Deformation prediction method and device based on prompt coding and adaptive feature fusion

Abstract

The invention discloses a deformation prediction method and device based on prompt coding and adaptive feature fusion, and the method comprises the following steps: obtaining technical parameters of a part, and carrying out the conversion processing of the technical parameters, and obtaining a deformation probability data set of the part; performing parameter coding on the deformation probability data set, and extracting feature data of technical parameters; according to the type of the technical parameter, adaptively fusing the feature data to obtain fused feature data of the part; inputting the fused feature data into a preset prediction module to obtain a prediction result, the prediction result comprising a limit value of the deformation data and an error mean variance; and comparing the deformation data of the current part with the prediction result, and adjusting the heat treatment process data of the current part, so that the prediction precision and efficiency of the hot working deformation of the part are improved, and the deformation prediction error of the part is reduced.

Description

Technical Field

[0001] The present invention relates to the field of manufacturing technology, and in particular to a deformation prediction method and device based on hint coding and adaptive feature fusion. Background Art

[0002] During the heat treatment of die-cast parts, the deformation of the workpiece after heat treatment is highly uncertain due to the complex geometry, material properties, and variable heat treatment process parameters. This deformation not only affects the processing accuracy and performance of the parts, but also poses challenges to product production efficiency and cost control. Traditional deformation prediction methods, such as finite element analysis, although mature in theory, often face limitations in practical applications such as high computational costs, strong parameter dependence, and difficulty in efficiently processing multimodal data. Therefore, the development of an efficient, intelligent, and accurate deformation prediction method and the use of the prediction method to optimize the closed-loop verification of process parameters during the processing process have become urgent needs in the industry. Summary of the Invention

[0003] To solve the above problems, the purpose of the present invention is to provide a deformation prediction method, device and storage medium based on prompt coding and adaptive feature fusion, so as to improve the prediction accuracy and efficiency of the thermal processing deformation of parts and reduce the deformation prediction error of parts.

[0004] The technical solution adopted by the present invention to solve the problem is:

[0005] In a first aspect, an embodiment of the present application provides a deformation prediction method based on prompt coding and adaptive feature fusion, the method comprising: acquiring technical parameters of a part, and converting the technical parameters to obtain a deformation probability data set of the part, wherein the technical parameters include: shape data, material data, heat treatment process data, and deformation data of the part; parameter encoding the deformation probability data set to extract feature data of the technical parameters; adaptively fusing the feature data according to the type of the technical parameters to obtain fused feature data of the part; inputting the fused feature data into a preset prediction module to obtain a prediction result, wherein the prediction result includes the limit value and error mean variance of the deformation data; comparing the deformation data of the current part with the prediction result, and adjusting the heat treatment process data of the current part.

[0006] In a second aspect, an embodiment of the present application provides a deformation prediction device based on prompt coding and adaptive feature fusion, comprising: an acquisition module for acquiring technical parameters of a part and converting the technical parameters to obtain a deformation probability dataset of the part, wherein the technical parameters include: shape data, material data, heat treatment process data, and deformation data of the part;

[0007] An encoding module is used to perform parameter encoding on the deformation probability data set and extract the characteristic data of the technical parameters; a fusion module is used to adaptively fuse the characteristic data according to the type of the technical parameters to obtain the fused characteristic data of the part; a prediction module is used to input the fused characteristic data into a preset prediction module to obtain a prediction result, wherein the prediction result includes the limit value and error mean variance of the deformation data; an adjustment module is used to compare the deformation data of the current part with the prediction result and adjust the heat treatment process data of the current part.

[0008] In a third aspect, an embodiment of the present application provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, it implements the deformation prediction method based on prompt coding and adaptive feature fusion as described above.

[0009] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the deformation prediction method based on hint coding and adaptive feature fusion as described above is implemented.

[0010] In an embodiment of the present application, technical parameters of a part are acquired and converted to obtain a deformation probability data set of the part, wherein the technical parameters include: shape data, material data, heat treatment process data, and deformation data of the part; parameter encoding is performed on the deformation probability data set to extract characteristic data of the technical parameters; characteristic data are adaptively fused according to the type of technical parameters to obtain fused characteristic data of the part; the fused characteristic data are input into a preset prediction module to obtain a prediction result, wherein the prediction result includes a limit value and an error mean variance of the deformation data; the deformation data of the current part is compared with the prediction result, and the heat treatment process data of the current part is adjusted to improve the prediction accuracy and efficiency of the thermal processing deformation of the part and reduce the deformation prediction error of the part.

[0011] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become obvious from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 Flowchart of a deformation prediction method based on hint coding and adaptive feature fusion according to an embodiment of the present invention;

[0013] Figure 2 for Figure 1 Flowchart of step S1000;

[0014] Figure 3 for Figure 1Flowchart of another embodiment of step S1000;

[0015] Figure 4 for Figure 1 Flowchart of step S2000;

[0016] Figure 5 for Figure 4 Flowchart of step S2300;

[0017] Figure 6 for Figure 1 Flowchart of step S3000;

[0018] Figure 7 for Figure 1 Flowchart of step S4000;

[0019] Figure 8 This is a structural diagram of a deformation prediction device based on hint coding and adaptive feature fusion according to an embodiment of the present invention;

[0020] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0022] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limitations on the present invention.

