A deep learning-based process tolerance classification method and system

Abstract

The application discloses a process tolerance classification method and system based on deep learning, which processes process tolerance text data based on a classification model file trained based on deep learning, automatically determines the category of a quality inspection process to which the process tolerance text data belongs, can improve classification accuracy and classification efficiency, and reduces consumption of human and material resources. Meanwhile, the application can form an automatic process tolerance classification processing flow, improve work efficiency, and provides a convenient operation flow for the mechanical processing industry.

Description

Technical Field

[0001] This invention belongs to the field of process tolerance classification, specifically relating to a process tolerance classification method and system based on deep learning. Background Technology

[0002] With the popularization of quality management systems in factories of the machinery industry, a large amount of text data containing the requirements for the inspection of tolerances in the machining process of products has been accumulated in the quality management system. The lengths vary and the descriptions vary from person to person. For example: (1) "Circular arc R6-10" means that the circular arc is measured using a radius gauge to see if it meets the standard. The quality inspection process should be "circular arc measurement" and the category should be "circular arc". "Outer circle Ra1.6" means that the roughness of the outer circle needs to be checked to see if it meets the standard. The quality inspection process should be "roughness inspection" and the category should be "roughness". (2) "M5-6H drilling depth 7" means that the hole depth needs to be measured using a vernier caliper. The quality inspection process should be "depth measurement" and the category should be "depth". "M5-6H tapping depth 5" means that the internal thread depth needs to be measured using a thread plug gauge. The quality inspection process should be "thread depth measurement" and the category should be "thread depth".

[0003] Metrology personnel must classify and schedule quality inspection procedures according to testing requirements. Currently, the method used is manual identification and classification, but this requires highly experienced metrology personnel, is time-consuming and labor-intensive, has low efficiency, and is prone to errors. Therefore, how to accurately and efficiently classify process tolerance text data (i.e., how to determine the category of the quality inspection procedure to which process tolerance text data belongs) is an urgent problem that needs to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method and system for classifying process tolerances, so as to accurately and efficiently classify process tolerance text data.

[0005] The deep learning-based process tolerance classification method of this invention includes:

[0006] Step 1: Segment the input n process tolerance text data to be classified into words, and obtain a set of n words that correspond one-to-one with the n process tolerance text data.

[0007] Step 2: Call the saved index dictionary and perform vector transformation on the n word set to obtain a numerical character vector of the n word set.

[0008] Step 3: Padded with equal lengths the numerical character vectors of the n word sets to obtain a sequence of n word set vectors of equal length.

[0009] Step 4: Call the saved classification model file to perform classification calculations on the n word set vector sequences to obtain n classification result matrices that correspond one-to-one with the n process tolerance text data; wherein, each classification result matrix is ​​composed of the probability floating-point values ​​of each label, and the classification model file is obtained through deep learning.

[0010] Step 5: Label A i The corresponding category is used as the classification result of the i-th process tolerance text data; where i takes all integers from 1 to n, A i The label with the highest probability floating-point value in the classification result matrix corresponding to the i-th process tolerance text data.

[0011] Preferably, in step one, the method for segmenting the input n process tolerance text data to be classified is as follows:

[0012] Based on the saved user dictionary and stop word list, the input n process tolerance text data to be classified are segmented using a Chinese word segmentation library to obtain n word sets that correspond one-to-one with the n process tolerance text data. Among them, the user dictionary consists of specialized vocabulary in the field of process tolerance, and the stop word list is the set of stop words that appear in the process tolerance text corpus.

[0013] Preferably, in step four, the method for obtaining the classification model file includes:

[0014] Step S1: Process the existing process tolerance text data in the quality management system database to obtain a sample dataset array; wherein, the sample dataset array includes process tolerance text data and labels corresponding to the category (quality inspection process) to which the process tolerance text data belongs.

[0015] Step S2: Segment the process tolerance text data in the sample dataset array to obtain the word set of the sample dataset array.

[0016] Step S3: Using the word set of the sample dataset array and the labels in the sample dataset array corresponding to the category of the process tolerance text data, create the training set, validation set, and test set required for deep learning.

[0017] Step S4: Build a deep learning training model.

[0018] Step S5: Substitute the training set, validation set, and test set into the deep learning training model. After training and testing, the classification model file is obtained.

