Apparatus and method for measuring the morphology of a deposition layer in additive manufacturing based on image-based intelligent reconstruction

Abstract

The application discloses a kind of based on image intelligent reconstruction's additive manufacturing deposition layer topography measuring device and method, and collects molten pool area image and vibration data;The molten pool area image and vibration data collected are input into deep self-encoder and are handled, and reconstruction image is generated by deep self-encoder;And using PSNR and SSIM assesses the quality and stability of reconstruction image;If image quality does not reach preset standard, the system will calculate the best illumination angle and camera position required to improve image quality;The angle and position of LED light and camera are adjusted by hollow shaft motor, ensure that light path evenly covers molten pool surface, and all possible light dark areas are captured;If image quality reaches preset standard, output reconstruction image.The application effectively solves the challenge faced by deposition layer topography test in laser additive manufacturing process by combining improved monitoring equipment and deep learning technology, improves test quality and reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to a laser additive manufacturing deposition layer morphology high-precision measurement device and method based on image intelligent reconstruction, belonging to the technical field of laser additive manufacturing. BACKGROUND

[0002] In the process of laser additive manufacturing, the accurate testing of the surface morphology of the deposition layer is crucial for ensuring the manufacturing quality. However, a series of challenges will be faced in the experimental process. First, there may be dark areas that light cannot reach in some areas, which is mainly due to the characteristics of laser irradiation. The propagation of laser will be affected by the reflection and refraction of the material surface, causing some areas to be directly irradiated by light, thus forming dark areas. However, these dark areas may contain critical surface detail information such as microstructure, cracks or defects, but cannot be captured by traditional monitoring equipment. Second, the vibration of the manufacturing equipment itself and the surrounding environment is also an important influencing factor. In the process of laser additive manufacturing, the movement of the laser head, the injection of the material and the operation of other mechanical parts will cause slight vibration. These vibrations appear as image jitter and blur in video data, affecting the stability and clarity of the test data. Vibration phenomena may be caused by a variety of factors, including the accuracy of the manufacturing equipment, the adhesion of the material, the stability of the workbench and the interference of the surrounding environment. SUMMARY

[0003] The purpose of the present application is to solve the problems of insufficient light and equipment vibration in the process of laser additive manufacturing. The present application provides a laser additive manufacturing deposition layer morphology measurement method based on image intelligent reconstruction. By integrating a variable angle camera and a deep autoencoder technology, and combining a vibration elimination method, the effects of light and vibration can be effectively eliminated, thereby realizing high-precision testing of the surface morphology of the deposition layer.

[0004] Technical solution: In order to achieve the above purpose, the technical solution adopted by the present application is as follows:

[0005] A laser additive manufacturing deposition layer morphology measurement method based on image intelligent reconstruction, by improving the monitoring equipment of laser metal deposition, the light in the dark area during the experiment can be supplemented, so that the surface details of these key areas can also be accurately captured. Secondly, by using deep learning technology, the collected video data can be processed and optimized to reduce the influence of vibration, thereby improving the stability and clarity of the test data. The present application effectively solves the challenges faced by the deposition layer morphology test in the process of laser additive manufacturing by combining the improved monitoring equipment and deep learning technology, improves the test quality and reliability, and specifically includes the following steps:

[0006] Step 1, collect the images of the molten pool area and the vibration data.

[0007] Step 2 involves inputting the acquired molten pool region image and vibration data into a deep autoencoder for processing, thereby generating a reconstructed image. The deep autoencoder includes an encoder, a decoder, and a loss function. The encoder uses a convolutional neural network structure to convert the input image and vibration data into a low-dimensional representation, and the decoder utilizes this low-dimensional representation to reconstruct a stable and accurate image.

[0008] Step 3: Evaluate the quality and stability of the reconstructed image using PSNR and SSIM. If the image quality does not meet the preset standards, the system will calculate the optimal lighting angle and camera position required to improve the image quality.

[0009] Step 4: Adjust the angle and position of the LED lights and camera using the hollow shaft motor to ensure that the light path evenly covers the surface of the molten pool and captures all possible areas of light and shadow.

[0010] Step 5: If the image quality meets the preset standard, output the reconstructed image.

