Monitoring method and system and platform for monitoring laser cladding process of metal matrix composite

Abstract

The application discloses a monitoring method, system and platform for monitoring a laser cladding process of a metal matrix composite material. The monitoring method adopts a full-depth separable convolutional neural network based on a CNN framework, and the establishment method comprises the following steps: a plurality of full-depth separable convolutional modules are established by using a depth separable convolution and channel cleaning, and a network model is established by stacking different numbers. Wherein, the input channels of each full-depth separable convolutional module are equally divided, two branches are subjected to depth separable convolution, different size convolution kernels are used to extract features, and two-dimensional dimensions are formed; two-dimensional dimensions are spliced by using a Concat operation, and then a channel cleaning operation is performed to realize feature fusion between different channels. The application greatly shortens the parameter quantity and inference time while ensuring high precision, greatly saves the calculation time and resources, and lays a foundation for designing a real-time control system of MMC additive manufacturing.

Description

TECHNICAL FIELD

[0001] The present application relates to a monitoring method in the field of additive manufacturing process monitoring, a monitoring system using the monitoring method, and an experimental platform with the monitoring system, in particular to a monitoring method for monitoring metal matrix composites in the laser cladding process based on full-depth separable convolution, a monitoring system using the monitoring method, and an experimental platform with the monitoring system. BACKGROUND

[0002] With the increasing complexity and harshness of the service environment of product parts, adding hard ceramic particles to the base material to form ceramic reinforced metal matrix composites (CRMMC) can further improve the mechanical properties such as hardness and wear resistance of the product formed by laser directed energy deposition. However, during the deposition process, ceramic particles will be aggregated and dissolved to varying degrees, affecting the stability of the CRMMC forming quality, and the dissolution and aggregation of ceramic particles are closely related to the characteristics of the molten pool. Therefore, monitoring the molten pool state generated during the deposition process of CRMMC is an effective means to improve the performance of the final product.

[0003] Related research shows that the molten pool image contains important process information and can be used for defect identification and quality monitoring of parts. At present, among the processing methods of molten pool images, machine learning technology is being applied more and more widely. However, traditional machine learning algorithms such as BP neural network, SVM and random forest require manual extraction of simple features such as the length, width and temperature of the molten pool. Based on the above features, a prediction relationship between process parameters and molten pool state or forming quality is established. This method has obvious defects: the features extracted manually are relatively simple and the process is time-consuming, resulting in low model training efficiency, and the application scenario is single, which makes it difficult to meet the real-time requirements brought by high-frequency acquisition of molten pool state monitoring in actual production process, and it is impossible to realize industrial real-time monitoring of the molten pool state.

[0004] With the rise of deep learning algorithms, an important branch of deep learning algorithms, convolutional neural networks (CNN), has been widely used in molten pool image recognition processing due to its fast data processing and image feature extraction capabilities, and does not need to manually extract features like traditional machine learning. SUMMARY

[0005] Solving the real-time demand of molten pool state in actual industry is an application problem of online monitoring of additive manufacturing which needs to be solved urgently, and therefore the application provides a monitoring method for monitoring metal matrix composites in a laser cladding process based on full-depth separable convolution, a monitoring system adopting the monitoring method, and an experimental platform with the monitoring system, wherein the monitoring method is based on a CNN framework to establish a high-precision lightweight full-depth separable neural network model (FD-Net), and the infrared image of the molten pool collected by an infrared thermal imager is sent into the model to realize rapid and accurate identification of the molten pool state of the CRMMC.

[0006] To achieve the above object, the application adopts the following technical solutions:

[0007] A monitoring method for monitoring metal matrix composites in a laser cladding process based on full-depth separable convolution, the monitoring method comprising the following steps:

[0008] S1, capturing real-time laser cladding dynamic video of metal matrix composites in a laser cladding process at continuous time;

[0009] S2, decomposing the real-time laser cladding dynamic video into a plurality of real-time laser cladding images of individual frames;

[0010] S3, sequentially sending each real-time laser cladding image into a network model of a neural network to identify the laser cladding state of the metal matrix composites in the laser cladding process in real time through the network model;

[0011] The neural network is a full-depth separable convolution neural network FD-Net established based on a CNN framework, and the establishment method of the neural network comprises the following steps:

