A metal additive manufacturing time prediction system and method based on graphical processing

Abstract

The present invention relates to the field of metal additive manufacturing, and in particular to a metal additive manufacturing time prediction system and method based on graphical processing, wherein the system comprises a data extraction and storage module, a preprocessing module, and a processing time prediction module; the data extraction and storage module comprises a data acquisition unit and a data storage unit; the preprocessing module is used to preprocess data information to obtain preprocessed data, and associate the preprocessing information of each layer of all training parts with the corresponding processing time information to form a training database; the processing time prediction module is used to predict the processing time of the parts to be processed according to the training database constructed by the preprocessing data module. The present invention can effectively solve the problem in the prior art that the cross-sections of parts are different and the time prediction of layer manufacturing is difficult, effectively reducing the difficulty of prediction and improving the accuracy of prediction.

Description

Technical Field

[0001] The present invention relates to the field of metal additive manufacturing, and in particular to a metal additive manufacturing time prediction system and method based on graphical processing. Background Art

[0002] Rapid additive manufacturing is a general term for rapid prototyping technologies that are directly driven by digital models to rapidly manufacture three-dimensional physical entities of arbitrary complex shapes. Rapid additive manufacturing is based on the manufacturing concept of "layered manufacturing and layer-by-layer stacking". It is a green and intelligent "additive" manufacturing, which is different from the traditional "subtractive" manufacturing model. Additive manufacturing integrates virtual design and digital manufacturing. Based on the three-dimensional model constructed on the computer, special metal materials, non-metallic materials or medical biomaterials are stacked layer by layer through software and CNC systems in the manner of extrusion, sintering, melting, light curing, spraying, etc. to produce physical objects. Rapid additive manufacturing can produce products in a short time. Compared with traditional machining machine tools and mold manufacturing, it has many advantages such as low cost, short cycle, simple modification, and stable size.

[0003] Current additive manufacturing technology generally roughly estimates the manufacturing time of a part based on data such as the part's volume, height, XY axis length, and projection area. However, in the actual additive manufacturing process, the placement of parts, the direction and sequence of scan lines, and fluctuations in the working time of auxiliary mechanisms will all affect the manufacturing time of the part. If manufacturing is performed according to the software's preset time, there will be differences between the finished product and the design model. For example, in the metal additive manufacturing process, due to the existence of stress, the processing area needs to be divided into many small areas and scanned in a jumpy manner, which greatly increases the difficulty of time prediction. The time prediction accuracy is low, resulting in a large difference between the finished product and the design model, a low product qualification rate, and a high production cost. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a metal additive manufacturing time prediction system based on graphical processing, comprising a data extraction and storage module, a preprocessing module, a processing time prediction module and a data output module.

[0005] The data extraction and storage module includes a data acquisition unit and a data storage unit. The data acquisition unit is used to import training parts or parts to be processed, and collect data information of the training parts or parts to be processed and processing time information of the training parts. The data information includes scanning path information of each layer of the training parts or parts to be processed. The format of the data information is, for example, a picture; the data storage unit is used to store the data information and time information collected by the data acquisition unit.

[0006] The preprocessing module is used to preprocess the data information to obtain preprocessed data, and is also used to associate the preprocessing information of each layer of all training parts with the corresponding processing time information to form a training database.

[0007] The processing time prediction module is used to predict the processing time of the part to be processed according to the training database constructed by the preprocessing data module, and the processing time includes the processing time of each layer of the part to be processed and the total processing time.

[0008] The processing time prediction module includes at least one module training unit, which is used to train the preprocessing data and processing time of the currently imported training parts and obtain a corresponding prediction model. The prediction model predicts the processing time and total processing time of each layer of the part to be processed based on the preprocessing data of the part to be processed.

[0009] The module training unit includes a convolutional neural model training unit and an SVM model training unit, wherein the convolutional neural model training unit is used to train the preprocessing data and processing time of the training parts to obtain a convolutional network model;

[0010] The SVM model training unit is used to perform training according to the preprocessing data and processing time of the training parts to obtain the SVM model.

[0011] Preferably, the convolutional neural model training unit and the convolutional network model both include an input layer, a first convolutional layer, a rectified linear unit layer, a local response normalization layer, a second convolutional layer, a fully connected layer and a Softmax layer which are sequentially connected in series from front to back.