[0011] Preferably, the loss function is:

[0012]

[0013] Among them, L total Let λ(t) be the loss value, and L be the weight parameter that is dynamically adjusted according to the training process. recon To reconstruct the error loss, L stablity For stability loss, η is used to adjust the contribution of nonlinear interaction terms to the total loss, α is the reconstruction error adjustment coefficient, β is the stability adjustment coefficient, γ is the error and stability correlation adjustment coefficient, δ is the deep feature interaction information adjustment coefficient, P(z|I) is the conditional probability distribution of the latent representation of the original image I after encoding, and P(z|K) is the conditional probability distribution of the latent representation of the reconstructed image K after encoding.

[0014] Preferably, the formula for calculating PSNR is:

[0015]

[0016] Where PSNR represents peak signal-to-noise ratio, MAX1 represents the maximum value of the color of the image point, f(V) represents the influence function of the depth autoencoder on image quality, V is the vibration amplitude data received from the depth autoencoder, I represents the original image, K represents the reconstructed image, m represents the image height, and n represents the image width.

[0017] Preferably, the formula for calculating SSIM is:

[0018]

[0019] where SSIM(x, y) represents the structural similarity of images x and y, μ x is the mean value of image x, μ y is the mean value of image y. σ x is the standard deviation of image x, σ y is the standard deviation of image y, σ xy is the covariance. c1, c2 are constants used to maintain stability, g(V) is the mean value coefficient for adjusting the denominator, and h(V) is the variance term coefficient for adjusting the denominator.

[0020] Preferably, the encoder comprises a vibration feature encoder and an image feature encoder, the vibration feature encoder comprising a first input layer one, a first RNN layer one, a first RNN layer two, a first full connection layer one, and a first output feature vector layer connected in sequence. The vibration feature encoder comprises a second input layer, a second convolutional layer one, a second pooling layer one, a second convolutional layer two, a second pooling layer two, a second convolutional layer three, a second full connection layer, and a second output feature vector layer connected in sequence.

[0021] Preferably, the decoder comprises a vibration feature decoder and an image feature decoder, the vibration feature decoder comprising a third input layer one, a third full connection layer one, a third RNN layer one, a third RNN layer two, and a third output layer connected in sequence. The vibration feature decoder comprises a fourth input layer, a fourth full connection layer, a fourth deconvolutional layer one, a fourth up-sampling layer one, a fourth deconvolutional layer two, a fourth up-sampling layer two, a fourth deconvolutional layer three, and a fourth output layer connected in sequence.

[0022] Preferably, according to the evaluation result, the parameters of the encoder and decoder are adjusted through a feedback adjustment mechanism, and the optimized parameters are used to further improve the image processing process.

[0023] Another object of the present application is to provide an image-based intelligent reconstruction additive manufacturing deposition layer topography measurement device for implementing the image-based intelligent reconstruction additive manufacturing deposition layer topography measurement method, which comprises a laser head, a hollow shaft motor, a measured substrate, LED lamps, a laser emitter, a remelted molten pool topography monitoring sensor, and a variable angle camera system, the variable angle camera system comprises a side monitoring CCD and a line laser monitoring CCD, the laser head is installed on the top of the device, the hollow shaft motor is installed on the laser head, the line laser monitoring CCD, the LED lamps, the laser emitter, and the side monitoring CCD are installed on the hollow shaft motor, and the hollow shaft motor is used to drive the line laser monitoring CCD, the LED lamps, the laser emitter, and the side monitoring CCD to rotate around the laser head.

[0024] Preferably, the LED lamps are two and arranged at 180°.

[0025] Preferably, the hollow shaft motor is arranged at the lower part of the laser head and is kept a certain distance from the nozzle.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] 1. The image is stabilized by the deep autoencoder, so that even in the presence of vibration and other interference, a stable image can still be output. Such image output can be directly used for display and further processing, ensuring the stability of the image and the accuracy of the processing.

[0028] 2. By integrating the variable angle camera and the deep autoencoder technology, the present application can eliminate dark areas that cannot be covered by light, and at the same time process vibration interference, so as to obtain high-quality and stable deposition layer surface morphology data, and improve the accuracy and reliability of the test.

[0029] 3. By using the variable angle camera, the present application can obtain image data of the deposition layer surface at different angles, providing more comprehensive information for analysis, and helping to more accurately understand the morphological characteristics of the deposition layer. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is the front view of the device designed by the present application.