[0019] Preferably, in step S1, the method for processing the existing process tolerance text data in the quality management system database to obtain a sample dataset array includes:

[0020] Step S101: Use the numerical serial number as the label corresponding to the category (of the quality inspection process) to form a label-category relationship table.

[0021] Step S102: Organize the existing process tolerance text data in the quality management system database and the category to which the process tolerance text data belongs. Classify the process tolerance text data belonging to the same category. Based on the label-category relationship table, bind the process tolerance text data with the label corresponding to the (category to which the process tolerance text data belongs) to generate a labeled process tolerance text dataset array.

[0022] Step S103: Shuffle the process tolerance text data in the labeled process tolerance text dataset array to obtain the sample dataset array.

[0023] Preferably, in step S2, the method for segmenting the process tolerance text data in the sample dataset array to obtain the word set of the sample dataset array includes:

[0024] Step S201: Load the user dictionary; wherein, the user dictionary consists of specialized terms in the field of process tolerances.

[0025] Step S202: Load the stop word list; wherein, the stop word list is a set of stop words appearing in the process tolerance text corpus.

[0026] Step S203: Based on the user dictionary and stop word list, use the Chinese word segmentation library to segment the process tolerance text data in the sample dataset array to obtain the word set of the sample dataset array.

[0027] Preferably, in step S3, the method for creating the training set, validation set, and test set required for deep learning includes:

[0028] Step S301: Load the index dictionary.

[0029] Step S302: Based on the index dictionary, perform vector transformation on the word set of the sample dataset array to obtain a numerical character vector of the word set of the sample dataset array, and save the index dictionary.

[0030] Step S303: Pad the numerical character vectors of the word set in the sample dataset array with equal length to obtain a sequence of word set vectors of equal length in the sample dataset array; at the same time, convert the labels in the sample dataset array corresponding to the category to which the process tolerance text data belongs into binary matrices.

[0031] Step S304: Combine the word set vector sequence and the corresponding binary matrix of the sample dataset array to form a data packet.

[0032] Step S305: Divide the data packet into the training set, validation set and test set required for deep learning according to a preset ratio.

[0033] Preferably, the preset ratio is 6:2:2; that is, 60% of the data packets are used as the training set, 20% of the data packets are used as the validation set, and 20% of the data packets are used as the test set.

[0034] Preferably, in step S4, the method for building the deep learning training model includes:

[0035] Step S401: Select a convolutional neural network model as the training model.

[0036] Step S402: Add an embedding layer.

[0037] Step S403: Add convolutional and pooling layers.

[0038] Step S404: Add a fully connected layer to complete the deep learning training model setup.

[0039] Preferably, in step S5, the method for substituting the training set, validation set, and test set into the deep learning training model, and obtaining the classification model file after training and testing, includes:

[0040] Step S501: Set the optimizer, loss function, and accuracy to be used during training.

[0041] Step S502: Set the batch size and number of iterations.

[0042] Step S503: Input the training set, validation set, and test set into the deep learning training model. Use the training set for deep learning training, use the validation set to adjust the model parameters, and use the test set to test the model's generalization ability.

[0043] Step S504: Check the accuracy of the deep learning training model. When the accuracy of the deep learning training model is greater than or equal to the set accuracy, stop training and save the trained model as a classification model file.

[0044] The deep learning-based process tolerance classification system of the present invention includes a memory and a processor; the memory is used to store machine-readable programs; the processor is used to call the machine-readable programs to execute the above-described deep learning-based process tolerance classification method.

[0045] This invention utilizes a classification model file trained using deep learning to process process tolerance text data, automatically determining the category of the quality inspection process to which the process tolerance text data belongs (i.e., automatically classifying the process tolerance text data), thereby improving classification accuracy and efficiency while reducing the consumption of human and material resources. Furthermore, this invention can be used to create an automated process tolerance classification process, improving work efficiency and providing a convenient operating procedure for the machining industry. Attached Figure Description

[0046] Figure 1 This is a flowchart of the deep learning-based process tolerance classification method in this embodiment.

[0047] Figure 2 This is a flowchart illustrating the method for obtaining the classification model file in this embodiment.

[0048] Figure 3 This is a diagram of the human-computer interaction interface of the process tolerance classification system in this embodiment. Detailed Implementation

[0049] like Figure 1 , Figure 2 As shown, the deep learning-based process tolerance classification method in this embodiment is developed using the Python language. Its process tolerance classification method includes:

[0050] Step 1: Segment the input n process tolerance text data to be classified into words, and obtain a set of n words that correspond one-to-one with the n process tolerance text data.