[0012] The convolutional network model is used to predict the classification category of each layer and output the corresponding classification probability based on the preprocessed data of the parts to be processed. The SVM model is used to perform regression prediction based on the preprocessed data and classification probability of each layer of the parts to be processed to obtain the processing time of each layer.

[0013] The output module is used to output the processing time of each layer and the total processing time predicted by the processing time prediction module.

[0014] According to an embodiment of the present invention, the acquisition unit includes screenshot software, a camera, a mobile phone, a computer or other software or device with imaging function to image each layer of the scanning path, such as screenshot software or a camera.

[0015] According to an embodiment of the present invention, the acquisition unit further comprises a timer, and the timer is used to count the processing time of each layer of the training part.

[0016] According to an embodiment of the present invention, the preprocessing module includes a preprocessing unit, a classification unit and a balancing unit.

[0017] According to an embodiment of the present invention, the preprocessing unit is used to adjust the format of the data information to be the same, preferably to perform binary segmentation on the graphics, and to perform bicubic difference scaling to obtain a binary image of a size of 1200 pixels*1200 pixels.

[0018] Preferably, the preprocessing unit also includes correcting and denoising the data information.

[0019] According to an embodiment of the present invention, the correction includes correcting image distortion and cutting out the actual printing area, and the denoising includes performing three dilation and three erosion operations on the binary image, for example, the dilation and erosion are both performed using a 3*3 pixel window.

[0020] Preferably, the classification unit is used to classify the preprocessed data information according to the shape of the graphics in the data information, for example, into four categories: solid, hollow, multi-contour dispersion, and slender.

[0021] Preferably, the balancing unit is used to adjust the quantities of different types of data information to be balanced, for example, adjusting the quantities of four types of data information to 1:1:1:1.

[0022] According to an embodiment of the present invention, the balancing unit adjusts the quantity through a random oversampling algorithm, and the random oversampling algorithm is used to increase the quantity of data of a smaller class, reduce the imbalance between different classes of data, and obtain preprocessed data with a balanced quantity of data of different classes.

[0023] The present invention also provides a method for predicting metal additive manufacturing time based on the above system, comprising the following steps:

[0024] Step 1: Collect data information and time information of training parts, and collect data information of parts to be processed.

[0025] Step 2: Import the data information and time information into the convolutional neural model training unit for training to obtain a convolutional neural model; input the data information and time information into the convolutional neural model, and output the classification category and the corresponding probability; use the data information, time information, classification category and the corresponding probability as SVM training data, import the SVM training data into the support vector model for training to obtain the SVM model, import the data information of the parts to be processed into the convolutional neural model and the SVM model to obtain the predicted processing time of each layer and the total processing time.

[0026] According to an embodiment of the present invention, step 1 of collecting data information of the training parts and the parts to be processed includes: using an imaging device to collect scanning path data of each layer of the training parts and the parts to be processed and storing them as pictures.

[0027] The imaging device has the definition as described above.

[0028] According to an embodiment of the present invention, the following step is also included between step 1 and step 2: preprocessing the data information of each layer.

[0029] According to an embodiment of the present invention, the pretreatment comprises the following steps:

[0030] Step a: Correct image distortion and crop the actual printing area.

[0031] Step b: Binarize and segment the image of the printing area, and use bicubic difference to scale it to obtain a binary image with consistent size. For example, the size of the binary image is 1200 pixels*1200 pixels.

[0032] Step c: perform at least one dilation and at least one erosion operation on the binary image to remove noise and obtain preprocessed data. Preferably, the number of dilations is greater than or equal to 2, and the number of erosions is greater than or equal to 2, for example, three dilations and three erosions are performed.

[0033] Preferably, in step c, both the dilation and the erosion are performed using a 3*3 pixel window.

[0034] According to an embodiment of the present invention, after the preprocessing and before step 2, the following step is also included: increasing the amount of minority class data in the preprocessed data until the amount of data of all categories is balanced.

[0035] According to an embodiment of the present invention, the step of increasing the amount of minority class data in the preprocessed data is: determining the amount of each type of preprocessed data, using a random oversampling method, randomly copying the minority class data, reducing the imbalance between different types of preprocessing, and thereby increasing sensitivity to the minority class.

