Quality prediction method for laser melt injection composite coating based on multi-stage hybrid fusion

Abstract

The present application relates to a kind of quality prediction methods of laser melt injection composite coating based on multi-level hybrid fusion.The prediction method first establishes the laser melt injection process map of ceramic reinforced metal matrix composite coating, simultaneously collects the paired image data of infrared camera image and high-speed camera image of composite coating under multiple laser melt injection process parameters, and corresponding paired image data is grouped according to process map, then laser melt injection multi-level hybrid fusion framework combining data layer and feature layer is constructed, finally laser melt injection quality prediction model is trained based on paired image data and multi-level hybrid fusion framework, and the state of ceramic reinforced metal matrix composite coating in laser melt injection process is predicted in real time using the trained quality prediction model.The prediction method can integrate and fuse complementary molten pool and splash features to realize the quality prediction of laser melt injection ceramic reinforced metal matrix composite coating, and significantly improve the prediction accuracy of model.

Description

Technical Field

[0001] This invention relates to the field of material manufacturing monitoring and prediction technology, and in particular to a quality prediction method for laser-injected composite coatings based on multi-level hybrid fusion, as well as a system applying this quality prediction method. Background Technology

[0002] Traditional subtractive manufacturing is limited to processing monolithic alloy products, making it difficult to improve the functional performance of localized areas while maintaining plasticity and relatively low cost. However, with the development of high-energy beam technology, laser fusion casting offers a new approach to manufacturing multi-material products with locally customized physical properties. As the demand for complex materials with special functions increases, selecting ceramic materials corresponding to the requirements and injecting them onto the surface of metal parts can form ceramic-reinforced metal matrix composite coatings. Localized ceramic-reinforced metal matrix composite coatings are particularly important in industries with high demand, such as aerospace, space exploration, energy development, and biomedicine. However, the adoption and application of laser-fused ceramic-reinforced metal matrix composite coatings in these industries has been slow, and current process reliability, stability, and repeatability are the main obstacles to the further application of laser-fused composite coatings.

[0003] From a materials perspective, the fabrication of laser-injected composite coatings requires controlling the temperature of the molten pool to avoid thermal decomposition of ceramic particles and ensure the melting of the metal matrix. Compared to high-energy beam processing of single alloy materials, ceramic-reinforced metal matrix composite coatings, with their more complex material systems, exhibit greater uncertainty. From a processing perspective, variations in processing boundary conditions due to part structural design, uneven powder feeding, and inconsistent laser processing equipment can all affect the state of ceramic particles and the solidification process of the molten pool. These uncertainties lead to defects in part quality and performance, resulting in significant additional costs and material waste during post-processing inspection. Therefore, real-time monitoring and predictive evaluation of the manufacturing quality of laser-injected composite coatings are crucial for promoting their large-scale application.

[0004] With the advancement of computer information processing technology and the rise of intelligent manufacturing technology, combining multi-source sensors with deep learning in the field of high-energy beam processing has become a new research hotspot. The combination of deep learning and multi-source monitoring data can not only improve the robustness of the monitoring system but also capture complementary information from different modalities. While deep learning's multi-layered networks offer flexibility in achieving multi-source monitoring data fusion, a unified data fusion architecture is still lacking, and the current selection of fusion structures is often based on intuition, making the selection of the optimal fusion architecture for specific problems more challenging. Currently, there is a lack of effective solutions for monitoring quality defects in laser-injected ceramic-reinforced metal matrix composite coatings. How to effectively utilize multi-source sensor data and select appropriate fusion structures to highlight the complementarity and robustness of the fusion architecture is a key issue in solving the multi-modal fusion problem in laser injection process monitoring. Summary of the Invention

[0005] Therefore, it is necessary to address the technical problem of the lack of effective solutions for monitoring the quality defects of laser-injected ceramic-reinforced metal matrix composite coatings in the existing technology. This invention provides a quality prediction method for laser-injected composite coatings based on multi-level hybrid fusion.

[0006] This invention discloses a quality prediction method for laser-injected composite coatings based on multi-level hybrid fusion, used for real-time prediction of the quality of ceramic-reinforced metal matrix composite coatings. The prediction method includes the following steps:

[0007] S1. Establish a laser melting process map for ceramic-reinforced metal matrix composite coatings.

[0008] S2. Simultaneously acquire paired image data of infrared camera images and high-speed camera images of composite coatings under multiple sets of laser melting process parameters, and group the corresponding paired image data according to the process map.