[0031] Figure 2 is the side view of the device designed by the present application.

[0032] Figure 3 is the structural framework diagram of the autoencoder proposed by the present application.

[0033] Figure 4 is the structural framework diagram of the encoder part of the present application.

[0034] Figure 5 is the structural framework diagram of the decoder part of the present application.

[0035] Figure 6 is the flow chart of the quality evaluation module of the present application.

[0036] Figure 7 is the flow chart of the feedback adjustment mechanism of the present application. DETAILED DESCRIPTION

[0037] The present application will be further illustrated below in conjunction with the drawings and specific embodiments, and it should be understood that these examples are only used to illustrate the present application and not to limit the scope of the present application. After reading the present application, those skilled in the art can make various equivalent modifications to the present application, which all fall within the scope defined by the appended claims.

[0038] The application discloses a kind of based on image intelligent reconstruction's additive manufacturing deposition layer topography measuring device, for realizing the based on image intelligent reconstruction's additive manufacturing deposition layer topography measuring method, including laser head 100, hollow shaft motor 200, measured substrate 300, LED lamp 202, laser emitter 203 and remelting pool topography monitoring sensor 400 and variable angle camera system, the variable angle camera system includes side monitoring CCD 204, line laser monitoring CCD 202, laser head 100 is installed on the top of equipment, is responsible for emitting laser to melt powder, forms melt pool 301 and creates deposition layer 302 on substrate.The variable angle camera system is composed of a side monitoring CCD and a laser monitoring CCD, is connected with laser head by hollow shaft motor, to capture melt pool image from multiple angles.The hollow shaft motor 200 is installed on the laser head 100, the line laser monitoring CCD 201, LED lamp 202, laser emitter 203, side monitoring CCD 204 are installed on the hollow shaft motor 200, the line laser emitter is arranged 90 ° with line laser monitoring CCD, ensures that line laser is accurately directed to melt pool area.The hollow shaft motor 200 is used to drive line laser monitoring CCD 201, LED lamp 202, laser emitter 203, side monitoring CCD 204 to rotate around laser head 100.

[0039] The illumination device is composed of two LED lights, which are connected with the laser head by the hollow shaft motor. The two LED lights are arranged at 180° to uniformly illuminate the melt pool area. An 808 filter is installed in front of the lens, and the illumination light path points to the melt pool.

[0040] The hollow shaft motor is arranged at the lower part of the laser head and maintains a certain distance from the nozzle, and is used to adjust the angle and position of the illumination system and the camera system. The hollow shaft motor can drive the illumination device, the variable angle camera system and the line laser generator to rotate around the laser head. By improving the illumination device of the laser head, the dark area that cannot be reached by light during the laser additive manufacturing process is eliminated. The vibration amplitude data during the experiment is collected. The collected image data is processed and optimized using a deep autoencoder to reduce the influence of vibration on the test data. The stable image is monitored and output in real time.

[0041] The angle and position of the two LED lights and the variable angle camera are adjusted by the hollow shaft motor to ensure that the light is covered from multiple angles and comprehensively, and the image is captured.

[0042] The back of the measured substrate is provided with a remelting pool topography monitoring sensor, and a computer is further included. The vibration amplitude of the camera and the monitoring platform during the monitoring process is obtained by subtracting the height of the cooled deposition layer.

[0043] During the manufacturing process, the hollow shaft motor 200 can drive the LED light 202, the side monitoring CCD 204, the line laser monitoring CCD 202, and the laser emitter 203 to rotate around the laser head 100.

[0044] The remelted pool morphology monitoring sensor employs an ultrasonic probe located on the back side of the substrate to detect the three-dimensional morphology of the molten pool, thereby achieving high-precision deposition layer morphology testing.

[0045] The camera system of the variable-angle vision measurement device can dynamically adjust the position and angle of the camera to capture all possible areas of light and shadow, as well as detailed information.

[0046] During testing, the laser head is first activated to form the deposition layer, while the illumination and camera systems are simultaneously activated. The illumination system is adjusted via a hollow shaft motor to ensure uniform illumination of the molten pool area, reducing dark areas caused by insufficient lighting. The camera system captures images from different angles for subsequent image analysis.