[0036] Preferably, the ratio of the data amounts of all categories is 1:1.

[0037] Preferably, before step 2, the method further includes the following steps: randomly dividing the training database into a training set, a validation set and a test set, wherein the ratio of the data volume in the training set, the validation set and the test set is 6:2:2.

[0038] According to an embodiment of the present invention, in step 2, the data information and the time information are imported into the convolutional neural model training unit for training, and obtaining the convolutional network model includes the following steps:

[0039] Step 21: inputting the associated data in the training database into the convolutional neural model training unit for training to obtain a convolutional network model, wherein the output layer of the convolutional network model outputs the classification category of the training data and the probability corresponding to each classification category;

[0040] Step 22: Add the classification categories and corresponding probabilities to the training database and train them as the training data set of the support vector machine model to obtain the SVM model.

[0041] According to an embodiment of the present invention, step 21 comprises the following steps:

[0042] Step 211: Select a group of multi-scale sub-images from the associated data of the training set of the training database (data corresponding to a training part) to form multi-scale data (including images of all layers of the training part) as input data of the convolutional neural model training unit, freeze the various layers of the convolutional neural model training unit or define them as untrainable, only train the weights of the new classifier layer, use bilinear interpolation to adjust both the small image blocks and the large image blocks to 227 pixels × 227 pixels, and use transfer learning to perform frozen training to obtain a frozen model.

[0043] Step 212: Define each layer in the frozen model as trainable, input the associated data of the training database test set and the validation set into the frozen model in sequence for training, and obtain a convolutional network model. The Softmax layer of the convolutional network model outputs the classification category of the training data and the probability of each classification category.

[0044] Preferably, the sub-image is composed of a number of image blocks of two different sizes, the small image block is an image area of ​​100 pixels × 100 pixels, and the large image block is an image area of ​​1200 pixels × 1200 pixels; bilinear interpolation is used to adjust both the small image block and the large image block to 227 pixels × 227 pixels.

[0045] Preferably, the convolutional neural model training unit adopts a pre-trained AlexNet architecture convolutional neural network to form a model, and the convolutional neural network model includes an input layer, a first convolutional layer, a rectified linear unit layer (ReLU layer), a local response normalization layer (LRN layer), a second convolutional layer, a fully connected layer (FC layer) and a Softmax layer connected in series from front to back.

[0046] Preferably, the new classifier layer includes an FC layer and a Softmax layer.

[0047] Preferably, the moving stride of the first convolutional layer is 4, and the moving stride of the second convolutional layer is 6.

[0048] According to an embodiment of the present invention, the convolutional neural network model training data includes the following steps:

[0049] Step 311: Use the first convolutional layer of the convolutional neural network model unit to operate on the full depth of each layer of the adjusted image to obtain the first convolutional layer output.

[0050] Step 312: Input the output of the first convolutional layer to the rectified linear unit layer to obtain the output of the ReLU layer, and input the output of the ReLU layer to the local response normalization layer to obtain normalized data.

[0051] Step 313: Use the second convolutional layer to operate on the normalized data to obtain the second convolutional layer output. Then the second convolutional layer output is input into a fully connected (FC) layer.

[0052] Step 314: Use the Softmax layer to convert the output of the FC layer into the probability of each classification category. The classification category with the largest response has the highest probability, and the sum of the probabilities of all classifications is equal to 1.

[0053] According to an embodiment of the present invention, the operation of the first convolution layer includes extracting low-level features, such as lines, edges, etc., and the operation of the second convolution layer includes extracting high-level features, such as area, angle, etc.

[0054] According to an embodiment of the present invention, step 22 comprises the following steps:

[0055] Step 221: The classification categories and corresponding probabilities output by the convolutional network model are added to the layer data of the training database as feature data of each layer to obtain an extended database.

[0056] Step 222: Input the data in the extended database as a training set into the support vector machine model, train the support vector machine model, and adjust the support vector machine model parameters according to the error value between the predicted manufacturing time of the support vector machine model and the corresponding actual manufacturing time in the expanded database to obtain a trained SVM model.

[0057] The present invention also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the method described is implemented.

[0058] The present invention also provides a device, comprising the storage medium and a processor, wherein the processor implements the steps of the method when executing the computer program on the storage medium.