[0009] S3. Construct a multi-level hybrid fusion framework for laser fusion injection that combines a data layer and a feature layer. The construction method is as follows:

[0010] S31. Extract the molten pool feature region from the infrared camera image and the spatter feature region from the high-speed camera image using feature extraction processing methods.

[0011] S32. The molten pool feature region image is fused with the corresponding spatter feature region image to obtain a fused image.

[0012] S33. Perform data augmentation on the molten pool feature region image, the spatter feature region image, and the fused image, and use the three types of images as inputs respectively. Train and recognize the model through a variety of advanced feature extractors, and then determine the optimal feature extractor corresponding to the three types of image inputs.

[0013] S34. The molten pool feature region image, the spatter feature region image and the fused image are simultaneously used as input to achieve data layer fusion. The corresponding optimized feature extractors are used to perform feature layer fusion before entering the pooling layer and the fully connected layer, thereby obtaining the laser fusion multi-level hybrid fusion framework.

[0014] S4. A laser fusion quality prediction model is trained based on paired image data and a multi-level hybrid fusion framework, and the trained quality prediction model is used to predict the state of ceramic-reinforced metal matrix composite coatings in real time during the laser fusion process.

[0015] As a further improvement to the above scheme, in S1, the method for establishing the laser melting process map includes the following steps:

[0016] Based on the analysis of the dissolution state of ceramic particles inside the coating, mesoscopic surface roughness, and macroscopic cracking, the forming quality state of laser-injected composite coatings is divided into four categories: particle adhesion, pellets and cracks, normal melting, and excessive particle decomposition.

[0017] As a further improvement to the above scheme, in S31, the method for extracting molten pool features from infrared camera images includes the following steps:

[0018] (1) Extract the region of interest by taking the coordinates corresponding to the highest temperature in the infrared camera image as the center.

[0019] (2) Perform nonlocal mean filtering on the infrared camera image of the region of interest to eliminate isolated noise caused by splashing.

[0020] (3) The extracted molten pool region image is magnified by the bilinear interpolation algorithm.

[0021] (4) The emissivity is corrected by using the temperature gradient method and the temperature image is repaired to obtain the characteristic region of the molten pool.

[0022] As a further improvement to the above scheme, in S31, the method for extracting splash features from high-speed camera images includes the following steps:

[0023] (1) Extract the region of interest by taking the centroid of the arc in the high-speed camera image as the center.

[0024] (2) Perform median filtering to denoise the high-speed camera image from which the region of interest is extracted.

[0025] (3) Convert the grayscale image into a binary image using the Otsu global thresholding method.

[0026] (4) Extract the arc region and the laser head reflection region by area search method, and extract the remaining part in the image as the splash feature region.

[0027] As a further improvement to the above scheme, in S32, a pixel-weighted fusion method is used for fusion processing.

[0028] As a further improvement to the above scheme, in S33, data augmentation performs the same operation on image data acquired at the same time, including the following steps:

[0029] (1) Randomly select half of the images and add 3% salt and pepper noise.

[0030] (2) Randomly select half of the image and flip it horizontally.

[0031] (3) Rotate each image randomly by ±20 degrees, then crop and fill the rotated images to their original size.

[0032] (4) Add a random offset of 20% for width and a random offset of 20% for height to each image in sequence.

[0033] As a further improvement to the above scheme, the types of advanced feature extractors in S33 include: ResNet18, VGG16, InceptionV3 and MobileNetV3.

[0034] As a further improvement to the above scheme, in S33, the preferred feature extractor is determined by the average prediction time and the model prediction efficiency, specifically including the following steps:

[0035] (1) Remove feature extractors whose average prediction time is higher than the camera acquisition frequency.

[0036] (2) Calculate the model prediction efficiency of each feature extractor The calculation formula is:

[0037]

[0038] In the formula, m This indicates the total number of training sessions. t This indicates the average image processing time. n i For the first i Total number of samples for each training session. n TPi and n TFi The first i The number of true positive and false positive samples in the training.

[0039] (3) Select the feature extractor with the highest model prediction efficiency as the corresponding preferred feature extractor.

[0040] As a further improvement to the above scheme, in S2, the acquisition frequencies of the infrared camera and the high-speed camera are adjusted to be consistent before acquiring images from the infrared camera and the high-speed camera.