[0023] In the description of the present invention, "several" means one or more, "many" means more than two, "greater than," "less than," and "exceed" are understood to exclude the number itself, while "above," "below," and "within" are understood to include the number itself. The use of "first" and "second" in the description is solely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, implicitly specifying the number of the indicated technical features, or implicitly specifying the order of the indicated technical features.

[0024] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0025] The embodiments of the present invention involve a deformation prediction method, device and storage medium based on prompt coding and adaptive feature fusion, which obtains the technical parameters of the part and converts the technical parameters to obtain the deformation probability data set of the part, wherein the technical parameters include: shape data, material data, heat treatment process data, and deformation data of the part; parameter encoding is performed on the deformation probability data set to extract the feature data of the technical parameters; according to the type of technical parameters, the feature data is adaptively fused to obtain the fused feature data of the part; the fused feature data is input into a preset prediction module to obtain a prediction result, wherein the prediction result includes the limit value and error mean variance of the deformation data; the deformation data of the current part is compared with the prediction result, and the heat treatment process data of the current part is adjusted to improve the prediction accuracy and efficiency of the thermal processing deformation of the part and reduce the deformation prediction error of the part.

[0026] In practical applications, heat treatment is a key process in the production of parts. Specifically, compared to other machining processes, metal heat treatment generally does not alter the shape or overall chemical composition of a workpiece. Instead, it imparts or improves the workpiece's performance by altering its internal microstructure or surface chemical composition. This process is characterized by improving the workpiece's intrinsic quality, which is generally not visible to the naked eye. In order to achieve the desired mechanical, physical, and chemical properties in metal workpieces, in addition to the appropriate selection of materials and various forming processes, heat treatment processes are often essential. Steel is the most widely used material in the machinery industry. Its complex microstructure can be controlled through heat treatment, making heat treatment of steel a major component of metal heat treatment. Additionally, aluminum, copper, magnesium, titanium, and their alloys can also have their mechanical, physical, and chemical properties altered through heat treatment to achieve different performance characteristics.

[0027] When parts are just produced, there may be some problems with their internal structure, such as residual stress, uneven grain size, etc., which will affect the strength, hardness, toughness and other properties of the die-casting. Through heat treatment, residual stress can be eliminated, the internal structure can be made more uniform, the comprehensive performance of the die-casting can be improved, and the die-casting can be made more durable and less prone to damage. Therefore, the heat treatment process of parts is a core link in the die-casting manufacturing process. For die-castings with different shape data and different material data, the optimal process parameters corresponding to the heat treatment process are different. The deformation prediction methods in the existing technology, such as finite element analysis, although mature in theory, often face the limitations of high computational cost, strong dependence on parameters, and difficulty in efficiently processing multi-modal data in actual applications, and cannot perform efficient, intelligent and accurate deformation prediction of parts during the production process.

[0028] Based on the above, an embodiment of the present invention provides a deformation prediction method, device and storage medium based on prompt coding and adaptive feature fusion, which obtains the technical parameters of the part and converts the technical parameters to obtain the deformation probability data set of the part, wherein the technical parameters include: shape data, material data, heat treatment process data, and deformation data of the part; parameter encoding is performed on the deformation probability data set to extract the feature data of the technical parameters; according to the type of technical parameters, the feature data is adaptively fused to obtain the fused feature data of the part; the fused feature data is input into a preset prediction module to obtain the prediction result, wherein the prediction result includes the limit value and error mean variance of the deformation data; the deformation data of the current part is compared with the prediction result, and the heat treatment process data of the current part is adjusted to improve the prediction accuracy and efficiency of the thermal processing deformation of the part and reduce the deformation prediction error of the part.

[0029] See Figure 1 , Figure 1 FIG. 4 shows the process of the deformation prediction method based on hint coding and adaptive feature fusion provided by an embodiment of the present invention. Figure 1 As shown, the deformation prediction method based on hint coding and adaptive feature fusion in an embodiment of the present invention includes the following steps:

[0030] Step S1000: Acquire technical parameters of a part and convert the technical parameters to obtain a deformation probability data set of the part, wherein the technical parameters include: shape data, material data, heat treatment process data, and deformation data of the part.

[0031] It's understandable that during the heat treatment of parts, different shape data, such as mass, height, volume, and surface area, as well as different material data, must be considered. Due to their varying geometric properties, parts with different shape data experience varying heating and cooling effects under the same heat treatment process data, leading to differences in their deformation data. Furthermore, material data, which refers to the physical, chemical, and mechanical properties and characteristics of a material, is crucial for the selection and use of part materials. Parts with different material data exhibit varying performance in terms of physical, chemical, mechanical, optical, thermal, magnetic, and acoustic parameters.

