A two-stage fusion-based cutter wear state classification method for disc-type milling cutters of a gear milling machine

Abstract

The application provides a two-stage fusion based on sound and vibration signals for a milling cutter wear state classification method of a disc type milling cutter of a gear milling machine, relates to the fields of multi-source information fusion and fault diagnosis, and mainly comprises the following steps: 1) four noise sensors are used to collect noise signals in four directions, and the correlation degrees of the noise signals in different directions are obtained through a gray B-type correlation degree method, so that the sound signals after fusion are obtained through data stage fusion; 2) the vibration signals of the main shaft of the gear milling machine are collected, and the collected vibration signals and the fused sound signals are preprocessed and subjected to time-frequency feature extraction; a Relieff-mRMR joint algorithm is used to optimize the extracted characteristic values, and the characteristic values with greater influence are selected to construct a data set; 3) a BiLSTM is used to extract the features of the data set, and then an XGBoost is used to classify the extracted characteristic vectors, so that the classification results of different wear states of the disc type milling cutter are obtained. The two-stage fusion method is used to fuse the sound and vibration signals, so that the characteristics of tool wear can be more effectively captured, and the wear state of the disc type milling cutter of the gear milling machine can be judged.

Description

Technical Field

[0001] This invention relates to the fields of monitoring various parameters of CNC machine tools, data fusion, and fault diagnosis, specifically to a method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals. Background Technology

[0002] With the rapid development of modern manufacturing, CNC gear milling machines are playing an increasingly important role in gear processing. As one of the core components of a CNC gear milling machine, the wear condition of the cutting tool directly affects processing quality, efficiency, and cost. Tool wear is inevitable during processing, and real-time monitoring and classification of wear conditions are crucial for ensuring processing accuracy, extending tool life, reducing production costs, and improving production safety. Therefore, the classification and monitoring of gear milling machine tool wear conditions is essential.

[0003] Multi-source signal fusion technology is an important technique in the field of information processing, aiming to extract effective and reliable information from multiple different sources to improve the accuracy of decision-making and system performance. With the rapid development of sensor technology, network communication technology, and artificial intelligence, multi-source signal fusion technology has shown great application potential in various fields such as military, aerospace, transportation, medicine, and environmental monitoring. In modern technological applications, a single sensor or information source often cannot provide complete and comprehensive information. The fusion of multi-source information can compensate for the shortcomings of a single information source, providing richer and more accurate data, thereby supporting more complex decision-making processes. Through multi-level fusion of multi-source signals, the accuracy of tool wear state classification can be further improved. Therefore, we propose a method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals. Summary of the Invention

[0004] To address and resolve the above problems, this invention provides a method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals. This method can more accurately assess the wear state of disc milling cutters and effectively reduce the problem of low gear machining accuracy caused by tool wear.

[0005] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0006] A method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals includes the following steps:

[0007] A method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals mainly includes the following steps:

[0008] S1: Four noise sensors are used to collect noise signals from four directions, and the correlation degree of signals from different directions is obtained by using the gray B-type correlation method. Data-level fusion is then performed to obtain the fused sound signal.

[0009] S2: An accelerometer is used to collect vibration signals of the gear milling machine spindle. Then, time-frequency domain features are extracted from the fused sound signal and vibration signal. The extracted features are optimized using the ReliefF-mRMR joint algorithm, and the feature vectors with the greatest influence are selected to construct the dataset.

[0010] S3: Use BiLSTM to extract features from the dataset, transform the extracted features to form a feature vector that XGBoost can recognize, and then use XGBoost for classification to obtain the classification results of different wear states of the milling machine tool.

[0011] In step S1, four noise sensors are used to collect noise signals from four directions. The correlation degree of the signals from different directions is obtained through the gray B-type correlation method. Data-level fusion is then performed to obtain the fused sound signal. The main steps include:

[0012] S1.1: The noise sensors are evenly distributed around the spindle of the gear milling machine spindle box, and the collected data is arranged in a time series.

[0013] S1.2: Using one point as a reference sequence and the other three points as comparison sequences, the correlation degree between the points in the reference sequence and the other points is obtained; the grey B-type correlation degree method is used, and the discrete signal is defined as y i (k),y j (k), k = 1, 2, 3..., n, First calculate the displacement difference:

[0014]

[0015] Calculate the speed difference again:

[0016] Next, calculate the acceleration difference:

[0017] The formula for the grey type B correlation degree is:

[0018]

[0019] Where X i As a reference sequence, X j For comparison sequences;

[0020] S1.3 After obtaining the gray B-type correlation degree of each point relative to other points, the correlation degree energy of the measurement point can be calculated, and then the sound fusion signal can be obtained by using the relationship between correlation degree and weight.

