A deep learning-based casting internal defect detection method

Abstract

The present application belongs to the field of full-process automation and intellectualization of casting production, and provides a casting internal defect detection method based on deep learning, which comprises: image shooting for detection, HDR image enhancement processing, one-key image enhancement processing and mask image fusion, industrial X-ray AI network detection, post-processing operation and data statistics. The present application avoids the defects of relying on personnel experience and poor precision in conventional defect detection by constructing an industrial X-ray AI network. Through HDR image enhancement processing, one-key image enhancement processing and mask image fusion, the influence of the complex structure of the detection device, the defects of the shooting equipment itself and the noise points is reduced.

Description

Technical Field

[0001] This invention relates to the field of automation and intelligentization of the entire casting production process, and in particular to a method for detecting internal defects in castings based on deep learning. Background Technology

[0002] The metal industry is one of the important basic industries of manufacturing. Complex metal components are often formed as a whole using precision casting. During the casting process, defects such as shrinkage cavities, porosity, inclusions, and cracks often appear inside the castings. Hot isostatic pressing and welding have limited ability to eliminate these internal defects. Therefore, non-destructive testing of castings based on X-ray images has become a necessary process to ensure product quality.

[0003] However, traditional X-ray defect detection in castings, or visual inspection of casting X-ray images, is easily affected by the skill level and physical and mental state of the inspectors. This makes it difficult to guarantee the efficiency, accuracy, reliability, and stability of the inspection. Furthermore, the storage and retrieval of inspection information is extremely time-consuming and labor-intensive, severely hindering the development of automation and intelligence in the entire casting production process. X-ray image-based non-destructive testing has become a widely used inspection technology. However, in traditional X-ray imaging techniques, the clarity of the obtained images is usually limited due to the different materials of the inspected workpiece, the complexity of the structure, the superposition of the workpiece's structure and scattered X-rays, defects in the imaging equipment itself, and the influence of noise. This often results in low contrast, blurred key features, and other drawbacks. Even if the image contains rich information, it is difficult to observe directly with the human eye. Traditional visible light imaging can form a massive digital image database, and its acquisition and collection methods are relatively convenient and fast. However, acquiring X-ray images of castings requires dedicated X-ray inspection equipment, and the diverse types of castings make collection and acquisition more difficult. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a deep learning-based method for detecting internal defects in castings. By constructing an industrial X-ray AI network, it avoids the reliance on human experience and poor accuracy in conventional defect detection. Through HDR image enhancement processing, one-click image enhancement processing, and mask image fusion, it reduces the impact of complex detection device structures, defects in the imaging equipment itself, and noise.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A deep learning-based method for detecting internal defects in castings, comprising:

[0007] Acquire images of the casting workpiece to be inspected to obtain a 16-bit X-ray image of the workpiece to be inspected;

[0008] The 16-bit X-ray image to be detected is converted into an 8-bit image using HDR image enhancement processing to obtain the preprocessed image to be detected;

[0009] The image to be trained and preprocessed is subjected to one-click image enhancement processing and fused with a mask image generated according to the ROI detection region parameters set by the user to obtain the preprocessed data to be detected.

[0010] The pre-processed data to be detected is input into a pre-trained industrial X-ray AI network for detection to obtain detection data;

[0011] The detection data is post-processed to obtain post-processed data; the post-processing operations include: original image defect display and image prediction;

[0012] The post-processed data is statistically analyzed to obtain the output results, and the defect data in the output results is saved according to the file type.

[0013] The training process of the industrial X-ray AI network includes:

[0014] Acquire existing images of casting workpieces to obtain 16-bit X-ray images to be trained;

[0015] The 16-bit X-ray image to be trained is converted into an 8-bit image using HDR image enhancement processing to obtain the preprocessed image to be trained.

[0016] The preprocessed image to be trained is subjected to one-click image enhancement processing to obtain preprocessed data to be trained; the one-click image enhancement processing includes: brightness adjustment, contrast adjustment, filtering, and sharpening;

[0017] The preprocessed data to be trained is used to generate images using an AI network generation algorithm based on GAN network to obtain generated data;

[0018] The preprocessed data to be trained and the generated data are combined and subjected to super-resolution enhancement operation based on the SR algorithm to obtain SR data;

[0019] The preprocessed data to be trained and the generated data are labeled and shared and paired respectively, and the SR data is labeled with defects to obtain the original paired dataset in binary image format;

[0020] The original paired dataset is subjected to data augmentation operations to obtain a generalized dataset; the data augmentation operations include: rotation, scaling, cropping, and noise addition.