[0032] It is understandable that the heat treatment process parameters corresponding to die-casting parts with different shape data and material coefficients during the heat treatment process are different, such as the optimal heating temperature, optimal heating time, optimal holding time, optimal cooling rate and time. In the embodiment of the present application, by obtaining parts with good mechanical properties after the heat treatment process, that is, parts with high strength and small deformation after hot working, the corresponding shape data, material data, and heat treatment process data are recorded and constructed as the input data of the data set. The deformation amount after the heat treatment process is recorded and constructed as the deformation probability data set of the parts.

[0033] See Figure 2 , Figure 2 FIG. 5 is a schematic diagram showing a specific implementation process of another embodiment of the above step S1000. Figure 2 As shown, step S1000 at least includes the following steps:

[0034] Step S1100: extract shape data from the three-dimensional model of the part.

[0035] It should be noted that in order to accurately obtain the shape data of a part, it is necessary to obtain a three-dimensional model file of the part, or scan the part to obtain a real-time three-dimensional model. In actual applications, a CAD file or a three-dimensional scanner is used to obtain shape data. Specifically, the obtained three-dimensional model data is converted into corresponding two-dimensional views from the six angles of up, down, left, right, front, and back in three-dimensional space to obtain the corresponding geometric shape image data of the part, that is, the shape data of the part. It should be noted that extracting the shape data required in the embodiment of this application through a three-dimensional model file of the part or scanning the part belongs to the prior art and will not be elaborated here.

[0036] Step S1200: Obtain material properties of the part and extract keywords to obtain material data.

[0037] It's understandable that part material properties, one of the most fundamental material properties, can influence their physical properties. These properties primarily include density, melting point, and thermal expansion coefficient. We collected basic die-casting material properties, including alloy composition, density, and thermal expansion coefficient. We performed keyword extraction on the text data and constructed key material attribute tags to obtain a textual description of the casting material's properties, i.e., the material data.

[0038] Step S1300: Collect heat treatment process conditions of parts to obtain heat treatment process data.

[0039] Heat treatment, as it's understood, refers to a metal heat processing process in which a material is heated, held, and cooled in a solid state to achieve the desired microstructure and properties. Metal heat treatment is a key process in mechanical manufacturing. Compared to other machining processes, heat treatment generally does not alter the shape or overall chemical composition of a workpiece. Instead, it imparts or improves the workpiece's performance by altering its internal microstructure or surface chemical composition.

[0040] It should be noted that the heat treatment process of a part involves several key process parameters, including treatment temperature, duration, and cooling rate. These key process parameters are stored as numerical heat treatment process data to obtain the part's heat treatment process data. In actual applications, key process parameters such as treatment temperature, duration, and cooling rate can be read directly from the processing equipment via instructions. This is prior art and will not be discussed here.

[0041] Step S1400: Obtain deformation information of the part after heat treatment, perform statistics on the deformation information, and obtain deformation data.

[0042] It is understandable that after a part has completed heat treatment, its dimensions will change compared to its shape data before heat treatment, as heat treatment will alter the workpiece's internal microstructure or the chemical composition of its surface, thereby imparting or improving its performance. To control the dimensional changes before and after heat treatment within a preset range, it is necessary to collect statistical information on the part's deformation. Specifically, the deformation data of die-cast parts after heat treatment is obtained from actual production and experiments. Statistical calculations are performed on the deformation data, and statistical distribution information such as the mean μ, standard deviation δ, and range [min, max] of the deformation data are extracted.

[0043] See Figure 3 , Figure 3 FIG. 5 is a schematic diagram showing a specific implementation process of another embodiment of the above step S1000. Figure 3 As shown, step S1000 at least includes the following steps:

[0044] Step S1500: Clean the technical parameters to obtain cleaning technical parameters.

[0045] It is understood that data cleaning refers to the process of identifying and correcting errors, inconsistencies or missing values ​​in data through technical means to improve data quality. The core goal of data cleaning is to process dirty data, such as duplicate records, outliers, inconsistent formats, etc., so that it reaches accurate, complete, consistent and reliable standards, providing a high-quality data foundation for subsequent analysis or decision-making. In the embodiment of the present application, data cleaning includes missing value processing and outlier processing. For the missing values ​​in the deformation probability data set of the part, the mean filling and interpolation methods are randomly used to complete the data. For example, when the three-dimensional model data in the above step S1100 is missing, 6 two-dimensional images of different perspectives are randomly generated based on the three-dimensional space. In addition, when the material data in the above step S1200 is missing, a material with properties similar to the part is used for copy filling. In addition, outlier processing is mainly aimed at the values ​​in the heat treatment process data in the above step S1300, for example: abnormal processing temperature, processing temperature exceeding the safe range of the material is regarded as abnormal data, and extreme outliers are eliminated to ensure the rationality and consistency of the technical parameters.

[0046] Step S1600: converting the shape data in the cleaning technical parameters into two-dimensional image information, and converting the material data in the cleaning technical parameters into text data in a key-value format.

[0047] It's understandable that to improve the processing efficiency of shape and material data, shape data needs to be converted into two-dimensional images, and material data into key-value text data. Specifically, geometric shape data is uniformly converted into two-dimensional images with a resolution of 512×512. Material data is processed to remove stop words and punctuation and then uniformly converted into key-value text data, making it suitable for subsequent feature extraction and model training. The data type of heat treatment process data remains unchanged.

