An on-line spectral quantitative detection method and device for arc wire additive manufacturing

Abstract

The present invention relates to the technical field of porosity monitoring in arc additive manufacturing, and particularly to an online spectral quantitative detection method for arc wire additive manufacturing. The Alexnet network model is improved to obtain a lightweight Alexnet network model and trained. A spectrometer is started to collect the arc spectral signals generated during the manufacturing of aluminum alloy components by the arc additive manufacturing device as the input of the lightweight Alexnet network model, and the lightweight Alexnet network model outputs the porosity degree of the corresponding printed aluminum alloy component. After the training of the lightweight Alexnet network model of the present invention is completed, the spectrometer records the changes in the arc spectral signals during the arc additive manufacturing process and converts them from a one-dimensional array to a two-dimensional array through python as the input of the lightweight Alexnet to realize the online detection of porosity defects during the aluminum alloy arc additive manufacturing process, timely predict the generation of pores and take remedial measures in advance to reduce pores, and obtain high-quality aluminum alloy components.

Description

Technical Field

[0001] The present invention relates to the technical field of porosity monitoring in arc additive manufacturing, and particularly relates to an online spectral quantitative detection method for arc wire additive manufacturing. Background Art

[0002] Porosity is the most common defect in the process of arc additive manufacturing. Taking the arc additive manufacturing of aluminum alloy as an example, in the process of arc additive manufacturing of aluminum alloy, the welding wire melts to form liquid aluminum. A large amount of hydrogen can be dissolved in the liquid aluminum. The molten metal in the molten pool cools and solidifies rapidly, while solid aluminum hardly dissolves hydrogen. Therefore, the hydrogen existing in the liquid aluminum remains in the solidified aluminum alloy component without time to escape, forming pores. Pores will not only reduce the strength and hardness of the printed component, resulting in a decline in the mechanical properties of the product, but also lead to a reduction in the fatigue life of the printed component, making the product prone to cracks or even fractures. In addition, pores will also affect the corrosion resistance of the component.

[0003] Currently, the monitoring of pores in the process of arc additive manufacturing of aluminum alloy is achieved through off-line detection methods after processing. After processing, methods such as CT are used to detect the internal pores of the printed component. These off-line detection methods have disadvantages such as large time consumption, high cost, and inability to predict and avoid the generation of pores, which hinder the further development of arc additive manufacturing of aluminum alloy. The applicant hereby proposes an online spectral quantitative detection method for arc wire additive manufacturing. Summary of the Invention

[0004] To solve the deficiencies of the existing technology, the present invention provides an online spectral quantitative detection method for arc wire additive manufacturing.

[0005] The technical solution of the present invention is as follows:

[0006] The present invention provides an online spectral quantitative detection method for arc wire additive manufacturing, including the following steps:

[0007] S1: Improve the Alexnet network model to obtain a lightweight Alexnet network model;

[0008] S2: Train the lightweight Alexnet network model;

[0009] S3: Start the arc additive manufacturing device, and collect the arc spectral signal during the manufacturing process of the arc additive manufacturing device by a spectrometer. The arc spectral signal is a one-dimensional array;

[0010] S4: Reconstruct and splice the arc spectral signal from a one-dimensional array into a two-dimensional array by using the resize function provided by python;

[0011] S5: Use the arc spectrum signal reconstructed and spliced into a two-dimensional array by Python as the input of the lightweight Alexnet network model, and the lightweight Alexnet network model will output the porosity degree of the corresponding printed aluminum alloy component.

[0012] Further, step S1 includes:

[0013] Replace the last three ordinary convolutional layers in the Alexnet network model with three GHPA modules;

[0014] Use the solubility curve of hydrogen in the metal material as the activation function;

[0015] Use DW convolution to replace the ordinary convolution of the first two ordinary convolutional layers in the Alexnet network model.

[0016] Further, step S2 includes:

[0017] S201: Input the two features of electron density and electron temperature into the lightweight Alexnet network model manually;

[0018] S202: Integrate the artificial features with the features learned by the lightweight Alexnet network model through deep learning;

[0019] S203: Use the output features of the ordinary convolutional layer as the input of the fully connected layer to realize the prediction of electron density and electron temperature;

[0020] S204: Use the features output by the third GHPA module as the input of another fully connected layer to realize the prediction of porosity;

[0021] S205: Finally, use the difference between the predicted electron temperature, electron density, porosity and the actual values as the loss function, and the specific loss function is:

[0022] loss = a1(T e_p - T e ) 2 + a2(D e_p - D e ) 2 + a3(p _p - p) 2

[0023] where a1, a2, and a3 are the coefficients of the three different losses, taking 0.1, 0.1, and 0.8 in this algorithm, Te_p, Te, De_p, De, p_P, and p are the predicted electron temperature, electron temperature, predicted electron density, electron density, predicted porosity, and porosity respectively.

