Surface migration prediction-based forging defect identification data processing method and system

Abstract

The application discloses a forging defect identification data processing method and system based on surface migration prediction, relates to the technical field of defect identification, and comprises the following steps: determining an image difference vector of a single forging according to image data and a pre-generated surface migration prediction model, and determining a temperature vector of the single forging according to temperature time sequence data; constructing an image-temperature correlation vector corresponding to the single forging according to the image difference vector and the temperature vector, and constructing an image-temperature correlation matrix corresponding to a plurality of forgings according to the image-temperature correlation matrix; decomposing the image-temperature correlation matrix to extract a main feature space; for each forging, determining a corresponding identification index according to the corresponding image-temperature correlation vector and the main feature space; and jointly analyzing identification indexes of the plurality of forgings to generate a defect identification result. The application has the effect of improving the reliability of forging defect identification.

Description

Technical Field

[0001] This application relates to the technical field of defect identification, and in particular to data processing methods, systems, equipment and media for forging defect identification based on surface migration prediction. Background Technology

[0002] In the manufacturing industry, accurate identification and causal analysis of surface defects in forgings are crucial for ensuring the safety and reliability of components during service. Currently, two main types of technologies are used in industry for defect detection: one is image analysis based on machine vision, which acquires images of the forging surface and uses algorithms to locate and classify defects. While this method has the advantages of automation and non-contact operation, its judgment relies entirely on surface morphological features and cannot reveal the process-related causes of defects. The other is a comprehensive assessment based on human experience, where technicians combine surface observation with process knowledge to infer the causes of defects. This method is limited by the experience level of personnel and has inherent limitations such as strong subjectivity, low efficiency, and difficulty in solidifying and passing on knowledge.

[0003] The current technological system suffers from a significant "data silo" phenomenon: visual inspection systems only process image information, while process monitoring systems only record parameters such as temperature and pressure. These two types of data are isolated at the analysis and decision-making levels. This fragmentation makes it difficult for production personnel to quickly and accurately correlate defect characteristics with specific process parameter anomalies when surface defects are discovered. This hinders the ability to trace the root cause from a "quality phenomenon" to a "process ailment," severely restricting precise control of the production process and closed-loop management of quality issues. Therefore, developing an intelligent method that can deeply integrate multi-source information and achieve accurate defect identification and process root cause tracing has become an urgent need to overcome the current bottlenecks in quality control. Summary of the Invention

[0004] To improve the reliability of forging defect identification, this application provides a forging defect identification data processing method, system, equipment, and medium based on surface migration prediction.

[0005] In the first aspect, this application provides a data processing method for forging defect identification based on surface migration prediction, which adopts the following technical solution:

[0006] A data processing method for forging defect identification based on surface migration prediction includes:

[0007] Image data and temperature time-series data of multiple forgings in the same batch are acquired. Based on the image data and a pre-generated surface migration prediction model, the image difference vector of a single forging is determined, and the temperature vector of the single forging is determined based on the temperature time-series data. The temperature time-series data is used to represent the temperature change of each feature point of the forging during the processing.

[0008] The image temperature correlation vector corresponding to the single forging is constructed based on the image difference vector and the temperature vector, and the image temperature correlation matrix corresponding to the multiple forgings is constructed based on the image temperature correlation matrix.

[0009] The image temperature correlation matrix is ​​decomposed to extract the main feature space;

[0010] For each forging, a corresponding identification index is determined based on the corresponding image temperature correlation vector and the main feature space;

[0011] The identification indicators of the multiple forgings are jointly analyzed to generate defect identification results, wherein the defect identification results include surface defect types and temperature anomaly patterns that are causally related to the surface defect types.

[0012] By employing the above technical solution, image data and temperature time-series data of multiple forgings in the same batch are acquired. Based on the image data and a pre-generated surface migration prediction model, the image difference vector of a single forging is determined, and the temperature vector of a single forging is determined based on the temperature time-series data. The temperature time-series data represents the temperature changes at various feature points of the forging during processing. Then, an image-temperature correlation vector corresponding to a single forging is constructed based on the image difference vector and the temperature vector. Furthermore, an image-temperature correlation matrix is ​​constructed for multiple forgings based on the image-temperature correlation matrix. The image-temperature correlation matrix is ​​then decomposed to extract the main feature space. Finally, for... For each forging, corresponding identification indicators are determined based on the corresponding image temperature correlation vector and main feature space. Then, the identification indicators of multiple forgings are jointly analyzed to generate defect identification results. The defect identification results include surface defect types and temperature anomaly patterns that are causally related to the surface defect types. In the above method, on the one hand, temperature data is used to corroborate the image identification results, effectively overcoming the limitation of relying solely on visual features being easily interfered with. On the other hand, process anomalies are traced back through defect features, forming a complete evidence chain of "phenomenon-root cause". Through this cross-modal correlation analysis, the accuracy and reliability of defect identification are improved.

[0013] Optionally, the step of constructing the image temperature correlation vector corresponding to the single forging based on the image difference vector and the temperature vector includes:

[0014] For each forging, the forging is taken as the target forging. The image difference vector of the target forging is semantically embedded to obtain the image difference embedding feature, and the temperature vector of the target forging is semantically embedded to obtain the temperature embedding feature.

[0015] The image difference embedding features are obtained by performing high-level semantic extraction on the image difference embedding features through the first high-level semantic extraction sub-model;

[0016] The temperature embedding features are obtained by performing high-level semantic extraction on the second high-level semantic extraction sub-model.

[0017] By using a semantic analysis model, cross-domain semantic encoding is performed on the image difference features and the temperature features to obtain the corresponding image temperature correlation vector.

[0018] By adopting the above technical solution, in order to construct the image temperature correlation vector, for each forging, the forging is taken as the target forging. The image difference vector of the target forging is semantically embedded to obtain the image difference embedding feature, and the temperature vector of the target forging is semantically embedded to obtain the temperature embedding feature. Then, through the first high-level semantic extraction sub-model, the image difference embedding feature is subjected to high-level semantic extraction to obtain the image difference feature. Then, through the second high-level semantic extraction sub-model, the temperature embedding feature is subjected to high-level semantic extraction to obtain the temperature feature. Finally, through the semantic analysis model, the image difference feature and the temperature feature are cross-domain correlated semantically encoded to obtain the corresponding image temperature correlation vector.

[0019] Optionally, the step of performing high-level semantic extraction on the image difference embedding features through the first high-level semantic extraction sub-model to obtain image difference features includes:

[0020] The image difference embedding features are fed into a first high-level semantic extraction sub-model, wherein the first high-level semantic extraction sub-model includes a first feature enhancement channel and a second feature enhancement channel set in parallel, and the first feature enhancement channel and the second feature enhancement channel are respectively configured with independent dimension transformation layers and feature activation layers.

[0021] Autocorrelation analysis is performed on the image difference embedding features to extract image difference contextual association information;

[0022] The first image difference deep feature is generated by performing deep feature transformation on the image difference context association information through the first feature enhancement channel;

[0023] The image difference context association information is transformed into a second image difference depth feature by using the second feature enhancement channel;

[0024] Cross-attention integration is performed on the first image difference depth feature and the second image difference depth feature to obtain the corresponding image difference feature.