[0048] Step S1700: Summarize the two-dimensional image information, text data, heat treatment process data, and deformation data to obtain a deformation probability data set.

[0049] It's understandable that by combining the 2D image information, text data, heat treatment process data, and deformation data obtained in the above steps, a high-quality dataset of part deformation after heat treatment is constructed, namely the deformation probability dataset. Each data item in the deformation probability dataset includes 2D image information V, material property text data T, and process parameter numerical data N.

[0050] Step S2000: Parameter encoding is performed on the deformation probability data set to extract characteristic data of technical parameters.

[0051] It is understandable that since the deformation probability dataset obtained in the above steps contains different types of data, feature extraction is required for different types of data. The deformation probability dataset is parameter-encoded, and the feature data of the technical parameters is extracted, mainly for the two-dimensional image information, text data, and heat treatment process data of the parts involved in the heat treatment process of the parts to construct features. In machine learning, pattern recognition, and image processing, feature extraction starts with an initial set of measurement data and establishes features designed to provide information and non-redundant features, thereby facilitating subsequent learning and generalization steps, and bringing better interpretability during the model training process and data analysis process.

[0052] See Figure 4 , Figure 4 FIG. 5 is a schematic diagram showing a specific implementation process of another embodiment of the above step S2000. Figure 4 As shown, step S2000 includes at least the following steps:

[0053] Step S2100: Input the two-dimensional image information into the visual transformer and output a high-dimensional feature vector of the shape data.

[0054] The Vision Transformer (VIT) is a neural network model based on the Transformer architecture, specifically designed for computer vision tasks. Traditional computer vision models, such as convolutional neural networks (CNNs), have achieved great success in image processing, but they have limitations, such as poor ability to model long-range dependencies. ViT addresses these issues by introducing the Transformer's attention mechanism, achieving excellent results on several vision tasks.

[0055] It should be noted that when using ViT to extract features from two-dimensional image information, firstly, the V i ∈R H×W×C , the input image is divided into N fixed-size P×P image blocks (patches), each patch is flattened and linearly projected into a high-dimensional feature space V k ∈R D , where D is the embedding dimension. Position encoding is added to preserve the spatial layout information of each patch in the image to form the image block embedding matrix V i =[v1,v2,v3,…,v N ]∈R D, and processes the image blocks through the self-attention mechanism to capture the global and local features in the image, and outputs the high-dimensional feature representation vector F of the shape data v ∈R D , as shown in the following formula:

[0056] V i =[v1,v2,v3,…,v N ], F v =ViT(T i )

[0057] Step S2200: Input text data into a bidirectional encoder and output text features of the material data.

[0058] It is understandable that the bidirectional encoder (Bidirectional Encoder Representations from Transformers, BERT) is a pre-trained model based on the Transformer architecture, mainly used for language representation learning. BERT uses an encoder-only Transformer structure, and simultaneously captures the left and right contextual information of words through a multi-head self-attention mechanism, which solves the limitations of traditional unidirectional models. BERT is pre-trained through a masked language model and next sentence prediction tasks to learn language representation. The BERT model is composed of multiple layers of Transformer encoders stacked together, which can simultaneously consider contextual information and achieve bidirectional language understanding. This bidirectional feature enables BERT to perform well in various NLP tasks and become the base model for many downstream tasks.

[0059] It should be noted that BERT is used to extract features from text data. Specifically, the text data T i As input to the BERT model, the BERT model encodes each word of the text through the self-attention mechanism, captures the relationship between words, and then generates text features of the text data. In practical applications, given an input text Where L is the length of the text, D t is the dimension of text embedding. The text data is segmented and mapped into the word embedding space to obtain the input matrix in, Represents the embedding vector of each word in the text. Then, the text vector T k The input is sent to the Transformer model, and the dynamic representation of each word is generated by combining the sentence context information through a multi-layer self-attention mechanism to obtain text features. The specific formula is as follows:

[0060] T k =[t1,t2,t3,…,t L]

[0061] F t =BERT(T k )

[0062] Step S2300: Input the heat treatment process data into the prompt learning-based encoder and output the text features of the heat treatment process data.

[0063] It is understandable that the role of prompts in the learning model is mainly to provide the model with the context of the input information and the parameter information of the input model. The core idea of ​​the prompt model is to introduce a prompt-based learning mechanism in the pre-training stage, so that the model can quickly adapt to different tasks based on the given prompt information. This mechanism greatly improves the flexibility and scalability of the model and can adapt the model to various complex scenarios through simple prompt adjustments. Therefore, prompt learning is a new fine-tuning strategy that guides the model to learn specific tasks by adding specific prompt information to the heat treatment process data. This strategy allows the prompt information to be adjusted to adapt to different tasks without changing the model structure. By introducing the prompt learning mechanism, the prompt model achieves rapid fine-tuning based on the pre-trained model.