[0024] Furthermore, the input layer of the lightweight Alexnet network model can accept images with an input feature size of 64×64.

[0025] The present invention also provides an on-line spectral quantitative detection device for arc wire additive manufacturing, which adopts the above-mentioned on-line spectral quantitative detection method for arc wire additive manufacturing. The device includes an arc additive manufacturing device, a spectrometer and an external computer. The spectrometer is fixedly installed on the welding robotic arm of the arc additive manufacturing device through a fixture and is connected to the external computer. The spectrometer probe is vertically oriented towards the side of the welding wire; the external computer is installed with python for reconstructing and splicing arc spectral signals and a lightweight Alexnet network model for analyzing and predicting the porosity degree of aluminum alloy components.

[0026] Furthermore, the distance between the spectrometer probe and the central area of the welding wire is 2.5 - 3.5 cm.

[0027] Furthermore, the fixture is a crab claw clamp.

[0028] The beneficial effects achieved by the present invention are as follows:

[0029] After the lightweight Alexnet network model of the present invention is trained, the spectrometer records the changes in the arc spectral signals during the arc additive manufacturing process and converts them from a one-dimensional array to a two-dimensional array through python as the input of the lightweight Alexnet, so as to realize the on-line detection of pore defects during the aluminum alloy arc additive manufacturing process, predict the generation of pores in time and take remedial measures in advance to reduce pores, and obtain high-quality aluminum alloy components.

[0030] Compared with the Alexnet network model, the lightweight Alexnet network model in the present invention reduces the number of parameters by 6% and the number of floating-point operations by 50% while maintaining the same accuracy, improving the calculation speed. Description of the Drawings

[0031] Figure 1 It is a schematic diagram of the processing flow of the present invention.

[0032] Figure 2 It is a schematic diagram of the technical route of the present invention.

[0033] Figure 3 It is a schematic diagram of the monitoring device of the present invention.

[0034] Figure 4 It is a structural diagram of the improved Alexnet network of the present invention.

[0035] Figure 5 It is a schematic diagram of per-channel convolution of the present invention. Detailed Embodiments

[0036] To facilitate the understanding of the present invention by those skilled in the art, the specific embodiments of the present invention will be described below with reference to the accompanying drawings. The embodiments described in the present invention are only a part of the embodiments, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0037] As Figures 1-4 shown, an on-line spectral quantitative detection device for arc wire additive manufacturing in this embodiment of the present invention includes an arc additive manufacturing device, a spectrometer (no specified model, and the example code is written according to the ATP2400 spectrometer of Opto-Tech Co., Ltd.), and an external computer. The spectrometer is fixedly installed on the welding robotic arm of the arc additive manufacturing device through a fixture and connected to the external computer. The spectrometer probe is vertically oriented towards the side of the welding wire, realizing the synchronous movement of the spectrometer and the welding robotic arm and collecting arc spectral signals. The external computer is installed with python for reconstructing and splicing arc spectral signals and a lightweight Alexnet network model for analyzing and predicting the porosity of aluminum alloy components.

[0038] The distance between the spectrometer probe and the central area of the welding wire is 2.5 - 3.5 cm.

[0039] The fixture is preferably a crab claw clamp.

[0040] The present invention proposes an on-line spectral quantitative detection method for arc wire additive manufacturing based on the above-mentioned arc wire additive manufacturing device. This method mainly uses a spectrometer to collect arc spectral signals generated during the manufacturing process of aluminum alloy components by the arc wire additive manufacturing device. Python reconstructs and splices the spectral signals into a two-dimensional array and predicts the porosity of the aluminum alloy components during the manufacturing process through a lightweight Alexnet. The method mainly includes the following steps:

[0041] S1: Improve the Alexnet network model to obtain a lightweight Alexnet network model.

[0042] Specifically, the improvement of the Alexnet network model includes the following steps:

[0043] S101: Replace the last three ordinary convolutional layers in the Alexnet network model with three GHPA modules; the original Alexnet has 8 layers, and its specific structure is successively 5 convolutional layers, 2 fully connected hidden layers, and 1 fully connected output layer; the specific structure of the replaced Alexnet is successively 2 convolutional layers, 3 GHPA modules, 2 fully connected hidden layers, and 1 fully connected output layer; where the GHPA (Global Hierarchical Pooling Attention) module uses the HPA (Hadamard product attention mechanism), by grouping the input features, performing HPA operations on different axes, and extracting features from multiple perspectives; thereby improving the feature extraction ability of the lightweight Alexnet network model, reducing the number of parameters, and making its calculation speed faster.