[0047] The captured image data and real-time measured vibration data are input into a deep autoencoder. The autoencoder first cleans and normalizes the data through a preprocessing layer, and then processes it through the encoder section, including the encoding of vibration and image features. After feature fusion in the latent space, the decoder reconstructs the image and vibration features to produce a stabilized image.

[0048] The quality assessment module evaluates the quality of the reconstructed image, assessing whether the current lighting conditions are sufficient based on the image quality metrics output by the autoencoder. If the image quality does not meet the preset standards, it calculates the optimal lighting angle and camera position required to improve the image quality. Based on the evaluation results, a feedback adjustment mechanism adjusts the encoder and decoder parameters, and the optimized parameters are used to further improve the image processing.

[0049] A method for measuring the morphology of additive manufacturing deposition layers based on intelligent image reconstruction, such as... Figures 3-7 As shown, the specific steps include:

[0050] Step 1: Acquire images and vibration data of the molten pool area.

[0051] Step 2: Input the collected images of the molten pool area and vibration data into a depth autoencoder for processing, and generate a reconstructed image through the depth autoencoder.

[0052] Deep autoencoders optimize the testing accuracy of sedimentary layer morphology by integrating image and vibration data. The input layer of the autoencoder receives image data and vibration amplitude. The stable image generation process of the deep autoencoder involves processing the input image data and vibration amplitude using a pre-trained deep autoencoder to output a stable image.

[0053] In real-time monitoring, the current image data and vibration amplitude are continuously obtained and input into the deep autoencoder, and the stabilized image is directly output for display and further processing.

[0054] The deep autoencoder includes an encoder, a decoder, and a loss function. The encoder uses a convolutional neural network structure to convert the input image data and vibration data into low-dimensional representations. This process focuses on capturing complex image features to ensure accurate encoding of data. The encoder part is composed of two sub-modules: a convolutional neural network for extracting image features and a recurrent neural network for processing vibration data to capture time-series vibration features. These features are combined in a feature fusion layer to ensure effective integration of image and vibration information. As shown in Figure 4 The encoder includes a vibration feature encoder and an image feature encoder. The vibration feature encoder includes a first input layer, a first RNN layer, a second RNN layer, a first full connection layer, and a first output feature vector layer. The vibration feature encoder includes a second input layer, a second convolutional layer, a second pooling layer, a second convolutional layer, a second pooling layer, a second convolutional layer, a second full connection layer, and a second output feature vector layer.

[0055] The decoder uses these low-dimensional representations to reconstruct stable and accurate images. Advanced image restoration techniques are used to ensure the quality and detail fidelity of the reconstructed images. The decoder part uses a transposed convolutional network to reconstruct images from the fused features, while adjusting the reconstruction strategy based on vibration data to suppress distortions caused by vibration. As shown in Figure 5 The decoder includes a vibration feature decoder and an image feature decoder. The vibration feature decoder includes a third input layer, a third full connection layer, a third RNN layer, a fourth RNN layer, and a third output layer. The vibration feature decoder includes a fourth input layer, a fourth full connection layer, a fourth deconvolutional layer, a fourth upsampling layer, a fourth deconvolutional layer, a fourth upsampling layer, a fourth deconvolutional layer, and a fourth output layer.

[0056] The loss function includes mean square error to evaluate the difference between the reconstructed image and the original image, and total variation regularization to enhance the stability of the image, thereby generating stable and real image data. Through this method, the autoencoder can effectively recover clear and high-quality images from images affected by vibration, improving the accuracy and reliability of the deposition layer topography test.

[0057] In the generation process of stable images, in order to optimize this process, a comprehensive loss function is designed, including reconstruction error loss and stability loss. The reconstruction error is evaluated by mean square error, which focuses on minimizing the difference between the reconstructed image and the original image; the stability loss uses the total variation of the image to measure, which aims to reduce the image jitter and blur caused by vibration, so as to ensure the smooth output of the image. Its purpose is to optimize the parameters of the deep autoencoder and generate stable and real image data. The loss function is composed of two parts: reconstruction error loss L recon and stability loss L stablity . In order to fully consider the independence and mutual influence of the two losses, a dynamic weight adjustment strategy is introduced: L total = λ(t) L recon + (1- λ(t)) L stablity , where λ(t) is a weight parameter that is dynamically adjusted according to the training process. In order to further enhance the self-adjusting ability of the model, the invention also introduces a combination containing a nonlinear term: L interation = α L recon 2 + β L stablity 2 + γ L recon L stablity . Where α, β, and γ are adjustment coefficients used to control the influence strength of each loss term in the autoencoder, so that the model can enhance the stability while ensuring the reconstruction quality, which is optimized through experiments. Finally, the weight adjustment and nonlinear interaction effect are combined to form a comprehensive loss function:

[0058] L total = λ(t) L recon + (1- λ(t)) L stablity + η (α L recon 2 + β L stablity 2 + γ L recon L stablity ),

[0059] where η is used to adjust the contribution size of the nonlinear interaction term to the total loss, which is adjusted through experiments.

[0060] In addition, in order to further improve the quality of image reconstruction, deep feature interaction information is introduced: where P(z|I) and P(z|K) are the conditional probability distributions of the latent representations of the original image I and the reconstructed image K after encoding, respectively, so as to obtain the final loss function:

[0061]

[0062]

[0063] where L total is the loss value, λ(t) is a weight parameter dynamically adjusted according to the training process, L recon is the reconstruction error loss, L stablity is the stability loss, η is a parameter used to adjust the contribution size of the nonlinear interaction term to the total loss, which is adjusted through experiments, α is the reconstruction error adjustment coefficient, β is the stability adjustment coefficient, γ is the error and stability related adjustment coefficient, δ is the depth feature interaction information adjustment coefficient, P(z|I) is the conditional probability distribution of the latent representation of the original image I after encoding, P(z|K) is the conditional probability distribution of the latent representation of the reconstructed image K after encoding.

[0064] By optimizing this composite loss function, the model can generate more stable and high-quality image data. This method ensures that the generated images are both realistic and stable, effectively reflecting the detailed features of the deposition layer surface.

[0065] The data set of the deep autoencoder comes from real-time monitoring data and image acquisition during the laser metal deposition process. Specifically, during the processing, the monitoring equipment records the images of the molten pool and the deposition layer, and at the same time, by differentiating the height of the cooled deposition layer, the vibration amplitude can be indirectly obtained. The captured images are paired with the corresponding vibration amplitudes, and the real-time measured vibration amplitudes are labeled for each set of image data. Finally, the processed image data and the corresponding vibration data are integrated into a training set for the training of the deep autoencoder.

[0066] The specific steps to eliminate vibration are: first, use a variable angle camera to capture images from multiple angles, and simultaneously collect vibration data through a vibration sensor. These data are preprocessed and input into the deep autoencoder, where the encoder part converts the image and vibration data into low-dimensional feature representations. In the decoder stage, these low-dimensional representations are used to reconstruct the image, while a comprehensive loss function is applied to optimize the stability and quality of the reconstructed image. This process includes dynamic weight adjustment and nonlinear dependence to finely control the trade-off between image reconstruction and stability.

[0067] Step 3, as shown in Figure 6 and 7 , the quality and stability of the reconstructed image are evaluated using PSNR and SSIM. If the image quality does not meet the preset standard, the system will calculate the optimal lighting angle and camera position required to improve the image quality.

[0068] The specific steps for eliminating the light dark zone are as follows: first, the image of the molten pool area is captured by the variable view camera system, and the relevant vibration data is collected by the line laser. Then, the image and vibration data are input into the autoencoder for processing, and the autoencoder generates a reconstructed image and evaluates its quality and stability using PSNR and SSIM. According to the image quality indicators output by the autoencoder, the system evaluates the current lighting conditions. If the image quality does not meet the preset standard, the system will calculate the optimal lighting angle and camera position required to improve the image quality. Then, the angle and position of the LED light and camera are adjusted by the hollow shaft motor to ensure that the light path uniformly covers the molten pool surface and captures all possible light dark zones. This process is repeated until the image quality is optimal. This integrated feedback control mechanism with an autoencoder not only improves the automation of the test, but also ensures the accuracy and repeatability of the deposited layer topography test under various operating conditions.