[0012] S31, establishing a plurality of full-depth separable convolution modules by using depth separable convolution and channel cleaning,

[0013] S32, stacking the plurality of full-depth separable convolution modules in different numbers to establish the network model, and the model represents the full-depth separable convolution neural network FD-Net established based on the CNN framework;

[0014] The establishment method of each full-depth separable convolution module comprises the following steps:

[0015] S311, equally dividing the input channels of each full-depth separable convolution module, performing depth separable convolution on two branches respectively, using convolution kernels of different sizes to extract features, and forming two routes of dimensions;

[0016] S312, splicing the two routes of dimensions by using a Concat operation;

[0017] S313, the two-dimension after splicing is subjected to channel cleaning operation, feature fusion between different channels is realized, and the full-depth separable convolution module is formed.

[0018] As a further improvement of the above scheme, in S311, the full-depth separable convolution module uses 3*3 and 7*7 size convolution kernels for feature extraction, respectively.

[0019] As a further improvement of the above scheme, in S32, the network model is established by stacking the full-depth separable convolution modules according to the number ratio of 1:1:3:1.

[0020] As a further improvement of the above scheme, the network model needs to be trained, optimized and tested before use:

[0021] Carrying out laser direct energy deposition experiments of the metal matrix composite under multiple groups of different laser energy inputs, and collecting corresponding multiple groups of experimental laser cladding dynamic videos;

[0022] Each group of experimental laser cladding dynamic videos is decomposed into multiple single-frame experimental laser cladding images;

[0023] All experimental laser cladding images are preprocessed and data enhanced, and image labels are added, thereby establishing an original data set;

[0024] All experimental laser cladding images in the original data set are randomly divided into a training set, a validation set and a test set according to a target ratio;

[0025] The training set, the validation set and the test set are respectively sent into the network model for training, optimization and testing, and the best model parameters are saved, thereby updating the network model.

[0026] Further, when the network model is trained, optimized and tested:

[0027] After model convolution, ReLU activation function activation and BN normalization stacking operation, the molten pool state image features in each experimental laser cladding image of the original data set are continuously extracted in depth, and high-dimensional abstract output feature maps are obtained;

[0028] The output feature map is unfolded into a one-dimensional vector through the full connection layer of the network model, and then input into the softmax classification layer of the network model, and the loss is calculated by using the loss function;

[0029] The weight parameters of the network model are continuously updated by the first-order gradient random optimization algorithm ADAM, so as to minimize the loss function and help the network model complete training and optimization;

[0030] whether to save the network model according to the performance of the network model on the verification set; if it is judged to save the network model, the weight of the network model is saved if the accuracy on the verification set is the current optimal result every training cycle, otherwise not;

[0031] The test set is input into the network model, the saved model parameters of the network model are loaded, and the final performance of the network model is judged according to the test result.

[0032] Further, the preprocessing is that Gaussian filtering, contour detection and image cropping are performed on each original experimental laser cladding image by using a Python-OpenCV tool, and the ROI of the corresponding image is extracted.

[0033] Further, the data enhancement operation is that offset, flip and affine transformation data enhancement operations are performed on the preprocessed image.

[0034] Further, the image label is that 0-n corresponding to the energy input size level E1-En is taken as the data set label, and n is a positive integer.

[0035] The application also provides a monitoring system for monitoring the laser cladding process of the metal matrix composite material based on the full-depth separable convolution, and the monitoring system comprises:

[0036] A real-time dynamic video acquisition module is used to capture real-time laser cladding dynamic video of the metal matrix composite material in the laser cladding process under continuous time;

[0037] A real-time laser cladding image module is used to decompose the real-time laser cladding dynamic video into a plurality of single-frame real-time laser cladding images;

[0038] A neural network model is used to identify the laser cladding state of the metal matrix composite material in the laser cladding process in real time for each real-time laser cladding image;

[0039] The neural network is a full-depth separable convolution neural network FD-Net based on a CNN framework; and the establishment method of the neural network comprises the following steps:

[0040] S31, a plurality of full-depth separable convolution modules are established by using depth separable convolution and channel cleaning,

[0041] S32, the plurality of full-depth separable convolution modules are stacked in different numbers to establish the network model, and the model represents a full-depth separable convolution neural network FD-Net based on a CNN framework;

[0042] The method for establishing each full-depth separable convolution module comprises the following steps.