[0025] By adopting the above technical solution, in order to obtain image difference features, the image difference embedding features are fed into a first high-level semantic extraction sub-model. The first high-level semantic extraction sub-model includes a first feature enhancement channel and a second feature enhancement channel set in parallel. The first feature enhancement channel and the second feature enhancement channel are respectively configured with independent dimension transformation layers and feature activation layers. Then, autocorrelation analysis is performed on the image difference embedding features to extract image difference context association information. Then, deep feature transformation is performed on the image difference context association information through the first feature enhancement channel to generate the first image difference deep feature. Then, deep feature transformation is performed on the image difference context association information through the second feature enhancement channel to generate the second image difference deep feature. Finally, cross-attention integration is performed on the first image difference deep feature and the second image difference deep feature to obtain the corresponding image difference feature.

[0026] Optionally, the step of performing high-level semantic extraction on the temperature embedding features through the second high-level semantic extraction sub-model to obtain temperature features includes:

[0027] The temperature embedding feature is fed into the second high-level semantic extraction sub-model, wherein the second high-level semantic extraction sub-model includes a third feature enhancement channel and a fourth feature enhancement channel set in parallel, and the third feature enhancement channel and the fourth feature enhancement channel are respectively configured with independent dimension transformation layer and feature activation layer;

[0028] Autocorrelation analysis is performed on the temperature embedding features to extract temperature context association information;

[0029] The temperature context association information is subjected to deep feature transformation through the third feature enhancement channel to generate a first temperature deep feature.

[0030] The temperature context association information is subjected to deep feature transformation through the fourth feature enhancement channel to generate a second temperature deep feature.

[0031] Cross-attention integration is performed on the first temperature depth feature and the second temperature depth feature to obtain the corresponding temperature feature.

[0032] By adopting the above technical solution, in order to obtain temperature features, the temperature embedding features are fed into a second high-level semantic extraction sub-model. The second high-level semantic extraction sub-model includes a third feature enhancement channel and a fourth feature enhancement channel set in parallel. The third feature enhancement channel and the fourth feature enhancement channel are respectively configured with independent dimension transformation layers and feature activation layers. Then, autocorrelation analysis is performed on the temperature embedding features to extract temperature context association information. Then, the temperature context association information is subjected to deep feature transformation through the third feature enhancement channel to generate a first temperature deep feature. Then, the temperature context association information is subjected to deep feature transformation through the fourth feature enhancement channel to generate a second temperature deep feature. Finally, the first temperature deep feature and the second temperature deep feature are integrated by cross-attention to obtain the corresponding temperature feature.

[0033] Optionally, the step of performing cross-domain semantic encoding on the image difference features and the temperature features using a semantic analysis model to obtain the corresponding image temperature association vector includes:

[0034] The image difference features and the temperature features are loaded into the semantic analysis model, wherein the semantic analysis model includes multiple semantic analysis sub-models;

[0035] The image difference features and temperature features are correlated and mined using each of the semantic analysis sub-models to obtain the corresponding intermediate correlation features;

[0036] The intermediate correlation features are fused to obtain the image temperature correlation vector.

[0037] By adopting the above technical solution, in order to obtain the corresponding image temperature correlation vector, image difference features and temperature features are loaded into the semantic analysis model. The semantic analysis model includes multiple semantic analysis sub-models. Then, the image difference features and temperature features are correlated and mined through each semantic analysis sub-model to obtain the corresponding intermediate correlation features. Then, the intermediate correlation features are fused to obtain the image temperature correlation vector.

[0038] Optionally, the step of performing correlation mining on the image difference features and the temperature features through each of the semantic analysis sub-models to obtain the corresponding intermediate correlation features includes:

[0039] For each of the semantic analysis sub-models, the semantic analysis sub-model is used as the target semantic analysis sub-model, and the image difference features and the temperature features are loaded into the target semantic analysis sub-model. The target semantic analysis sub-model has a built-in first projection matrix and a second projection matrix. The first projection matrix is ​​used to transform the image difference features from the current semantic space to the target semantic space, and the second projection matrix is ​​used to transform the temperature features from the current semantic space to the target semantic space.

[0040] The image difference features are obtained by performing a linear transformation on the image difference features using the first projection matrix;

[0041] The temperature feature is linearly transformed using the second projection matrix to obtain the temperature transformation feature;

[0042] Determine the feature correlation score between the image difference transformation feature and the temperature transformation feature, and standardize the feature correlation score to obtain a standardized score;

[0043] The image difference transformation feature or the temperature transformation feature is weighted according to the standardized score to obtain the corresponding intermediate correlation feature.

[0044] By adopting the above technical solution, in order to obtain the corresponding intermediate association features, for each semantic analysis sub-model, the semantic analysis sub-model is used as the target semantic analysis sub-model. Image difference features and temperature features are loaded into the target semantic analysis sub-model. The target semantic analysis sub-model has a built-in first projection matrix and a second projection matrix. The first projection matrix is ​​used to transform the image difference features from the current semantic space to the target semantic space, and the second projection matrix is ​​used to transform the temperature features from the current semantic space to the target semantic space. Then, the image difference features are linearly transformed using the first projection matrix to obtain image difference transformation features, and the temperature features are linearly transformed using the second projection matrix to obtain temperature transformation features. Then, the feature correlation score between the image difference transformation features and the temperature transformation features is determined, and the feature correlation score is standardized to obtain a standardized score. Then, the image difference transformation features or the temperature transformation features are weighted according to the standardized score to obtain the corresponding intermediate association features.

[0045] Optionally, the step of decomposing the image temperature correlation matrix and extracting the main feature space includes:

[0046] The image temperature correlation matrix is ​​centered to determine the corresponding covariance matrix;

[0047] The covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues;

[0048] The eigenvalues ​​are sorted from largest to smallest, and the eigenvectors corresponding to the top K eigenvalues ​​are extracted as principal component directions. The top K principal component directions span the principal feature space, where the value of K is a positive integer determined by the cumulative variance contribution rate.

[0049] By adopting the above technical solution, in order to extract the main feature space, the image temperature correlation matrix is ​​centered to determine the corresponding covariance matrix. Then, the covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues. The eigenvalues ​​are then sorted from largest to smallest, and the eigenvectors corresponding to the first K eigenvalues ​​are extracted as principal component directions. The main feature space is spanned by the first K principal component directions, where the value of K is a positive integer determined by the cumulative variance contribution rate.

[0050] Optionally, the step of determining the corresponding recognition index based on the corresponding image temperature correlation vector and the main feature space includes:

[0051] The image temperature correlation vector is projected onto the main feature space to obtain the corresponding reconstructed vector;

[0052] The reconstruction error between the image temperature correlation vector and the reconstruction vector is determined, and the distance metric from the image temperature correlation vector to the main feature space is determined, wherein the reconstruction error is determined by Euclidean distance or cosine similarity, and the distance metric is determined by Mahalanobis distance.

[0053] The reconstruction error and the distance metric are used as identification indicators.

