Numerical control machining process generation method based on PVFNN neural network model

Abstract

The invention discloses a numerical control machining process generation method based on a PVFNN neural network, and the method mainly comprises the steps: 1) data preprocessing: collecting and processing the 3D shape, material and quality information of a part; 2) constructing and segmenting a data set to form a training, verification and test set; 3) constructing a PVFNN neural network model, integrating a 3D convolutional neural network, a point cloud neural network and a full connection network, and realizing feature fusion through a feature fusion layer and an MLP; 4) performing model training and verification, using a training set to train the model, and using a test set to evaluate performance; 5) process generation: performing model processing on new part information to generate a processing process; the method is used for automatically learning the capability of the discrete manufacturing process from design and manufacturing data, the accuracy and efficiency of numerical control machining process generation are improved, and the problem that in the prior art, due to the influence of manufacturing process capability knowledge and manufacturing experience, a user often cannot select the optimal manufacturing process is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of numerical control machining, and in particular to a numerical control machining process generation method based on a PVFNN neural network model. Background Art

[0002] The future of manufacturing is rapidly shifting from mass production to mass customization, with personalized products enabled by decentralized production systems. This trend includes cloud manufacturing services, which encapsulate geographically distributed manufacturing resources online and make them accessible to users via the internet. Such services allow users to upload a computer-aided design (CAD) model of a part design to an online platform, select a manufacturing process, and receive cost and delivery time estimates almost instantly. After the user accepts the estimate, the part design is manufactured by one or more suppliers in the service provider network and delivered to the user. One limitation of such cloud manufacturing services is that the manufacturing process selection is done manually by the user. Since the correct selection of a process requires a deep understanding of it, and these capabilities are usually acquired through experience, users of these services may not have the knowledge of manufacturing process capabilities required to select the most appropriate process to realize their design, resulting in suboptimal selection.

[0003] Deep learning has made rapid progress in recent years. The ability of machine learning algorithms to learn the manufacturing process and obtain different shapes and required component qualities of successfully manufactured parts from design and production data is an attractive solution. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a CNC machining process generation method based on the PVFNN neural network model. By performing multimodal fusion of 3D convolutional neural network, point cloud hierarchical feature learning network, and residual neural network, and combining the attention mechanism with a simple multi-layer perceptron (MLP), traditional machining and finishing are automatically learned to obtain the optimal machining plan, thereby solving the problem in the prior art that users often cannot choose the optimal manufacturing process due to the influence of manufacturing process capability knowledge and manufacturing experience.

[0005] In order to solve the above technical problems, the present invention provides a technical solution: a method for generating a CNC machining process based on a PVFNN neural network model, characterized in that:

[0006] Step S1: Data collection:

[0007] Collect 3D shape data including parts, as well as material and quality information related to each part, and pre-process the collected part information;

[0008] Step S2: Dataset construction and segmentation:

[0009] The processed part 3D shape information, as well as the quantitative description and quality information of the corresponding part materials form a data set, and the data set is divided into a training set, a validation set, and a test set;

[0010] Step S3: PVFNN neural network model construction:

[0011] The PVFNN neural network model adopts a parallel network architecture of a 3D convolutional backbone network, a Pointnet++ backbone network, and a fully connected neural network for extracting quality information. The 3D convolutional backbone network, the Pointnet++ backbone network, and the fully connected neural network integrate features through a feature fusion layer, and the voxelized 3D shape data features and point cloud data features are spliced, and then pass through a multi-head attention mechanism layer and a fully connected layer.

[0012] Step S4: Model training and validation:

[0013] The PVFNN neural network model is trained and verified by constructing a training set and a validation set, learning how to generate the corresponding processing technology for the part data, and finally the model is tested using the test set;

[0014] Step S5: Process generation:

[0015] After the PVFNN neural network model passes the test, the new part information is input into the PVFNN neural network model, which will process the part information and generate a processing technology suitable for processing the part.

[0016] Furthermore, in step S1, the process of preprocessing the part information is as follows:

[0017] 3D shape information of parts: firstly processed by voxelization and point cloudization, then the voxelized shape information is converted into a 100×100×100 binary array, and the point clouded shape information is converted into 1024 points by farthest point sampling;

[0018] Part material and quality information: The dimensional tolerances defined by material properties are composed of a 1×7 array.