[0043] S311, input channels of each full-depth separable convolution module are equally divided, two branches are respectively subjected to depth separable convolution, different size convolution kernels are used to extract features, and two dimensions are formed;

[0044] S312, the two dimensions are spliced by using a Concat operation;

[0045] S313, the two spliced dimensions are subjected to a channel cleaning operation to realize feature fusion between different channels, and the full-depth separable convolution module is formed.

[0046] The application also provides an experimental platform, which comprises an energy deposition system and a monitoring system. The energy deposition system comprises an L-DED workbench, a base arranged on the workbench, a laser head for providing laser energy to a metal matrix composite on the base so that the metal matrix composite is subjected to laser cladding, and an infrared thermal imager for capturing real-time laser cladding dynamic video of the metal matrix composite in a laser cladding process at continuous time.

[0047] Compared with the prior art, the application has the beneficial effects that the full-depth separable module is established, feature fusion between different channels is realized by using two different size convolution kernels, and a full-depth separable convolution neural network is established by stacking according to a certain quantity ratio, so that the parameter quantity and inference time are greatly shortened while ensuring high precision, the calculation time and resources are greatly saved, and a basis is laid for designing a real-time control system of MMC additive manufacturing. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A structural schematic diagram of an experimental platform provided for an embodiment of the application.

[0049] Figure 2 For Figure 1 A monitoring flowchart of a monitoring method for monitoring a metal matrix composite in a laser cladding process based on full-depth separable convolution, which is adopted by a monitoring system of an experimental platform.

[0050] Figure 3 For Figure 2 A structural schematic diagram of a full-depth separable convolution module adopted by the monitoring method.

[0051] Figure 4 For adopting Figure 3A structural schematic diagram of a full-depth separable convolution neural network FD-Net based on a full-depth separable convolution module.

[0052] Figure 5 For Figure 4 A flowchart of training, tuning and testing based on the full-depth separable convolution neural network.

[0053] Figure 6 With Figure 5 Similarly, a flowchart of training, tuning and testing in the experimental operation. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical scheme and advantages of the present application clearer, further detailed description will be made below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0055] Please refer to Figure 1 It is a structural schematic diagram of an experimental platform provided by the embodiment of the present application, which includes an energy deposition system and a monitoring system. The energy deposition system is used to provide a laser cladding platform for a metal matrix composite, and the monitoring system is used to monitor the metal matrix composite subjected to laser cladding by the energy deposition system, i.e. to monitor various molten states of the metal matrix composite in the laser cladding process.

[0056] The energy deposition system mainly includes an L-DED workbench 1, a substrate 2, a laser head 3, a powder feeder 4, a protective gas 5, an infrared thermal imager 6 and a computer 7.

[0057] The workbench 1 can adopt an L-DED workbench of a 506R large environmental box laser cladding equipment produced by China Hui Rui Company, and the substrate 2 is arranged on the workbench 1. The laser head 3 is used to provide laser energy for the metal matrix composite on the substrate 2, so that the metal matrix composite is subjected to laser cladding. The powder feeder 4 can adopt a Hui Rui T2 double-cylinder powder feeder, and the infrared thermal imager 6 can adopt an FLIRA655sc model infrared thermal imager. In order to better capture the molten pool image, the infrared thermal imager 6 is placed at a horizontal distance of 900 mm and a vertical distance of 450 mm from the workbench 1. The resolution of the original image collected by the visual sensing system of the infrared thermal imager 6 is 640 pixels x 480 pixels, the collection frequency is 50 frames / s, and the radiation rate is 0.25. The computer 7 is used for displaying and storing the collected images and identifying the molten pool state.

[0058] The monitoring system adopts a monitoring method based on full-depth separable convolution to identify the laser cladding state of the metal matrix composite in the laser cladding process. Please refer to Figure 2 It is a structural schematic diagram of a full-depth separable convolution neural network FD-Net based on a full-depth separable convolution module. Figure 1A monitoring method for monitoring a metal matrix composite in a laser cladding process based on a full-depth separable convolution is adopted by a monitoring system of a test platform. The monitoring method comprises: S1, capturing real-time laser cladding dynamic video of the metal matrix composite in the laser cladding process at continuous time; S2, decomposing the real-time laser cladding dynamic video into a plurality of real-time laser cladding images of single frames; and S3, sequentially sending each real-time laser cladding image into a network model of a neural network to identify a laser cladding state of the metal matrix composite in the laser cladding process in real time through the network model.