[0064] F n =MLP(N×Prompt n )

[0065] It should be noted that through the encoder based on prompt learning, deep representation of process data is achieved, and its potential correlation with deformation data is automatically mined, which improves the adaptability of the prediction model to diversified process data and reduces the dependence on manual parameter adjustment in traditional methods.

[0066] See Figure 5 , Figure 5 FIG. 2 shows a schematic diagram of a specific implementation process of another embodiment of the above step S2300. Figure 5 As shown, step S2300 includes at least the following steps:

[0067] Step S2310: configure a multi-dimensional vector for each type of heat treatment process data, and initialize it in a random manner to obtain a process vector.

[0068] It is understandable that the heat treatment process data is numerical data that describes the process conditions of the heat treatment, such as temperature, time, and cooling rate. The embodiment of the present application constructs a coding mechanism based on learnable prompts for numerical data, designs a 1024-dimensional vector for each type of process data, and initializes it randomly with an initialization value range of (0,1). The vector is represented as Prompt n.

[0069] Step S2320: Combine the process vector and the heat treatment process data to obtain associated parameter features.

[0070] It is understandable that the numerical characteristics of heat treatment process data are compared with Prompt n The features are combined to form a more semantically relevant feature representation of the associated parameters, as shown in the following formula:

[0071]

[0072] Here, i represents the serial number of a specific numerical parameter in the process data, and the embodiment of the present application mainly includes three numerical parameters: temperature, time, and cooling rate.

[0073] Step S2330: Input the associated parameter features into the multilayer perceptron and output the process features.

[0074] It is understood that the Multilayer Perceptron (MLP) is a feedforward artificial neural network model that maps multiple input data sets to a single output data set. By inputting the associated parameter features into the MLP, the process feature F of the process data can be obtained. n , as shown in the following formula:

[0075] F n =MLP(N×Prompt n )

[0076] Step S2400: Summarize high-dimensional feature vectors, text features, and process features to obtain feature data.

[0077] It is understandable that in order to facilitate data transmission and processing, the high-dimensional feature vector F obtained in the above steps is v , text features F t and process characteristics F n Perform aggregation processing to obtain feature data.

[0078] Step S3000: Adaptively fuse feature data according to the type of technical parameters to obtain fused feature data of the part.

[0079] It is understandable that adaptive feature fusion is a technology in deep learning that dynamically adjusts the feature fusion process, aiming to improve the generalization ability and performance of the model by optimizing feature representation. By dynamically optimizing the fusion process, the robustness of the model in complex tasks is significantly improved. The future direction of adaptive feature fusion also includes a more lightweight fusion layer design and enhanced cross-modal versatility. The high-dimensional feature vector F obtained in the above steps is v , text features F tand process characteristics F n Through the type parameter feature adaptive interactive fusion module, interactive fusion is performed to achieve information complementarity of multi-type data and enhance the synergy of multimodal data.

[0080] See Figure 6 , Figure 6 FIG. 5 is a schematic diagram showing a specific implementation process of another embodiment of the above step S3000. Figure 6 As shown, step S3000 includes at least the following steps:

[0081] Step S3100: Perform high-dimensional feature mapping on the high-dimensional feature vector, text features, and text features to obtain mapping features.

[0082] It can be understood that by taking the high-dimensional feature vector F v , text features F t and process characteristics F n Input into the multi-layer perceptron for mapping, so as to analyze and process the multi-type parameter features. v ,F t ,F n There is a problem of inconsistent dimensions. For different types of parameter type features, corresponding multi-layer perceptron mapping layers are set to convert the multi-type parameter features F v ,F t ,F n Projecting into a unified high-dimensional feature space facilitates the subsequent cross-attention layer to realize the interaction between different types of parameters and obtain mapping features. Specifically, the high-dimensional feature mapping process is shown in the following formula:

[0083] F′ v =MLP′ v (F v )

[0084] F′ t =MLP′ t (F t )

[0085] F′ n =MLP′ n (F n )

[0086] Step S3200: Input the mapped features into the cross-attention module to obtain cross features.

[0087] It is understandable that the embodiment of the present application constructs three parallel cross-attention modules, each module contains two layers of cross-attention mechanisms, and the difference between the three cross-attention modules is that the input query (Query), key (Key), and value (Value) are different. The cross-attention mechanism allows different types of parameter features to be jointly modeled in the same space, thereby capturing the association between different types of parameters. In this process, the high-dimensional feature vectors, text features, and process features of the parts interact with each other as queries, keys, and values, respectively. The calculation process of the overall cross-attention layer is shown in the following formula:

[0088]

[0089] Step S3300: Fuse cross features by feature concatenation to obtain preliminary fused features.

[0090] It is understandable that in order to better combine the high-dimensional feature vector, text features and process features of the parts and obtain more abundant parameter information of the heat treatment process of the parts, the embodiment of the present application preliminarily obtains the corresponding fusion feature F by feature splicing. cat , the entire splicing process is shown in the following formula:

[0091]

[0092] Step S3400: Combine the preliminary fusion features and the adaptive correlation matrix, input them into the feedforward network, and obtain fusion feature data.