[0021] The generalization dataset is randomly divided into a validation dataset and a training dataset according to a preset matching ratio.

[0022] The pre-trained dataset obtained by combining the validation dataset and the training dataset is input into the W-Net network and the model is trained using a GPU to obtain the trained industrial X-ray AI network.

[0023] Preferably, the objects identified by the industrial X-ray AI network include: pores, inclusions, looseness, cracks, and shrinkage cavities.

[0024] Preferably, the parameter control method of the industrial X-ray AI network includes: regularization technology.

[0025] Preferably, it further includes:

[0026] When process parameters change, the defect data is updated, and the updated defect data is used to retrain the industrial X-ray AI network.

[0027] Preferably, it further includes:

[0028] When the performance indicators of the monitored data exceed the preset warning threshold, a warning message is sent to the target object.

[0029] Preferably, it further includes:

[0030] Introduce an anomaly detection mechanism and a feedback loop mechanism;

[0031] The output results are analyzed a second time using the aforementioned anomaly detection mechanism to obtain defect omission data;

[0032] The feedback loop mechanism is used to collect feedback data from target users, and the feedback data is used to optimize the industrial X-ray AI network.

[0033] The present invention discloses the following technical effects:

[0034] This invention provides a deep learning-based method for detecting internal defects in castings. By constructing an industrial X-ray AI network, it solves the problems of reliance on human experience and poor accuracy in conventional defect detection, and achieves precise detection of defect locations on target detection devices. Through HDR image enhancement processing, one-click image enhancement processing, and mask image fusion, it solves the problems of complex detection device structure, defects in the imaging equipment itself, and noise, and achieves enhancement of defect areas and image clarity. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A schematic diagram of the deep learning-based casting internal defect detection process provided in an embodiment of the present invention;

[0037] Figure 2 This is a flowchart of the AI ​​model file training process provided in an embodiment of the present invention;

[0038] Figure 3 This is a flowchart of the AI ​​model detection process provided in an embodiment of the present invention;

[0039] Figure 4 is a schematic diagram of the network structure provided in the embodiment of the present invention. Figure 4(a) is the first sub-figure and Figure 4(b) is the second sub-figure. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The purpose of this invention is to provide a deep learning-based method for detecting internal defects in castings. By constructing an industrial X-ray AI network, it avoids the reliance on human experience and poor accuracy in conventional defect detection. Through HDR image enhancement processing, one-click image enhancement processing, and mask image fusion, it reduces the impact of complex detection device structures, defects in the imaging equipment itself, and noise.

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Figure 1 This is a schematic diagram of a deep learning-based casting internal defect detection process provided in an embodiment of the present invention. Figure 2 This is a flowchart of the AI ​​model file training process provided in an embodiment of the present invention. Figure 3 The AI ​​model detection flowchart provided in the embodiments of the present invention is as follows: Figure 1 , Figure 2 as well as Figure 3 As shown, this invention provides a deep learning-based method for detecting internal defects in castings, comprising:

[0044] Step 100: Acquire an image of the casting workpiece to be inspected to obtain a 16-bit X-ray image of the workpiece to be inspected;

[0045] Step 200: Use HDR image enhancement processing to convert the 16-bit X-ray image to be detected into an 8-bit image to obtain the preprocessed image to be detected;

[0046] Step 300: Perform one-click image enhancement processing on the preprocessed image to be trained and fuse it with the mask image generated according to the ROI detection region parameters set by the user to obtain the preprocessed data to be detected;

[0047] Step 400: Input the pre-processed data to be detected into the pre-trained industrial X-ray AI network for detection to obtain detection data;

[0048] Step 500: Perform post-processing operations on the detection data to obtain post-processed data; post-processing operations include: original image defect display and image prediction;

[0049] Step 600: Perform data statistics on the post-processed data to obtain the output results, and save the defect data in the output results according to the file type; the output results include: XML files, JSON files, Excel files, YAML files, etc.

[0050] The training process for an industrial X-ray AI network includes:

[0051] Acquire existing images of casting workpieces to obtain 16-bit X-ray images to be trained;

[0052] The 16-bit X-ray image to be trained is converted into an 8-bit image using HDR image enhancement processing to obtain the preprocessed image to be trained.