[0059] When the monitoring method is applied, it can be applied in the form of software, such as being designed as a program running independently, installed on a computer terminal, which can be a computer, a smart phone, a smart chip, etc. It can also be designed as an embedded program running, installed on a computer terminal, such as a single-chip microcomputer. When the monitoring method is applied, it can also be designed as a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps of the monitoring method. The monitoring method is applied through a computer readable storage medium in the form of software, such as being designed as a program running independently on a computer readable storage medium, which can be a U disk, designed as a U disk, and designed as a program starting the whole method through external triggering through a U disk.

[0060] When the monitoring method is implemented, a corresponding monitoring system can be set. The monitoring system comprises: a real-time dynamic video acquisition module for capturing real-time laser cladding dynamic video of the metal matrix composite in the laser cladding process at continuous time, executing S1 step of the monitoring method; a real-time laser cladding image module for decomposing the real-time laser cladding dynamic video into a plurality of real-time laser cladding images of single frames, executing S2 step of the monitoring method; and a network model of a neural network for identifying a laser cladding state of the metal matrix composite in the laser cladding process in real time for each real-time laser cladding image, executing S3 step of the monitoring method.

[0061] In this embodiment, the neural network is a full-depth separable convolution neural network FD-Net based on a CNN framework. The establishment method of the neural network mainly comprises two steps of S31 and S32.

[0062] S31, a plurality of full-depth separable convolution modules are established by using a depth separable convolution and a channel cleaning. The establishment method of each full-depth separable convolution module mainly comprises three steps of S311, S312 and S313.

[0063] S311, the input channels of each full depth separable convolution module are equally divided, and two branches are respectively subjected to depth separable convolution, and different size convolution kernels are used to extract features to form two dimensions. In S311, the full depth separable convolution module respectively uses 3x3 and 7x7 size convolution kernels for feature extraction.

[0064] S312, the two dimensions are spliced by using the Concat operation.

[0065] S313, the two dimensions after splicing are subjected to channel cleaning operation to realize feature fusion between different channels to form the full depth separable convolution module.

[0066] S32, the plurality of full depth separable convolution modules are stacked in different numbers to establish the network model, and the model represents a full depth separable convolution neural network FD-Net based on a CNN framework. In S32, the full depth separable convolution modules are stacked according to a number ratio of 1:1:3:1 to establish the network model.

[0067] In this embodiment, in order to consider the real-time requirement of high-frequency acquisition of L-DED in actual industry, the full depth separable convolution module is proposed, as shown in Figure 3 The input channels are first equally divided, two branches are respectively subjected to depth separable convolution, 3x3+1x1 and 7x7+1x1 convolution kernels are used to extract features, then the dimensions of the two branches are spliced, and finally the channel cleaning operation is performed. The significance of this design is that each branch obtains feature fusion between different channels, and the parameter amount is significantly reduced; then the Concat is used to speed up the inference speed of the model. Finally, after the channel cleaning, the information exchange between different channels is strengthened, and more rich features are fused.

[0068] The full depth separable convolution modules are stacked according to the number ratio of 1, 1, 3, and 1 to establish the FD-Net network model, as shown in Figure 4 The model structure generally extracts deep abstract features of the molten pool infrared image through a series of full depth separable convolution modules to obtain high-level features, and finally calculates the probability of belonging to each class through the full connection layer and the softmax classification layer, so as to correctly identify the molten pool state.

[0069] The calculation formula of the softmax function is: In the formula, i is the value of the i-th element, m is the total number of elements, and sigma(·) is the function output value.

[0070] Before the network model is used, it needs to be trained, optimized and tested, please refer to Figure 5which is a flow chart of the training, tuning and testing process of the network model before use. The pre-use processing method comprises the following steps:

[0071] (1) A series of laser direct energy deposition experiments of the metal matrix composite under different laser energy inputs are carried out, and a plurality of sets of experimental laser cladding dynamic videos corresponding to the experiments are collected.