[0093] It is understandable that the long-range dependencies between different modalities are further captured through the self-attention mechanism, thereby enhancing the semantic associations between shape data, material data, and heat treatment process data. By calculating the correlation matrix between global features, the model can effectively extract complementary information and reduce the limitations of single modal information. At the same time, the self-attention mechanism can automatically weight the importance of features, allowing the model to pay more attention to key shape data, material data, and heat treatment process data related to deformation prediction, and better establish complex interactive relationships between features. The specific formula is as follows:

[0094] F self =SlefAttention(F cat )

[0095] It is understandable that a feed-forward network (FFN) is added after the self-attention mechanism to further improve the model's expressiveness and nonlinear modeling capabilities. The feed-forward network enhances feature expression through nonlinear transformations, while combining element-by-element operations to improve the model's adaptability to complex data patterns and robustness to data features, effectively compensating for the shortcomings of the self-attention mechanism in modeling local details. The specific formula is as follows:

[0096] F fuse =FFN(F self )

[0097] It should be noted that by setting an adaptive feature fusion strategy, enhancing the synergy of multimodal data, and making full use of the complementary information of different types of data in the preliminary fusion features, the model's ability to understand complex data patterns is improved, and a comprehensive analysis from microscopic material properties to macroscopic geometric shapes is achieved.

[0098] Step S4000: Input the fused feature data into a preset prediction module to obtain a prediction result, wherein the prediction result includes the limit value and error mean variance of the deformation data.

[0099] It is understandable that after obtaining the fused feature data through the above steps S1000-S3000, it is input into the preset prediction module to obtain the prediction result. Specifically, the prediction module is provided with two prediction heads (Prediction Head) to realize the prediction of the limit value and error mean variance of the statistical probability of the heat treatment deformation variable of the part. Among them, the prediction head is the key module in the neural network model responsible for converting the results after feature extraction into the final prediction output. Its core function is to map abstract features to specific results according to task requirements, such as classification, detection, time series prediction, etc. The prediction head is usually located at the end of the model and consists of several fully connected layers, activation functions or specific task modules, which are used to convert the features extracted by the backbone network into structured outputs.

[0100] It should be noted that the embodiment of the present application constructs a joint prediction method of limit value and error mean variance, which is more in line with industrial needs, provides deformation range information, and improves model robustness and interpretability of prediction results.

[0101] See Figure 7 , Figure 7 FIG. 4 is a schematic diagram showing a specific implementation process of another embodiment of the above step S4000. Figure 7 As shown, step S4000 at least further includes the following steps:

[0102] Step S4100: Input the fused feature data into a prediction module equipped with a multi-layer perceptron to obtain a limit value of the deformation data.

[0103] It can be understood that the fusion feature data is input into the prediction module with a multi-layer perceptron, and a prediction module MLP containing two multi-layer perceptrons is set up. min-max This allows for prediction of the limits of part heat treatment deformation, i.e., predicting the upper and lower limits of part deformation. By directly quantifying the range of deformation, it better meets the application requirements of industrial reality and provides a useful reference range for process designers. Specifically, the predicted output of the limit is shown in the following formula:

[0104]

[0105] Step S4200: Input the fused feature data into a prediction module equipped with a multi-layer perceptron to obtain the error mean variance of the deformation data.

[0106] It can be understood that, when the fusion feature data is input into the prediction module with a multi-layer perceptron, a prediction module MLP including two multi-layer perceptrons is set up. μ-δ The prediction of the mean and variance of the deformation error during heat treatment of parts is achieved. By predicting the mean and variance of the deformation error, the robustness to abnormal data is improved, and the reliability and interpretability of the prediction model are further improved. Specifically, the prediction output of the mean and variance of the deformation data error is shown in the following formula:

[0107]

[0108] Step S4300: Combine the limit value and the error mean variance to output the prediction result.

[0109] It is understandable that the final prediction result includes the statistical limit of the deformation variable and the error mean variance, as shown in the following formula:

[0110]

[0111] Step S5000: Compare the deformation data of the current part with the prediction result, and adjust the heat treatment process data of the current part.

[0112] It is understood that in actual production, by comparing the current part's deformation data with the prediction results output by the prediction model, the current part's heat treatment process data can be accurately corrected, reducing the part's heat treatment deformation, thus achieving a closed-loop verification mechanism. In actual applications, sending instructions to the heat treatment equipment to modify the current part's heat treatment process data based on the prediction results is a prior art and will not be further described here.

[0113] It should be noted that by comparing the current part's deformation data with the predicted results and adjusting the current part's heat treatment process data, this method breaks through the limitations of traditional point-value deformation prediction and manual parameter adjustment, achieving a deep integration of deformation statistical distribution and process parameter optimization. Dynamically adjusting process data based on model prediction results not only improves optimization efficiency but also significantly enhances adaptability and robustness, providing an intelligent heat treatment control solution for parts.