[0044] Specifically, the GHPA module has only linear complexity, and its algorithm pseudocode is as follows:

[0045]

[0046] Evenly split the input features into 4 groups (x1, x2, x3, x4) along the channel dimension, perform HPA on the height-width, channel-height, and channel-width axes of x1, x2, x3, while x4 only uses DW on the feature map; finally, connect the four groups along the channel dimension and use another DW to integrate information from different perspectives.

[0047] S102: Use the solubility curve of hydrogen in the metal material as the activation function; during the arc additive manufacturing process, the generation of pores is mainly caused by the huge difference in the solubility of hydrogen between the liquid and solid states of the metal material. Embedding this physical information as the activation function into the lightweight Alexnet network model can improve the performance (accuracy) and interpretability of the lightweight Alexnet network model, and at the same time reduce the dependence of the lightweight Alexnet network model on data.

[0048] S103: Use DW convolution to replace the ordinary convolution of the first two ordinary convolutional layers in the Alexnet network model; further reduce the number of parameters of the model and reduce the calculation time of the model.

[0049] The DW (Depthwise Separable) convolution is a convolution operation that divides the standard ordinary convolution operation into two steps: depth convolution and pointwise convolution; by combining depth convolution and pointwise convolution, DW convolution can reduce the number of parameters and the amount of calculation. Compared with the standard ordinary convolution operation, DW convolution only needs to perform a small-scale convolution operation at each position, thereby providing more efficient calculation and a more lightweight model design, and achieving more efficient feature extraction and model compression.

[0050] As Figure 5 shown, the per-channel convolution: a convolutional kernel only performs convolution on one channel to obtain a feature map.

[0051] The pointwise convolution: a common 1×1 convolution is used to achieve the transformation of the channel dimension.

[0052] The metal material is aluminum alloy, etc.;

[0053] S2: Train the lightweight Alexnet network model;

[0054] Specifically, the training of the lightweight Alexnet network model includes the following steps:

[0055] S201: Input the two features of electron density and electron temperature into the lightweight Alexnet network model manually;

[0056] S202: Fuse the artificial features with the features learned by deep learning of the lightweight Alexnet network model:

[0057] The features learned by deep learning are specifically a 1152-dimensional feature vector directly extracted by the encoder part of the lightweight Alexnet network model, which increases the feature learning ability.

[0058] S203: Use the output features of the ordinary convolutional layer as the input of the fully connected layer to achieve the prediction of electron density and electron temperature;

[0059] S204: Use the features output by the third-layer GHPA module as the input of another fully connected layer to achieve the prediction of porosity;

[0060] S205: Finally, use the difference between the predicted electron temperature, electron density, porosity and the actual values as the loss function, and its specific loss function is:

[0061] loss=a1(T e_p -T e ) 2 +a2(D e_p -D e ) 2 +a3(p _p -p) 2

[0062] Among them, a1, a2, and a3 are coefficients of three different losses, which are taken as 0.1, 0.1, and 0.8 in this algorithm. Te_p, Te, De_p, De, p_P, and p are the predicted electron temperature, electron temperature, predicted electron density, electron density, predicted porosity, and porosity, respectively. Embedding artificial features into the lightweight Alexnet network model through auxiliary tasks can better learn the discriminative features and greatly improve the prediction accuracy of the lightweight Alexnet network model.

[0063] S3: starting the arc additive manufacturing device, and using a spectrometer to collect arc spectrum signals during the manufacturing process of the arc additive manufacturing device, where the arc spectrum signals are a one-dimensional array;

[0064] S4: The arc spectrum signal is reconstructed and concatenated from a one-dimensional array into a two-dimensional array by using the resize function provided by Python. A lightweight Alexnet network model is implemented to automatically extract features.

[0065] S5: The arc spectrum signal reconstructed and spliced ​​into a two-dimensional array by python is used as the input of the lightweight Alexnet network model, and the lightweight Alexnet network model will output the porosity of the corresponding printed aluminum alloy component.

[0066] After the lightweight Alexnet network model of the present invention is trained, the spectrometer records the changes in the arc spectrum signal during the arc additive manufacturing process and converts it from a one-dimensional array to a two-dimensional array through python as the input of the lightweight Alexnet, so as to realize online detection of porosity defects in the aluminum alloy arc additive manufacturing process, and can timely predict the generation of porosity and take remedial measures in advance to reduce the porosity, thereby obtaining high-quality aluminum alloy components.