[0019] Furthermore, in step S2, the training set, validation set and test set are divided in a random manner according to a certain ratio.

[0020] Furthermore, in step S3, the 3D convolution backbone network is composed of an input layer, three convolution layers, two residual block layers, three batch normalization layers, three dropout layers, three maximum pooling layers and a concatenation layer; the Pointnet++ backbone network is composed of two set abstraction layers; the fully connected layer is composed only of Fc1: (7, 7); the feature fusion layer adopts vector concatenation fusion; the multi-head attention mechanism layer has an embedding dimension of 256 and 4 heads.

[0021] Furthermore, in step S3, the PFVFNN model uses residual blocks to alleviate the gradient vanishing problem to enhance feature transfer.

[0022] Furthermore, in step S3, a RELU activation function and a Dropout regularization layer are added after each convolution of the PFVFNN model to improve the generalization ability of the model and prevent overfitting.

[0023] Furthermore, in step S4, during the model training process, the model is optimized using the cross entropy loss function and the AdamW optimizer to ensure the accuracy and robustness of the model in the classification task.

[0024] In order to solve the above technical problems, another technical solution provided by the present invention is: a NC machining process generation system based on PVFNN for implementing the above method, characterized in that it includes a data acquisition module, a data processing module, a model building module, and a process generation module, wherein:

[0025] Data acquisition module: collects 3D shape data of parts, as well as material and quality information related to each part, and sends the collected part information to the data processing module;

[0026] Data processing module: obtain the part information sent by the data acquisition module, first voxelize and point cloud the 3D shape information of the part, then convert the voxelized shape information into a 100×100×100 binary array, and convert the point cloud shape information into 1024 points through farthest point sampling; the part material and quality information are formed into a 1×7 array by the dimensional tolerance defined by the material characteristics, and the processed part 3D shape information, as well as the quantitative description and quality information of the corresponding part material are formed into a data set, and the data set is divided into a training set, a validation set, and a test set;

[0027] Model construction module: The PVFNN neural network model adopts a parallel network architecture of a 3D convolutional backbone network, a Pointnet++ backbone network, and a fully connected neural network for extracting quality information. The 3D convolutional backbone network, the Pointnet++ backbone network, and the fully connected neural network integrate features through a feature fusion layer, and splice the voxelized 3D shape data features and point cloud data features, and then pass through a multi-head attention mechanism layer and a fully connected layer; the PVFNN neural network model is trained and verified through the constructed training set and verification set, and how to generate the corresponding processing technology for the part data is learned, and finally the model is tested using a test set;

[0028] Process generation module: Input new part information into the PVFNN neural network model, which will process the part information and generate a processing technology suitable for processing the part.

[0029] The beneficial effects of the present invention are:

[0030] 1. This application uses multimodal fusion of 3D convolutional neural networks, point cloud hierarchical feature learning networks, and residual neural networks, and combines the attention mechanism with a simple multi-layer perceptron (MLP) to construct a process generation model. By inputting part information including the 3D shape of the part, a quantitative description of the part material, and part quality information, it can automatically learn and generate the processing technology for processing the part. It can comprehensively consider various factors that affect the processing technology, reduce the workload of manual selection of processing technology, and improve the accuracy of prediction and processing efficiency.

[0031] 2. The PVFNN neural network model in this application uses the Pointnet++ backbone network to extract global features from point cloud data. The features extracted by PointNet++ can be fused with the features extracted by CNN to form a more comprehensive part representation, which helps to improve the accuracy of the model's prediction of the processing technology. The fused features are used to splice the voxelized 3D shape data features and the point cloud data features to improve the classification performance of the model.

[0032] 3. In this application, the PFVFNN model adds a RELU activation function and a Dropout regularization layer after each convolution to improve the generalization ability of the model and prevent overfitting, and uses the cross entropy loss function and AdamW optimizer to optimize the model to ensure the accuracy and robustness of the model in classification tasks. It can be applied to different types of parts and processing processes. By providing the corresponding input data, the model can adaptively learn to generate the corresponding processing technology.

[0033] In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only four of the drawings of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 A flowchart of the application method;

[0036] Figure 2 It is the structural diagram of the PVFNN neural network model;

[0037] Figure 3 Confusion matrix for manufacturing process prediction on the test dataset;

[0038] Figure 4 This is a structural block diagram of the device of this application. DETAILED DESCRIPTION

[0039] Embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although certain embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as being limited to the embodiments described herein, which are instead provided for a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not intended to limit the scope of protection of the present invention.

