters based on deep learning, characterized in that, Includes the following steps: Step 1: Collect raw stamping data for automotive stamping processes; Step 2: Obtain standard stamping data by preprocessing the raw stamping data; Step 3: Perform spatiotemporal separation and feature transformation on the standard stamping data to obtain the spatial stamping feature tensor and the temporal stamping feature tensor; Step 4: Input the spatial stamping feature tensor and the temporal stamping feature tensor into the improved ConvMixer network, perform spatiotemporal feature fusion and stamping parameter prediction configuration, and output the stamping parameter prediction vector; the improved ConvMixer network includes a dual-channel convolution module, a spatial-temporal co-modeling module, a physical feature constraint module, a dynamic convolution gating module, and a process parameter mapping module; Step 5: Anomaly detection and correction are performed on the stamping parameter prediction vector using an improved interquartile range method and weighted correction to obtain the optimized stamping parameter vector; the improved interquartile range method introduces the standard deviation of the prediction error sequence to adaptively adjust the anomaly factor; Step Six: Configure the automotive stamping process production equipment based on the stamping parameter optimization vector; Step 7: During the production process, continuously collect stamping feedback data and incrementally learn the improved ConvMixer network.
2. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, The raw stamping data includes the pressure, temperature, displacement, speed, stamping equipment operating status data, material thickness, material type, production cycle, ambient temperature, ambient humidity, and process variation data generated during the stamping process.
3. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, Step two specifically includes: The raw stamping data is time-aligned according to the set time steps; The time-aligned raw stamping data is smoothed by using a moving average filtering method to obtain denoised stamping data. The z-score method is used to identify and remove outliers in the denoised stamping data, and the median filling method is used to fill in the removed outliers to obtain the de-outlier stamping data. The missing data in the abnormal stamping data is filled by linear interpolation to obtain complete stamping data; Standard stamping data is obtained by standardizing the completed stamping data using the min-max normalization method.
4. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, Step three specifically includes: Standard stamping data is spatiotemporally separated according to spatial and temporal characteristics to obtain spatial stamping data and temporal stamping data. Among them, spatial stamping data includes pressure, temperature, displacement, stamping equipment operating status data, material thickness and material type; time-series stamping data includes production cycle, ambient temperature, ambient humidity and process variation data; The spatial stamping data is divided according to spatial location, and the spatial stamping data of each spatial location is transformed by 3x3 convolution and linear mapping to obtain the spatial stamping feature tensor. The time-series stamping data is sorted according to the time step order, and the time-series stamping data of each time step is transformed by one-dimensional convolution and linear mapping to obtain the time-series stamping feature tensor.
5. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, Step four specifically includes: The dual-channel convolution module includes a spatial convolution module and a multi-scale temporal convolution module, which respectively generate a convolutional stamping spatial feature tensor and a multi-scale stamping temporal feature tensor. In the spatial-temporal collaborative modeling module, the convolutional stamping spatial feature tensor and the multi-scale stamping temporal feature tensor are spatiotemporally fused through bidirectional cross-attention operation to obtain the spatiotemporal collaborative stamping feature tensor. In the physical characteristic constraint module, the physical characteristics of the automotive stamping process are queried from the set materials science database. The physical characteristics of the materials include elastic modulus, yield strength and temperature-stress relationship. The material physical features are transformed into material physical feature tensors through a trainable physical embedding mapping matrix; Based on the physical laws of automotive stamping process, the physical feature constraint matrix is initialized, and the physical feature constraint matrix is multiplied with the material physical feature tensor to obtain the material physical constraint tensor. The spatiotemporal co-stamping feature tensor and the material physical constraint tensor are added element by element to obtain the physical constraint stamping feature tensor. In the dynamic convolution gating module, the physical constraint stamping feature tensor is channel-compressed through 1×1 convolution to obtain the compressed stamping feature tensor. The compressed stamping feature tensor is linearly mapped through a fully connected layer and normalized using the Sigmoid function to obtain a dynamic gating signal. Based on the dynamic gating signal, the element-wise multiplication of the physical constraint stamping feature tensor is performed to obtain the gated stamping feature tensor. The gated stamping feature tensor is convolved using a 3×3 dynamic convolution to obtain the convolutional gated stamping feature tensor. In the process parameter mapping module, the convolutional gated stamping feature tensor is flattened in the spatial and channel dimensions to obtain the flattened stamping feature vector; the flattened stamping feature vector is then linearly mapped through a fully connected layer to obtain the stamping parameter prediction vector.
6. The method for configuring automotive stamping process parameters based on deep learning according to claim 5, characterized in that, The dual-channel convolution module includes a spatial convolution module and a multi-scale temporal convolution module, which respectively generate convolutional stamping spatial feature tensors and multi-scale stamping temporal feature tensors, specifically including: In the spatial convolution module, the spatial stamping feature tensor is sequentially passed through 3×3 convolution and 5×5 convolution to obtain the initial stamping feature tensor; The convolutional punching feature tensor is downsampled by a 2×2 pooling kernel to obtain the downsampled punching feature tensor. The spatial features of the downsampled stamping feature tensor are extracted by 3×3 convolution to obtain the convolutional stamping spatial feature tensor; In the multi-scale temporal convolution module, the temporal stamping feature tensor is subjected to one-dimensional convolution operation through temporal convolution with kernels of 1×3, 1×5 and 1×7, respectively, to obtain local stamping feature tensor, medium-scale stamping feature tensor and long-scale stamping feature tensor, respectively. The local stamping feature tensor, the mesoscale stamping feature tensor, and the long-scale stamping feature tensor are concatenated in the channel dimension, and the channel dimension is reduced by 1×1 convolution to obtain the multi-scale stamping temporal feature tensor.