[0067] The specific settings of the lightweight Alexnet network model in the present invention (taking the ATP2400 spectrometer as an example) are as follows:

[0068] Its input layer can accept input feature images of size 64×64.

[0069] The convolutional layer, pooling layer and CHPA module structure and specific parameters are shown in the following table:

[0070]

[0071] The main task fully connected layer structure and specific parameters are shown in the following table:

[0072] Serial number Input dimension Output dimension Drop out ratio Dropl None None 0.5 FCl 1154 2048 None Drop1 None None 0.5 FC2 2048 2048 None FC3 2048 2 None

[0073] The auxiliary fully connected layer structure and specific parameters are shown in the following table:

[0074] Serial number Input dimension Output dimension Drop out ratio Dropl None None 0.5 FCl 1152 2048 None Drop1 None None 0.5 FC2 2048 2048 None FC3 2048 2 None

[0075] The performance comparison between the Alexnet network model and the lightweight Alexnet network model is shown in the following table:

[0076]

[0077] In the lightweight Alexnet network model of the present invention, without changing the accuracy, the number of parameters is reduced by 6%, and the number of floating-point operations is reduced by 50%, improving the calculation speed.

[0078] The embodiments of the present invention described above do not limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.

Claims

1. An online spectral quantitative detection method for arc fuse additive manufacturing, characterized in that: S1: Improve the Alexnet network model to obtain a lightweight Alexnet network model, and replace the last three ordinary convolutional layers in the Alexnet network model with three-layer GHPA modules; S2: Train the lightweight Alexnet network model; S3: starting the arc additive manufacturing device, and using a spectrometer to collect arc spectrum signals during the manufacturing process of the arc additive manufacturing device, where the arc spectrum signals are a one-dimensional array; S4: The arc spectrum signal is reconstructed and concatenated from a one-dimensional array into a two-dimensional array by using the resize function provided by Python; S5: The arc spectrum signal reconstructed and spliced ​​into a two-dimensional array by python is used as the input of the lightweight Alexnet network model, and the lightweight Alexnet network model will output the porosity of the corresponding printed aluminum alloy component.

2. The method for online spectral quantitative detection of arc fuse additive manufacturing according to claim 1, characterized in that In S1: the solubility curve of hydrogen in metal materials is used as the activation function.

3. The method for online spectral quantitative detection of arc fuse additive manufacturing according to claim 1 or 2, characterized in that In S1: DW convolution is used to replace the ordinary convolution of the first two layers of ordinary convolution layers in the Alexnet network model.

4. The method for online spectral quantitative detection of arc fuse additive manufacturing according to claim 1, characterized in that In S2: S201: The two features of electron density and electron temperature are manually input into the lightweight Alexnet network model; S202: Combine the manually input features with the features learned by the lightweight Alexnet network model; S203: Use the output features of the common convolutional layer as the input of the fully connected layer to predict the electron density and electron temperature; S204: Use the features output by the third-layer GHPA module as the input of another fully connected layer to predict the porosity; S205: Finally, the difference between the predicted electron temperature, electron density, porosity and the actual value is used as the loss function. The specific loss function is: loss=a1(T e_p -T e ) 2 +a2(D e_p -D e ) 2 +a3(p _p -p) 2 Among them, a1, a2, and a3 are coefficients of three different losses, which are taken as 0.1, 0.1, and 0.8 in the algorithm. Te_p, Te, De_p, De, p_P, and p are the predicted electron temperature, electron temperature, predicted electron density, electron density, predicted porosity, and porosity, respectively.

5. The method for online spectral quantitative detection of arc fuse additive manufacturing according to claim 4, characterized in that: The input layer of the lightweight Alexnet network model can accept images with an input feature size of 64×64.

6. An arc fuse additive manufacturing online spectral quantitative detection device used in an arc fuse additive manufacturing online spectral quantitative detection method according to any one of claims 1 to 5, characterized in that: The invention comprises an arc additive manufacturing device, a spectrometer and an external computer. The spectrometer is fixedly mounted on a welding robot arm of the arc additive manufacturing device through a clamp and is connected to the external computer. The probe of the spectrometer is perpendicular to the side of the welding wire. The external computer is installed with a python for reconstructing spliced ​​arc spectrum signals and a lightweight Alexnet network model for analyzing and predicting the porosity of aluminum alloy components.

7. The arc fuse additive manufacturing online spectral quantitative detection device according to claim 6, characterized in that: The distance between the spectrometer probe and the center area of ​​the welding wire is 2.5-3.5 cm.

8. The arc fuse additive manufacturing online spectral quantitative detection device according to claim 7, characterized in that: The clamp is a crab claw clamp.

Images