[0040] Example

[0041] like Figure 1 As shown, a method for generating a CNC machining process based on a PVFNN neural network model comprises the following steps:

[0042] Step S1: Data collection:

[0043] Collect 3D shape data including parts, as well as material and quality information related to each part, and pre-process the collected part information;

[0044] Step S2: Dataset construction and segmentation:

[0045] The processed part 3D shape information, as well as the quantitative description and quality information of the corresponding part materials form a data set, and the data set is divided into a training set, a validation set, and a test set;

[0046] Step S3: PVFNN neural network model construction:

[0047] The PVFNN neural network model adopts a parallel network architecture of a 3D convolutional backbone network, a Pointnet++ backbone network, and a fully connected neural network for extracting quality information. The 3D convolutional backbone network, the Pointnet++ backbone network, and the fully connected neural network integrate features through a feature fusion layer, and the voxelized 3D shape data features and point cloud data features are spliced, and then pass through a multi-head attention mechanism layer and a fully connected layer.

[0048] Step S4: Model training and validation:

[0049] The PVFNN neural network model is trained and verified by constructing a training set and a validation set, learning how to generate the corresponding processing technology for the part data, and finally the model is tested using the test set;

[0050] Step S5: Process generation:

[0051] After the PVFNN neural network model passes the test, the new part information is input into the PVFNN neural network model, which will process the part information and generate a processing technology suitable for processing the part.

[0052] In step S1, the 3D shape of the part is stored in stl format, and the quantitative description of the part material includes elastic modulus [GPa], elongation at break [%], yield strength [MPa], Brinell hardness [HB] and thermal diffusivity [m3 / s]. The quantitative description information of the part is determined by the ASM International Manual Committee. The quality information of the part includes the dimensional tolerance of the part [mm] and the arithmetic mean surface roughness of the part area [Sa]. The acquisition of the part information can be manually imported by personnel.

[0053] The process of part information preprocessing is as follows:

[0054] 3D shape information of the part: firstly processed by voxelization and point cloud, and then the voxelized shape information is converted into a 100×100×100 binary array as input 1, and the point clouded shape information is converted into 1024 points as input 2 through farthest point sampling;

[0055] Part material and quality information: The dimensional tolerance [mm] defined by the material properties is a 1×7 array as input three.

[0056] Material properties include elastic modulus [GPa], elongation at break [%], yield strength [MPa], Brinell hardness [HB], thermal diffusivity [m3 / s], and surface roughness [μm].

[0057] In step S2, the training set, validation set and test set are divided in a random manner according to a certain ratio.

[0058] Training set: usually 60%-80% of the dataset. Used to train the model so that the model can learn the characteristics of the data; Validation set: usually 10%-20% of the dataset. Used to evaluate the performance of the model, adjust hyperparameters, and select the best model during the training process; Test set: usually 10%-20% of the dataset. Used to evaluate the generalization ability of the model after the model training is completed to ensure that the model performs well on unseen data.

[0059] In step S3, the 3D convolution backbone network is composed of an input layer, three convolution layers, two residual block layers, three batch normalization layers, three dropout layers, three maximum pooling layers and a concatenation layer; the Pointnet++ backbone network is composed of two set abstraction layers; the fully connected layer is composed of only Fc1: (7, 7); the feature fusion layer adopts vector concatenation fusion; the multi-head attention mechanism layer has an embedding dimension of 256 and 4 heads.

[0060] The PFVFNN model uses residual blocks to alleviate the gradient vanishing problem and enhance feature transfer. The PFVFNN model adds a RELU activation function and a Dropout regularization layer after each convolution to improve the generalization ability of the model and prevent overfitting.

[0061] In the present invention, the ReLU function is used as the activation function of each convolutional layer and the fully connected layer, the Softmax function is used as the activation function of the output layer, and the classification cross entropy function is used as the loss function during training of this model, and the optimizer uses the AdamW optimizer.

