A deep learning-based tool wear detection method

Abstract

The application belongs to the technical field of deep learning neural networks, and particularly relates to a tool wear detection method based on deep learning, which comprises the following steps: data set construction, data preprocessing, data set division, model construction, model training and model evaluation, the data set construction collects three original time domain signals of force signals, vibration signals and acoustic emission signals in the machining process to construct a data set; the data preprocessing converts the data set into a form available for training and identification by a deep learning model and performs normalization processing; the data set division divides the data set into a training set, a validation set and a test set; the model construction uses CNN, GRU and LSTM networks to construct the model and performs joint analysis through feature fusion; the model training trains the model using specified training parameters to obtain an identification model; and the model evaluation evaluates the identification effect of the model using F1-Score.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep learning neural networks, and particularly relates to a tool wear detection method based on deep learning. BACKGROUND

[0002] Due to serious instrument measurement obstruction, the direct detection method has poor effect, so the indirect detection method is widely used. The indirect detection mainly uses sensors to collect force signals, vibration signals and acoustic emission signals in the machining process, and through preprocessing and sensitive feature extraction, a machine learning model is used to monitor tool wear. This method is limited by the quality of feature extraction, and the results of traditional machine learning methods are unstable, and there are great differences between different models. Different feature extraction methods have a great influence on the final results.

[0003] Reasons for problems or defects: the existing tool wear detection methods are mainly machine learning methods. In order to extract features closely related to the tool wear state, researchers not only need to be good at observation and discovery, but also need certain skills and experience. Moreover, the extracted features have weak interpretability and universality, sensitive features are easily missed, and the recognition effect cannot meet the actual demand. SUMMARY

[0004] In view of the above problems that the technical model of the tool wear has poor universality, the application provides a tool wear detection method based on deep learning.

[0005] In order to solve the above technical problems, the technical scheme adopted by the application is:

[0006] A tool wear detection method based on deep learning, comprising the following steps:

[0007] S100, data set construction: collecting three kinds of original time domain signals of force signals, vibration signals and acoustic emission signals in the machining process, and constructing the data set for training and testing;

[0008] S200, data preprocessing: converting the data set into a form available for deep learning model training and recognition, and performing normalization processing;

[0009] S300, data set division: dividing the data set into a training set, a validation set and a test set;

[0010] S400, model construction: using CNN, GRU and LSTM networks to construct the model, and performing joint analysis through feature fusion;

[0011] S500, model training: training the model by using specified training parameters to obtain a tool wear condition recognition model;

[0012] S600, model evaluation: using F1-score to evaluate the model recognition effect.

[0013] In the S100 data set construction, the data set is the tool wear competition data of the American PHM Association in 2010, the experimental machine tool is -TechRFM760 high-speed numerical control milling machine, the experimental tool is three-blade tungsten carbide ball milling cutter, the cutting material is stainless steel HRC-52, the force signal, vibration signal and acoustic emission signal original time domain signal in the machining process are collected through the force sensor, acceleration sensor and acoustic emission sensor, the signal sampling frequency is 50KHz, each time the tool is cut along the X direction to cut 108mm, which is recorded as a cutting stroke, each tool cuts 315 strokes, and the wear amount of the tool is recorded after each cutting stroke, a total of 945 times, forming 945 samples.

[0014] In the S200 data preprocessing, the force signal, vibration signal and acoustic emission signal original time domain signal are respectively constructed into standard npy data set, and the wear amount data set is constructed as data label, the force signal, vibration signal and acoustic emission signal original time domain signal data are normalized to the range of (0, 1), and the normalization method is Min-Max normalization, Where X is all data points of the entire stroke, x is the data point to be normalized, and x' is the value of the normalized data point.

[0015] In the S300 data set division, the data is divided into training set, validation set and test set in the ratio of 7:1:2, wherein the training set is used for parameter iterative training of the model, the validation set is used for verifying whether the model has learned all effective features and has been completely trained, and the test set is used for evaluating the recognition effect of the model.