[0054] By adopting the above technical solution, in order to determine the recognition index, the image temperature correlation vector is projected into the main feature space to obtain the corresponding reconstruction vector. Then, the reconstruction error between the image temperature correlation vector and the reconstruction vector is determined, and the distance metric from the image temperature correlation vector to the main feature space is determined. The reconstruction error is determined by Euclidean distance or cosine similarity, and the distance metric is determined by Mahalanobis distance. Then, the reconstruction error and the distance metric are used as the recognition index.

[0055] Optionally, the step of jointly analyzing the identification indicators of the multiple forgings to generate defect identification results includes:

[0056] Cluster analysis was performed on the identification indicators of the multiple forgings to obtain multiple clusters;

[0057] Based on the distribution characteristics of the clusters, identify normal forging clusters and at least one abnormal forging cluster;

[0058] For each abnormal forging cluster, extract the common features of image differences of the forgings in the abnormal forging cluster, determine the corresponding surface defect type, and extract the common temperature anomaly pattern of the forgings in the abnormal forging cluster.

[0059] Establish a causal relationship mapping between the surface defect types and the common temperature anomaly patterns;

[0060] Based on the causal relationship mapping, defect identification results are generated.

[0061] By adopting the above technical solution, in order to achieve joint analysis of identification indicators, cluster analysis is performed on the identification indicators of multiple forgings to obtain multiple clusters. Then, based on the distribution characteristics of the clusters, normal forging clusters and at least one abnormal forging cluster are identified. For each abnormal forging cluster, the common features of image differences of forgings within the abnormal forging cluster are extracted to determine the corresponding surface defect type. The common temperature anomaly pattern of forgings within the abnormal forging cluster is also extracted. Then, a causal relationship mapping between surface defect type and temperature anomaly common pattern is established. Finally, based on the causal relationship mapping, defect identification results are generated.

[0062] Optionally, the step of generating the surface migration prediction model includes:

[0063] Obtain a pre-trained deep convolutional neural network model and use the deep convolutional neural network model as the base model, wherein the deep convolutional neural network model is pre-trained on a large-scale general image dataset;

[0064] The network structure of the base model is adjusted to obtain the adjusted base model, wherein the network structure adjustment includes replacing the top classification layer;

[0065] Historical forging images are acquired, and the adjusted base model is trained by transfer learning based on the historical forging image training set to obtain the trained base model. The historical forging image training set contains multiple forging image samples with labeled defect types, and the training process adopts a hierarchical training strategy.

[0066] After training is completed, the output feature vectors of all normal forging samples in the training set are extracted from the last hidden layer of the trained base model, and the mean vector of all normal forging samples is determined according to the output feature vectors to obtain the feature center of normal forgings.

[0067] The model parameters of the trained base model are solidified with the feature centers of the normal forging to form a surface migration prediction model.

[0068] By employing the above technical solution, in order to generate a surface migration prediction model, a pre-trained deep convolutional neural network (DCNN) model is obtained and used as the base model. The DCNN model is pre-trained on a large-scale general image dataset. Then, the network structure of the base model is adjusted to obtain an adjusted base model, including replacing the top classification layer. Next, historical forging images are acquired, and the adjusted base model is trained using transfer learning based on the historical forging image training set. The historical forging image training set contains forging image samples with multiple labeled defect types, and a hierarchical training strategy is used during training. After training, the output feature vectors of all normal forging samples in the training set are extracted from the last hidden layer of the trained base model. The mean vector of all normal forging samples is determined based on the output feature vectors to obtain the normal forging feature centers. Finally, the model parameters of the trained base model and the normal forging feature centers are solidified into a surface migration prediction model.

[0069] Optionally, the step of determining the mean vector value corresponding to all normal forging samples based on the output feature vector to obtain the feature center of the normal forging includes:

[0070] Principal component analysis is performed on each of the output feature vectors to determine the principal component space;

[0071] Each of the output feature vectors is projected onto the principal component space to obtain the corresponding output feature projection vector;

[0072] The mean value of each output feature projection vector is calculated to obtain the feature center of the normal forging.

[0073] By adopting the above technical solution, in order to obtain the feature center of a normal forging, principal component analysis is performed on each output feature vector to determine the principal component space. Then, each output feature vector is projected onto the principal component space to obtain the corresponding output feature projection vector. Finally, the mean of each output feature projection vector is calculated to obtain the feature center of the normal forging.

[0074] Optionally, the hierarchical training strategy includes freezing the bottom-level network parameters, training the newly added classification layer and the top-level network parameters, unfreezing all network parameters, and performing global fine-tuning training with a reduced learning rate.

[0075] Secondly, this application also provides a data processing system for forging defect identification based on surface migration prediction, which adopts the following technical solution:

[0076] A data processing system for forging defect identification based on surface migration prediction includes:

[0077] An initial vector generation module is used to acquire image data and temperature time-series data of multiple forgings in the same batch, determine the image difference vector of a single forging based on the image data and a pre-generated surface migration prediction model, and determine the temperature vector of the single forging based on the temperature time-series data, wherein the temperature time-series data is used to represent the temperature change status of each feature point of the forging during the processing.

[0078] The feature construction module is used to construct an image temperature correlation vector corresponding to the single forging based on the image difference vector and the temperature vector, and to construct an image temperature correlation matrix corresponding to the multiple forgings based on the image temperature correlation matrix;

[0079] The feature space extraction module is used to decompose the image temperature correlation matrix and extract the main feature space;

[0080] The identification index generation module is used to determine the corresponding identification index for each forging based on the corresponding image temperature correlation vector and the main feature space.

[0081] The identification result generation module is used to jointly analyze the identification indicators of the multiple forgings and generate defect identification results. The defect identification results include surface defect types and temperature anomaly patterns that are causally related to the surface defect types.

[0082] In summary, this application includes at least the following beneficial technical effects: acquiring image data and temperature time-series data of multiple forgings in the same batch; determining the image difference vector of a single forging based on the image data and a pre-generated surface migration prediction model; and determining the temperature vector of a single forging based on the temperature time-series data, wherein the temperature time-series data is used to represent the temperature change status of each feature point of the forging during processing; then constructing an image-temperature correlation vector corresponding to a single forging based on the image difference vector and the temperature vector; constructing an image-temperature correlation matrix corresponding to multiple forgings based on the image-temperature correlation matrix; and then decomposing the image-temperature correlation matrix to extract the main feature space. Then, for each forging, the corresponding identification index is determined based on the corresponding image temperature correlation vector and main feature space. The identification indexes of multiple forgings are then jointly analyzed to generate defect identification results. The defect identification results include surface defect types and temperature anomaly patterns that are causally related to the surface defect types. In the above method, on the one hand, temperature data is used to corroborate the image identification results, effectively overcoming the limitation of relying solely on visual features being easily interfered with. On the other hand, process anomalies are traced back through defect features, forming a complete evidence chain of "phenomenon-root cause". Through this cross-modal correlation analysis, the accuracy and reliability of defect identification are improved. Attached Figure Description

[0083] Figure 1 This is a schematic diagram of the overall process of an embodiment of this application.