[0051] Specifically, based on the saved user dictionary and stop word list, a Chinese word segmentation library (such as jieba) is used to segment the input n pieces of process tolerance text data (which exist in the quality management system according to time span) to be classified, resulting in n word sets corresponding one-to-one with the n pieces of process tolerance text data. The user dictionary consists of specialized vocabulary in the field of process tolerance, and the stop word list is a set of stop words appearing in the process tolerance text corpus. The user dictionary (userdict.txt) contains words such as: temperature, coaxiality, bounce, cylindricity, symmetry, etc., which can assist in reasonable word segmentation and avoid splitting proper nouns. The stop word list (stopwords.txt) contains stop words such as modal particles and punctuation marks, which are not helpful for training. Removing these words reduces the interference of irrelevant data on the corpus features.

[0052] Step 2: Use the saved index dictionary to perform vector conversion on the n-word set, obtaining a numeric character vector for the n-word set. For example, use the pickle.load() function to call the saved index dictionary, and use the tokenizer.texts_to_sequences() function to perform vector conversion on the n-word set, obtaining a numeric character vector for the n-word set.

[0053] Step 3: Pad the numeric character vectors of the n word sets to the same length to obtain a sequence of n word set vectors of equal length. For example, use the pad_sequences() function to pad the data to the same length.

[0054] Step 4: Call the saved classification model file to perform classification calculations (predictions) on the n word set vector sequences, obtaining n classification result matrices that correspond one-to-one with the n process tolerance text data. For example, use the `load_model()` function to call the saved classification model file (HDF5 format) to perform classification calculations on the n word set vector sequences, and finally use the `model.predict()` function to obtain the classification result matrices. The classification model file is obtained through deep learning. Each classification result matrix consists of the probability floating-point values ​​of each label; the probability floating-point value of any label is greater than 0 and less than 1, and the sum of the probability floating-point values ​​of all labels in each classification result matrix equals 1.

[0055] like Figure 2 As shown, the methods for obtaining classification model files include:

[0056] Step S1: Process the existing process tolerance text data in the quality management system database to obtain a sample dataset array. Specifically, this includes:

[0057] Step S101: Use numerical serial numbers (e.g., 0, 1, 2, 3…19) as labels corresponding to the categories of quality inspection processes to form a label-category relationship table (e.g., Table 1).

[0058] Label category Label category Label category Label category 0 Aperture depth 5 Cleanliness 10 bounce 15 Concentricity 1 roughness 6 Cylindricity 11 Symmetry 16 Straightness 2 Thread depth 7 coaxiality 12 Taper 17 Engraving operation 3 Aperture roundness 8 Verticality 13 Position 18 flatness 4 Roundness 9 depth 14 Parallelism 19 arc

[0059] Table 1

[0060] Step S102: Organize the existing process tolerance text data in the quality management system database and the categories to which the process tolerance text data belongs. Classify the process tolerance text data belonging to the same category, and based on the above label-category relationship table, use the preprocess_text() function to bind the process tolerance text data with the label corresponding to the (category to which the process tolerance text data belongs) to generate a labeled process tolerance text dataset array, ensuring the accuracy and uniqueness of the category to which the process tolerance text data belongs.

[0061] Step S103: Use the random.shuffle() function to shuffle the process tolerance text data in the labeled process tolerance text data set array to obtain a sample data set array. The sample data set array includes process tolerance text data and labels corresponding to the categories to which the process tolerance text data belongs.

[0062] Step S2: Segment the process tolerance text data in the sample data set array to obtain a word set of the sample data set array. Specifically, it includes:

[0063] Step S201: Load the user dictionary. The user dictionary consists of special vocabulary in the field of process tolerance. The vocabulary included in the user dictionary (userdict.txt) is such as temperature, coaxiality, runout, cylindricity, symmetry, etc. The user dictionary can assist in reasonable word segmentation and avoid splitting proper nouns.

[0064] Step S202: Load the stop word list. The stop word list is a set of stop words that appear in the process tolerance text corpus. The stop words included in the stop word list (stopwords.txt) are such as modal particles and punctuation marks like 'le' and 'ye' and other vocabulary characters that are not helpful for training. Removing these vocabulary characters can reduce the interference of irrelevant data on the corpus features.