[0016] In the S400 model construction, the model takes the force signal, vibration signal and acoustic emission signal as input, and the wear degree of the tool as output, and uses CNN, GRU and LSTM to construct the network, after the data is input into the network, the three inputs are first extracted through the CNN network, and the data dimension is improved, each input corresponding CNN is 3 layers, the convolution kernel size of 3 layers of CNN is 3, and the step is 1, after each convolution, ReLU is used for activation, which helps the network to better fit the features.

[0017] The ReLU formula in the S400 model construction is f(x) = max(0, x), where x is the input feature value, f(x) is the feature value after ReLU activation, and then one Max Pooling and one Batch Normalization are performed to process the data and prevent gradient disappearance. The data dimension is increased to 8, 16, and 32 through three convolutions. After the CNN extracts the features, a GRU network with 32 hidden nodes is used to further extract the time domain features of the data. GRU has two gates, a reset gate and an update gate. The specific calculation method is: z = sigma(x t U z +s t-1 W z ), r = sigma(x t U r +s t-1 W r ), h = tanh(x t U h +(S t-1 *r)W h , s t = (1-z)*h+z*s t-1 , where t represents the time step, x t is the t-th component of the input sequence, W is the corresponding weight, U represents the update gate, z represents the output of the update gate, sigma represents the sigmoid function, r represents the output result of the reset gate, h is the hidden information of the GRU unit, and GRU network is used to extract features. After that, three 32-dimensional feature vectors are obtained, and the Concate method is used to combine the three feature vectors.

[0018] In the S400 model construction, after feature fusion, the LSTM network is input. LSTM uses a gating unit to control which information should be remembered and which information should be forgotten. This is achieved through three gates. The first gate is the forget gate, which uses a sigmoid layer to determine the information to be deleted from the LSTMCell state. f t = sigma(W f ·[h t-1 ,x t ]+b f ); The second gate is the input gate, which uses a sigmoid layer to determine the value to be updated and a tanh layer to define the new updated value. i t = sigma(W f ·[h t-1 , x t ]+b i ), c t = tanh(W c ·[h t-1 , xt ]+b c ), the third gate output gate determines the output of the current LSTMCell state, o t = sigma (W o · [h t-1 , x t ]+b o ), h t =o t ·tanh (c t ), wherein x t is the input sequence at time t, h t-1 is the past hidden state, b f , b i , b c , b o is the bias vector of each layer, and the LSTM network outputs the predicted tool wear degree after identifying and analyzing the data features.

[0019] In the S500 model training, after the network is built, the network parameters are trained using the training set data, Adam is used as the optimizer, the initial learning rate is 0.01, the learning rate is attenuated by 60% every 100 epochs, the batchsize size is 32, the loss function uses the cross-entropy loss function, 300 epochs are set for training, and the model loss value does not decrease for 20 consecutive epochs, then the training is stopped, the model is saved, and the training completed model is trained again using the validation set, if the model loss value does not decrease any more, it indicates that the model has reached the best state, the model training is completed, if the loss continues to decrease, the model learning rate is reduced, the model is trained again using the training set data, the above steps are repeated until the model reaches the best state.

[0020] In the S600 model evaluation, the trained model is used to classify and predict the test set data, the prediction results are compared with the corresponding labels, the identification effect is evaluated, the evaluation method is F1-score, the higher the F1-Score value, the better the identification effect, Wherein, F1 is F1-score, A is accuracy, R is recall rate, TP is the number of positive class determined as positive class, FP is the number of negative class determined as positive class, FN is the number of positive class determined as negative class, and TN is the number of negative class determined as negative class.