[0072] A series of ceramic reinforced metal matrix composite laser direct energy deposition experiments carried out in this step are carried out under different energy inputs to ensure that different molten pool states are obtained. In the present application, tungsten carbide (WC) is selected as the reinforcing phase of the L-DED process material. In order to obtain the image data set corresponding to different molten pool states, the laser energy input is linearly increased to realize the difference of the molten pool state, and the remaining parameters remain unchanged, as shown in Table 1.

[0073] Table 1 L-DED process parameters

[0074]

[0075] (2) Each set of experimental laser cladding dynamic video is decomposed into a plurality of single-frame experimental laser cladding images.

[0076] In this embodiment, the infrared thermal imager 6 is located on the side of the substrate 2, and synchronously captures the molten pool dynamic video at continuous time, decomposes the video into single-frame molten pool infrared images as the original data set.

[0077] (3) All experimental laser cladding images are pre-processed and data enhanced, and image labels are added, thereby establishing the original data set. The pre-processing is that the Python-OpenCV tool is used to perform Gaussian filtering, contour detection and image cropping on each original experimental laser cladding image to extract the ROI of the corresponding image. The data enhancement operation is that the pre-processed image is subjected to offset, flip and affine transformation data enhancement operation. The image label is that 0-n corresponding to the energy input size level E1-En is used as the data set label, and n is a positive integer. In this embodiment, 0-9 corresponding to the energy input size (E1-E9) is used as the data set label.

[0078] (4) All experimental laser cladding images in the original data set are randomly divided into a training set, a validation set and a test set according to a target proportion.

[0079] Since the collected molten pool infrared image includes a lot of invalid background, in order to highlight the effective area of the molten pool in the image, in the embodiment, the original image is cropped by ROI, and the size of the cropped image is 73 pixels x 23 pixels. Table 2 is a pre-processed 4500 group image data set including images and labels, and is divided into a training set, a validation set and a test set according to a ratio of 6:2:2, and the model is trained, optimized and tested, as shown in Table 2.

[0080] Table 2 molten pool image data set

[0081] Melt pool class Training set Validation set Test set Label E1 300 100 100 0 E2 300 100 100 1 E3 300 100 100 2 E4 300 100 100 3 E5 300 100 100 4 E6 300 100 100 5 E7 300 100 100 6 E8 300 100 100 7 E9 300 100 100 8 Total 2700 900 900 —

[0082] The model training module is mainly based on a full-depth separable convolutional neural network, which uses a depth separable convolution instead of a traditional convolution, greatly reduces the parameter quantity, improves the inference speed of the model, and can meet the high-frequency acquisition requirements of L-DED actual industry.

[0083] In order to further improve the generalization ability of the model, the pre-processed training set image is subjected to data enhancement operations such as rotation, flipping, translation and affine transformation, and then converted into a three-dimensional matrix and sent into the network for training.

[0084] (5) The training set, the validation set and the test set are respectively sent into the network model for training, optimization and testing, and the best model parameters are saved, thereby updating the network model.

[0085] Wherein, when the network model is trained, optimized and tested:

[0086] After model convolution, ReLU activation function activation and BN normalization stacking operation, the molten pool state image features in each experimental laser cladding image of the original data set are continuously extracted in depth, and a high-dimensional abstract output feature map is obtained;

[0087] The output feature map is expanded into a one-dimensional vector through the full connection layer of the network model, and then input into the softmax classification layer of the network model, and the loss is calculated by using the loss function;

[0088] The weight parameters of the network model are continuously updated by the first-order gradient random optimization algorithm ADAM, so as to minimize the loss function and help the network model complete training and optimization;

[0089] Whether to save the network model is determined by the performance of the network model on the validation set; if it is determined to save the network model, the weight of the network model is saved if the accuracy on the validation set is the current optimal result after each training cycle, otherwise it is not saved;

[0090] The test set is sent into the network model, the model parameters of the saved network model are loaded, and the final performance of the network model is judged according to the test result.