[0065] Step S203: Based on the user dictionary and the stop word list, use a Chinese word segmentation library (such as jieba) to segment the process tolerance text data in the sample data set array to obtain a word set of the sample data set array (such as Table 2).

[0066]

[0067] Table 2

[0068] Step S3: Use the word set of the sample data set array and the labels corresponding to the categories to which the process tolerance text data belongs in the sample data set array to make training sets, validation sets, and test sets required for deep learning. Specifically, it includes:

[0069] Step S301: Load the index dictionary.

[0070] Step S302: Based on the index dictionary, perform vector transformation on the word set of the sample dataset array to obtain a numeric character vector of the word set of the sample dataset array, and save the index dictionary (cnndict.pickle) using the persistence module (pickle.dump) for later use. For example, use the tokenizer.texts_to_sequences() function to perform vector transformation on the word set of the sample dataset array to obtain a numeric character vector of the word set of the sample dataset array, which is convenient for feature calculation.

[0071] Step S303: Pad the numerical character vectors of the word set in the sample dataset array with equal length to obtain a sequence of word set vectors of equal length in the sample dataset array; at the same time, convert the labels in the sample dataset array corresponding to the category to which the process tolerance text data belongs into binary matrices.

[0072] Since the word sets in the sample dataset array have varying lengths after transformation, the `pad_sequences()` function can be used to pad them to the same length. Simultaneously, the `to_categorical()` function converts the labels in the sample dataset array corresponding to the categories of the process tolerance text data into binary matrices, which is equivalent to converting the original labels into one-hot encoded forms. For example, in a partial sample dataset array, the label corresponding to the category of the process tolerance text data is label = [0,1,2,3,4,5,6,7,8]. The nine labels are converted to obtain the binary matrix of the labels as follows: label = [[1 0 0 0 0 0 0 0 0][0 1 0 0 0 0 0 0][0 0 1 0 0 0 0 0 0][0 0 0 1 0 0 0 0 0][0 0 0 1 0 0 0 0][0 0 0 0 1 0 0 0][0 0 0 0 0 1 0 0 0][0 0 0 0 0 10 0][0 0 0 0 0 0 1 0][0 0 0 0 0 0 0 1]].

[0073] Step S304: Combine the word set vector sequence and the corresponding binary matrix of the sample dataset array to form a data packet.

[0074] Step S305: Divide the data packet into the training set, validation set and test set required for deep learning according to a preset ratio.

[0075] In this embodiment, the preset ratio is 6:2:2; that is, 60% of the data packets are used as the training set, 20% as the validation set, and 20% as the test set. The training set will be used extensively during the model building process, and learning and training will be carried out through repeated iterative calculations. The validation set is used to adjust the model parameters to determine whether overfitting exists, and the test set is used to detect the model's generalization ability.

[0076] Step S4: Build a deep learning training model. This is developed based on the TensorFlow or PyTorch deep learning framework.

[0077] Specifically, it includes:

[0078] Step S401: Select a convolutional neural network model as the training model.

[0079] The principle of deep learning is to process and fit the input information through multi-layer neural networks. The results of each layer are superimposed to obtain a weight matrix model that best maps the input to the corresponding output. Deep learning algorithms can be trained and self-learned based on the accumulation of data, continuously improving the accuracy of separation. Common deep learning algorithms include Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and Gated Recurrent Unit (GRU). This embodiment uses a CNN model. A CNN model mainly consists of convolutional layers, max pooling layers, and fully connected layers. Convolutional layers extract text features, max pooling layers downsample the data to reduce dimensionality and avoid overfitting, and finally, fully connected layers perform classification and output the results.

[0080] As another embodiment, a Long Short-Term Memory (LSTM) network model can be used. LSTM is an improved version of the Recurrent Neural Network (RNN), which can address the problem that RNNs cannot handle long-range dependencies. However, its training time is longer than that of a Convolutional Neural Network (CNN). Alternatively, a Gated Recurrent Unit (GRU) model can be used. GRU is a simplified version of the LSTM algorithm. With optimized hyperparameters, it can achieve classification results with fewer parameters and is less prone to overfitting. However, its training time is also longer than that of a CNN.

[0081] Step S402: Add an embedding layer. Use the `model.add()` function to add an embedding layer. This layer is a connection layer combining the data matrix and the hidden layer of the neural network. It maps each numerical value in the two-dimensional process tolerance text data vector sequence in the training set to a three-dimensional vector, using a three-dimensional vector to represent a word. The embedding layer takes a two-dimensional tensor as input, with shape (batch size, input length) and outputs a three-dimensional tensor with shape (batch size, input length, output dimension)).