[0084] Figure 2 This is a schematic diagram of the system structure of this application. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0086] This application discloses a data processing method for identifying defects in forgings based on surface migration prediction.

[0087] Reference Figure 1 A data processing method for forging defect identification based on surface migration prediction includes:

[0088] Step S11: Obtain image data and temperature time series data for multiple forgings in the same batch; determine the image difference vector of a single forging based on the image data and the pre-generated surface migration prediction model; and determine the temperature vector of a single forging based on the temperature time series data.

[0089] Among them, temperature time series data is used to represent the temperature change of various feature points of the forging during the processing.

[0090] It should be noted that in step S11, surface image data of the forging is acquired using an industrial camera, while temperature time-series data during the forging process is collected using sensors such as thermocouples. Step S11 does not directly use the original image, but instead uses a pre-trained surface migration prediction model to analyze the image. The core function of this model is to compare the current forging image with its learned standard model of "normal forgings," thereby outputting a quantized image difference vector. This vector encodes the degree and direction of deviation of the forging's surface features from the normal state, and initially vectorizes the temperature time-series data to generate a temperature vector that represents its key change patterns.

[0091] Step S12: Construct an image temperature correlation vector for a single forging based on the image difference vector and temperature vector, and construct an image temperature correlation matrix for multiple forgings based on the image temperature correlation matrix.

[0092] It should be noted that in step S12, for each forging, the corresponding image difference vector and temperature vector are fused to construct a unified image temperature correlation vector. This fused vector contains both information such as "what anomaly has occurred on the surface" and "what is the temperature process of the forging". Subsequently, these correlation vectors of all forgings in the same batch are combined or spliced ​​to construct an image temperature correlation matrix. This matrix macroscopically depicts the distribution of all products in the "surface-process" joint feature space of the entire production batch.

[0093] Step S13: Decompose the image temperature correlation matrix and extract the main feature space.

[0094] It should be noted that in step S13, the most important change patterns, namely the main feature space, can be extracted from this complex dataset through matrix factorization. This space is composed of several main principal component directions, which represent the dominant "image-temperature" correlation feature patterns shared by the vast majority of forgings in the current production batch. It can be understood as a mathematical representation of the "health status" or "normal production mode" of the entire batch. Any forging that conforms to this mainstream pattern is considered normal.

[0095] Step S14: For each forging, determine the corresponding identification index based on the corresponding image temperature correlation vector and main feature space.

[0096] It should be noted that in step S14, the image temperature correlation vector of each forging is projected into the main feature space obtained in step S13. If a forging is normal, its features should be well interpreted or reconstructed by the main feature space. If it is abnormal, its features will not be effectively represented by the space. By calculating the "reconstruction error" between its original vector and the vector reconstructed in the main feature space, or by calculating its distance metric from the feature space itself, a quantitative identification index can be obtained. The larger the index value, the further the forging deviates from the normal pattern of the batch, and the greater the suspicion of defects.

[0097] Step S15: Perform joint analysis on the identification indicators of multiple forgings to generate defect identification results.

[0098] The defect identification results include surface defect types and temperature anomaly patterns that are causally related to the surface defect types.

[0099] It should be noted that in step S15, the identification indicators for each forging are not considered in isolation. Instead, the indicators of all forgings are jointly analyzed to identify which forgings belong to the normal group and which form an abnormal group. For the identified abnormal group, the system will backtrack and analyze the common image features (to determine the surface defect type) and common temperature patterns (to determine the temperature anomaly pattern) of these forgings. Finally, through stable statistical correlation, a causal relationship mapping is established between specific defect types and specific temperature anomaly patterns. Therefore, the defect identification result not only includes "what kind of defect it is," but also provides "what kind of temperature change caused this defect."

[0100] In the above implementation, image data and temperature time-series data of multiple forgings in the same batch are acquired. Based on the image data and a pre-generated surface migration prediction model, the image difference vector of a single forging is determined, and the temperature vector of a single forging is determined based on the temperature time-series data. The temperature time-series data represents the temperature changes at various feature points of the forging during processing. Then, an image-temperature correlation vector corresponding to a single forging is constructed based on the image difference vector and the temperature vector. An image-temperature correlation matrix corresponding to multiple forgings is constructed based on the image-temperature correlation matrix. The image-temperature correlation matrix is ​​then decomposed to extract the main feature space. For each forging, a corresponding identification index is determined based on the corresponding image-temperature correlation vector and the main feature space. The identification indices of multiple forgings are then jointly analyzed to generate defect identification results. These defect identification results include surface defect types and temperature anomaly patterns that are causally related to the surface defect types. In this method, on the one hand, temperature data is used to corroborate the image identification results, effectively overcoming the limitation of relying solely on visual features being easily interfered with. On the other hand, by tracing back process anomalies through defect features, a complete "phenomenon-root cause" evidence chain is formed. This cross-modal correlation analysis improves the accuracy and reliability of defect identification.

[0101] As a further implementation of the method, the step of constructing an image temperature correlation vector corresponding to a single forging based on the image difference vector and the temperature vector includes:

[0102] Step S21: For each forging, take the forging as the target forging, perform semantic embedding on the image difference vector of the target forging to obtain image difference embedding features, and perform semantic embedding on the temperature vector of the target forging to obtain temperature embedding features.

[0103] Step S22: High-level semantic extraction is performed on the image difference embedding features through the first high-level semantic extraction sub-model to obtain the image difference features.

[0104] Step S23: The temperature embedding features are extracted using the second high-level semantic extraction sub-model to obtain the temperature features.

[0105] Step S24: Using a semantic analysis model, cross-domain semantic encoding is performed on image difference features and temperature features to obtain the corresponding image temperature correlation vector.

[0106] It should be noted that from steps S21 to S24, the image difference vector and temperature vector are mapped to a unified semantic space through semantic embedding, eliminating the semantic gap between different modal data. Then, a specially designed sub-model is used to perform deep semantic extraction on image features, mining discriminative high-level semantic information from surface texture features. For the temporal characteristics of temperature data, key pattern features in the temperature change process are extracted through a dedicated sub-model to capture the deep patterns of process parameters. Then, a semantic bridge is established between the image domain and the temperature domain. Through the encoding mechanism, deep interaction and information complementarity of the two different modal features are realized, thereby achieving deep integration of image surface information and temperature process information, providing a feature representation basis for defect identification.

[0107] In the above implementation, in order to construct the image temperature correlation vector, for each forging, the forging is taken as the target forging. The image difference vector of the target forging is semantically embedded to obtain the image difference embedding feature, and the temperature vector of the target forging is semantically embedded to obtain the temperature embedding feature. Then, the image difference embedding feature is extracted using a first high-level semantic extraction sub-model to obtain the image difference feature. Then, the temperature embedding feature is extracted using a second high-level semantic extraction sub-model to obtain the temperature feature. Finally, the image difference feature and the temperature feature are cross-domain correlated semantically encoded using a semantic analysis model to obtain the corresponding image temperature correlation vector.