[0041] This invention also discloses a quality prediction system for laser-injected composite coatings based on multi-level hybrid fusion, which applies the aforementioned quality prediction method. The quality prediction system includes an image acquisition module and a data processing module.

[0042] The image acquisition module includes an infrared camera and a high-speed camera, which are used to acquire infrared images and high-speed images of the composite coating under multiple sets of laser melting process parameters, respectively, and form paired image data.

[0043] The data processing module is used to establish a laser sintering process map of ceramic-reinforced metal matrix composite coatings, and group the corresponding paired image data according to the process map. Then, the paired image data and the constructed laser sintering multi-level hybrid fusion framework are used to train a laser sintering quality prediction model, and the trained quality prediction model is used to predict the state of ceramic-reinforced metal matrix composite coatings in real time during the laser sintering process.

[0044] Compared with the prior art, the technical solution disclosed in this invention has the following beneficial effects:

[0045] This invention proposes a method for predicting the quality of laser-injected ceramic-reinforced metal matrix composite coatings based on infrared and high-speed camera images. This method integrates and fuses complementary molten pool and spatter features to predict the quality of laser-injected ceramic-reinforced metal matrix composite coatings, significantly improving the model's prediction accuracy. Specifically, a multi-level hybrid fusion framework combining data-level and feature-level fusion is established. The three types of input data in the proposed fusion framework are both complementary and redundant. Complementary data allows the feature extraction network to consider global information between the molten pool and spatter, while redundant data enhances the reliability of the extracted features and the predicted results.

[0046] Furthermore, the proposed multi-level hybrid fusion framework does not require specifying a feature extraction network. Hybrid fusion allows the use of different feature extraction networks for each channel, and a more suitable feature extraction network can be selected as the input data and device conditions change, thereby achieving greater flexibility. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the structure of a laser melting and casting experimental platform provided in a preferred embodiment of the present invention;

[0048] Figure 2 This is a flowchart of the quality prediction method for laser-injected composite coatings based on multi-level hybrid fusion according to the present invention;

[0049] Figure 3 These are typical macro- and micro-state diagrams of four types of laser-injected WC-reinforced metal matrix composite coatings in embodiments of the present invention;

[0050] Figure 4 This is a process diagram of laser-injected WC-enhanced 316L metal-based composite coating in an embodiment of the present invention;

[0051] Figure 5 This is a schematic diagram illustrating the process of extracting molten pool features from an infrared camera image in an embodiment of the present invention;

[0052] Figure 6 This is a schematic diagram illustrating the process of extracting splash features from a high-speed camera image in an embodiment of the present invention;

[0053] Figure 7 This is a schematic diagram illustrating the process of fusing images from a high-speed camera and an infrared camera.

[0054] Figure 8 This is a schematic diagram of a multi-stage hybrid fusion framework for laser melting in an embodiment of the present invention. Detailed Implementation

[0055] 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.

[0056] It should be noted that when a component is said to be "installed on" another component, it can be directly on the other component or it may be in a component that is centered on it. When a component is said to be "set on" another component, it can be directly set on the other component or it may also be in a component that is centered on it. When a component is said to be "fixed to" another component, it can be directly fixed to the other component or it may also be in a component that is centered on it.

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0058] This embodiment provides a quality prediction method for laser-injected composite coatings based on multi-level hybrid fusion, which is used to predict the quality of ceramic-reinforced metal matrix composite coatings in real time.

[0059] This embodiment describes the laser infusion of WC particles onto a 316L steel substrate, and the quality prediction of the resulting WC particle-reinforced 316L metal-based composite coating. Please refer to [link to relevant documentation]. Figure 1 The hardware platform used in this embodiment includes a laser generator 1, an argon gas cylinder 2, a powder feeder 3, an environmental chamber 4, a laser beam 5, a laser head 6, a processing platform 7, a substrate 8, a ceramic-reinforced metal-based composite coating 9, and a quality prediction system applying the above-described quality prediction method. This prediction system includes an image processing module and a data processing module. The image processing module may include an infrared camera 11 and a high-speed camera 10, and the data processing module may include a data acquisition unit 12 and a computer 13 with prediction software installed.

[0060] Please see Figure 2 The prediction method includes the following steps, namely S1 to S4.

[0061] S1. Establish a laser melting process map for ceramic-reinforced metal matrix composite coatings.

[0062] S2. Simultaneously acquire paired image data of infrared camera images and high-speed camera images of composite coatings under multiple sets of laser melting process parameters, and group the corresponding paired image data according to the process map.