[0059] Beneficial Effects

[0060] 1. The metal additive manufacturing time prediction method based on graphical processing of the present invention uses screenshots or cameras to collect the scanning paths of each layer of the printed part, extracts and classifies the data through a multi-scale convolutional neural network, adds the classification categories and probabilities to the database, and finally uses the trained SVM model to predict the manufacturing time of each layer, and at the same time obtains the manufacturing time of the entire part. Through the above method, the manufacturing time of each layer and the manufacturing time of the entire part can be quickly predicted when the part is not processed.

[0061] 2. The method of the present invention utilizes screenshots or cameras to collect scanning path data of each layer of the printed part, thereby increasing the compatibility of the method, making the method applicable not only to open source additive manufacturing equipment, but also to additive manufacturing equipment with relatively closed control software.

[0062] 3. Through random oversampling (ROS), the data of minority classes are randomly copied to reduce the imbalance between multiple classes, so as to increase the sensitivity to minority classes. Therefore, a more accurate model can be trained with less training data, which reduces the amount of training data used and improves the efficiency of prediction.

[0063] 4. The method of the present invention utilizes a method of first classification and then regression prediction, which effectively solves the problem in the prior art that the cross-sections of parts are different and the time prediction of layer manufacturing is difficult, effectively reduces the difficulty of prediction and improves the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of a metal additive manufacturing time prediction method based on data fusion according to an embodiment of the present invention;

[0065] Figure 2 This is a structural block diagram of a metal additive manufacturing time prediction system structure based on data fusion according to an embodiment of the present invention;

[0066] Figure 3 This is a common classification of image data in the embodiment of the present invention. DETAILED DESCRIPTION

[0067] The system and method of the present invention will be further described in detail below in conjunction with specific embodiments. It should be understood that the following embodiments are only exemplary descriptions and explanations of the present invention and should not be construed as limiting the scope of protection of the present invention. All technologies implemented based on the above content of the present invention are included in the scope of protection that the present invention is intended to protect.

[0068] Example 1

[0069] like Figure 1 As shown, a metal additive manufacturing time prediction method based on data fusion includes the following steps:

[0070] Step 1: Use screenshots or cameras to collect the scanning path data of each layer of the training part, store it as a picture, and use a timer to collect the manufacturing time of each layer of the training part.

[0071] Specifically, the scanning path data of each layer of the training part is collected by screenshots and stored as pictures. For additive manufacturing equipment that cannot install screenshot software, a camera is used to collect the scanning path data of each layer and store it as pictures.

[0072] According to the shape of the training parts, the image data is divided into four categories: solid, hollow, multi-contour scattered, and slender.

[0073] This data acquisition method has high compatibility, making it applicable to most additive manufacturing equipment.

[0074] Step 2: Import images of the scanning path data of each layer of the training part, import the corresponding processing parameters (processing time), preprocess the image data of each layer, use random oversampling (ROS) to increase the number of minority class data in the preprocessed data, reduce the imbalance between different classes of data, make the ratio of each class close to 1:1:1:1, and obtain balanced image data. Associate each balanced image data with the corresponding scanning time to obtain associated data. All associated data constitute the training database.

[0075] Preprocess the image data of each layer, including the following steps:

[0076] Step 21: Correct image distortion and crop the actual printing area.

[0077] Step 22: Binarize and segment the image of the printing area, and scale it using bicubic interpolation to obtain a binary image of 1200 pixels * 1200 pixels, and record the scaling ratio.

[0078] Step 23: Perform three dilation and three erosion operations on the binary image to remove noise and obtain preprocessed data. The dilation and erosion are both performed using a 3*3 pixel window.

[0079] Using random oversampling (ROS) to increase the amount of minority class data in preprocessed data includes: determining the amount of each type of preprocessed data, using random oversampling (ROS), and using random replication of minority class data to reduce the imbalance between different types of preprocessing, thereby increasing sensitivity to the minority class.

[0080] This step makes the image data have a uniform size through image preprocessing, removes the noise in the image, and obtains effective image information. By changing the class imbalance, a model with higher accuracy can be trained with less training data, reducing the amount of training data used and improving the efficiency of prediction.

