A tool wear state monitoring method based on VWC-Transformer transfer learning under variable working conditions

Abstract

The application discloses a tool wear state monitoring method based on VWC-Transformer transfer learning under variable working conditions, which comprises the following steps: step 1) obtaining tool variable working condition historical data based on multi-sensor information; step 2) performing time feature extraction and normalization processing on each working condition data; step 3) inputting the above feature sequence into a VWC-Transformer tool wear prediction model for training and optimization; step 4) collecting a small amount of real-time signals under different working condition parameter change conditions by a sensor respectively and performing data preprocessing; step 5) using a pre-training-fine-tuning transfer learning method to migrate the trained VWC-Transformer model to a target domain, completing a migration task, and reusing and fine-tuning the model; and step 6) predicting the tool wear state by using the migrated model, so as to judge whether the tool needs to be replaced. The application introduces a multi-working condition parameter introduction layer to build a VWC-Transformer prediction model, so that the VWC-Transformer prediction model has the ability of learning and training variable working condition parameters.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical processing, in particular to a variable working condition tool wear state monitoring method based on VWC-Transformer transfer learning. BACKGROUND

[0002] As an important execution end of mechanical manufacturing, the tool wear state has an important influence on the machining precision and quality of the workpiece. As a consumable item, it has the characteristics of short service life and easy damage, which directly affects the efficiency of the equipment and even the entire production system.

[0003] At present, the mainstream method on the market is a tool wear state monitoring method based on a large amount of data. Through various sensors, relevant information in the mechanical processing process is collected, and various signal processing technologies and intelligent algorithm analysis means are used to mine the relationship between the tool wear state and the sensor signal, so as to realize the identification of the tool wear state. If the tool prediction model trained under the historical working condition is applied to the new working condition data set, the model accuracy will be reduced. If the model is retrained using the data set under the new working condition, a large amount of new working condition data needs to be collected. For variable working condition processing in actual production, this model training method not only causes heavy training tasks, but also affects the real-time response of the monitoring system. Moreover, most of them are offline operations and are for single working conditions, but the working conditions in actual processing are often complex and variable. If the working condition parameters change, it is difficult to accurately and efficiently predict the tool wear state. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a variable working condition tool wear state monitoring method based on VWC-Transformer transfer learning in view of the defects in the prior art.

[0005] The technical solution adopted by the present application to solve the technical problem is: a variable working condition tool wear state monitoring method based on VWC-Transformer transfer learning, comprising the following steps:

[0006] Step 1) Obtain tool variable working condition historical data based on multi-sensor information; including vibration signal, cutting force signal, acoustic emission signal data, which are composed of source domain data;

[0007] The variable working condition is the change of the working condition parameters, including the working condition parameters of spindle speed, feed rate, and back engagement amount;

[0008] Step 2) Time feature extraction and normalization processing are performed on each working condition data, specifically as follows:

[0009] 2.1) slice along the time dimension; obtain the variance, mean and maximum value of each part of the data to form a shorter time sequence;

[0010] 2.2) time feature extraction using a short-time Fourier transform-based feature extraction method;

[0011] The feature extraction uses a short-time Fourier transform-based feature extraction method, which further reduces the time series data amount of the model input and improves the information capacity of a single time step while retaining rich information;

[0012] The short-time Fourier transform formula is as follows:

[0013]

[0014] In the formula, h(τ-t) is a window function, f is a frequency (Hz), and t is a time (s);

[0015] Step 3) input the above feature sequence into the VWC-Transformer tool wear prediction model for training and optimization;

[0016] The VWC-Transformer prediction model structure constructed in step 3) includes, in sequence:

[0017] An input embedding layer;

[0018] A position encoding layer; the position encoding layer encodes the signal by position encoding to obtain the sequence order;

[0019] A multi-head attention mechanism; the multi-head self-attention mechanism is used to encode the sequence to extract effective information;

[0020] A residual connection and normalization layer;

[0021] A feedforward network layer; the feedforward network layer introduces nonlinear modeling capability to the model;

[0022] A residual connection and normalization layer;

[0023] A multi-condition parameter input layer; the multi-condition parameter input layer is used to introduce variable condition parameter features. Since the variable condition parameters are represented by discrete values rather than continuous values, the distance relationship between the feature values is measured by the numerical size. After data standardization processing of the multi-condition parameters, the concatenate connection function is used to map the machining parameters to the multi-condition parameter layer in the form of an array.