[0114] To quantify the performance of the deformation prediction model, three key evaluation metrics were used: mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE). MAE measures the average absolute difference between the predicted and true values; MSE measures the average squared error, reflecting the degree of error dispersion. The dataset used consists of both real-world data and Abaqus simulation data. The real-world data includes deformation data of die-cast parts collected during heat treatment experiments. This data reflects deformation conditions under real-world production environments and accounts for 10% of the dataset (approximately 200 records). Furthermore, Abaqus simulation data, generated using finite element analysis software, simulates the deformation behavior of die-cast parts during heat treatment, covering a wide range of process parameters and operating conditions, and accounts for 90% of the dataset (approximately 1,800 records). By combining real-world and simulated data, this dataset is both representative of real-world operating conditions and provides a sufficient sample size for model training and comparative analysis. Linear regression, support vector machine, decision tree, and the prediction method of this application were used for training, and cross-validation was used to evaluate model performance. The MAE, MSE, and RMSE indicators were used to quantitatively compare the accuracy of deformation prediction of each model. The lower the value of these indicators, the closer the prediction result is to the actual value. The comparative experimental results are shown in Table 1 below:

[0115] Comparison Method MAE MSE RMSE Linear regression 3.16 1.3 3.63 Support Vector Machine 7.02 6.2 7.9 Decision Tree 4.82 4.2 6.48 This application prediction method 1.89 0.68 1.62

[0116] Table 1: Performance comparison of various prediction methods

[0117] As can be seen from Table 1, the prediction method of the present application is significantly superior to the traditional machine learning method in terms of deformation prediction accuracy. By introducing deep learning hint coding and adaptive feature fusion, the prediction method of the present application can more accurately capture the intrinsic relationship between multimodal data and achieve high-precision prediction of deformation probability. This method uses a mixture of actual collected data and Abaqus simulation data to construct a rich data set that is both representative of real working conditions and covers a variety of process conditions, thereby greatly improving the generalization ability of the model. In addition, through the closed-loop verification mechanism, the deformation prediction results can provide real-time feedback and guide the optimization of process parameters to achieve dynamic control of deformation during heat treatment. This closed-loop control strategy effectively reduces product scrap and rework due to excessive deformation, and improves overall production efficiency and product consistency.

[0118] See also Figure 8 , Figure 8 This is a structural diagram of a deformation prediction device 600 based on hint coding and adaptive feature fusion provided in an embodiment of the present application. The entire process of the deformation prediction method based on hint coding and adaptive feature fusion provided in an embodiment of the present application involves the following modules in the deformation prediction device based on hint coding and adaptive feature fusion: an acquisition module 610, an encoding module 620, a fusion module 630, a prediction module 640 and an adjustment module 650.

[0119] The acquisition module 610 is used to acquire technical parameters of the part and convert the technical parameters to obtain a deformation probability data set of the part, wherein the technical parameters include: shape data, material data, heat treatment process data, and deformation data of the part;

[0120] The encoding module 620 is used to perform parameter encoding on the deformation probability data set and extract characteristic data of the technical parameters;

[0121] A fusion module 630 is used to adaptively fuse feature data according to the type of technical parameters to obtain fused feature data of the part;

[0122] Prediction module 640, used to input the fused feature data into a preset prediction module to obtain a prediction result, wherein the prediction result includes the limit value and error mean variance of the deformation data;

[0123] The adjustment module 650 is used to compare the deformation data of the current part with the prediction result and adjust the heat treatment process data of the current part.

[0124] It should be noted that the information interaction, execution process, etc. between the modules of the above-mentioned device are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0125] Figure 9 An electronic device 700 provided in an embodiment of the present application is shown. The electronic device 700 includes but is not limited to:

[0126] Memory 701, used for storing programs;

[0127] The processor 702 is configured to execute the program stored in the memory 701. When the processor 702 executes the program stored in the memory 701, the processor 702 is configured to execute the above-mentioned deformation prediction method based on hint coding and adaptive feature fusion.

[0128] The processor 702 and the memory 701 may be connected via a bus or other means.

[0129] Memory 701, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer executable programs, such as the deformation prediction method based on hint coding and adaptive feature fusion described in any embodiment of this application. Processor 702 implements the deformation prediction method based on hint coding and adaptive feature fusion by executing the non-transitory software program and instructions stored in memory 701.

[0130] The memory 701 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store the execution of the above-mentioned deformation prediction method based on prompt coding and adaptive feature fusion. In addition, the memory 701 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 701 may optionally include a memory remotely located relative to the processor 702, and these remote memories may be connected to the processor 702 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0131] The non-transient software programs and instructions required to implement the above-mentioned deformation prediction method based on hint coding and adaptive feature fusion are stored in the memory 701. When executed by one or more processors 702, the deformation prediction method based on hint coding and adaptive feature fusion provided in any embodiment of the present application is executed.

[0132] An embodiment of the present application further provides a storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the above-mentioned deformation prediction method based on hint coding and adaptive feature fusion.

[0133] In one embodiment, the storage medium stores computer-executable instructions, which are executed by one or more control processors 702, for example, by a processor 702 in the above-mentioned electronic device 700, so that the above-mentioned one or more processors 702 can execute the deformation prediction method based on hint coding and adaptive feature fusion provided in any embodiment of the present application.