[0021] Compared with the prior art, the present application has the beneficial effects that:

[0022] The model proposed in the present application extracts signal features from the original time domain signal, identifies the tool wear state through separate extraction and fusion of the features, not only proves the effectiveness of the convolution network on the time domain signal, but also has excellent detection effect on the tool wear degree, and has good generalization ability and strong robustness. Attached Figure Description

[0023] Figure 1 Flowchart of the main steps of this invention;

[0024] Figure 2 The network structure diagram of this invention; Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] A deep learning-based method for tool wear detection, such as Figure 1 As shown, it includes the following steps:

[0027] S100. Dataset Construction: Collect three types of raw time-domain signals—force signals, vibration signals, and acoustic emission signals—during the processing to construct training and testing datasets.

[0028] S200, Data Preprocessing: Convert the dataset into a format suitable for training and recognition by deep learning models, and perform normalization processing;

[0029] S300, Dataset Partitioning: Divide the dataset into training set, validation set and test set;

[0030] S400, Model Building: The model is built using CNN, GRU, and LSTM networks, and joint analysis is performed through feature fusion.

[0031] S500, Model Training: The model is trained using specified training parameters to obtain a tool wear identification model;

[0032] S600, Model Evaluation: The F1-Score is used to evaluate the model's recognition performance.

[0033] Furthermore, in step S100, the dataset used is the tool wear competition data from the 2010 US PHM Association, and the experimental machine tool is... -Tech RFM760 high-speed numerical control milling machine, the experimental tool is a three-blade tungsten carbide ball head milling cutter, the cutting material is stainless steel HRC-52, the force signal, vibration signal and acoustic emission signal original time domain signals in the processing process are collected through the force measuring instrument, acceleration sensor and acoustic emission sensor, the signal sampling frequency is 50KHz, each time the tool is cut along the X direction for 108mm, which is recorded as one cutting stroke, each tool is cut for 315 strokes, and the wear of the tool is recorded after each cutting stroke, a total of 945 times, forming 945 samples.

[0034] Further, in the data preprocessing of step S200, the force signal, vibration signal and acoustic emission signal original time domain signals are respectively constructed into standard npy data sets, and a wear data set is constructed as a data label. The data amplitude will affect the network training speed and recognition effect, therefore, the force signal, vibration signal and acoustic emission signal original time domain signal data are normalized to the range of (0, 1) to accelerate the network convergence and prevent model overfitting. The normalization method is Min-Max normalization. Wherein X is all data points of the entire stroke, x is the data point to be normalized, and x' is the value of the normalized data point.

[0035] Further, in the data set division of step S300, the data is divided into a training set, a validation set and a test set in a ratio of 7:1:2. The training set is used for parameter iterative training of the model, the validation set is used to verify whether the model has learned all effective features and has been completely trained, and the test set is used to evaluate the recognition effect of the model.

[0036] Further, in the model construction of step S400, as shown in Figure 2 The model takes the force signal, vibration signal and acoustic emission signal as input and the wear degree of the tool as output. CNN, GRU and LSTM are used to construct the network. After the data is input into the network, the three inputs are first extracted through the CNN network to improve the data dimension. Each input corresponds to a 3-layer CNN, and the convolution kernel size of the 3-layer CNN is 3 and the step is 1. After each convolution, ReLU is used for activation to help the network better fit the features.

[0037] Further, in the step S400 model construction, the ReLU formula is f(x) = max(0, x), where x is the input feature value, and f(x) is the feature value after ReLU activation. Then, one Max Pooling and one Batch Normalization are performed to process the data and prevent gradient disappearance. The data dimension is increased to 8, 16, and 32 through three convolutions. After the CNN extracts the features, a GRU network with 32 hidden nodes is used to further extract the time domain features of the data. GRU has two gates, a reset gate and an update gate. Intuitively, the reset gate determines how to combine new input information with previous memory, and the update gate defines the amount of previous memory saved to the current time step. These two gate vectors determine which information can be output as the gated recurrent unit. The special feature of these two gate mechanisms is that they can save information in long sequences without being cleared over time or removed because it is not related to prediction. The specific calculation method is: z = σ(x t U z +s t-1 W z ),

[0038] r = σ(x t U r +s t-1 W r ), h = tanh(x t U h + (s t-1 *r)W h ),

[0039] s t = (1-z)*h+z*s t-1 , where t represents the time step, x t is the t-th component of the input sequence, W is the corresponding weight, U represents the update gate, z represents the output of the update gate, σ represents the sigmoid function, r represents the output result of the reset gate, h is the hidden information of the GRU unit, and GRU network is used to extract features. After the feature extraction is completed, three 32-dimensional feature vectors are obtained. The Concate method is used to combine the three feature vectors.