[0108] As a further implementation of the method, the step of performing high-level semantic extraction on the image difference embedding features through a first high-level semantic extraction sub-model to obtain the image difference features includes:

[0109] Step S31: The image difference embedding features are fed into the first high-level semantic extraction sub-model. The first high-level semantic extraction sub-model includes a first feature enhancement channel and a second feature enhancement channel set in parallel. The first feature enhancement channel and the second feature enhancement channel are respectively configured with independent dimension transformation layers and feature activation layers.

[0110] Step S32: Perform autocorrelation analysis on the image difference embedding features to extract the image difference context association information.

[0111] Step S33: Perform deep feature transformation on the image difference context association information through the first feature enhancement channel to generate the first image difference deep feature.

[0112] Step S34: Perform deep feature transformation on the image difference context association information through the second feature enhancement channel to generate the second image difference deep feature.

[0113] Step S35: Perform cross-attention integration on the first image difference depth features and the second image difference depth features to obtain the corresponding image difference features.

[0114] It should be noted that from step S31 to step S35, a dual-channel parallel processing architecture is adopted. Through high-level semantic extraction sub-models with different dimensional transformations and activation function combinations, autocorrelation analysis is used to mine the structured information inside the features in order to capture the interdependencies between feature elements. Then, independent feature enhancement channels are used to perform deep transformation of contextual information from different dimensions to generate complementary feature representations. Furthermore, the dual-channel outputs are adaptively fused through an attention mechanism to highlight important features and suppress noise interference, thereby improving the discrimination ability and robustness of image defect features.

[0115] In the above implementation, in order to obtain image difference features, the image difference embedding features are fed into a first high-level semantic extraction sub-model. The first high-level semantic extraction sub-model includes a first feature enhancement channel and a second feature enhancement channel set in parallel. The first feature enhancement channel and the second feature enhancement channel are respectively configured with independent dimension transformation layers and feature activation layers. Then, autocorrelation analysis is performed on the image difference embedding features to extract image difference context association information. Then, deep feature transformation is performed on the image difference context association information through the first feature enhancement channel to generate a first image difference deep feature. Then, deep feature transformation is performed on the image difference context association information through the second feature enhancement channel to generate a second image difference deep feature. Finally, cross-attention integration is performed on the first image difference deep feature and the second image difference deep feature to obtain the corresponding image difference feature.

[0116] As a further implementation of the method, the step of performing high-level semantic extraction on the temperature embedding features through a second high-level semantic extraction sub-model to obtain the temperature features includes:

[0117] Step S41: The temperature embedding feature is fed into the second high-level semantic extraction sub-model, wherein the second high-level semantic extraction sub-model includes a third feature enhancement channel and a fourth feature enhancement channel set in parallel, and the third feature enhancement channel and the fourth feature enhancement channel are respectively configured with independent dimension transformation layer and feature activation layer.

[0118] Step S42: Perform autocorrelation analysis on the temperature embedding features to extract temperature contextual information.

[0119] Step S43: Perform deep feature transformation on the temperature context association information through the third feature enhancement channel to generate the first temperature deep feature.

[0120] Step S44: Perform deep feature transformation on the temperature context association information through the fourth feature enhancement channel to generate the second temperature deep feature.

[0121] Step S45: Perform cross-attention integration on the first temperature depth feature and the second temperature depth feature to obtain the corresponding temperature feature.

[0122] It should be noted that from step S41 to step S45, a dedicated processing architecture is designed for the temporal characteristics of temperature data. The multi-dimensional features of temperature changes are captured through a dual-channel structure. Then, the time-dependent patterns in the temperature sequence are analyzed to uncover the deep-seated patterns of process parameter changes. Subsequently, the temperature context information is deeply transformed from different perspectives to extract temperature feature representations with physical meaning. Finally, the temperature features from multiple perspectives are integrated to form a comprehensive and accurate description of temperature process features.

[0123] In the above implementation, in order to obtain temperature features, the temperature embedding features are fed into a second high-level semantic extraction sub-model. The second high-level semantic extraction sub-model includes a third feature enhancement channel and a fourth feature enhancement channel configured in parallel. The third feature enhancement channel and the fourth feature enhancement channel are respectively configured with independent dimension transformation layers and feature activation layers. Then, autocorrelation analysis is performed on the temperature embedding features to extract temperature context association information. Then, the temperature context association information is subjected to deep feature transformation through the third feature enhancement channel to generate a first temperature deep feature. Then, the temperature context association information is subjected to deep feature transformation through the fourth feature enhancement channel to generate a second temperature deep feature. Finally, the first temperature deep feature and the second temperature deep feature are integrated by cross-attention to obtain the corresponding temperature feature.

[0124] As a further implementation of the method, the step of performing cross-domain semantic encoding on image difference features and temperature features through a semantic analysis model to obtain the corresponding image temperature correlation vector includes:

[0125] Step S51: Load the image difference features and temperature features into the semantic analysis model, wherein the semantic analysis model includes multiple semantic analysis sub-models.

[0126] Step S52: Through each semantic analysis sub-model, the image difference features and temperature features are correlated and mined to obtain the corresponding intermediate correlation features.

[0127] Step S53: Fuse the intermediate correlation features to obtain the image temperature correlation vector.

[0128] It should be noted that from step S51 to step S53, the idea of ​​ensemble learning is adopted. Multiple semantic analysis sub-models provide multiple feature association perspectives. Each sub-model works in parallel and independently, mining the potential association between image and temperature features from different dimensions. Then, the findings of each sub-model are integrated to form comprehensive and robust cross-modal association features, thereby improving the accuracy and stability of feature fusion.

[0129] In the above implementation, in order to obtain the corresponding image temperature correlation vector, the image difference features and temperature features are loaded into the semantic analysis model. The semantic analysis model includes multiple semantic analysis sub-models. Then, the image difference features and temperature features are correlated and mined through each semantic analysis sub-model to obtain the corresponding intermediate correlation features. Then, the intermediate correlation features are fused to obtain the image temperature correlation vector.

[0130] As a further implementation of the method, the step of performing correlation mining on image difference features and temperature features through each semantic analysis sub-model to obtain the corresponding intermediate correlation features includes:

[0131] Step S61: For each semantic analysis sub-model, the semantic analysis sub-model is used as the target semantic analysis sub-model. Image difference features and temperature features are loaded into the target semantic analysis sub-model. The target semantic analysis sub-model has a first projection matrix and a second projection matrix built in. The first projection matrix is ​​used to transform the image difference features from the current semantic space to the target semantic space, and the second projection matrix is ​​used to transform the temperature features from the current semantic space to the target semantic space.

[0132] Step S62: The image difference features are linearly transformed using the first projection matrix to obtain the image difference transformation features.

[0133] Step S63: Perform a linear transformation on the temperature features using the second projection matrix to obtain the temperature transformation features.

[0134] Step S64: Determine the feature correlation score between the image difference transformation feature and the temperature transformation feature, and standardize the feature correlation score to obtain the standardized score.

[0135] Step S65: Weight the image difference transformation features or temperature transformation features according to the standardized scores to obtain the corresponding intermediate correlation features.