7. The method for configuring automotive stamping process parameters based on deep learning according to claim 5, characterized in that, The spatial-temporal collaborative modeling module fuses the convolutional stamping spatial feature tensor and the multi-scale stamping temporal feature tensor spatiotemporally through bidirectional cross-attention operations to obtain a spatiotemporal collaborative stamping feature tensor, specifically including: Calculate the dot product between the convolutional stamping spatial feature tensor and the multi-scale stamping temporal feature tensor, and perform softmax normalization in the temporal dimension to obtain the spatial-temporal attention tensor. Calculate the dot product between the multi-scale stamping temporal feature tensor and the convolutional stamping spatial feature tensor, and perform softmax normalization in the spatial dimension to obtain the temporal-spatial attention tensor. The spatial-temporal attention tensor is element-wise multiplied with the multi-scale stamping temporal feature tensor to obtain the weighted temporal stamping feature tensor. The temporal-spatial attention tensor is then element-wise multiplied with the convolutional stamping spatial feature tensor to obtain the weighted spatial stamping feature tensor. The weighted temporal stamping feature tensor is added to the weighted spatial stamping feature tensor to obtain the weighted spatiotemporal stamping feature tensor; The spatiotemporal co-stamping feature tensor is obtained by performing a 3×3 dynamic convolution operation on the weighted spatiotemporal stamping feature tensor.
8. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, Step five specifically includes: Define the historical time step interval, and obtain the historical stamping parameter optimization vector for each historical time step based on the historical time step interval; The absolute error between the predicted stamping parameter vector and the historical optimized stamping parameter vector at each historical time step is calculated to obtain the prediction error vector sequence; wherein, the prediction error vector sequence is composed of the prediction error sequence of each predicted value of stamping parameter; Anomaly detection is performed on the prediction error sequence of each stamping parameter prediction value using the improved interquartile range method. The predicted stamping parameter values in the stamping parameter prediction vector are determined, and the prediction errors with anomalies in the prediction error sequence are taken as historical prediction errors. The predicted stamping parameters that exhibit abnormalities are then weighted and corrected, specifically as follows: Set a regularization constant, add the standard deviation of the prediction error sequence to the regularization constant, and take the reciprocal to obtain the weighted correction coefficient; Calculate the product of the weighted correction coefficient and the historical prediction error to obtain the optimization correction term; add the predicted stamping parameter value to the corresponding optimization correction term to obtain the optimized stamping parameter value; The stamping parameter optimization vector is obtained by replacing the corresponding stamping parameter prediction value in the stamping parameter prediction vector with the optimized stamping parameter value.
9. The method for configuring automotive stamping process parameters based on deep learning according to claim 8, characterized in that, The anomaly detection method, which uses an improved interquartile range (IQR) to perform anomaly detection on the prediction error sequence of each stamping parameter prediction value, specifically includes: Calculate the 25th percentile and 75th percentile of the prediction error sequence for each stamping parameter, and obtain the 1st quartile and 3rd quartile. Calculate the difference between the 3rd quartile and the 1st quartile to obtain the interquartile range; Calculate the standard deviation of the prediction error sequence and use the ratio of the standard deviation to the interquartile range as the outlier factor. Multiply the anomaly factor by the interquartile range to obtain the anomaly correction term; Calculate the difference between the first quartile and the anomaly correction term to obtain the lower limit of anomalies; calculate the sum of the third quartile and the anomaly correction term to obtain the upper limit of anomalies. If the current prediction error in the prediction error sequence is less than the lower limit of abnormality or greater than the upper limit of abnormality, it is determined that the corresponding predicted value of the stamping parameter in the stamping parameter prediction vector of the prediction error sequence is abnormal, and the prediction error with abnormality in the prediction error sequence is taken as the historical prediction error.
10. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, Step seven specifically includes: The stamping feedback data includes stamping equipment operating status data, material thickness, material type, ambient temperature, ambient humidity, and process variation data; The stamping feedback data is merged with the original stamping data, and after data preprocessing, spatiotemporal separation and feature transformation, it is input into the improved ConvMixer network. The model parameters of the improved ConvMixer network are incrementally updated by gradient descent. The model parameters include: spatial convolution kernel weights and spatial feature channel weights of the spatial convolution module; temporal convolution kernel weights and temporal feature channel weights of the multi-scale temporal convolution module; spatial-temporal attention tensors and temporal-spatial attention tensors in the spatial-temporal co-modeling module; trainable physical embedding mapping matrix in the physical feature constraint module; fully connected layer weights and dynamic convolution kernel weights of the dynamic convolution gating module; and fully connected layer weights of the process parameter mapping module.