[0082] Step S403: Add a convolutional layer (Conv1D Layer) and a pooling layer (MaxPooling1D Layer).

[0083] Add convolutional layers. One-dimensional convolution (Conv1d) is used for text data, performing convolution only on the width, not the height. The convolution kernel window slides along the sentence length direction to perform the convolution operation. Key settings include the word vector dimension (in_channels), the number of channels generated by the convolution (out_channels), the kernel size (kernel_size), the stride (stride=1), the number of layers to pad each edge of the input with zeros (padding=0), and the choice of neuron activation function.

[0084] A pooling layer is added after each convolutional layer, using the MaxPooling1D rule to select the maximum value in a region as the pooled value for that region. The pooling layer serves to downsample, i.e., reduce feature dimensionality to prevent overfitting and achieve a non-linear feature extraction process.

[0085] Step S404: Add a fully connected layer (Dense Layer) to complete the deep learning training model construction.

[0086] A fully connected layer can connect text data and corresponding labels based on labels.shape, which is logically equivalent to the function: y = Activation(Wx + bias). The input x is an a-dimensional vector, the weights W is a b*a matrix, Activation represents the activation function, bias represents the bias, and the output y is a b-dimensional vector; that is, the input matrix is ​​transformed by a linear transformation plus a non-linear transformation to point to the corresponding output y.

[0087] Step S5: Substitute the training set, validation set, and test set into the deep learning training model. After training and testing, a classification model file is obtained. It is developed based on the TensorFlow or PyTorch deep learning framework. Specifically, it includes:

[0088] Step S501: Set the optimizer, loss function, and accuracy to be used during training; for example, use the model.compile() function to set the optimizer, loss function, and accuracy to be used during training.

[0089] Step S502: Set the batch size (batch_size) and the number of iterations (epochs).

[0090] Step S503: Input the training set, validation set, and test set into the deep learning training model. Use the training set for deep learning training, use the validation set to adjust the model parameters, and use the test set to test the model's generalization ability.

[0091] Step S504: Check the accuracy of the deep learning training model. When the accuracy is greater than or equal to the set accuracy (e.g., 0.98), stop training and save the trained model as a classification model file. For example, use `model.save()` to save the trained model as an HDF5 format classification model file (e.g., CNN.h5). This classification model file consists of two parts: folders (Groups) and dataset arrays (Datasets). Each dataset array can be divided into two parts: raw data (rawdatavalues) and metadata. These two parts work together to map process tolerance text data to labels (corresponding categories).

[0092] Step 5: Label A i The corresponding category is used as the classification result of the i-th process tolerance text data; where i takes all integers from 1 to n, A i The label with the highest probability floating-point value in the classification result matrix corresponding to the i-th process tolerance text data.

[0093] The entire processing procedure from steps one to five above is encapsulated in the backend, and the frontend human-computer interaction interface is as follows: Figure 3 As shown.

[0094] This embodiment also provides a deep learning-based process tolerance classification system, including a memory and a processor. The memory is used to store machine-readable programs; the processor is used to call the machine-readable programs and execute the aforementioned deep learning-based process tolerance classification method.

Claims

1. A deep learning-based process tolerance classification method, characterized in that, The application comprises the following steps: Step 1: Tokenizing the input n pieces of process tolerance text data to be classified to obtain n word sets corresponding to the n pieces of process tolerance text data; Step 2: Calling the saved index dictionary to convert the n word sets into numerical character vectors to obtain numerical character vectors of the n word sets; Step 3: Padding the numerical character vectors of the n word sets to obtain n word set vector sequences with equal lengths; Step 4: Calling the saved classification model file to perform classification calculation on the n word set vector sequences to obtain n classification result matrices corresponding to the n pieces of process tolerance text data; each classification result matrix is composed of possible floating point values of each label, and the classification model file is obtained through deep learning; Step five, label A i a corresponding category as a classification result of the i-th process tolerance text data; wherein i takes all integers from 1 to n in turn, A i is the label corresponding to the largest possible floating-point value in the classification result matrix corresponding to the i-th process tolerance text data. In step 4, the method for obtaining the classification model file comprises the following steps: Step S1: Processing the existing process tolerance text data in the quality management system database to obtain a sample data set array; the sample data set array comprises process tolerance text data and labels corresponding to the categories to which the process tolerance text data belongs; Step S2: Tokenizing the process tolerance text data in the sample data set array to obtain word sets of the sample data set array; Step S3: Using the word sets of the sample data set array and the labels corresponding to the categories to which the process tolerance text data belongs in the sample data set array to make a training set, a validation set and a test set required for deep learning; Step S4: Building a deep learning training model; Step S5: inputting the training set, the validation set and the test set into the deep learning training model to obtain the classification model file through training and testing.