[0024] A ReLU activation function;

[0025] A fully connected layer;

[0026] Step 4) The sensor collects a small amount of real-time signals under different working condition parameter change conditions respectively and carries out data preprocessing;

[0027] Step 5) The pre-training-fine-tuning transfer learning method is used to transfer the trained VWC-Transformer model to the target domain, complete the transfer task, and reuse and fine-tune the model;

[0028] Step 6) The transferred model is used to predict the tool wear state, so as to judge whether the tool needs to be replaced.

[0029] According to the above scheme, in the step 3), the multi-working condition parameter input layer includes: an input layer taking the working condition parameters as input; and a feature layer connecting multiple input parameters together through a concatenate function, and then combining the feature layer as part of the input layer to form the multi-working condition parameter input layer.

[0030] According to the above scheme, the position encoding layer encodes the signal and obtains the sequence order, and the formula is:

[0031]

[0032]

[0033] In the formula, PE is an encoding function, pos is a position, d model is the feature dimension number of each time step data, q is the qth element of the position information encoding feature vector, and 2q∈[0,d model ).

[0034] According to the above scheme, after the position encoding, the multi-head self-attention mechanism is used to encode the sequence, establish the connection between different time steps in the input data, extract different time steps in the time sequence input data, and filter out important information related to each other from a large amount of input data;

[0035] The function of scaling click attention is:

[0036]

[0037] The multi-head attention output is as follows:

[0038] MultiHead(Q,K,V)=Concatenate(head1,…,head h ,…,head H )W C

[0039]

[0040] When Q=K=V=X′ iAt that time, the h-th head of multi-head self-attention is:

[0041]

[0042] In the formula, W C ∈R HN×M For learnable matrices, MultiHead represents a multi-head attention mechanism, Concatenate indicates concatenation, and head... h Representing the h-th head, X′ i Let M be the input matrix, N be the dimension of the key matrix, V be the value matrix, which is the matrix obtained by transforming a set of values ​​of dimension N, Q be the query matrix, which is the matrix obtained by transforming the query and key of dimension M, and K be the key matrix.

[0043] According to the above scheme, step 5) is as follows:

[0044] There are three types of cutting parameters, and a source domain D is established for each type of cutting parameter. S To target domain D T The transfer task, and the transfer tasks from the source domain to the target domain are established by the simultaneous changes of various cutting parameters, that is, a total of 7 transfer tasks are established for transfer learning.

[0045] in,

[0046] Migration Task D S1 →D T1 spindle speed V c1 →V c2 The feed rate f1 and depth of cut C1 remain unchanged.

[0047] Migration Task D S2 →D T2 For the feed rate f1→f2 to change, the spindle speed V c1 The depth of cut C1 remains unchanged;

[0048] Migration Task D S3 →D T3 For the depth of cut C1→C2, the spindle speed V c1 The feed rate f1 remains constant;

[0049] Migration Task D S1 →D T1 +D S2 →D T2 When the spindle speed and feed rate change simultaneously, the depth of cut remains constant.

[0050] Migration Task D S1 →DT1 +D S3 →D T3 For spindle speed and back engagement amount change at the same time, the feed amount remains unchanged;

[0051] D S2 →D T2 +D S3 →D T3 For feed and back engagement amount change at the same time, the spindle speed remains unchanged;

[0052] D S1 →D T1 +D S2 →D T2 +D S3 →D T3 For spindle speed, feed, back engagement amount change at the same time;

[0053] After introducing the multi-working condition parameter layer, the migration of the VWC-Transformer model is as follows:

[0054] First, the source domain data is input into the model to generate the regression loss of the source domain sample;

[0055] The data collected by the target domain is input into the model to generate the regression loss of the target domain sample;

[0056] Then, the features output by the source domain and the target domain are extracted and input into the maximum mean difference MMD calculated by the edge distribution adaptation principle to increase the similarity between the two, and the loss function Loss converges to the optimal value, and Loss is:

[0057] Loss=MSE source +MSE target +λMMD 2 loss

[0058] In the formula, MSE source is the regression loss of the source domain sample;MSE target is the regression loss of the target domain sample;

[0059] MMD 2 loss is the adaptive loss;λ is the penalty coefficient;Loss is the final loss;

[0060] Freeze all layers, fine-tune the last fully connected layer, combine the sensor signals extracted by the VWC-Transformer before the fully connected layer, and retrain the weight parameters of the entire network layer.