[0021] Assume the grey B-type correlation energy between the sound signal at each point and the signals at the other three points is E. ij Then the correlation signal energy of the i-th point is E i :

[0022]

[0023] S1.4 Since the correlation degree is directly proportional to the weight, the sum of the weight values ​​is 1. The specific relationship is as follows:

[0024] E1:E2:E3:E4=P1:P2:P3:P4

[0025] P1+P2+P3+P4=1;

[0026] P1, P2, P3, and P4 are the weights of the four sound test points. After obtaining the weights of each point, weighted fusion is performed to obtain the sound fusion signal Nf at the data level.

[0027] Nf = P1X1 + P2X2 + P3X3 + P4X4.

[0028] In step S2, an accelerometer is used to collect vibration signals from the spindle of the gear milling machine. Then, time-frequency domain features are extracted from the fused sound and vibration signals. The extracted features are optimized using the ReliefF-mRMR joint algorithm. The dataset is constructed by selecting the feature vectors with the greatest influence. The main steps are as follows:

[0029] S2.1 uses an accelerometer placed at the spindle position. After the tool goes through a full life cutting process, the vibration signal of each cut is collected. The vibration signals of the three wear states are marked as 1, 2, and 3 by the tool wear curve.

[0030] S2.2 extracts the time-frequency domain features of the fused sound and vibration signals, including time-domain features: mean, root mean square, standard deviation, waveform factor, skewness, kurtosis, peak value, peak factor, and impulse factor; and frequency-domain features: power spectrum mean, frequency centroid, mean square frequency, and time-frequency domain features, namely: wavelet packet energy.

[0031] S2.3 First, the time-frequency features are filtered using ReliefF, and a sample R is randomly selected from the sample set S. i Search and R i k nearest neighbors H of the same class of samples j and the k nearest neighbors M of different class sample sets j The weight W of each feature is updated according to the following formula:

[0032]

[0033] In the formula, m is the number of iterations; diff(A,X,Y) represents the distance between two objects on feature A; k is the number of nearest neighbor samples; P(c) represents the prior probability of target class c;

[0034] Based on W(A), useless feature sets are eliminated, and 15 feature subsets are selected from the 26 feature sets;

[0035] S2.4 From the selected 15 feature subsets, the feature subsets are further filtered using the mRMR algorithm. If x and y are two continuous random variables, P(x) and P(y) are the probabilities of x and y respectively, and the joint probability is P(x,y), then the formula for the mutual information (MI) between them is as follows:

[0036]

[0037] The formula for the m RMR principle is expressed as follows:

[0038]

[0039] In the formula, C represents the target category; S represents the feature set; is the number of features; x i x j Represent the i-th and j-th features;

[0040] The mRMR principle is denoted as maxΦ(D,R). Using the Mutual Information Quotient (MIQ) standard Φ(D,R) = D / R, a forward sequence search method is employed to find approximately optimal fault features. Assume X is the original feature set, and there are already m-1 feature subsets S. m-1 In the remaining features {XS m-1 The selection of the m-th feature must satisfy the following:

[0041]

[0042] Based on the results obtained from the m RMR principle, each feature is sorted from high to low, and the optimal feature classification subset is selected to construct the dataset.

[0043] In step S3, BiLSTM is used to extract features from the dataset. The extracted features are then transformed to form a feature vector that XGBoost can recognize. XGBoost is then used for classification to obtain classification results for different wear states of the milling machine tool. The main steps include:

[0044] S4.1 For the constructed dataset, a bidirectional long short-term recurrent neural network (BiLSTM) is used to extract features from the dataset. The final output calculation formula of BiLSTM is shown below:

[0045] hi =f1(w1x i +w2h i-1 );

[0046] H i =f2(w3x i +w5H i+1 );

[0047] Y i =f3(w4h i +w6H i );

[0048] Where, x i (i = 1, 2, ..., t) corresponds to the time input data, h i (i = 1, 2, ..., t) represents the hidden state of the LSTM during the forward iteration, H i (i = 1, 2, ..., t) represents the hidden state of the LSTM during the backward iteration, Y i (i = 1, 2, ..., t) represents the corresponding output data, w i (i = 1, 2, ..., 6) represents the weights of each layer, where f1, f2, f3 correspond to the activation functions of different layers;

[0049] S4.2 reconstructs the feature vectors extracted by BiLSTM and then enters the Extreme Gradient Boosting Decision Tree (XGBoost) for tool wear state classification.

[0050] The present invention provides a method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals, which has the following beneficial effects:

[0051] Compared to single-sensor assessment of tool wear conditions, multi-sensor fusion offers higher accuracy in classifying tool wear conditions. Furthermore, compared to single-level fusion, which only fuses feature layers, two-level fusion further improves assessment accuracy. Classification using XGBoost achieves higher accuracy than classification using a softmax classifier. Attached Figure Description

[0052] Figure 1 This is a distribution diagram of sensor measurement points.

[0053] Figure 2 This is a schematic diagram of the data acquisition system.

[0054] Figure 3 This is a flowchart of the diagnostic model.

[0055] Figure 4 This is a diagram of the BiLSTM network structure. Detailed Implementation

[0056] The specific embodiments of the present invention will be described with reference to the accompanying drawings and examples.