[0040] Further, in the step S400 model construction, after the feature fusion is completed, the LSTM network is input. LSTM uses a gating unit to control which information should be remembered and which information should be forgotten. This is achieved through three gates. The first gate is the forget gate, which uses a sigmoid layer to determine the information to be deleted from the LSTMCell state. f t = σ(W f ·[h t-1 ,x t ]+bf ); the second gate is an input gate, using a sigmoid layer to determine the value to be updated, using a tanh layer to define the new updated value, i t = σ(W f · [h t-1 , x t ]+b i ), c t =tanh(W c · [h t-1 , x t ]+b c ), the third gate output gate determines the output of the LSTMCeIl state, o t = σ(W o · [h t-1 , x t ]+b o ), h t =o t · tanh(c t ), where x t is the input sequence at time t, h t-1 is the past hidden state, b f , b i , b c , b o are the bias vectors of each layer, and the LSTM network outputs the predicted tool wear degree after identifying and analyzing the data features.

[0041] Further, in step S500, after the network is built, the network parameters are trained using the training set data, Adam is used as the optimizer, the initial learning rate is 0.01, the learning rate is attenuated by 60% every 100 epochs, the batchsize size is 32, the loss function uses the cross-entropy loss function, and the training is set to 300 epochs. If the model loss value does not decrease for 20 consecutive epochs, stop training, save the model, and use the validation set to train the trained model again. If the model loss value does not decrease, it means that the model has reached the best state, and the model training is completed. If the loss continues to decrease, reduce the learning rate of the model, and use the training set data to train the model again. Repeat the above steps until the model reaches the best state.

[0042] Further, in step S600, the trained model is used to classify and predict the test set data, the prediction results are compared with their corresponding labels, and the recognition effect is evaluated. The evaluation method is F1-score. The higher the F1-Score value, the better the recognition effect, Wherein, F1 is F1-score, A is accuracy, R is recall, TP is the number of positive class determined as positive class, FP is the number of negative class determined as positive class, FN is the number of positive class determined as negative class, and TN is the number of negative class determined as negative class.

[0043] The above only describes the preferred embodiments of the present application in detail, but the present application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application, and all the changes shall be included in the protection scope of the present application.

Claims

1. A deep learning-based tool wear detection method, characterized by: Comprise the following steps: S100, data set construction: collect the force signal, vibration signal and acoustic emission signal 3 original time domain signals in the processing process, and construct the training and test data set; In the S100 data set construction, the data set is the competition data of tool wear in 2010 of American PHM Association, the experimental machine tool is Röders-Tech RFM760 high-speed numerical control milling machine, the experimental tool is three-blade tungsten carbide ball milling cutter, the cutting material is stainless steel HRC-52, the force signal, vibration signal and acoustic emission signal original time domain signal in the processing process are collected by force sensor, acceleration sensor and acoustic emission sensor, the signal sampling frequency is 50KHz, each time the tool is cut along the X direction to cut 108mm, which is recorded as a cutting stroke, each tool cuts 315 strokes, and the wear of the tool flank surface of each cutting edge is recorded after each cutting stroke, a total of 945 times are collected, forming 945 samples; S200, data preprocessing: convert the data set into a form that can be used for training and identification by deep learning model, and perform normalization processing; In the S200 data preprocessing, standard npy data sets are respectively constructed for the force signal, the vibration signal and the acoustic emission signal original time domain signals, and a wear amount data set is constructed as a data label. The force signal, the vibration signal and the acoustic emission signal original time domain signal data are normalized to the range of (0, 1) by Min-Max normalization, wherein X is all data points of the entire stroke, and x is a data point to be normalized, is a value of the normalized data point. S300, data set division: divide the data set into training set, validation set and test set; In the S300 data set division, the data is divided into training set, validation set and test set in the ratio of 7:1:2, wherein the training set is used for parameter iterative training of the model, the validation set is used to verify whether the model has learned all the effective features and has been completely trained, and the test set is used to evaluate the recognition effect of the model; S400, model construction: use CNN, GRU and LSTM network to construct the model, and perform joint analysis through feature fusion; In the S400 model construction, the model takes the force signal, vibration signal and acoustic emission signal as input, and the tool wear degree as output, uses CNN, GRU and LSTM to construct the network, after the data is input into the network, the three inputs are first extracted through the CNN network, and the data dimension is improved, each input corresponding CNN is 3 layers, the convolution kernel size of 3 layers of CNN is 3, and the step is 1, after each convolution, ReLU is used for activation, which helps the network to better fit the features; S500, model training: train the model using specified training parameters to obtain a tool wear condition recognition model; S600, model evaluation: use F1-Score to evaluate the recognition effect of the model.