[0063] In this embodiment, the 506R large environmental chamber laser cladding equipment manufactured by Pfizer is used. The surface of the 316L steel substrate is sandblasted, polished, degreased, cleaned, and dried with cold air. Cast WC powder with a particle size of 50-150 mesh is loaded into the Pfizer T2 double-cylinder automatic powder feeder 3. The ceramic insulating plate and the 316L steel substrate are placed on the worktable in sequence.

[0064] Laser casting experiments were conducted on the WC-reinforced metal matrix composite coating under multiple sets of different laser process parameters, as shown in Table 1. Based on the analysis of the dissolution state of ceramic particles within the composite coating, mesoscopic surface roughness, and macroscopic cracking, the forming quality of the laser-cast composite coating was categorized into four types: particle adhesion, agglomeration and cracking, normal melting, and excessive particle decomposition. Typical macroscopic and microscopic states of the coating are shown below. Figure 3As shown. The specific characteristics are: (1) Particle adhesion: When there is a lack of energy input during the laser melting process, the substrate cannot form a continuous molten pool, and only a small number of particles adhere to the substrate surface; (2) Agglomeration and cracks: When the amount of powder fed for reinforcing particles exceeds the capacity limit of the substrate to form a molten pool, the particles agglomerate, the coating surface is rough, and multiple transverse through cracks are present; (3) Normal melting: When the laser energy input matches the amount of powder fed, the material surface is smooth and crack-free, and the internal reinforcing particles are evenly distributed; (4) Excessive particle decomposition: Under excessive energy input, the decomposition of ceramic particles has an adverse effect on the mechanical properties of the laser melting coating. Based on the above process experiments and analysis, a process map of laser melting WC-reinforced 316L metal matrix composite coating is established, as follows. Figure 4 As shown.

[0065] Table 1 Laser melt injection process parameters

[0066]

[0067] During the laser fusion process described above, corresponding infrared camera images and high-speed camera images were acquired. The selected infrared camera and high-speed camera models and parameters are shown in Table 2. Due to the high exposure frequency of the high-speed camera, downsampling was used to maintain the same acquisition frequency (0.02s) as the IR camera. For each set of process parameters, 1000 data pairs were collected from the high-speed camera and infrared camera. 70% of the images collected from the experiment were used to train the classification model, and the remainder were used for validation. That is, the training dataset and the validation dataset contained 8400 and 3600 data pairs, respectively.

[0068] Table 2 Camera Models and Data Acquisition Parameters

[0069]

[0070] S3. Construct a multi-level hybrid fusion framework for laser fusion injection that combines a data layer and a feature layer. The construction method is as follows:

[0071] S31. Extract the molten pool feature region from the infrared camera image and the spatter feature region from the high-speed camera image using feature extraction processing methods.

[0072] Among them, such as Figure 5 As shown, the method for extracting molten pool features from infrared camera images includes the following steps:

[0073] (1) Extract the region of interest (ROI) (80×40 pixels) centered on the coordinates of the highest temperature in the infrared camera image.

[0074] (2) Perform nonlocal mean filtering on the infrared camera image of the region of interest to eliminate isolated noise caused by splashing.

[0075] (3) The extracted molten pool region image is magnified 10 times by using the bilinear interpolation algorithm.

[0076] (4) The emissivity is corrected by using the temperature gradient method and the temperature image is repaired to obtain the characteristic region of the molten pool.

[0077] like Figure 6 As shown, the method for extracting splash features from high-speed camera images includes the following steps:

[0078] (1) Extract the region of interest (650×400 pixels) centered on the centroid of the arc in the high-speed camera image.

[0079] (2) Median filtering (3×3) is applied to the high-speed camera image from which the region of interest is extracted to remove noise.

[0080] (3) Convert the grayscale image into a binary image using the Otsu global thresholding method.

[0081] (4) Extract the arc region and the laser head reflection region by area search method, and extract the remaining part in the image as the splash feature region.

[0082] S32. The molten pool feature region image and the corresponding spatter feature region image are fused to obtain a fused image. Specifically, the preprocessed high-speed camera image and the infrared camera image are normalized and resized, and a pixel-weighted fusion method is used to form the fused image. The fusion process is as follows: Figure 7 As shown.

[0083] S33. Perform data augmentation on the molten pool feature region image, the spatter feature region image, and the fused image, and use the three types of images as inputs respectively. Train and recognize the model through a variety of advanced feature extractors, and then determine the optimal feature extractor corresponding to the three types of image inputs.