[0053] One-click image enhancement processing is performed on the preprocessed image to be trained, resulting in preprocessed data for training. One-click image enhancement processing includes: brightness adjustment, contrast adjustment, filtering, and sharpening.

[0054] An AI network generation algorithm based on GAN network is used to generate images from the preprocessed data to be trained, resulting in generated data.

[0055] The pre-processed data and generated data are combined and super-resolution enhancement operations based on the SR algorithm are performed to obtain SR data;

[0056] The pre-processed data and generated data were labeled and shared and paired separately, and the SR data were labeled with defects to obtain the original paired dataset in binary image format.

[0057] Data augmentation operations are performed on the original paired dataset to obtain a generalized dataset; data augmentation operations include rotation, scaling, cropping, and noise addition.

[0058] The generalization dataset is randomly divided into a validation dataset and a training dataset according to a preset matching ratio.

[0059] The pre-trained dataset, obtained by combining the validation dataset and the training dataset, is input into the W-Net network for model training using a GPU, resulting in a trained industrial X-ray AI network.

[0060] Specifically, the objects identified by the industrial X-ray AI network include: pores, inclusions, looseness, cracks, and shrinkage cavities.

[0061] Preferably, the parameter control method for industrial X-ray AI networks includes: regularization technology.

[0062] Optionally, it also includes:

[0063] When process parameters change, the defect data is updated, and the updated defect data is used to retrain the industrial X-ray AI network.

[0064] Preferably, it further includes:

[0065] When the performance indicators of the monitored data exceed the preset warning threshold, an early warning message is sent to the target object.

[0066] Furthermore, it also includes:

[0067] Introduce an anomaly detection mechanism and a feedback loop mechanism;

[0068] The output results are analyzed a second time using an anomaly detection mechanism to obtain defect omission data;

[0069] Feedback data from target users is collected using a feedback loop mechanism, and the feedback data is then used to optimize the industrial X-ray AI network.

[0070] Specifically, the preprocessing method for casting image data acquisition is as follows: First, the acquired 16-bit X-ray imaging image is processed using a high dynamic range enhancement preprocessing algorithm (HDR algorithm) to display the detailed information of the defect location in the image; second, to address the problem of image noise introduced during HDR algorithm enhancement, super-resolution image enhancement technology (SR) is used to effectively reduce image noise while enhancing image clarity; finally, the 16-bit image is converted into an 8-bit image using an image normalization method and saved for subsequent image processing operations and to improve image processing speed.

[0071] Further, the process of establishing the training and validation sets for casting image data is as follows: First, the pre-processed casting workpiece images are manually generalized (due to the scarcity of industrial image data, data generalization is necessary to increase the number of defective images). This operation is mainly achieved through a virtual generation algorithm constructed using generative adversarial networks in AI technology. Then, the original pre-processed casting workpiece images and the image data generated by the virtual generation algorithm are merged. Finally, the dataset composed of the merged image data is labeled with image defects using ImageJ software. The labeled images are binary reference images. The labeled dataset is then divided into training sample datasets, validation sample datasets, and test sample datasets according to actual needs.

[0072] Preferably, the core network algorithm design of the AI ​​industrial network model for detecting internal defects in castings produced by different processes (the scope of casting identification includes defects such as porosity, inclusions, looseness, cracks, and shrinkage cavities) is shown in Figures 4(a) and 4(b). The designed AI industrial network model mainly utilizes skip connections and residual connections to perform deep feature extraction of casting images through three cascaded U-shaped networks. Simultaneously, channel attention and spatial attention mechanisms are integrated during the skip connection process to address the problem of small defects being easily missed. Furthermore, multiple activation functions are combined to achieve more accurate X-ray image segmentation of castings, increasing the network's fitting ability and improving its robustness and accuracy. The specific implementation process is as follows: First, a first U-shaped network extracts low-frequency feature information from the casting image, mainly involving four downsampling operations and four upsampling operations. Second, a second U-shaped network captures mid-frequency information from the casting image, also involving four downsampling operations and four upsampling operations. To avoid increasing network parameters due to continuous stacking, bridging and residual connections are used to cascade the two networks. Finally, a third U-shaped network is used to obtain fine image details, i.e., high-frequency component information, also involving four downsampling operations and four upsampling operations. These three U-shaped networks can extract different features from the image at different frequency levels, thus accelerating the convergence speed of the entire model network. Furthermore, channel attention and spatial attention mechanisms are integrated during the bridging and residual connections to improve the image segmentation accuracy of the entire network model.