[0062] The activation function used in the PVFNN neural network model is the ReLU function, which is expressed as:

[0063] ReLu(x)=max(x,0) (1)

[0064] The output of the PVFNN neural network model is expressed using the Softmax function:

[0065]

[0066] Where X = (x 1, …,x K ) represents the K neuron values ​​of the output layer, where K is the number of class labels. The output of the Softmax function σ i(X) represents the probability that the input data belongs to the i-th class. In the proposed network architecture, the final output is a vector of size 6, consisting of σ1(X), σ2(X), σ3(X), σ4(X), σ5(X), σ6(X), with the maximum σ j The j-th class label of the (X) value is the predicted class label, i.e., the manufacturing process required for the part to be produced is predicted.

[0067] The loss function used in the training of the PVFNN neural network model is the classification cross entropy function, which is expressed as:

[0068]

[0069] Where D is a training dataset containing n K types of training data, d l is the lth training data, σ i (d l ) is the model prediction training data d l The probability of belonging to the i-th class, y i is the training data d l The true probability of belonging to the i-th class; if the training data belongs to the i-th class, then y i =1; if it does not belong to the i-th category, then y i =0.

[0070] Finally, the six processes of milling, turning, drilling, milling and grinding, turning and grinding, and drilling and boring are selected according to the probability of classification.

[0071] like Figure 2 As shown in the figure, a multimodal PVFNN neural network model: The PVFNN neural network model uses 3D convolution and residual blocks to extract local spatial features of voxel data, extracts geometric features of point cloud data through the Set Abstraction module, and fuses additional input features. The model uses a multi-head self-attention mechanism to capture the global correlation between different modal features, and combines the feature fusion module to efficiently integrate multimodal features. After feature extraction, the dimension is gradually reduced through a series of fully connected layers, and finally Softmax is used to achieve multi-classification tasks. The model improves training stability through residual connections, batch normalization and other mechanisms, combines local and global features, and is suitable for multimodal tasks of processing three-dimensional data and time series data.

[0072] The main process of the multimodal PVFNN neural network model algorithm is described as follows.

[0073]

[0074]

[0075] In the technical implementation, special attention is paid to the computational efficiency and practicality of the model. For example, by introducing layer normalization and residual connections, not only the convergence speed in the training process is accelerated, but also the stability of the model in the network structure is ensured, avoiding the gradient vanishing problem.

[0076] Formula example: The calculation formula of the self-attention mechanism is expressed as:

[0077] Q=H i W Q ,K=H i W K ,V=H i W V (4)

[0078] Among them, H i is the input representation of the i-th layer, Q: query matrix, represented by the input H i Multiply by the query weight matrix W Q Get, K: key matrix, represented by the input H i Multiply by the key weight matrix W K Get, V: value matrix, represented by the input H i Multiply by the value weight matrix W V Get,WQ,WK,WV: weight matrices of query, key and value.

[0079]

[0080] Among them, QK T : Multiply the query matrix Q by the transpose of the key matrix K to get the attention score. k : The dimension of the key vector, usually a scaling factor to prevent the dot product value from being too large. softmax: Apply the softmax function to normalize the attention score. Attention(Q,K,V): Get the weighted value matrix.

[0081] The operation of each layer of the PVFNN neural network model is as follows: In the model structure, the convolution kernels of the three convolution layers all use 3×3×3 convolution kernels, and the number of input and output channels are (1,32), (32,64) and (64,64) respectively. The three maximum pooling layers are all set to 2 to halve the feature map, and the number of channels of the normalization layer is set to 32, 64 and 64 respectively. The dropout rate of the three dropout layers is set to 0.5. The two residual blocks are placed after the first three-dimensional convolution layer and the second three-dimensional convolution layer respectively. The number of input and output channels of the residual block is the number of channels of the previous convolution output. Two three-dimensional convolutions are performed inside the residual block, and the convolution kernel is set to 3×3×3 in size, and the padding is set to 1. Point cloud data processing is achieved through two Setabstraction operations of PointNet++. Feature fusion is achieved by splicing voxels, part material and quality information and point cloud features, and then the features are projected to 256-dimensional space through a linear layer and passed through a multi-head attention mechanism layer, which has an embedding dimension of 256 and 4 heads. Finally, the features pass through a series of fully connected layers, including Fc2: (74, 256), Fc3: (256, 128), Fc4: (128, 64), Fc5: (64, 6), and the classification results are output through the Softmax layer.