[0136] It should be noted that from step S61 to step S65, the features of different modalities are mapped to a unified semantic space through the projection matrix to solve the feature alignment problem. Then, the main structure of the features is preserved by linear transformation, realizing dimensional unification and feature enhancement. Then, the statistical correlation between features is calculated, and adaptive feature fusion is achieved based on the correlation weight to highlight the combination of important features.

[0137] In the above implementation, in order to obtain the corresponding intermediate association features, for each semantic analysis sub-model, the semantic analysis sub-model is used as the target semantic analysis sub-model. Image difference features and temperature features are loaded into the target semantic analysis sub-model. The target semantic analysis sub-model has a built-in first projection matrix and a second projection matrix. The first projection matrix is ​​used to transform the image difference features from the current semantic space to the target semantic space, and the second projection matrix is ​​used to transform the temperature features from the current semantic space to the target semantic space. Then, the image difference features are linearly transformed using the first projection matrix to obtain image difference transformation features. Then, the temperature features are linearly transformed using the second projection matrix to obtain temperature transformation features. Then, the feature correlation score between the image difference transformation features and the temperature transformation features is determined, and the feature correlation score is standardized to obtain a standardized score. Then, the image difference transformation features or the temperature transformation features are weighted according to the standardized score to obtain the corresponding intermediate association features.

[0138] As a further implementation of the method, the step of decomposing the image temperature correlation matrix and extracting the main feature space includes:

[0139] Step S71: Center the image temperature correlation matrix to determine the corresponding covariance matrix.

[0140] Step S72: Perform eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues.

[0141] Step S73: Sort the eigenvalues ​​from largest to smallest, extract the eigenvectors corresponding to the top K eigenvalues ​​as principal component directions, and span the principal feature space by the top K principal component directions, where the value of K is a positive integer determined by the cumulative variance contribution rate.

[0142] In the above implementation, in order to extract the main feature space, the image temperature correlation matrix is ​​centered to determine the corresponding covariance matrix. Then, the covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues. The eigenvalues ​​are then sorted from largest to smallest, and the eigenvectors corresponding to the first K eigenvalues ​​are extracted as principal component directions. The main feature space is spanned by the first K principal component directions, where the value of K is a positive integer determined by the cumulative variance contribution rate.

[0143] As a further implementation of the method, the step of determining the corresponding recognition index based on the corresponding image temperature correlation vector and the main feature space includes:

[0144] Step S81: Project the image temperature correlation vector into the main feature space to obtain the corresponding reconstruction vector.

[0145] Step S82: Determine the reconstruction error between the image temperature correlation vector and the reconstruction vector, and determine the distance metric from the image temperature correlation vector to the main feature space. The reconstruction error is determined by Euclidean distance or cosine similarity, and the distance metric is determined by Mahalanobis distance.

[0146] Step S83: Use reconstruction error and distance metric as identification indicators.

[0147] In the above implementation, in order to determine the recognition index, the image temperature correlation vector is projected into the main feature space to obtain the corresponding reconstruction vector. Then, the reconstruction error between the image temperature correlation vector and the reconstruction vector is determined, and the distance metric from the image temperature correlation vector to the main feature space is determined. The reconstruction error is determined by Euclidean distance or cosine similarity, and the distance metric is determined by Mahalanobis distance. Then, the reconstruction error and the distance metric are used as the recognition index.

[0148] As a further implementation of the method, the step of jointly analyzing the identification indicators of multiple forgings to generate defect identification results includes:

[0149] Step S91: Perform cluster analysis on the identification indicators of multiple forgings to obtain multiple clusters.

[0150] Step S92: Based on the distribution characteristics of the clusters, identify normal forging clusters and at least one abnormal forging cluster.

[0151] Step S93: For each abnormal forging cluster, extract the common features of image differences of the forgings in the abnormal forging cluster, determine the corresponding surface defect type, and extract the common temperature anomaly pattern of the forgings in the abnormal forging cluster.

[0152] Step S94: Establish a causal relationship mapping between surface defect types and common temperature anomaly patterns.

[0153] Step S95: Based on the causal association mapping, generate the defect identification result.

[0154] It should be noted that from step S91 to step S95, natural grouping patterns in the data are automatically discovered through unsupervised learning. Then, normal and abnormal sample groups are distinguished based on statistical characteristics. Common defect patterns and temperature anomalies are mined from the abnormal groups, and a mapping relationship between defect phenomena and process causes is constructed to form a quality analysis report with diagnostic value. This realizes the whole chain of quality diagnosis from anomaly detection to root cause analysis, providing decision support for process optimization and defect identification.

[0155] In the above embodiments, in order to achieve joint analysis of identification indicators, cluster analysis is performed on the identification indicators of multiple forgings to obtain multiple clusters. Then, based on the distribution characteristics of the clusters, normal forging clusters and at least one abnormal forging cluster are identified. For each abnormal forging cluster, the common features of image differences of forgings within the abnormal forging cluster are extracted to determine the corresponding surface defect type. The common temperature anomaly pattern of forgings within the abnormal forging cluster is also extracted. Then, a causal relationship mapping between surface defect type and temperature anomaly pattern is established. Finally, based on the causal relationship mapping, defect identification results are generated.

[0156] As a further implementation of the method, the step of generating the surface migration prediction model includes:

[0157] Step S100: Obtain the pre-trained deep convolutional neural network model and use the deep convolutional neural network model as the base model. The deep convolutional neural network model is pre-trained on a large-scale general image dataset.

[0158] Step S101: Adjust the network structure of the base model to obtain the adjusted base model, wherein the network structure adjustment includes replacing the top classification layer.

[0159] Step S102: Obtain historical forging images, and perform transfer learning training on the adjusted base model based on the historical forging image training set to obtain the trained base model. The historical forging image training set contains forging image samples with multiple labeled defect types, and the training process adopts a hierarchical training strategy.

[0160] It should be noted that the layered training strategy includes freezing the parameters of the bottom layer network, training the parameters of the newly added classification layer and the top layer network, unfreezing all network parameters, and performing global fine-tuning training with a reduced learning rate.

[0161] Step S103: After training is completed, extract the output feature vectors of all normal forging samples in the training set at the last hidden layer of the trained basic model, and determine the mean vector of all normal forging samples based on the output feature vectors to obtain the feature center of normal forgings.

[0162] Step S104: Solidify the model parameters of the trained base model and the feature center of the normal forging into a surface migration prediction model.

[0163] In the above implementation, to generate a surface migration prediction model, a pre-trained deep convolutional neural network (DCNN) model is obtained and used as the base model. The DCNN model is pre-trained on a large-scale general image dataset. The base model's network structure is then adjusted, including replacing the top classification layer. Historical forging images are then acquired, and the adjusted base model is trained using transfer learning based on this historical forging image training set. The historical forging image training set contains forging image samples with multiple labeled defect types, and a hierarchical training strategy is employed. After training, the output feature vectors of all normal forging samples in the training set are extracted from the last hidden layer of the trained base model. The mean vector of all normal forging samples is determined based on the output feature vectors, yielding the normal forging feature centers. Finally, the model parameters of the trained base model and the normal forging feature centers are solidified into a surface migration prediction model.