[0073] Furthermore, the post-processing of casting image data is as follows: First, to improve the detection rate of defect locations in casting images while reducing the false detection rate, a mask image is created to mask the areas (for example, to prevent the background from being treated as a defect point, a mask image template can be created to mask the background area of ​​the image, or this method can also be used to mask areas that are frequently falsely detected). Then, the binary image predicted by the AI ​​industrial network model is combined with the mask image and annotated on the original image (usually in red or yellow). Finally, the defect area or size can be further processed according to the parameter settings in the XML or JSON file (the parameter values ​​are generally determined by the user). In this way, based on the set threshold parameters, it can be finally determined which of the annotated areas belong to OK and which belong to NG. After that, the statistical results are written into the XML or JSON file, and the annotated defect image is output.

[0074] Preferably, in the later stages of ADR algorithm optimization, when image collection becomes relatively difficult, improving ADR accuracy can be achieved through various strategies and techniques:

[0075] (a) Data augmentation and composition:

[0076] Data augmentation: New training samples are generated by performing operations such as rotation, flipping, scaling, cropping, and color transformation on existing images. This not only increases the diversity of the dataset but also helps the model generalize better to different situations.

[0077] Synthetic data: This method utilizes techniques such as Generative Adversarial Networks (GANs) to generate new, high-quality image data from a small number of real samples. This approach can significantly expand the dataset without increasing the actual collection cost.

[0078] (b) Transfer learning:

[0079] Leveraging pre-trained models: Starting with a model pre-trained on a large dataset (such as one pre-trained on ImageNet), fine-tuning it to adapt to a task-specific dataset. This allows you to utilize the powerful feature representations already learned by the pre-trained model, improving accuracy on new tasks.

[0080] (c) Semi-supervised and unsupervised learning:

[0081] Semi-supervised learning: This method utilizes partially labeled data and a large amount of unlabeled data for training. By combining supervised and unsupervised learning methods, useful information can be extracted from unlabeled data to assist in model training.

[0082] Unsupervised learning: Through unsupervised learning methods such as clustering and autoencoders, potential structures and features are extracted from unlabeled data, and then these features are used to improve the performance of the model.

[0083] (d) Model compression and distillation: Model compression: By using techniques such as pruning and quantization, the size of the model is reduced while maintaining or improving its performance. This helps to train the model better on a limited dataset.

[0084] Knowledge distillation: Distilling knowledge from large, complex models into small, simple models. In this way, small models with good performance can be trained on smaller datasets.

[0085] (e) Regularization and optimization algorithms:

[0086] Regularization: Using L1, L2 regularization, or other forms of regularization techniques prevents the model from overfitting. This helps improve the model's generalization ability on limited datasets.

[0087] Optimization algorithm: Select appropriate optimization algorithms (such as Adam, RMSprop, etc.) and learning rate scheduling strategies to improve the training efficiency and convergence of the model.

[0088] Ensemble learning and model fusion:

[0089] (f) Ensemble learning: Train multiple different models and then combine their predictions to improve overall accuracy.

[0090] Model fusion: This involves combining the outputs of different models (such as CNN, RNN, etc.) to fully utilize their respective advantages and improve overall performance.

[0091] Therefore, even in the later stages of ADR algorithm optimization, when image collection is difficult, the accuracy of ADR can still be improved through various strategies and techniques. The key is to make full use of existing resources, leverage the advantages of deep learning models, and conduct targeted optimizations and adjustments based on the specific problem.

[0092] Furthermore, when process parameters change, the ADR algorithm may indeed encounter defects that were previously unseen or less frequent, potentially increasing the likelihood of missed detections. To ensure the accuracy of ADR in such situations, the following strategies can be adopted:

[0093] (a) Real-time data updates and retraining:

[0094] When process parameters change, new defect data should be collected as soon as possible, and the AI ​​model should be retrained. This ensures that the model can adapt to new environments and situations, reducing the chance of missed detections.

[0095] To maintain the real-time performance and accuracy of the model, a continuous learning mechanism can be established, enabling the model to continuously learn and evolve from new data.