[0069] The formula for calculating PSNR is:

[0070]

[0071]

[0072] where PSNR represents the peak signal-to-noise ratio, MAX1 represents the maximum value of the image point color, f(V) represents the influence function of the deep autoencoder on image quality to automatically adjust the evaluation parameters, V is the vibration amplitude data received from the deep autoencoder, I represents the original image, K represents the reconstructed image, m represents the image height, and n represents the image width.

[0073] Preferably, the formula for calculating SSIM is:

[0074]

[0075] where SSIM(x, y) represents the structural similarity of images x and y, μ x is the mean value of image x, μ y is the mean value of image y. σ x is the standard deviation of image x, σ y is the standard deviation of image y, σ xy is the covariance. c1, c2 are constants used to maintain stability, g(V) is a coefficient for adjusting the mean value in the denominator, and h(V) is a coefficient for adjusting the variance term in the denominator. g(V) and h(V) are used to adjust the mean and variance terms in the denominator, reflecting the influence of vibration on image stability, which is obtained by statistical analysis of image data under different vibration conditions.

[0076] Step 4: Adjust the angle and position of the LED light and camera through the hollow shaft motor to ensure that the light path uniformly covers the surface of the molten pool and captures all possible light dark areas.

[0077] Step 5: If the image quality meets the preset standard, output the reconstructed image.

[0078] Step 6: According to the evaluation results, adjust the parameters of the encoder and decoder through the feedback adjustment mechanism, and the optimized parameters are used to further improve the image processing process.

[0079] First, use the variable angle camera to collect multiple images and use the vibration sensor to obtain the corresponding vibration amplitude data. Then, input these image data into the deep autoencoder for data preprocessing to ensure data quality and consistency. The encoder part includes a vibration feature encoder and an image feature encoder, which encode the input data into a low-dimensional representation. Then, feature fusion and dimension reduction are performed in the latent space to ensure that the data representation contains vibration and image features. The decoder part includes a vibration feature decoder and an image feature decoder to reconstruct the image and vibration features from the low-dimensional representation. During the generation of stable images, the quality assessment module uses PSNR and SSIM indicators to evaluate the quality of the reconstructed images. Subsequently, the feedback adjustment mechanism adjusts the parameters of the encoder and decoder according to the evaluation results. This process optimizes the parameters of the deep autoencoder to minimize the reconstruction error loss and stability loss, generating stable and realistic image data. The final stable image data can be used for image-based intelligent reconstruction of additive manufacturing deposition layer topography measurement, ensuring that the test data has high clarity and stability, accurately reflecting the topographic features of the deposition layer surface.

[0080] Through the improvement of the lighting device in the monitoring device, the purpose is to eliminate the dark areas that cannot be involved by light in the laser metal deposition process; a variable angle camera is installed in the test device to obtain deposition layer surface image data at multiple angles. The amplitude of the vibration during the experiment is tested, and a deep autoencoder is used to manufacture inverse radians to keep the corrected image level, solving the vibration problem during the experiment. By solving these problems, the surface topography of the deposition layer is tested with high precision.

[0081] This embodiment improves the lighting system of the monitoring device to solve the problem of dark areas under laser irradiation, and uses deep learning technology, especially deep autoencoder, to process and optimize the collected video data to reduce the impact of vibration on test data and improve the stability and clarity of test data. The advantages of this invention include solving the problems of insufficient light and vibration interference, improving the quality and reliability of test data.

[0082] The above merely describes the preferred embodiments of the present application, and it should be pointed out that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A method for measuring the morphology of a deposited layer in additive manufacturing based on image-based intelligent reconstruction, characterized in that, The method comprises the following steps: Step 1, collect the molten pool area image and vibration data; Step 2, input the collected molten pool area image and vibration data into a deep autoencoder for processing, and generate a reconstructed image through the deep autoencoder; the deep autoencoder comprises an encoder, a decoder and a loss function, the encoder uses a convolutional neural network structure, and is used for converting input image data and vibration data into a low-dimensional representation, and the decoder uses the low-dimensional representation to reconstruct a stable and accurate image; Step 3, and evaluate the quality and stability of the reconstructed image using PSNR and SSIM; If the image quality does not reach the preset standard, the system will calculate the optimal illumination angle and camera position required to improve the image quality; Step 4, adjust the angle and position of the LED light and camera through the hollow shaft motor, ensure that the light path uniformly covers the molten pool surface, and capture all possible light dark areas; Step 5, if the image quality reaches the preset standard, output the reconstructed image.