[0084] The data augmentation process involves: (1) randomly selecting half of the image and adding 3% salt and pepper noise; (2) randomly selecting half of the image and flipping it horizontally; (3) randomly rotating the image by ±20 degrees, then cropping and filling the rotated image to its original size; (4) randomly offsetting the width by 20%; and (5) randomly offsetting the height by 20%.

[0085] In this embodiment, three types of images are used as inputs, and the model is trained and recognized by four feature extractors: ResNet18, VGG16, InceptionV3, and MobileNetV3. After 10 cross-validation training experiments, it is shown that the average image processing time of the four feature extractors is less than the image acquisition frequency (0.02s). Based on the model prediction efficiency, the optimal feature extractors corresponding to the three types of image inputs are determined. The results are shown in Table 3. Finally, the optimized feature extractors for the molten pool feature region image, the splash feature region image, and the fused image are InceptionV3, MobileNetV3, and InceptionV3, respectively.

[0086] Table 3 Model efficiency based on images from different sensors

[0087]

[0088] In this embodiment, the preferred feature extractor is determined by the average prediction time and the model prediction efficiency, specifically including the following steps:

[0089] (1) Remove feature extractors whose average prediction time is higher than the camera acquisition frequency.

[0090] (2) Calculate the model prediction efficiency of each feature extractor The calculation formula is:

[0091]

[0092] In the formula, m This indicates the total number of training sessions. t This indicates the average image processing time. n i For the first i Total number of samples for each training session. n TPi and n TFi The first i The number of true positive and false positive samples in the training.

[0093] (3) Select the feature extractor with the highest model prediction efficiency as the corresponding preferred feature extractor.

[0094] S34. The molten pool feature region image, the spatter feature region image, and the fused image are simultaneously used as input to achieve data layer fusion. These are processed by the corresponding preferred feature extractors InceptionV3, MobileNetV3, and InceptionV3, respectively. Feature layer fusion is performed before entering the pooling layer and the fully connected layer to determine a multi-level hybrid fusion framework combining the data layer and the feature layer in laser fusion. Figure 8 As shown.

[0095] S4. Based on the collected data and the determined multi-level hybrid fusion framework, train a laser melting and casting quality prediction model, and use the quality prediction model to predict the quality status of WC-enhanced 316L metal matrix composite coating in real time during the laser melting and casting process.

[0096] To demonstrate the superiority of the proposed multi-level hybrid fusion framework, the performance of two types of single-sensor training models, a data-layer fusion training model (based on fused images), a feature-layer fusion training model, and a multi-level hybrid fusion training model on real-time datasets is shown in Table 4. Among these, the data-layer fusion training model, the feature-layer fusion training model, and the multi-level hybrid fusion model selected the preferred feature extractor. Observations show that, across the four accuracy metrics, the multi-level hybrid fusion model achieves improved accuracy compared to both feature-layer fusion and data-layer fusion models.

[0097] Table 4. Accuracy metrics for different fusion models

[0098]

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for predicting the quality of laser-injected composite coatings based on multi-level hybrid fusion, characterized in that, A method for real-time prediction of the quality of ceramic-reinforced metal-based composite coatings; the prediction method includes the following steps: S1. Establish a laser melting process map for the ceramic-reinforced metal matrix composite coating; S2. Simultaneously acquire paired image data of infrared camera images and high-speed camera images of the composite coating under multiple sets of laser melting process parameters, and group the corresponding paired image data according to the process spectrum; S3. Construct a multi-level hybrid fusion framework for laser fusion injection that combines a data layer and a feature layer. The construction method is as follows: S31. Extract the molten pool feature region from the infrared camera image and the spatter feature region from the high-speed camera image using a feature extraction processing method; S32. The molten pool feature region image and the corresponding spatter feature region image are fused to obtain a fused image; in S32, the fusion process is performed by a pixel-weighted fusion method. S33. Data augmentation is performed on the molten pool feature region image, the spatter feature region image, and the fused image. The three types of images are used as inputs, and the model is trained and recognized using multiple advanced feature extractors to determine the preferred feature extractor corresponding to the three types of image inputs. In S33, the preferred feature extractor is determined by the average prediction time and the model prediction efficiency. S34. The molten pool feature region image, the spatter feature region image and the fused image are simultaneously used as input to achieve data layer fusion. The data layers are then fused through the corresponding preferred feature extractors and before entering the pooling layer and the fully connected layer, thereby obtaining the laser molten injection multi-level hybrid fusion framework. S4. Train a laser fusion quality prediction model based on the paired image data and the multi-level hybrid fusion framework, and use the trained quality prediction model to predict the state of the ceramic-reinforced metal matrix composite coating during the laser fusion process in real time.