[0091] In the embodiment, referring to Figure 6 The training set is sent into the FD-Net model, the first-order gradient random optimization algorithm ADAM is used to continuously update the model weight parameters, so that the cross-entropy function is minimized, and whether the model is saved is judged by the performance of the model on the verification set; the criterion is that if the accuracy on the verification set is the current optimal result, the network model weight is saved to the pth format file, otherwise, the network model weight is not saved.

[0092] The test set is sent into the model, and the saved pth model weight file is loaded, the recognition accuracy of the test set molten pool state and the reasoning time are calculated, and the final performance of the model is determined;

[0093] The accuracy rate and the recognition time of each photo are selected as the evaluation indexes of the classification result of the model; if the accuracy rate is high, the classification effect of the model is good; if the recognition time is low, the recognition speed of the model is fast; table 3 is the performance of different models on the test set, and it can be found that, compared with the existing advanced model, the FD-Net has high recognition accuracy, shorter model reasoning time and smaller model memory occupation, and therefore can better meet the real-time demand of the molten pool state recognition of the MMC in the L-DED process, as shown in table 3.

[0094] Table 3 shows the performance of different models on the test set

[0095] Model name Accuracy (%) Time / ms Model memory / M FD-Net 96.55 3.01 1.91 ResNet34 96.11 4.10 81.30 MobileNetV2 87.78 3.67 8.76 ShuffleNet 79.33 3.22 1.47

[0096] The monitoring method is a metal matrix composite laser cladding process monitoring method based on full-depth separable convolution, and belongs to the field of process monitoring of additive manufacturing. First, infrared images of a molten pool of a ceramic reinforced metal matrix composite in a directional energy deposition process are collected by an infrared thermal imager, and an original image ROI is extracted, and then a series of data enhancement operations are performed on the cropped image. Secondly, a full-depth separable convolution neural network is established by using deep separable convolution combined with channel cleaning. The processed image is sent into the network model for training, the best model parameters are saved, and finally the model parameters are called in the test set to test the final performance of the model. The deep separable convolution is used instead of the traditional convolution combined with the channel cleaning operation in the present application, which reduces a large number of network parameters and saves the molten pool state recognition time on the basis of ensuring high precision of the model, and lays a foundation for realizing a real-time control system of the metal matrix composite additive manufacturing forming quality.

[0097] The above merely describes preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement and improvement within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A monitoring method for metal-based composites during laser cladding based on full-depth separable convolution, the monitoring method comprising the following steps: S1 captures real-time dynamic video of laser cladding of metal matrix composites during the laser cladding process over a continuous time period; S2, decompose the real-time laser cladding dynamic video into multiple individual frames of real-time laser cladding images; S3, each real-time laser cladding image is sequentially fed into a neural network model, and the laser cladding state of the metal matrix composite during the laser cladding process is identified in real time through the network model; The neural network is characterized by being a fully deep separable convolutional neural network (FD-Net) built upon the CNN framework; the method for building the neural network includes the following steps: S31 utilizes depthwise separable convolution and channel cleaning to establish multiple fully depthwise separable convolutional modules. S32, stack the multiple fully deep separable convolutional modules in different numbers to build the network model, the model representing a fully deep separable convolutional neural network FD-Net built based on the CNN framework; The method for establishing each fully depthwise separable convolutional module includes the following steps: S311 divides the input channels of each fully depthwise separable convolutional module into equal parts, performs depthwise separable convolution on the two paths respectively, and extracts features using convolutional kernels of different sizes, forming two-way dimensions; S312, use the Concat operation to concatenate the two dimensions; S313, After splicing, the two dimensions are cleaned by channel cleaning to achieve feature fusion between different channels and form the full-depth separable convolutional module; During training, tuning, and testing of the network model: Through model convolution, ReLU activation, and BN normalization stacking operations, the features of the molten pool state in each experimental laser cladding image of the original dataset are continuously extracted to obtain a high-dimensional abstract output feature map. This output feature map is then expanded into a one-dimensional vector through the fully connected layer of the network model and input into the softmax classification layer, where a loss function is calculated. The weight parameters of the network model are continuously updated using the first-order gradient stochastic optimization algorithm ADAM to minimize the loss function, thus helping the network model complete training and optimization. The performance of the network model on the validation set determines whether to save the network model. If saving is determined, the weights are saved after each training iteration if the accuracy on the validation set is the current optimal result; otherwise, they are not saved. The test set is then fed into the network model, loading the saved model parameters, and the final performance of the network model is determined based on the test results.