[0134] The embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0135] Those skilled in the art will appreciate that all or some of the steps and systems in the method disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, and the computer-readable medium can include computer storage media (or non-transitory media) and communication media (or temporary media). As known to those skilled in the art, the term computer storage media is included in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data) and is volatile and non-volatile, removable, and non-removable. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory, or other memory technology, CD-ROM, digital versatile disks (DVD), or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage, or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically includes computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

Claims

1. A deformation prediction method based on hint coding and adaptive feature fusion, characterized in that: The following steps are involved: Acquiring technical parameters of a part and converting the technical parameters to obtain a deformation probability data set of the part, wherein the technical parameters include: shape data, material data, heat treatment process data, and deformation data of the part; Performing parameter encoding on the deformation probability data set to extract characteristic data of the technical parameters; Adaptively fusing the feature data according to the type of the technical parameters to obtain fused feature data of the part; Inputting the fused feature data into a preset prediction module to obtain a prediction result, wherein the prediction result includes the limit value and error mean variance of the deformation data; The deformation data of the current part is compared with the prediction result, and the heat treatment process data of the current part is adjusted.

2. The deformation prediction method based on hint coding and adaptive feature fusion according to claim 1, characterized in that: The technical parameters of the parts are obtained, including: extracting the shape data from the three-dimensional model of the part; Obtaining material properties of the part and extracting keywords to obtain the material data; Collecting the heat treatment process conditions of the part to obtain the heat treatment process data; Obtain deformation information of the part after heat treatment, perform statistics on the deformation information, and obtain the deformation data.

3. The deformation prediction method based on hint coding and adaptive feature fusion according to claim 2, characterized in that: The converting process of the technical parameters includes: Performing cleaning processing on the technical parameters to obtain cleaning technical parameters; Converting the shape data in the cleaning technical parameters into two-dimensional image information, and converting the material data in the cleaning technical parameters into text data in a key-value format; The two-dimensional image information, the text data, the heat treatment process data and the deformation data are aggregated to obtain the deformation probability data set.

4. The deformation prediction method based on hint coding and adaptive feature fusion according to claim 3, characterized in that: The performing parameter encoding on the deformation probability data set and extracting characteristic data of the technical parameters includes: Inputting the two-dimensional image information into a visual converter and outputting a high-dimensional feature vector of the shape data; Inputting the text data into a bidirectional encoder and outputting text features of the material data; Inputting the heat treatment process data into a prompt learning-based encoder to output text features of the heat treatment process data; The high-dimensional feature vector, the text feature, and the process feature are aggregated to obtain the feature data.

5. The deformation prediction method based on hint coding and adaptive feature fusion according to claim 4, characterized in that: The step of inputting the heat treatment process data into an encoder based on prompt learning and outputting process features of the heat treatment process data comprises: configuring a multi-dimensional vector for each type of the heat treatment process data, and initializing it in a random manner to obtain a process vector; Combining the process vector with the heat treatment process data to obtain associated parameter features; The associated parameter features are input into a multilayer perceptron, and the process features are output.

6. The deformation prediction method based on hint coding and adaptive feature fusion according to claim 4, characterized in that: Adaptively fusing the feature data according to the type of the technical parameter includes: Performing high-dimensional feature mapping on the high-dimensional feature vector, the text feature, and the text feature to obtain mapping features; Inputting the mapped features into a cross attention module to obtain cross features; The cross features are fused by feature splicing to obtain preliminary fused features; The preliminary fusion features and the adaptive correlation matrix are combined and input into a feedforward network to obtain the fusion feature data.

7. The deformation prediction method based on hint coding and adaptive feature fusion according to claim 1, characterized in that: The step of inputting the fused feature data into a preset prediction module to obtain a prediction result includes: Inputting the fused feature data into the prediction module provided with a multi-layer perceptron to obtain a limit value of the deformation data; Inputting the fused feature data into the prediction module provided with a multi-layer perceptron to obtain the error mean variance of the deformation data; The prediction result is output by combining the limit value and the error mean variance.

8. A deformation prediction device based on hint coding and adaptive feature fusion, characterized in that: include: an acquisition module, configured to acquire technical parameters of a part and convert the technical parameters to obtain a deformation probability data set of the part, wherein the technical parameters include: shape data, material data, heat treatment process data, and deformation data of the part; An encoding module, configured to perform parameter encoding on the deformation probability data set and extract characteristic data of the technical parameters; A fusion module, configured to adaptively fuse the feature data according to the type of the technical parameters to obtain fused feature data of the part; A prediction module, configured to input the fused feature data into a preset prediction module to obtain a prediction result, wherein the prediction result includes a limit value and an error mean variance of the deformation data; An adjustment module is used to compare the deformation data of the current part with the prediction result and adjust the heat treatment process data of the current part.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for deformation prediction based on hint coding and adaptive feature fusion according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that A computer program is stored, and when the computer program is executed by a processor, the deformation prediction method based on hint coding and adaptive feature fusion according to any one of claims 1 to 7 is implemented.

Images