[0061] The beneficial effects of the present application are:

[0062] 1) The present application subdivides the features along the time dimension into variance, mean, maximum value, extracts useful time features through feature extraction of short-time Fourier transform, and prevents overfitting.

[0063] 2) The present application introduces a multi-working condition parameter introduction layer to build a VWC-Transformer prediction model, so that it has the ability to learn and train variable working condition parameters.

[0064] 3) The present application adopts a pre-training-fine-tuning transfer learning method to migrate the VWC-Transformer model to the target domain, the network structure of the model and the weights and biases of each layer of neurons are the same as the original model, and the two share model structure parameters, and the migration learning is performed on the 7 kinds of variable working condition migration tasks under the change of cutting force, spindle speed and back engagement amount parameters.

[0065] 4) The present application adopts a method combining source domain offline training and target domain online monitoring, fully utilizes historical data, and only needs to collect a small amount of sample signals under unknown working conditions online in actual processing, so as to efficiently and accurately monitor the tool wear state under multiple working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0066] The present application will be further described below in combination with the drawings and examples, wherein:

[0067] Figure 1 is a method flowchart of an embodiment of the present application;

[0068] Figure 2 is a source domain and target domain migration learning flowchart of an embodiment of the present application;

[0069] Figure 3 is a VWC-Transformer model structure diagram of an embodiment of the present application;

[0070] Figure 4 is a variable working condition parameter and migration task diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0072] As shown in Figure 1 and Figure 2 , the present application provides a variable working condition tool wear state monitoring method based on VWC-Transformer migration learning, comprising the following steps:

[0073] In the source domain offline training, obtain the tool variable working condition historical data based on multi-sensor information, including vibration signal, cutting force signal, acoustic emission signal, and compose the source domain data;

[0074] Extract time characteristics and normalize each working condition data, slice along the time dimension, obtain the variance, mean value and maximum value of each part of data to form a shorter time sequence;

[0075] The feature extraction method based on short-time Fourier transform is adopted, which further reduces the time sequence data amount of model input and improves the information capacity of single time step under the premise of retaining rich information. The short-time Fourier transform formula is as follows:

[0076]

[0077] In the formula, h(τ-t) is the window function, f is the frequency (Hz), and t is the time (s).

[0078] Among them, the normalized processing maps the feature data to [0, 1], and the formula is:

[0079]

[0080] X″=X ′ *(mx-mi)+mi

[0081] In the formula, max is the maximum value of a column, min is the minimum value of a column, X″ is the final result, mx and mi are specified interval values, and mx is 1 by default, and mi is 0.

[0082] After normalization, save it as a csv file, and make corresponding vectorization labels for data of different tool wear states. The vectorization label of the initial wear stage is [0, 0, 1], the vectorization label of the normal wear stage is [0, 1, 0], and the vectorization label of the severe wear stage is [1, 0, 0];

[0083] The above features are input into the Transformer tool wear prediction model for training and optimization, a variable working condition parameter introduction layer is added to the model, the variable working condition parameters are introduced, and a VWC-Transformer model is built, so that the model has the ability to learn and train the variable working condition parameters;

[0084] In the target domain online monitoring process, the sensors collect a small amount of signals under different working condition parameter change conditions and perform preprocessing;

[0085] The trained VWC-Transformer model is migrated to the target domain by using the pre-training-fine-tuning transfer learning method, the migration task is completed, the model is reused and fine-tuned, and the tool wear state is predicted to determine whether the tool needs to be replaced.

[0086] In the embodiment, the variable working conditions are mainly changes in cutting parameters, including spindle speed, feed rate, and back engagement amount. The coordinated changes between the parameters can constitute various working conditions. The sample number ratio of each working condition is 1, and each working condition accounts for 30% of all samples in the offline training samples and 70% of all samples in the online test samples. The time feature extraction obtains the variance (var), mean (mean), and maximum (max) of each part of data, and the entire sequence is subdivided along the time dimension to form shorter time sequences, which not only reduces the amount of sensor data, but also extracts useful time features to prevent overfitting.