2. The monitoring method for monitoring metal-based composites during laser cladding based on full-depth separable convolution as described in claim 1, characterized in that: In S311, the fully depth-separable convolutional module uses 3×3 and 7×7 convolutional kernels for feature extraction, respectively.

3. The monitoring method for monitoring metal-based composites during laser cladding based on full-depth separable convolution as described in claim 1, characterized in that: In S32, the network model is built by stacking the full-depth separable convolutional modules in a ratio of 1:1:3:

1.

4. The monitoring method for monitoring metal-based composites during laser cladding based on full-depth separable convolution as described in claim 1, characterized in that: The network model needs to be trained, optimized, and tested before it can be used. Laser direct energy deposition experiments were conducted on the metal matrix composite under multiple sets of different laser energy inputs, and corresponding dynamic videos of laser cladding were collected. Each set of experimental laser cladding dynamic videos was decomposed into multiple individual frames of experimental laser cladding images; All experimental laser cladding images were preprocessed and data augmented, and image labels were added to establish the original dataset. All experimental laser cladding images in the original dataset are randomly divided into training set, validation set and test set according to the target proportion; The training set, the validation set, and the test set are respectively fed into the network model for training, tuning, and testing, and the best model parameters are saved to update the network model.

5. The monitoring method for monitoring metal-based composites during laser cladding based on full-depth separable convolution as described in claim 4, characterized in that: The preprocessing involves using the Python-OpenCV tool to perform Gaussian filtering, contour detection, and image cropping on each original experimental laser cladding image, and extracting the ROI of the corresponding image.

6. The monitoring method for monitoring metal-based composites during laser cladding based on full-depth separable convolution as described in claim 4, characterized in that: The data augmentation operation is to perform offset, flip, and affine transformation data augmentation operations on the preprocessed image.

7. The monitoring method for monitoring metal-based composites during laser cladding based on full-depth separable convolution as described in claim 4, characterized in that: The image labels are: the energy input size levels E1~En, corresponding to 0~n, are used as dataset labels, where n is a positive integer.

8. A monitoring system employing the monitoring method for monitoring metal-based composites during laser cladding based on full-depth separable convolution as described in any one of claims 1 to 7, the monitoring system comprising: A real-time dynamic video acquisition module is used to capture real-time dynamic video of laser cladding of metal matrix composites during the laser cladding process over a continuous period of time. A real-time laser cladding image module is used to decompose the real-time laser cladding dynamic video into multiple individual frames of real-time laser cladding images; A neural network model is used to identify the laser cladding state of the metal matrix composite during the laser cladding process in real time for each real-time laser cladding image; The neural network is characterized by being a fully deep separable convolutional neural network (FD-Net) built upon the CNN framework; the method for building the neural network includes the following steps: S31 utilizes depthwise separable convolution and channel cleaning to establish multiple fully depthwise separable convolutional modules. S32, stack the multiple fully deep separable convolutional modules in different numbers to build the network model, the model representing a fully deep separable convolutional neural network FD-Net built based on the CNN framework; The method for establishing each fully depthwise separable convolutional module includes the following steps: S311 divides the input channels of each fully depthwise separable convolutional module into equal parts, performs depthwise separable convolution on the two paths respectively, and extracts features using convolutional kernels of different sizes, forming two-way dimensions; S312, use the Concat operation to concatenate the two dimensions; S313, after splicing the two dimensions, channel cleaning operation is performed to achieve feature fusion between different channels and form the full-depth separable convolution module.

9. An experimental platform, the experimental platform comprising: Energy deposition system, including: L-DED workbench (1); The substrate (2) is set on the workbench (1); Laser head (3), used to provide laser energy to the metal matrix composite on the substrate (2) so that the metal matrix composite can be laser cladding; and An infrared thermal imager (6) is used to capture real-time dynamic video of the laser cladding process of the metal matrix composite over a continuous period of time; and A monitoring system, wherein the monitoring system uses the monitoring method for monitoring metal substrate composites during the laser cladding process based on full-depth separable convolution as described in any one of claims 1 to 7 to identify the laser cladding state of the metal substrate composites during the laser cladding process.

Images