[0081] Step 3: Divide the training database into training set, validation set, and test set in a ratio of 6:2:2. Use the associated data in the training set to train the multi-scale convolutional neural network unit, freeze or define each layer of the convolutional neural model training unit as untrainable, and only train the weights of the new classifier layer. Use bilinear interpolation to adjust both small and large image blocks to 227 pixels × 227 pixels, and use transfer learning to perform frozen training to obtain a frozen model.

[0082] Step 4: Define each layer in the frozen model as trainable, input the associated data of the training database test set and the validation set into the frozen model in sequence for training, and obtain the convolutional network model. The Softmax layer of the convolutional network model outputs the classification category of the training data and the probability of each classification category.

[0083] Fine-tune the network layer parameters (for example, only adjust the classifier parameters) to improve performance, optimize the convolutional neural network unit using the test set and validation set, and the output layer of the convolutional network model outputs the classification category of the training data and the probability corresponding to each classification category.

[0084] In this embodiment, the multi-scale convolutional neural network unit selects a model composed of a pre-trained AlexNet architecture convolutional neural network, hereinafter referred to as a CNN model. The CNN model includes an input layer, a first convolutional layer, a rectified linear unit layer (ReLU layer), a local response normalization layer (LRN layer), a second convolutional layer, a fully connected layer (FC layer) and a Softmax layer connected in series from front to back.

[0085] Training a multi-scale convolutional neural network unit (in this embodiment, ) using the associated data in the training database includes the following steps:

[0086] Step 31: Select a set of multi-scale sub-images from the associated data of the training database to form the input data of the multi-scale convolutional neural network, the sub-images are composed of a number of image blocks of two different sizes, the small image block is an image area of ​​100 pixels × 100 pixels, and the large image block is an image area of ​​1200 pixels × 1200 pixels. The number of small image blocks is greater than or equal to 1000, and the number of large image blocks is greater than or equal to 10.

[0087] Step 32: Use bilinear interpolation to resize both the small image block and the large image block to 227 pixels × 227 pixels, and use transfer learning to train. Use the first convolutional layer of the convolutional neural network model to operate on the full depth of the layer (i.e., all three channels of the image). The moving stride of the first convolutional layer is 4, and the output of the first convolutional layer is obtained.

[0088] Step 33: Input the output of the first convolutional layer to the rectified linear unit (ReLU) layer to obtain the output of the ReLU layer, and input the output of the ReLU layer to the local response normalization (LRN) layer to obtain normalized data.

[0089] Step 34: The normalized data is operated using the second convolutional layer of the convolutional neural network model, and the movement stride of the second convolutional layer is 6 to obtain the second convolutional layer output; the first convolutional layer extracts low-level features such as lines, edges, etc., and the second convolutional layer extracts high-level features such as area, angle, etc. The second convolutional layer output is then input into a fully connected (FC) layer.

[0090] Step 35: Use the Softmax layer to convert the output of the FC layer into the probability of each classification category. The classification category with the largest response has the highest probability, and the sum of the probabilities of all categories is equal to 1.

[0091] The beneficial effects of this step are: through multi-scale sub-images, it is ensured that both detail information and global information can be fully extracted. By using transfer learning and pre-training models, better prediction results can be produced with less data. The application of LRN improves the detection sensitivity of small spatial features.

[0092] Step 4 specifically includes the following steps:

[0093] Step 41: The classification categories and corresponding probabilities output by the multi-scale convolutional neural network are added to the layer data of the training database as feature data of each layer to obtain an extended database.

[0094] Step 42: Input the data in the extended database as a training set into the support vector machine model, train the support vector machine model, and adjust the support vector machine model parameters according to the error value between the predicted manufacturing time of the support vector machine model and the corresponding actual manufacturing time in the expanded database to obtain the SVM model.

[0095] The present invention first classifies the layer data, adds the classified category probabilities into a database, and then uses these data to perform regression prediction on the manufacturing time of the layer, which can further improve the prediction accuracy and enhance the prediction stability.

[0096] Step 5: Input the data parameters of the part to be predicted into the trained convolutional network model and SVM model in turn to obtain the prediction results of the manufacturing time of each layer and the manufacturing time of the entire part. Specifically, import the collected data of the part to be predicted, the program runs automatically, and finally automatically outputs the estimated time of each layer of the part and the manufacturing time of the entire part.