[0087] As Figure 3 The structure of the VWC-Transformer prediction model is described.

[0088] In the embodiment, the preprocessed features are input into the embedding layer, the input samples are mapped to a high-dimensional space using entity embedding, and the embedding dimension is 64. Then, the signal is positionally encoded by the position encoding layer to obtain the sequence order. The formula is:

[0089]

[0090]

[0091] In the formula, PE is the encoding function, pos is the position, d model is the feature dimension number of each time step data, q is the qth element of the position information encoding feature vector, and 2q∈[0,d model ).

[0092] After position encoding, the sequence is encoded using the multi-head self-attention mechanism to establish the relationship between different time steps in the input data and extract different time steps in the time sequence input data. The key and query are mapped to the output. A set of values with a dimension of N can be converted into a corresponding matrix V as a whole. The query and key with a dimension of M can be converted into matrices Q and K as a whole.

[0093] The function of scaling the attention of clicks is:

[0094]

[0095] The multi-head attention output is as follows:

[0096] MultiHead(Q,K,V)=Concatenate(head1,…,head h ,…,head H W C

[0097]

[0098] When Q = K = V = X i ′ At that time, the h-th head of multi-head self-attention is:

[0099]

[0100] In the formula, W C ∈R HN×M This is a learnable matrix. MultiHead is a multi-head attention mechanism; Concatenate means concatenation; head h Representing the h-th head, X′ i Let M be the input matrix, and N be the dimension of the key matrix and N be the dimension of the value matrix. V, or value, is a matrix obtained by transforming a set of values ​​of dimension N. Q, or query, is a matrix obtained by transforming the query and key of dimension M. K, or key, is a matrix obtained by transforming the query and key of dimension M.

[0101] Among them, residual connections and normalization layers are used to avoid gradient explosion or gradient vanishing after connecting multiple Transformer layers, and to enhance the stability of the model and accelerate its convergence.

[0102] The feedforward network layer introduces nonlinear modeling capabilities to the model, enhancing its representational power. Its calculation is shown in the formula:

[0103] FFN(x) = max(0, xW1+b1)W2+b2

[0104] The multi-condition parameter input layer introduces variable-condition parameter features. Since these parameters are represented by discrete rather than continuous values, the distance between feature values ​​is measured numerically. After data standardization, the concatenate join function is used to import the processing parameters into the multi-condition parameter layer as an array mapping. This enables the model to learn from variable-condition parameters.

[0105] The multi-condition parameter input layer includes: an input layer that takes multi-condition parameters as input; and a feature layer that connects multiple input parameters together using the concatenate function, and then combines the feature layer as part of the input layer to form the multi-condition parameter input layer.

[0106] ReLU as an activation function not only speeds up the calculation of the model, but also increases the sparsity of the network.

[0107] The formula is as follows:

[0108] ReLU(x)=max(0,x)

[0109] The full connection layer adopts a softmax classifier to classify the tool state. The VWC-Transformer prediction model is obtained by sequentially connecting the modules of each layer.

[0110] From Figure 4 It can be seen that the cutting parameter changes and migration tasks based on variable working conditions are divided into 3 kinds, each of which changes 2 levels, of which 1 is selected as the source domain (D S ), that is, the data domain under the historical working condition; 1 is selected as the target domain (D T ), that is, the data domain under the new working condition. Under each cutting parameter, a migration task from the source domain to the target domain (D S →D T ) is established, and 4 migration tasks from the source domain to the target domain (D S →D T ) are established when each cutting parameter changes simultaneously, that is, a total of 7 migration tasks are established for migration learning.