[0096] (b) Enhanced model robustness:

[0097] During model training, it is important to focus on improving the model's robustness, making it more adaptable to changes in input data. Robustness can be enhanced through techniques such as data augmentation, regularization, and model ensemble.

[0098] (c) Threshold adjustment and confidence analysis:

[0099] Based on the new data distribution, adjust the classification or detection thresholds to ensure the model can more accurately identify new defects. For each prediction, calculate its confidence level and decide whether to accept the prediction based on the confidence level. Predictions with low confidence levels can undergo further manual review or analysis.

[0100] (d) Monitoring and early warning system:

[0101] Establish a monitoring and early warning system to monitor the model's performance and prediction results in real time. When frequent missed or false positives are detected, an early warning should be issued and appropriate measures taken promptly. The early warning function can be implemented by setting thresholds for performance metrics (such as accuracy and recall).

[0102] (e) Integration of expert knowledge and rules:

[0103] Incorporating expert knowledge and experience into the model can be achieved through rule engines or hybrid models. This can, to some extent, compensate for the model's shortcomings when dealing with new situations.

[0104] By collaborating with experts, we developed identification rules and strategies for new defects and integrated them into the ADR system.

[0105] (f) Anomaly detection and feedback loop:

[0106] An anomaly detection mechanism is introduced to identify data that significantly deviates from the normal pattern. When an anomaly is detected, further investigation and analysis can be triggered to determine if new defects exist. A feedback loop mechanism is established, enabling the system to continuously learn and improve from real-world applications. By collecting user feedback and new data, the model's performance and accuracy are continuously optimized.

[0107] Therefore, by employing methods such as real-time data updates, enhanced model robustness, threshold adjustment, monitoring and early warning, expert knowledge fusion, and anomaly detection and feedback loops, the probability of missed detections in ADRs after changes in process parameters can be effectively reduced, ensuring the accuracy of ADRs. At the same time, these methods also need to be flexibly adjusted and optimized based on actual application scenarios.

[0108] Optionally, when the ADR algorithm fails to accurately identify certain situations, some strategies and technical means can still be adopted to minimize this risk and improve the accuracy of the AI ​​algorithm.

[0109] (a) In-depth analysis of known defects:

[0110] Carefully studying major flaws that AI algorithms fail to detect, and understanding their characteristics and causes, helps identify potential blind spots and limitations in the algorithms. Analyzing these cases allows for the extraction of useful information to improve algorithms or design new feature representation methods.

[0111] (b) Utilizing unsupervised learning: Since labeled data is scarce, unsupervised learning methods can be considered to uncover latent structures and features within the data. For example, clustering algorithms can be used to discover hidden patterns in the data that may be related to defects.

[0112] Unsupervised learning can help algorithms better understand the distribution and characteristics of data, thereby improving their ability to identify new defects.

[0113] (c) Combining expert knowledge and rules:

[0114] Collaborating with domain experts allows their knowledge and experience to be translated into rules and constraints within algorithms. This helps address data insufficiency and improves the algorithm's ability to identify specific defects. By incorporating expert knowledge, algorithms can be provided with more comprehensive guidance, enabling them to make more accurate judgments when faced with complex and rare defects.

[0115] (d) Using transfer learning and pre-trained models:

[0116] Transfer learning allows knowledge learned from other tasks or domains to be transferred to the current task. It's advisable to use a model pre-trained on a related task as a starting point, fine-tuning it to adapt to the current data and task. Pre-trained models typically have already learned a large number of general feature representations, which helps improve performance on the current task.

[0117] (e) Anomaly detection and alarm mechanism:

[0118] Design an anomaly detection mechanism to monitor the algorithm's output and detect potential anomalies or outliers. When the algorithm's output deviates from expectations, an alarm mechanism can be triggered to promptly identify and address potential defects.

[0119] By combining statistical methods and machine learning techniques, anomaly detection can be effectively achieved, and the risk of missed detection can be reduced.

[0120] (f) Continuous monitoring and feedback loop:

[0121] Establish a continuous monitoring and feedback loop mechanism to collect and process new defect data in a timely manner. Even if it cannot be used for training immediately, this data can be used for analysis and algorithm improvement. By continuously collecting feedback and updating the algorithm, its ability to identify rare and exceptionally large defects can be gradually improved.