2. The method of claim 1, wherein, The loss function is: wherein, L total is a loss value, λ(t) is a weight parameter dynamically adjusted according to a training process, L recon is a reconstruction error loss, L stablity is a stability loss, η is a parameter for adjusting the contribution size of the nonlinear interaction term to the total loss, α is a reconstruction error adjustment coefficient, β is a stability adjustment coefficient, γ is an error and stability related adjustment coefficient, δ is a deep feature interaction information adjustment coefficient, and P(z|I) is a conditional probability distribution of the latent representation of the original image I after encoding, and P(z|K) is a conditional probability distribution of the latent representation of the reconstructed image K after encoding.

3. The method of claim 2, wherein the method further comprises: The calculation formula of PSNR is: Wherein, PSNR represents the peak signal-to-noise ratio, MAX1 represents the maximum value of the color of the image point, f(V) represents the influence function of the deep autoencoder on the image quality, V is the vibration amplitude data received from the deep autoencoder, I represents the original image, K represents the reconstructed image, m represents the image height, and n represents the image width.

4. The method of claim 3, wherein the method further comprises: The calculation formula of SSIM is: where SSIM(x, y) represents the structural similarity of images x and y, μ x is the mean value of image x, μ y is the mean value of image y; σ x is the standard deviation of image x, σ y is the standard deviation of image y, σ xy is the covariance; c1, c2 are constants used to maintain stability, g(V) is a coefficient for adjusting the mean value in the denominator, and h(V) is a coefficient for adjusting the variance term in the denominator.

5. The method of claim 4, wherein: The encoder comprises a vibration feature encoder and an image feature encoder, the vibration feature encoder comprises a first input layer one, a first RNN layer one, a first RNN layer two, a first full connection layer one and a first output feature vector layer connected in sequence; and the image feature encoder comprises a second input layer, a second convolutional layer one, a second pooling layer one, a second convolutional layer two, a second pooling layer two, a second convolutional layer three, a second full connection layer and a second output feature vector layer connected in sequence.

6. The method of claim 5, wherein: The decoder comprises a vibration feature decoder and an image feature decoder, the vibration feature decoder comprises a third input layer one, a third full connection layer one, a third RNN layer one, a third RNN layer two and a third output layer connected in sequence; and the image feature decoder comprises a fourth input layer, a fourth full connection layer, a fourth deconvolutional layer one, a fourth upsampling layer one, a fourth deconvolutional layer two, a fourth upsampling layer two, a fourth deconvolutional layer three and a fourth output layer connected in sequence.

7. The method of claim 6, wherein: According to the evaluation result, the parameters of the encoder and decoder are adjusted through a feedback adjustment mechanism, and the optimized parameters are used for further improving the image processing process.

8. An apparatus for measuring the morphology of a deposited layer in additive manufacturing based on image-based intelligent reconstruction, characterized in that it comprises: The method for measuring the deposition layer topography of the additive manufacturing based on the image intelligent reconstruction of claim 1 comprises a laser head (100), a hollow shaft motor (200), a measured substrate (300), an LED lamp (202), a laser emitter (203), a remelting pool topography monitoring sensor (400), and a variable angle camera system comprising a side monitoring CCD (204) and a line laser monitoring CCD (201). The laser head (100) is installed on the top of the equipment, the hollow shaft motor (200) is installed on the laser head (100), and the line laser monitoring CCD (201), the LED lamp (202), the laser emitter (203), and the side monitoring CCD (204) are installed on the hollow shaft motor (200). The hollow shaft motor (200) is used to drive the line laser monitoring CCD (201), the LED lamp (202), the laser emitter (203), and the side monitoring CCD (204) to rotate around the laser head (100).

9. The apparatus for measuring the morphology of a deposited layer in additive manufacturing based on image intelligent reconstruction according to claim 8, characterized in that: The LED lamp (202) has two, which are arranged at 180°.

10. The apparatus for measuring the morphology of a deposited layer in additive manufacturing based on image intelligent reconstruction according to claim 9, characterized in that: The hollow shaft motor is arranged on the lower part of the laser head and keeps a certain distance from the nozzle.

Images