2. The tool wear detection method based on deep learning according to claim 1, characterized in that: The ReLU formula in the S400 model construction is where x is the input feature value, f(x) is the feature value after ReLU activation, then Max Pooling and Batch Normalization are performed once, the data is processed to prevent gradient disappearance, and the data dimension is increased to 8, 16 and 32 respectively through three times of convolution. After the CNN extracts the features, a GRU network with 32 hidden nodes is used to further extract the time domain features of the data. GRU has two gates, a reset gate and an update gate. The specific calculation method is: , , , where t represents the time step, that is, the t-th component of the input sequence, W is the corresponding weight, U represents the update gate, z represents the output of the update gate, sigmoid function, r represents the output result of the reset gate, h is the hidden information of the GRU unit, and after the GRU network extracts the features, three 32-dimensional feature vectors are obtained. The three feature vectors are combined using the Concate method.

3. The tool wear detection method based on deep learning according to claim 2, characterized in that: In the S400 model construction, after the feature fusion is completed, the LSTM network is input, the LSTM uses a gating unit to control which information should be remembered and which information should be forgotten, which is realized through three gates, the first gate is a forgetting gate, a sigmoid layer is used to determine the information to be deleted from the LSTM Cell state, ; The second gate is the input gate, which uses a sigmoid layer to determine the value to be updated and a tanh layer to define the new update value, , , the third gate output gate decides the output of the LSTM Cell state, , , where is the input sequence at time t, is the past hidden state, is the bias vector of each layer. After the LSTM network analyzes the data features, it outputs the predicted tool wear degree.

4. The tool wear detection method based on deep learning according to claim 3, characterized in that: In the S500 model training, after the network is built, the network parameters are trained using the training set data, Adam is used as the optimizer, the initial learning rate is 0.01, the learning rate is attenuated by 60% every 100 epochs, the batch size is 32, the loss function uses the cross-entropy loss function, 300 epochs are set for training, the model loss value does not decrease for 20 consecutive epochs, the training is stopped, the model is saved, the training completed model is trained again using the validation set, if the model loss value does not decrease any more, it indicates that the model has reached the best state, the model training is completed, if the loss continues to decrease, the model learning rate is reduced, the model is trained again using the training set data, the above steps are repeated until the model reaches the best state.

5. The tool wear detection method based on deep learning according to claim 4, characterized in that: In the S600 model evaluation, the trained model is used to classify and predict the test set data, the prediction results are compared with the corresponding labels, the recognition effect is evaluated, the evaluation method is F1-score, the higher the F1-Score value, the better the recognition effect, , , wherein F1 is F1-score, A is accuracy, R is recall, TP is the number of positive class determined as positive class, FP is the number of negative class determined as positive class, FN is the number of positive class determined as negative class, and TN is the number of negative class determined as negative class.

Images