[0111] Among them, the migration task D S1 →D T1 is the spindle speed V c1 →V c2 change, the feed amount f1 and the back engagement amount C1 remain unchanged; the migration task D S2 →D T2 is the feed amount f1→f2 change, the spindle speed V c1 and the back engagement amount C1 remain unchanged; the migration task D S3 →D T3 is the back engagement amount C1→C2 change, the spindle speed V c1 and the feed amount f1 remain unchanged; the migration task D S1 →D T1 +D S2 →D T2 is the simultaneous change of spindle speed and feed amount, and the back engagement amount remains unchanged; the migration task D S1 →D T1 +D S3 →D T3 is the simultaneous change of spindle speed and back engagement amount, and the feed amount remains unchanged; D S2 →D T2 +D S3 →D T3For the feed rate and the back engagement amount to change simultaneously, the spindle speed remains unchanged; D S1 →D T1 +D S2 →D T2 +D S3 →D T3 For the spindle speed, the feed rate and the back engagement amount to change simultaneously.

[0112] After the multi-working condition parameter introduction layer, the edge distribution adaptation principle is added before the full connection layer to realize the migration of the VWC-Transformer model. First, the source domain data is input into the model to generate the regression loss of the source domain sample, and the output feature is extracted. The data collected in the target domain is input into the model to generate the regression loss of the target domain sample. Then the output features of the source domain and the target domain are extracted, and are input into the edge distribution adaptation principle to calculate the maximum mean difference (MMD) and increase the similarity between the two, and the loss function Loss converges to the optimal. Loss is:

[0113] Loss = MSE source + MSE target + lambda MMD 2 loss

[0114] In the formula, MSE source is the regression loss on the source domain sample; MSE target is the regression loss on the target domain sample;

[0115] MMD 2 loss is the adaptive loss; lambda is the penalty coefficient; and Loss is the final loss.

[0116] Since the similarity between the small amount of signals collected in the target domain and the source domain signal data is high, a fine-tuning method for freezing all layers and fine-tuning the last full connection layer is used. The sensor signal extracted by the VWC-Transformer is combined before the full connection layer, and the weight parameters of the entire network layer are retrained.

[0117] Finally, the input signal features are classified to determine whether the tool needs to be replaced.

[0118] It should be understood that those skilled in the art can make improvements or changes according to the above description, and all these improvements and changes shall belong to the protection scope of the appended claims of the present application.

Claims

1. A variable working condition tool wear state monitoring method based on VWC-Transformer transfer learning, characterized in that, The method comprises the following steps: Step 1) obtaining tool variable working condition historical data based on multi-sensor information; including vibration signal, cutting force signal, acoustic emission signal data, which are composed of source domain data; The variable working condition is the change of the working condition parameters, including the spindle speed, the feed rate and the back engagement amount; Step 2) time feature extraction and normalization processing are performed on each working condition data, and the specific steps are as follows: 2.1) slicing along the time dimension; obtaining the variance, mean and maximum value of each part of data to form a shorter time sequence; 2.2) time feature extraction is performed on the processed time sequence by using a short-time Fourier transform-based feature extraction method; The feature extraction adopts a short-time Fourier transform-based feature extraction method, which further reduces the time sequence data amount of the model input and improves the information capacity of a single time step under the premise of retaining rich information; The short-time Fourier transform formula is as follows: In the formula, h (tau-t) is a window function, f is a frequency, and t is a time; Step 3) the above feature sequence is input into a VWC-Transformer tool wear prediction model for training and optimization; The VWC-Transformer prediction model structure constructed in step 3) comprises the following in sequence: An input embedding layer; A position encoding layer; the position encoding layer encodes the signal by position to obtain the sequence order; A multi-head attention mechanism; the multi-head self-attention mechanism is used to encode the sequence to realize effective information extraction; A residual connection and normalization layer; A feedforward network layer; the feedforward network layer introduces nonlinear modeling capability for the model; A residual connection and normalization layer; A multi-working condition parameter input layer; the multi-working condition parameter input layer is used to introduce variable working condition parameter features; since the variable working condition parameters are represented by discrete values rather than continuous values, the distance relationship between the feature values is measured by the numerical size; after data standardization processing of the multi-working condition parameters, the processing parameters are mapped in the form of an array and introduced into the multi-working condition parameter layer by using a concatenate connection function; A ReLU activation function; A fully connected layer; Step 4) a small amount of real-time signals under different working condition parameter change conditions are collected by sensors and data preprocessing is performed; Step 5) a pre-training-fine-tuning transfer learning method is used to migrate the trained VWC-Transformer model to a target domain, complete the migration task, and reuse and fine-tune the model; The pre-training-fine-tuning transfer learning method is used to migrate the trained VWC-Transformer model to the target domain, and the specific steps are as follows: The cutting parameters are divided into three kinds, and one source domain D is established under each cutting parameter S to the target domain D T , and four migration tasks of the source domain to the target domain are established by mutual simultaneous change between each cutting parameter, that is, a total of seven migration tasks are established for migration learning; Wherein, Migration task D S1 → D T1 For spindle speed V c1 → V c2 Change, feed amount f1, back engagement amount C1 remain unchanged; Migration task D S2 → D T2 For the feed amount f1→f2 change, spindle speed V c1 , back engagement amount C1 remains unchanged; Migration task D S3 → D T3 For the amount of back engagement C1→C2 change, spindle speed V c1 , the feed amount f1 remains unchanged; Migration task D S1 → D T1 + D S2 → D T2 For the spindle speed and feed rate changes, the back engagement amount remains unchanged; Migration task D S1 → D T1 + D S3 → D T3 For the spindle speed and back to the amount of change, the amount of feeding remains unchanged; Migration task D S2 → D T2 + D S3 → D T3 For the feed and back engagement amount change simultaneously, spindle speed remains unchanged; Migration task D S1 → D T1 + D S2 → D T2 + D S3 → D T3 For spindle speed, feed rate, back engagement amount change simultaneously; Step 6) the tool wear state is predicted by using the migrated model, so as to judge whether the tool needs to be replaced.