[0097] In the case of transfer learning, the features of the trained CNN model are used as initialization to train a CNN for actual classification. The CNN is instantiated using pre-trained weights and applied to a new classifier (the classifier is used to classify graphics), such as the classification of four categories: solid, hollow, multi-contour dispersed, and slender in the present invention.

[0098] Before using data to train the CNN model, freeze or define the layers of the CNN model as untrainable so that the pre-trained weights are not updated again in the new training cycle, otherwise the previously learned features will be destroyed. Only the weights of the new classifier layer added to the CNN model are trained, such as only the weights of the FC layer and the Softmax layer in this embodiment.

[0099] After the first freeze training, the CNN model was fine-tuned based on the original model (only the weights of the FC layer and the Softmax layer were adjusted during training). After that, each layer of the CNN model was defined as trainable for feature extraction, and the data in the test set and validation set were input for the second training to obtain the convolutional neural model. Under this operation, the pre-trained features of the previous CNN model were updated to make them more relevant to the new classification problem and better able to classify part images.

[0100] Since the input layer of the pre-trained AlexNet architecture convolutional neural network model is 227 pixels × 227 pixels, when transfer learning is applied to the pre-trained CNN, the architecture of the CNN, including the size of the input layer, must remain unchanged. Therefore, bilinear interpolation is used to adjust the input image to 227 pixels × 227 pixels. The filter size used in the first convolutional layer is 11 pixels × 11 pixels × 3 pixels. Finally, the output of the first convolutional layer is 55 pixels × 55 pixels × 96 pixels. ReLU does not change the size of the data volume, that is, the output size of the ReLU layer is 55 pixels × 55 pixels × 96 pixels. The LRN layer also does not change the size of the data volume, that is, the output size of the LRN layer is 55 pixels × 55 pixels × 96 pixels. The size of the finally constructed FC layer is 1 pixel × 1 pixel × 4 pixels.

[0101] like Figure 2 As shown, the present invention also provides a metal additive manufacturing time prediction system based on data fusion, including a data extraction and storage module, a preprocessing module, a processing time prediction module, and a data output module.

[0102] The data extraction and storage module includes a data acquisition unit and a data storage unit. The data acquisition unit is used to import training parts or parts to be predicted, and use screenshots or cameras to collect scanning path data of each layer of the parts used for training, store them as pictures as data information, and use a timer to collect the manufacturing time of each layer of the training parts as time information, and store the data information and time information.

[0103] The part data preprocessing module is used to preprocess the image data of each layer, associate the time information with the preprocessed data information, and form a training database.

[0104] The processing time prediction module includes a multi-scale convolutional neural network model training unit and a support vector machine model training. The multi-scale convolutional neural network model training unit is used to train the multi-scale convolutional neural network model using the data in the database to obtain a neural network model. The support vector machine model training unit is used to train the support vector machine model using the data in the database to obtain a support vector machine model.

[0105] The neural network model and the support vector machine model are used to predict the manufacturing time of the predicted parts, output the manufacturing time of each layer, and output the total manufacturing time of the parts.

[0106] The data output module is used to output the predicted manufacturing time of each layer and the total manufacturing time of the parts to be predicted.

[0107] The above is an explanation of the embodiments of the present invention. However, the present invention is not limited to the above embodiments. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A metal additive manufacturing time prediction system based on graphical processing, characterized in that: It includes data extraction and storage module, preprocessing module and processing time prediction module; The data extraction and storage module includes a data acquisition unit and a data storage unit, wherein the data acquisition unit is used to import training parts or parts to be processed, and collect data information and time information of the training parts, and collect data information of the parts to be processed, wherein the data information includes scanning path information of each layer of the training parts or parts to be processed; The data storage unit is used to store the data information and time information collected by the data collection unit; The preprocessing module is used to preprocess the data information to obtain preprocessed data, and associate the preprocessed data of each layer of all training parts with the corresponding time information to form a training database; The processing time prediction module is used to predict the processing time of the part to be processed according to the training database constructed by the preprocessing module, and the processing time includes the processing time of each layer of the part to be processed and the total processing time; The processing time prediction module includes at least one module training unit, which is used to train the preprocessing data and processing time of the currently imported training part and obtain a corresponding prediction model, and the prediction model predicts the processing time and total processing time of each layer of the part to be processed according to the preprocessing data of the part to be processed; The module training unit includes a convolutional neural model training unit and an SVM model training unit, wherein the convolutional neural model training unit is used to train the data information and time information of the training parts to obtain a convolutional network model; the convolutional network model is used to predict the classification category of each layer and output the corresponding classification probability according to the data information and time information; The SVM model training unit is used to train data information, time information, classification categories and corresponding probabilities to obtain an SVM model; The SVM model and the convolutional network model are used to predict the processing time and the total processing time of each layer according to the data information of the part to be processed.