[0082] Input 1 is processed in three convolutional layers, input 2 is processed in the Pointnet++ backbone network to capture geometric features, and input 3 is processed in a fully connected network (fc_1) with 7 inputs and 7 outputs. The output information of these three layers is then input into the feature fusion layer and concatenated into a one-bit vector of size 74. This vector is then sent to another fully connected layer (fc_2) with an input of 74 and an output of 256. The feature vector then passes through a multi-head attention mechanism layer, which uses a 256-dimensional embedding vector and is configured with 4 attention heads. After passing through the attention layer, the feature vector is sent to a neural network consisting of three fully connected layers, with the input and output set to (256, 128), (128, 64), and (64, 6) respectively. Finally, the output is passed through the Softmax function as the activation function of the output layer.

[0083] In step S4, during the model training process, the model is optimized using the cross entropy loss function and the AdamW optimizer to ensure the accuracy and robustness of the model in the classification task.

[0084] Experiments and Results

[0085] This paper evaluates the prediction performance through accuracy, precision, recall and F1 score metrics.

[0086] Accuracy: refers to the ratio of the number of samples correctly predicted by the model to the total number of samples.

[0087]

[0088] Among them, TP stands for True Positives, TN stands for True Negatives, FP stands for False Positives, and FN stands for False Negatives.

[0089] Precision: refers to the proportion of samples predicted by the model to be positive that are actually positive.

[0090]

[0091] Recall: refers to the proportion of samples that are actually positive examples that are predicted by the model as positive examples. It focuses on the model's coverage ability, that is, the degree of recognition of positive examples.

[0092]

[0093] F1 Score: It is the harmonic mean of precision and recall, which comprehensively considers the accuracy and recall of the model.

[0094]

[0095] The experiment divides the processing technology into six categories.

[0096] A 70:30 split of the dataset was chosen for training and validation. During the training phase, the Adaptive Moment Estimation with Weight Decay Repair (AdamW) optimizer was used to determine the optimal weights due to its fast convergence speed and suitability for complex model training with high-dimensional data. A total of 1500 epochs were run, with the learning rate set to 10-5 and the batch size set to 30.

[0097] Finally, the trained model is used to test 720 unseen test data, as shown in Table 1.

[0098] Table 1 Number of test sets for six categories of machining processes

[0099]

[0100] The test accuracy reached 95%. The experimental results are shown in Table 2. Figure 3 shown.

[0101] Table 2 Experimental data of the six-classification test set of machining technology

[0102]

[0103] In order to solve the above technical problems, another technical solution provided by the present invention is: a numerical control machining process generation system based on a PVFNN neural network model for implementing the above method, characterized in that it includes a data acquisition module, a data processing module, a model building module, and a process generation module, wherein:

[0104] Data acquisition module: collects 3D shape data of parts, as well as material and quality information related to each part, and sends the collected part information to the data processing module;

[0105] Data processing module: obtain the part information sent by the data acquisition module, first voxelize and point cloud the 3D shape information of the part, then convert the voxelized shape information into a 100×100×100 binary array, and convert the point cloud shape information into 1024 points through farthest point sampling; the part material and quality information are formed into a 1×7 array by the dimensional tolerance defined by the material characteristics, and the processed part 3D shape information, as well as the quantitative description and quality information of the corresponding part material are formed into a data set, and the data set is divided into a training set, a validation set, and a test set;

[0106] Model construction module: The PVFNN neural network model adopts a parallel network architecture of a 3D convolutional backbone network, a Pointnet++ backbone network, and a fully connected neural network for extracting quality information. The 3D convolutional backbone network, the Pointnet++ backbone network, and the fully connected neural network integrate features through a feature fusion layer, and splice the voxelized 3D shape data features and point cloud data features, and then pass through a multi-head attention mechanism layer and a fully connected layer; the PVFNN neural network model is trained and verified through the constructed training set and verification set, and how to generate the corresponding processing technology for the part data is learned, and finally the model is tested using a test set;

[0107] Process generation module: Input new part information into the PVFNN neural network model, which will process the part information and generate a processing technology suitable for processing the part.

[0108] The above description of the various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, this application will not go into details.