2. The VWC-Transformer based transfer learning method for tool wear state monitoring under varying operating conditions according to claim 1, characterized in that, In step 3), the multi-working condition parameter input layer comprises: an input layer taking the working condition parameters as input; and a feature layer connecting multiple input parameters together by using a concatenate function, and then combining the feature layer as part of the input layer to form the multi-working condition parameter input layer.

3. The VWC-Transformer based transfer learning method for tool wear state monitoring under varying working conditions according to claim 1, characterized in that, In step 3), the position encoding layer encodes the signal by position to obtain the sequence order, and the formula is as follows: where PE is an encoding function, pos is the position, d model is the number of feature dimensions of each time step data, q is the qth element of the position information encoding feature vector, 2q∈[0,d model ).

4. The VWC-Transformer based transfer learning method for tool wear state monitoring under varying working conditions according to claim 1, characterized in that, In the step 3), the sequence is encoded by using a multi-head self-attention mechanism after position encoding, so as to establish the connection between different time steps in the input data, extract different time steps in the time sequence input data, and filter out important information associated with each other from a large amount of input data. The function of scaling the click attention is as follows: The multi-head attention output is as follows: MultiHead(Q, K, V) = Concatenate(head1,..., head h ,…,head H )W C When Q = K = V = X' i the hth head of the multi-headed self-attention is: wherein, W C ∈R HN×M is a learnable matrix, MultiHead is a multi-head attention mechanism, Concatenate represents concatenation, head h represents the hth head, X′ i is an input matrix, M is a key matrix dimension, N is a value matrix dimension, V is a matrix converted from a set of numerical values with dimension N, Q is a matrix converted from a query and a key with dimension M, and K is a matrix converted from a query and a key with dimension M.

5. The VWC-Transformer based transfer learning method for tool wear state monitoring under varying operating conditions according to claim 1, wherein, In the step 5), After introducing the multi-working-condition parameters, the migration of the VWC-Transformer model is as follows: First, the source domain data is input into the model to generate the regression loss of the source domain sample; The data collected by the target domain is input into the model to generate the regression loss of the target domain sample; Then, the features output by the source domain and the target domain are extracted, and the maximum mean difference MMD is calculated by inputting the features into the edge distribution adaptation principle, so as to increase the similarity between the two, and the loss function Loss converges to the optimal value, and the Loss is as follows: Loss = MSE source + MSE target + λMMD 2 loss In the formula, MSE source is the regression loss on the source domain sample; MSE target is the regression loss on the target domain sample; MMD 2 loss is the adaptive loss; λ is the penalty coefficient; Loss is the final loss; All layers are frozen, the last fully connected layer is fine-tuned, the sensor signal extracted by the VWC-Transformer is combined before the fully connected layer, and the weight parameters of the entire network layer are retrained.

Images