2. The metal additive manufacturing time prediction system based on graphical processing according to claim 1, characterized in that: The prediction system also includes a data output module, which is used to output the processing time of each layer and the total processing time predicted by the processing time prediction module.

3. The metal additive manufacturing time prediction system based on graphical processing according to claim 1, characterized in that: The convolutional neural model training unit and the convolutional network model both include an input layer, a first convolutional layer, a rectified linear unit layer, a local response normalization layer, a second convolutional layer, a fully connected layer and a Softmax layer which are sequentially connected in series from front to back.

4. The metal additive manufacturing time prediction system based on graphical processing according to claim 1, characterized in that: The data acquisition unit includes screen capture software, a camera, a mobile phone, a computer or other software or devices with imaging functions to image the scanning paths of each layer.

5. The metal additive manufacturing time prediction system based on graphical processing according to claim 1, characterized in that: The data acquisition unit also includes a timer, which is used to time the processing time of each layer of the training part.

6. The metal additive manufacturing time prediction system based on graphical processing according to any one of claims 1 to 5, characterized in that: The preprocessing module includes a preprocessing unit, a classification unit and a balancing unit; The pre-processing unit is used to adjust the format of the data information to be the same; The classification unit is used to classify the pre-processed data information according to the shape of the graphics in the data information; The balancing unit is used to adjust the quantity of different types of data information to be balanced.

7. The metal additive manufacturing time prediction system based on graphical processing according to claim 6, characterized in that: The classification unit is used to perform binary segmentation on the graphics, and bicubic interpolation scaling to obtain a binary image with a size of 1200 pixels*1200 pixels.

8. The metal additive manufacturing time prediction system based on graphical processing according to claim 6, characterized in that: The preprocessing unit also includes correcting and denoising the data information; the correction includes correcting image distortion and cutting out the actual printing area, and the denoising includes performing three expansion and three erosion operations on the binary image.

9. The metal additive manufacturing time prediction system based on graphical processing according to claim 6, characterized in that: The balancing unit adjusts the quantity through a random oversampling algorithm, and the random oversampling algorithm is used to increase the quantity of data of a class with a smaller quantity, reduce the imbalance between data of different classes, and obtain preprocessed data with a balanced quantity of data of different classes.

10. A method for predicting metal additive manufacturing time based on the system according to any one of claims 1 to 9, characterized in that: The steps include: Step 1: Collect data information and time information of training parts, and collect data information of parts to be processed; Step 2: Import the data information and time information into the convolutional neural model training unit for training to obtain a convolutional neural model; input the data information and time information into the convolutional neural model, and output the classification category and the corresponding probability; use the data information, time information, classification category and the corresponding probability as SVM training data, import the SVM training data into the support vector model for training to obtain the SVM model, import the data information of the parts to be processed into the convolutional neural model and the SVM model to obtain the predicted processing time of each layer and the total processing time.

11. The metal additive manufacturing time prediction method according to claim 10, characterized in that: Step 1 of collecting data information of the training parts and the parts to be processed includes: using an imaging device to collect scanning path data of each layer of the training parts and the parts to be processed and storing them as pictures.

12. The metal additive manufacturing time prediction method according to claim 11, characterized in that: The following steps are also included between step 1 and step 2: preprocessing the data information of each layer of the part to be processed.

13. The metal additive manufacturing time prediction method according to claim 10, characterized in that: The pre-processing comprises the following steps: Step a: Correct the image distortion and crop the actual printing area; Step b: Binarize and segment the image of the printing area, and scale it using bicubic interpolation to obtain a binary image of the same size; Step c: Perform at least one dilation and at least one erosion operation on the binary image to remove noise and obtain preprocessed data.