[0164] As a further implementation of the method, the step of determining the mean vector value corresponding to all normal forging samples based on the output feature vector to obtain the feature center of the normal forging includes:

[0165] Step S111: Perform principal component analysis on each output feature vector to determine the principal component space.

[0166] Step S112: Project each output feature vector onto the principal component space to obtain the corresponding output feature projection vector.

[0167] Step S113: Calculate the mean of each output feature projection vector to obtain the feature center of the normal forging.

[0168] In the above implementation, in order to obtain the feature center of a normal forging, principal component analysis is performed on each output feature vector to determine the principal component space. Then, each output feature vector is projected onto the principal component space to obtain the corresponding output feature projection vector. Finally, the mean of each output feature projection vector is calculated to obtain the feature center of the normal forging.

[0169] This application also discloses a data processing system for identifying forging defects based on surface migration prediction.

[0170] refer to Figure 2 A data processing system for forging defect identification based on surface migration prediction includes:

[0171] The initial vector generation module is used to acquire image data and temperature time-series data of multiple forgings in the same batch. Based on the image data and the pre-generated surface migration prediction model, it determines the image difference vector of a single forging and the temperature vector of a single forging based on the temperature time-series data. The temperature time-series data is used to represent the temperature change of each feature point of the forging during the processing.

[0172] The feature construction module is used to construct an image temperature correlation vector corresponding to a single forging based on the image difference vector and the temperature vector, and to construct an image temperature correlation matrix corresponding to multiple forgings based on the image temperature correlation matrix;

[0173] The feature space extraction module is used to decompose the image temperature correlation matrix and extract the main feature space.

[0174] The identification index generation module is used to determine the corresponding identification index for each forging based on the corresponding image temperature correlation vector and main feature space.

[0175] The identification result generation module is used to jointly analyze the identification indicators of multiple forgings and generate defect identification results. The defect identification results include surface defect types and temperature anomaly patterns that are causally related to the surface defect types.

[0176] The forging defect identification data processing system based on surface migration prediction of the present invention can implement any of the methods in the forging defect identification data processing method based on surface migration prediction, and the specific working process of the forging defect identification data processing system based on surface migration prediction of the present invention can refer to the corresponding process in the above-mentioned forging defect identification data processing method based on surface migration prediction.

[0177] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A data processing method for forging defect identification based on surface migration prediction, characterized in that, include: Image data and temperature time-series data of multiple forgings in the same batch are acquired. Based on the image data and a pre-generated surface migration prediction model, the image difference vector of a single forging is determined, and the temperature vector of the single forging is determined based on the temperature time-series data. The temperature time-series data is used to represent the temperature change of each feature point of the forging during the processing. Based on the image difference vector and the temperature vector, an image temperature correlation vector corresponding to the single forging is constructed, and based on the image temperature correlation vector, an image temperature correlation matrix corresponding to the multiple forgings is constructed. The image temperature correlation matrix is ​​decomposed to extract the main feature space; For each forging, a corresponding identification index is determined based on the corresponding image temperature correlation vector and the main feature space; The identification indicators of the multiple forgings are jointly analyzed to generate defect identification results, wherein the defect identification results include surface defect types and temperature anomaly patterns that are causally related to the surface defect types. The step of constructing the image temperature correlation vector corresponding to the single forging based on the image difference vector and the temperature vector includes: For each forging, the forging is taken as the target forging. The image difference vector of the target forging is semantically embedded to obtain image difference embedding features, and the temperature vector of the target forging is semantically embedded to obtain temperature embedding features. The image difference embedding features are obtained by performing high-level semantic extraction on the image difference embedding features through the first high-level semantic extraction sub-model; The temperature embedding features are obtained by performing high-level semantic extraction on the second high-level semantic extraction sub-model. By using a semantic analysis model, cross-domain semantic encoding is performed on the image difference features and the temperature features to obtain the corresponding image temperature association vector; The step of performing high-level semantic extraction on the image difference embedding features through the first high-level semantic extraction sub-model to obtain image difference features includes: The image difference embedding features are fed into a first high-level semantic extraction sub-model, wherein the first high-level semantic extraction sub-model includes a first feature enhancement channel and a second feature enhancement channel set in parallel, and the first feature enhancement channel and the second feature enhancement channel are respectively configured with independent dimension transformation layers and feature activation layers. Autocorrelation analysis is performed on the image difference embedding features to extract image difference contextual association information; The first image difference deep feature is generated by performing deep feature transformation on the image difference context association information through the first feature enhancement channel; The image difference context association information is transformed into a second image difference depth feature by using the second feature enhancement channel; Cross-attention integration is performed on the first image difference depth feature and the second image difference depth feature to obtain the corresponding image difference feature; The step of performing high-level semantic extraction on the temperature embedding features through the second high-level semantic extraction sub-model to obtain temperature features includes: The temperature embedding feature is fed into the second high-level semantic extraction sub-model, wherein the second high-level semantic extraction sub-model includes a third feature enhancement channel and a fourth feature enhancement channel set in parallel, and the third feature enhancement channel and the fourth feature enhancement channel are respectively configured with independent dimension transformation layer and feature activation layer; Autocorrelation analysis is performed on the temperature embedding features to extract temperature context association information; The temperature context association information is subjected to deep feature transformation through the third feature enhancement channel to generate a first temperature deep feature. The temperature context association information is subjected to deep feature transformation through the fourth feature enhancement channel to generate a second temperature deep feature. Cross-attention integration is performed on the first temperature depth feature and the second temperature depth feature to obtain the corresponding temperature feature.

2. The forging defect identification data processing method based on surface migration prediction according to claim 1, characterized in that, The step of performing cross-domain semantic encoding on the image difference features and the temperature features using a semantic analysis model to obtain the corresponding image temperature association vector includes: The image difference features and the temperature features are loaded into the semantic analysis model, wherein the semantic analysis model includes multiple semantic analysis sub-models; The image difference features and temperature features are correlated and mined using each of the semantic analysis sub-models to obtain the corresponding intermediate correlation features; The intermediate correlation features are fused to obtain the image temperature correlation vector.

3. The forging defect identification data processing method based on surface migration prediction according to claim 2, characterized in that, The step of performing correlation mining on the image difference features and the temperature features through each of the semantic analysis sub-models to obtain the corresponding intermediate correlation features includes: For each of the semantic analysis sub-models, the semantic analysis sub-model is used as the target semantic analysis sub-model, and the image difference features and the temperature features are loaded into the target semantic analysis sub-model. The target semantic analysis sub-model has a built-in first projection matrix and a second projection matrix. The first projection matrix is ​​used to transform the image difference features from the current semantic space to the target semantic space, and the second projection matrix is ​​used to transform the temperature features from the current semantic space to the target semantic space. The image difference features are obtained by performing a linear transformation on the image difference features using the first projection matrix; The temperature feature is linearly transformed using the second projection matrix to obtain the temperature transformation feature; Determine the feature correlation score between the image difference transformation feature and the temperature transformation feature, and standardize the feature correlation score to obtain a standardized score; The image difference transformation feature or the temperature transformation feature is weighted according to the standardized score to obtain the corresponding intermediate correlation feature.