2. The deep learning-based process tolerance classification method of claim 1, wherein: In step 1, the method for tokenizing the input n pieces of process tolerance text data to be classified comprises the following steps: Based on the saved user dictionary and the stop word table, the input n pieces of process tolerance text data to be classified are tokenized by using a Chinese tokenization library to obtain n word sets corresponding to the n pieces of process tolerance text data; the user dictionary is composed of special words in the process tolerance field, and the stop word table is a set of stop words appearing in the process tolerance text corpus.

3. The deep learning-based process tolerance classification method of claim 1, wherein: In step S1, the method for processing the existing process tolerance text data in the quality management system database to obtain a sample data set array comprises the following steps: Step S101: Using a digital serial number as a label corresponding to a category to form a label-category relationship table; Step S102: Combing the existing process tolerance text data in the quality management system database and the categories to which the process tolerance text data belongs, classifying the process tolerance text data belonging to the same category, and binding the process tolerance text data and the corresponding labels based on the label-category relationship table to generate a labeled process tolerance text data set array; Step S103: Shuffling the process tolerance text data in the labeled process tolerance text data set array to obtain the sample data set array.

4. The deep learning-based process tolerance classification method of claim 1, wherein: In step S2, the method for tokenizing the process tolerance text data in the sample data set array to obtain word sets of the sample data set array comprises the following steps: Step S201, loading a user dictionary; wherein the user dictionary is composed of special vocabulary in the process tolerance field; Step S202, loading a stop word table; wherein the stop word table is a set of stop words appearing in the process tolerance text corpus; Step S203, based on the user dictionary and the stop word table, using a Chinese word segmentation library to perform word segmentation on the process tolerance text data in the sample data set array, to obtain a word set of the sample data set array.

5. The deep learning-based process tolerance classification method of claim 1, wherein: In the step S3, the method for making the training set, the validation set and the test set required by deep learning includes: Step S301, loading an index dictionary; Step S302, performing vector conversion on the word set of the sample data set array according to the index dictionary, to obtain a numerical character vector of the word set of the sample data set array, and saving the index dictionary; Step S303, performing equal-length padding on the numerical character vector of the word set of the sample data set array, to obtain a word set vector sequence of the sample data set array with equal length; and converting the labels corresponding to the category to which the process tolerance text data belongs in the sample data set array into a binary matrix; Step S304, combining the word set vector sequence of the sample data set array and the corresponding binary matrix to form a data packet; Step S305, dividing the data packet into the training set, the validation set and the test set required by deep learning according to a preset ratio.

6. The deep learning-based process tolerance classification method of claim 5, wherein: The preset ratio is 6:2:

2.

7. The deep learning-based process tolerance classification method of claim 1, wherein: In the step S4, the method for building a deep learning training model includes: Step S401, selecting a convolutional neural network model as the training model; Step S402, adding an embedding layer; Step S403, adding a convolutional layer and a pooling layer; Step S404, adding a fully connected layer, to complete the building of the deep learning training model.

8. The deep learning-based process tolerance classification method of claim 1, wherein: In the step S5, the method for substituting the training set, the validation set and the test set into the deep learning training model, and training, testing to obtain a classification model file includes: Step S501, setting the optimizer, the loss function and the accuracy rate used during training; Step S502, setting the batch size and the number of iterations; Step S503, inputting the training set, the validation set and the test set into the deep learning training model, using the training set for deep learning training, using the validation set for model parameter adjustment, and using the test set for model generalization ability detection; Step S504, checking the accuracy rate of the deep learning training model, and stopping training when the accuracy rate is greater than or equal to the set accuracy rate, and saving the model obtained by training as a classification model file.

9. A process tolerance classification system based on deep learning, comprising a memory and a processor, characterized in that: The memory is used to store a machine readable program; The processor is used to call the machine readable program and execute the process tolerance classification method based on deep learning according to any one of claims 1 to 8.

Images