14. The metal additive manufacturing time prediction method according to claim 12, characterized in that: After the preprocessing and before step 2, the following step is also included: increasing the amount of minority class data in the preprocessed data until the amount of data of all categories is balanced.

15. The metal additive manufacturing time prediction method according to claim 14, characterized in that: The steps to increase the number of minority class data in the preprocessed data are: determine the number of each type of preprocessed data, use random oversampling method, randomly copy the minority class data, reduce the imbalance between different types of preprocessing, and thus increase the sensitivity to the minority class.

16. The metal additive manufacturing time prediction method according to any one of claims 10 to 15, characterized in that: In step 2, the data information and time information are imported into the convolutional neural model training unit for training, and the convolutional network model is obtained, which includes the following steps: Step 21: inputting the associated data in the training database into the convolutional neural model training unit for training to obtain a convolutional network model, wherein the output layer of the convolutional network model outputs the classification category of the training data and the probability corresponding to each classification category; Step 22: Add the classification categories and corresponding probabilities to the training database and train them as the training set of the support vector machine model to obtain the SVM model.

17. The metal additive manufacturing time prediction method according to claim 16, characterized in that: Step 21 includes the following steps: Step 211: Select a group of multi-scale sub-images from the associated data of the training set of the training database to form multi-scale data as input data of the convolutional neural model training unit, freeze each layer of the convolutional neural model training unit or define it as untrainable, only train the weight of the new classifier layer, use bilinear interpolation to adjust both the small image block and the large image block to 227 pixels × 227 pixels, use transfer learning to perform frozen training, and obtain a frozen model; Step 212: Define each layer in the frozen model as trainable, input the associated data of the training database test set and the validation set into the frozen model in sequence for training, and obtain a convolutional network model. The Softmax layer of the convolutional network model outputs the classification category of the training data and the probability of each classification category.

18. The metal additive manufacturing time prediction method according to claim 17, characterized in that: The sub-image is composed of a number of image blocks of two different sizes, the small image block is an image area of ​​100 pixels×100 pixels, and the large image block is an image area of ​​1200 pixels×1200 pixels; bilinear interpolation is used to adjust the small image block and the large image block to 227 pixels×227 pixels.

19. The metal additive manufacturing time prediction method according to any one of claims 10 to 15, characterized in that: The convolutional neural model training unit adopts a pre-trained AlexNet architecture convolutional neural network to form a model. The convolutional neural network model includes an input layer, a first convolutional layer, a rectified linear unit layer, a local response normalization layer, a second convolutional layer, a fully connected layer and a Softmax layer which are connected in series from front to back.

20. The metal additive manufacturing time prediction method according to any one of claims 10 to 15, characterized in that: The convolutional neural network model training data includes the following steps: Step 311: using the first convolutional layer of the convolutional neural network model unit to operate on the full depth of each layer of the adjusted image to obtain the first convolutional layer output; Step 312: input the output of the first convolutional layer to the rectified linear unit layer to obtain the output of the ReLU layer, and input the output of the ReLU layer to the local response normalization layer to obtain normalized data; Step 313: Use the second convolutional layer to operate on the normalized data to obtain the second convolutional layer output; then input the second convolutional layer output into a fully connected layer; Step 314: Use the Softmax layer to convert the output of the FC layer into the probability of each classification category. The classification category with the largest response has the highest probability, and the sum of the probabilities of all classifications is equal to 1.

21. The metal additive manufacturing time prediction method according to claim 16, characterized in that: Step 22 includes the following steps: Step 221: The classification categories and corresponding probabilities output by the convolutional network model are added to the layer data of the training database as feature data of each layer to obtain an extended database; Step 222: Input the data in the extended database as a training set into the support vector machine model, train the support vector machine model, and adjust the support vector machine model parameters according to the error value between the predicted manufacturing time of the support vector machine model and the corresponding actual manufacturing time in the expanded database to obtain a trained SVM model.

22. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the method according to any one of claims 10 to 21 is implemented.

23. A computer device comprising the storage medium of claim 22 and a processor, wherein the processor implements the steps of the method of any one of claims 10 to 21 when executing the computer program on the storage medium.

Images