4. The forging defect identification data processing method based on surface migration prediction according to claim 1, characterized in that, The step of decomposing the image temperature correlation matrix and extracting the main feature space includes: The image temperature correlation matrix is ​​centered to determine the corresponding covariance matrix; The covariance matrix is ​​decomposed into eigenvalues ​​to obtain multiple eigenvalues; The eigenvalues ​​are sorted from largest to smallest, and the eigenvectors corresponding to the top K eigenvalues ​​are extracted as principal component directions. The top K principal component directions span the principal feature space, where the value of K is a positive integer determined by the cumulative variance contribution rate.

5. The forging defect identification data processing method based on surface migration prediction according to claim 1, characterized in that, The step of determining the corresponding recognition index based on the corresponding image temperature correlation vector and the main feature space includes: The image temperature correlation vector is projected onto the main feature space to obtain the corresponding reconstructed vector; The reconstruction error between the image temperature correlation vector and the reconstruction vector is determined, and the distance metric from the image temperature correlation vector to the main feature space is determined, wherein the reconstruction error is determined by Euclidean distance or cosine similarity, and the distance metric is determined by Mahalanobis distance. The reconstruction error and the distance metric are used as identification indicators.

6. The forging defect identification data processing method based on surface migration prediction according to claim 1, characterized in that, The step of jointly analyzing the identification indicators of the multiple forgings to generate defect identification results includes: Cluster analysis was performed on the identification indicators of the multiple forgings to obtain multiple clusters; Based on the distribution characteristics of the clusters, identify normal forging clusters and at least one abnormal forging cluster; For each abnormal forging cluster, extract the common features of image differences of the forgings in the abnormal forging cluster, determine the corresponding surface defect type, and extract the common temperature anomaly pattern of the forgings in the abnormal forging cluster. Establish a causal relationship mapping between the surface defect types and the common temperature anomaly patterns; Based on the causal relationship mapping, defect identification results are generated.

7. The forging defect identification data processing method based on surface migration prediction according to claim 1, characterized in that, The steps for generating the surface migration prediction model include: Obtain a pre-trained deep convolutional neural network model and use the deep convolutional neural network model as the base model, wherein the deep convolutional neural network model is pre-trained on a large-scale general image dataset; The network structure of the base model is adjusted to obtain the adjusted base model, wherein the network structure adjustment includes replacing the top classification layer; Historical forging images are acquired, and the adjusted base model is trained by transfer learning based on the historical forging image training set to obtain the trained base model. The historical forging image training set contains multiple forging image samples with labeled defect types, and the training process adopts a hierarchical training strategy. After training is completed, the output feature vectors of all normal forging samples in the training set are extracted from the last hidden layer of the trained base model, and the mean vector of all normal forging samples is determined according to the output feature vectors to obtain the feature center of normal forgings. The model parameters of the trained base model are solidified with the feature centers of the normal forging to form a surface migration prediction model.

8. The forging defect identification data processing method based on surface migration prediction according to claim 7, characterized in that, The step of determining the mean vector value corresponding to all normal forging samples based on the output feature vector to obtain the feature center of the normal forging includes: Principal component analysis is performed on each of the output feature vectors to determine the principal component space; Each of the output feature vectors is projected onto the principal component space to obtain the corresponding output feature projection vector; The mean value of each output feature projection vector is calculated to obtain the feature center of the normal forging.

9. The forging defect identification data processing method based on surface migration prediction according to claim 7, characterized in that, The hierarchical training strategy includes freezing the bottom layer network parameters, training the newly added classification layer and the top layer network parameters, unfreezing all network parameters, and performing global fine-tuning training with a reduced learning rate.

10. A forging defect identification data processing system based on surface migration prediction, characterized in that, include: An initial vector generation module is used to acquire image data and temperature time-series data of multiple forgings in the same batch, determine the image difference vector of a single forging based on the image data and a pre-generated surface migration prediction model, and determine the temperature vector of the single forging based on the temperature time-series data, wherein the temperature time-series data is used to represent the temperature change status of each feature point of the forging during the processing. The feature construction module is used to construct an image temperature correlation vector corresponding to the single forging based on the image difference vector and the temperature vector, and to construct an image temperature correlation matrix corresponding to the multiple forgings based on the image temperature correlation vector; The feature space extraction module is used to decompose the image temperature correlation matrix and extract the main feature space; The identification index generation module is used to determine the corresponding identification index for each forging based on the corresponding image temperature correlation vector and the main feature space. The identification result generation module is used to perform joint analysis on the identification indicators of the multiple forgings and generate defect identification results, wherein the defect identification results include surface defect types and temperature anomaly patterns that are causally related to the surface defect types. The step of constructing the image temperature correlation vector corresponding to the single forging based on the image difference vector and the temperature vector includes: For each forging, the forging is taken as the target forging. The image difference vector of the target forging is semantically embedded to obtain image difference embedding features, and the temperature vector of the target forging is semantically embedded to obtain temperature embedding features. The image difference embedding features are obtained by performing high-level semantic extraction on the image difference embedding features through the first high-level semantic extraction sub-model; The temperature embedding features are obtained by performing high-level semantic extraction on the second high-level semantic extraction sub-model. By using a semantic analysis model, cross-domain semantic encoding is performed on the image difference features and the temperature features to obtain the corresponding image temperature association vector; The step of performing high-level semantic extraction on the image difference embedding features through the first high-level semantic extraction sub-model to obtain image difference features includes: The image difference embedding features are fed into a first high-level semantic extraction sub-model, wherein the first high-level semantic extraction sub-model includes a first feature enhancement channel and a second feature enhancement channel set in parallel, and the first feature enhancement channel and the second feature enhancement channel are respectively configured with independent dimension transformation layers and feature activation layers. Autocorrelation analysis is performed on the image difference embedding features to extract image difference contextual association information; The first image difference deep feature is generated by performing deep feature transformation on the image difference context association information through the first feature enhancement channel; The image difference context association information is transformed into a second image difference depth feature by using the second feature enhancement channel; Cross-attention integration is performed on the first image difference depth feature and the second image difference depth feature to obtain the corresponding image difference feature; The step of performing high-level semantic extraction on the temperature embedding features through the second high-level semantic extraction sub-model to obtain temperature features includes: The temperature embedding feature is fed into the second high-level semantic extraction sub-model, wherein the second high-level semantic extraction sub-model includes a third feature enhancement channel and a fourth feature enhancement channel set in parallel, and the third feature enhancement channel and the fourth feature enhancement channel are respectively configured with independent dimension transformation layer and feature activation layer; Autocorrelation analysis is performed on the temperature embedding features to extract temperature context association information; The temperature context association information is subjected to deep feature transformation through the third feature enhancement channel to generate a first temperature deep feature. The temperature context association information is subjected to deep feature transformation through the fourth feature enhancement channel to generate a second temperature deep feature. Cross-attention integration is performed on the first temperature depth feature and the second temperature depth feature to obtain the corresponding temperature feature.

Images