[0122] Because data collection is particularly difficult, these optimization strategies may not completely eliminate the risk of missed detections. Therefore, in practical applications, it is necessary to develop appropriate solutions based on specific circumstances and needs, and to comprehensively consider factors such as the accuracy, reliability, and efficiency of the algorithm.

[0123] The beneficial effects of this invention are as follows:

[0124] This invention avoids the shortcomings of conventional defect detection, which relies on human experience and has poor accuracy, by constructing an industrial X-ray AI network; and reduces the impact of complex detection device structure, defects in the imaging equipment itself, and noise through HDR image enhancement processing, one-click image enhancement processing, and mask image fusion.

[0125] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0126] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1.A method for detecting internal defects of a casting based on deep learning, characterized by, The method comprises the following steps: Collecting a cast workpiece image to be detected to obtain a 16-bit X-ray image to be detected; Converting the 16-bit X-ray image to be detected into an 8-bit image by using HDR image enhancement processing to obtain a pretreated image to be detected; Performing one-key image enhancement processing on the pretreated image to be detected and fusing a mask image generated according to a user-set ROI detection region parameter to obtain pretreated data to be detected; Inputting the pretreated data to be detected into a pre-trained industrial X-ray AI network for detection to obtain detection data; Performing post-processing operations on the detection data to obtain post-processed data; the post-processing operations comprise original image defect display and image prediction; Performing data statistics on the post-processed data to obtain output results, and saving defect data in the output results according to file types; The training process of the industrial X-ray AI network comprises the following steps: Collecting existing cast workpiece images to obtain 16-bit X-ray images to be trained; Converting the 16-bit X-ray images to be trained into 8-bit images by using HDR image enhancement processing to obtain pretreated images to be trained; Performing one-key image enhancement processing on the pretreated images to be trained to obtain pretreated data to be trained; the one-key image enhancement processing comprises brightness adjustment, contrast adjustment, filtering and sharpening; Performing image generation on the pretreated data to be trained by using an AI network generation algorithm based on a GAN network to obtain generated data; Performing combination and super-resolution enhancement operations based on a SR algorithm on the pretreated data to be trained and the generated data to obtain SR data; Performing label data sharing and combination pairing on the pretreated data to be trained and the generated data respectively, and performing defect labeling on the SR data to obtain original paired data sets in a binary graph format; Performing data enhancement operations on the original paired data sets to obtain generalized data sets; the data enhancement operations comprise rotation, scaling, shearing and noise addition; Randomly dividing the generalized data sets into a verification data set and a training data set according to a preset pairing ratio; Inputting a pre-trained data set obtained by combining the verification data set and the training data set into a W-Net network to perform model training by using a GPU to obtain the trained industrial X-ray AI network. The AI industrial network model uses a method of jump and residual connection, extracts the features of the casting image in depth through cascading three U-shaped networks, fuses channel attention and spatial attention mechanisms in the process of jump connection, and combines a variety of activation functions; the specific implementation process is as follows: first, the low-frequency feature information of the casting image is extracted through the first U-shaped network, which involves 4 down-sampling operations and 4 up-sampling operations; second, the medium-frequency information of the casting image is captured using the second U-shaped network, which also involves 4 down-sampling operations and 4 up-sampling operations, and in order to avoid the problem of increasing network parameters caused by continuous stacking, a bridging and residual connection method is used to cascade the two networks; finally, the third U-shaped network is used to mainly obtain the fine image detail part, i.e. high-frequency component information, which also involves 4 down-sampling and 4 up-sampling operations; The identification object of the industrial X-ray AI network includes: pores, slag inclusion, loose, cracks and shrinkage. 2.The method of claim 1, wherein, The parameter regulation method of the industrial X-ray AI network includes: regularization technology. 3.The method of claim 1, wherein, Further comprising: When the process parameters change, update the defect data, and retrain the industrial X-ray AI network using the updated defect data. 4.The method of claim 1, wherein, Further comprising: When the performance index of the monitoring data exceeds the preset warning threshold, send a warning message to the target object. 5.The method of claim 1, wherein, Further comprising: Introducing an anomaly detection mechanism and a feedback loop mechanism; Using the anomaly detection mechanism to perform secondary analysis on the output results to obtain defect omission data; Using the feedback loop mechanism to collect feedback data of the target user, and using the feedback data to optimize the industrial X-ray AI network.

Images