[0109] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for generating a CNC machining process based on a PVFNN neural network model, characterized in that: Step S1: Data collection: Collect 3D shape data including parts, as well as material and quality information related to each part, and pre-process the collected part information; Step S2: Dataset construction and segmentation: The processed part 3D shape information, as well as the quantitative description and quality information of the corresponding part materials form a data set, and the data set is divided into a training set, a validation set, and a test set; Step S3: PVFNN neural network model construction: The PVFNN neural network model adopts a parallel network architecture of a 3D convolutional backbone network, a Pointnet++ backbone network, and a fully connected neural network for extracting quality information. The 3D convolutional backbone network, the Pointnet++ backbone network, and the fully connected neural network integrate features through a feature fusion layer, and the voxelized 3D shape data features and point cloud data features are spliced, and then pass through a multi-head attention mechanism layer and a fully connected layer; Step S4: Model training and validation: The PVFNN neural network model is trained and verified by constructing a training set and a validation set, learning how to generate the corresponding processing technology for the part data, and finally the model is tested using the test set; Step S5: Process generation: After the PVFNN neural network model passes the test, the new part information is input into the PVFNN neural network model, which will process the part information and generate a processing technology suitable for processing the part.

2. The method for generating a numerical control machining process based on a PVFNN neural network model according to claim 1, characterized in that: In step S1, the process of preprocessing the part information is as follows: 3D shape information of parts: firstly processed by voxelization and point cloudization, then the voxelized shape information is converted into a 100×100×100 binary array, and the point clouded shape information is converted into 1024 points by farthest point sampling; Part material and quality information: The dimensional tolerances defined by material properties are composed of a 1×7 array.

3. A method for generating a numerical control machining process based on a PVFNN neural network model according to claim 2, characterized in that: In step S2, the training set, validation set and test set are divided in a random manner according to a certain ratio.

4. The method for generating a numerical control machining process based on a PVFNN neural network model according to claim 1, characterized in that: In step S3, the 3D convolution backbone network is composed of an input layer, three convolution layers, two residual block layers, three batch normalization layers, three dropout layers, three maximum pooling layers and a concatenation layer; the Pointnet++ backbone network is composed of two set abstraction layers; the fully connected layer is composed of only Fc1: (7, 7); the feature fusion layer adopts vector concatenation fusion; the multi-head attention mechanism layer has an embedding dimension of 256 and 4 heads.

5. The method for generating a numerical control machining process based on a PVFNN neural network according to claim 1, characterized in that: In step S3, the PFVFNN model uses residual blocks to alleviate the gradient vanishing problem and enhance feature transfer.

6. The method for generating a numerical control machining process based on a PVFNN neural network model according to claim 1, characterized in that: In step S3, a RELU activation function and a Dropout regularization layer are added after each convolution of the PFVFNN model to improve the generalization ability of the model and prevent overfitting.

7. The method for generating a numerical control machining process based on a PVFNN neural network model according to claim 1, characterized in that: In step S4, during the model training process, the model is optimized using the cross entropy loss function and the AdamW optimizer to ensure the accuracy and robustness of the model in the classification task.

8. A numerical control machining process generation system based on the PVFNN neural network model according to claims 1-7, characterized in that: It includes data acquisition module, data processing module, model building module and process generation module, among which: Data acquisition module: collects 3D shape data of parts, as well as material and quality information related to each part, and sends the collected part information to the data processing module; Data processing module: obtain the part information sent by the data acquisition module, first voxelize and point cloud the 3D shape information of the part, then convert the voxelized shape information into a 100×100×100 binary array, and convert the point cloud shape information into 1024 points through farthest point sampling; the part material and quality information are formed into a 1×7 array by the dimensional tolerance defined by the material characteristics, and the processed part 3D shape information, as well as the quantitative description and quality information of the corresponding part material are formed into a data set, and the data set is divided into a training set, a validation set, and a test set; Model construction module: The PVFNN neural network model adopts a parallel network architecture of a 3D convolutional backbone network, a Pointnet++ backbone network, and a fully connected neural network for extracting quality information. The 3D convolutional backbone network, the Pointnet++ backbone network, and the fully connected neural network integrate features through a feature fusion layer, and splice the voxelized 3D shape data features and point cloud data features, and then pass through a multi-head attention mechanism layer and a fully connected layer; the PVFNN neural network model is trained and verified through the constructed training set and verification set, and how to generate the corresponding processing technology for the part data is learned, and finally the model is tested using a test set; Process generation module: Input new part information into the PVFNN neural network model, which will process the part information and generate a processing technology suitable for processing the part.

Images