2. The method for quality prediction of laser-injected composite coatings based on multi-level hybrid fusion according to claim 1, characterized in that, In S1, the method for establishing the laser melting and injection process map includes the following steps: Based on the analysis of the dissolution state of ceramic particles inside the coating, mesoscopic surface roughness, and macroscopic cracking, the forming quality state of laser-injected composite coatings is divided into four categories: particle adhesion, pellets and cracks, normal melting, and excessive particle decomposition.

3. The method for quality prediction of laser-injected composite coatings based on multi-level hybrid fusion according to claim 1, characterized in that, In S31, the method for extracting molten pool features from the infrared camera image includes the following steps: (1) Extract the region of interest by taking the coordinates corresponding to the highest temperature in the infrared camera image as the center; (2) Perform nonlocal mean filtering on the infrared camera image from which the region of interest is extracted to eliminate isolated noise caused by splashing; (3) The extracted molten pool region image is magnified using the bilinear interpolation algorithm; (4) The emissivity is corrected using the temperature gradient method and the temperature image is repaired to obtain the characteristic region of the molten pool.

4. The method for quality prediction of laser-injected composite coatings based on multi-level hybrid fusion according to claim 1, characterized in that, In S31, the method for extracting splash features from the high-speed camera image includes the following steps: (1) Extract the region of interest centered on the centroid of the arc in the high-speed camera image; (2) Perform median filtering to denoise the high-speed camera image from which the region of interest has been extracted; (3) Convert the grayscale image into a binary image using the Otsu global thresholding method; (4) Extract the arc region and the laser head reflection region by area search method, and extract the remaining part in the image as the splash feature region.

5. The method for quality prediction of laser-injected composite coatings based on multi-level hybrid fusion according to claim 1, characterized in that, In S33, the data enhancement performs the same operation on image data acquired at the same time, including the following steps: (1) Randomly select half of the images and add 3% salt and pepper noise; (2) Randomly select half of the image and flip it horizontally; (3) Rotate each image randomly by ±20 degrees, then crop and fill the rotated images to their original size; (4) Add a random offset of 20% for width and a random offset of 20% for height to each image in sequence.

6. The method for quality prediction of laser-injected composite coatings based on multi-level hybrid fusion according to claim 1, characterized in that, In S33, the types of advanced feature extractors include: ResNet18, VGG16, InceptionV3, and MobileNetV3.

7. The method for quality prediction of laser-injected composite coatings based on multi-level hybrid fusion according to claim 1, characterized in that, In S33, determining the preferred feature extractor based on average prediction time and model prediction efficiency includes the following steps: (1) Remove feature extractors whose average prediction time is higher than the camera acquisition frequency; (2) Calculate the model prediction efficiency of each feature extractor η The calculation formula is: ； In the formula, m Indicates the total number of training iterations; t This indicates the average image processing time. n i For the first i Total number of samples for each training session; n TPi and n TFi The first i The number of true positive and false positive samples in the training session; (3) Select the feature extractor with the highest model prediction efficiency as the corresponding preferred feature extractor.

8. The method for quality prediction of laser-injected composite coatings based on multi-level hybrid fusion according to claim 1, characterized in that, In S2, before acquiring the images from the infrared camera and the high-speed camera, the acquisition frequencies of the infrared camera and the high-speed camera are adjusted to be consistent.

9. A quality prediction system for laser-injected composite coatings based on multi-level hybrid fusion, characterized in that, Its application is the quality prediction method as described in any one of claims 1 to 8; the quality prediction system includes: The image acquisition module includes an infrared camera and a high-speed camera, which are used to acquire infrared images and high-speed images of the composite coating under multiple sets of laser melting process parameters and form paired image data. The data processing module is used to establish a laser melting process map of the ceramic-reinforced metal matrix composite coating, group the corresponding paired image data according to the process map, and then use the paired image data and the constructed laser melting multi-level hybrid fusion framework to train a laser melting quality prediction model. The trained quality prediction model is then used to predict the state of the ceramic-reinforced metal matrix composite coating in real time during the laser melting process.

Images