[0057] As attached Figure 1 , attached Figure 2 and appendix Figure 3 As shown, a method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals includes the following steps:

[0058] S1: Four noise sensors are used to collect noise signals from four directions, and the correlation degree of signals from different directions is obtained by using the gray B-type correlation method. Data-level fusion is then performed to obtain the fused sound signal.

[0059] S2: An accelerometer is used to collect vibration signals of the gear milling machine spindle. Then, time-frequency domain features are extracted from the fused sound signal and vibration signal. The extracted features are optimized using the ReliefF-mRMR joint algorithm, and the feature vectors with the greatest influence are selected to construct the dataset.

[0060] S3: Use BiLSTM to extract features from the dataset, transform the extracted features to form a feature vector that XGBoost can recognize, and then use XGBoost for classification to obtain the classification results of different wear states of the milling machine tool.

[0061] One possible implementation involves using four noise sensors to collect noise signals from four directions, obtaining the correlation between the signals from different directions using the grey B-type correlation method, and performing data-level fusion to obtain the fused sound signal. This mainly includes the following steps:

[0062] S1.1: The noise sensors are evenly distributed around the spindle of the gear milling machine spindle box, and the collected data is arranged in a time series.

[0063] S1.2: Using one point as a reference sequence and the other three points as comparison sequences, the correlation degree between the points in the reference sequence and the other points is obtained. The grey B-type correlation degree method is used, defining the discrete signal as y... i (k),y j (k), k = 1, 2, 3..., n, First calculate the displacement difference:

[0064]

[0065] Calculate the speed difference again:

[0066]

[0067] Next, calculate the acceleration difference:

[0068] The formula for the grey type B correlation degree is:

[0069]

[0070] Where X i As a reference sequence, X j For comparison sequences.

[0071] S1.3 After obtaining the grey B-type correlation degree of each point relative to other points, the correlation energy of that measurement point can be calculated, and then the sound fusion signal can be obtained using the relationship between correlation degree and weight. Assume that the grey B-type correlation energy of the sound signal at each point with the signals of the other three points is E. ij Then the correlation signal energy of the i-th point is E i :

[0072]

[0073] S1.4 Since the correlation degree is directly proportional to the weight, the sum of the weight values ​​is 1. The specific relationship is as follows:

[0074] E1:E2:E3:E4=P1:P2:P3:P4

[0075] P1+P2+P3+P4=1;

[0076] P1, P2, P3, and P4 are the weights of the four sound test points. After obtaining the weights of each point, weighted fusion is performed to obtain the sound fusion signal Nf at the data level.

[0077] Nf = P1X1 + P2X2 + P3X3 + P4X4;

[0078] One possible implementation involves using an accelerometer to collect vibration signals from the spindle of a gear milling machine. Then, time-frequency domain features are extracted from the fused sound and vibration signals. The extracted features are then optimized using a ReliefF-mRMR joint algorithm, and the most influential feature vectors are selected to construct a dataset. This process mainly includes the following steps:

[0079] S2.1 uses an accelerometer placed at the spindle position. After the tool goes through a full life cutting process, the vibration signal of each cut is collected. The vibration signals of the three wear states are marked as 1, 2, and 3 by the tool wear curve.

[0080] S2.2 extracts the time-frequency domain features of the fused sound and vibration signals, including time-domain features (mean, root mean square, standard deviation, waveform factor, skewness, kurtosis, peak value, peak factor, impulse factor), frequency-domain features (power spectrum mean, frequency centroid, mean square frequency), and time-frequency domain features (wavelet packet energy).

[0081] S2.3 First, the time-frequency features are filtered using ReliefF, and a sample R is randomly selected from the sample set S. i Search and R i k nearest neighbors H of the same class of samples j and the k nearest neighbors M of different class sample sets j The weight W of each feature is updated according to the following formula:

[0082]

[0083] In the formula, m is the number of iterations; diff(A,X,Y) represents the distance between two objects on feature A; k is the number of nearest neighbor samples; and P(c) represents the prior probability of target class c.

[0084] Based on W(A), useless feature sets are eliminated, and 15 feature subsets are selected from the 26 feature sets.

[0085] S2.4 From the selected 15 feature subsets, the feature subsets are further filtered using the mRMR algorithm. If x and y are two continuous random variables, P(x) and P(y) are the probabilities of x and y respectively, and the joint probability is P(x,y), then the formula for the mutual information (MI) between them is as follows:

[0086]

[0087] The formula for the m RMR principle is expressed as follows:

[0088]

[0089] In the formula, C represents the target category; S represents the feature set; is the number of features; x i x j Let represent the i-th and j-th features. The mRMR principle is denoted as maxΦ(D,R). By using the mutual information quotient (MIQ) standard Φ(D,R) = D / R, a forward sequence search method is used to find the approximate optimal fault features. Assume X is the original feature set, and there are already m-1 feature subsets S. m-1 In the remaining features {XS m-1 The selection of the m-th feature must satisfy the following:

[0090]

[0091] Based on the results obtained from the mRMR principle, each feature is sorted from high to low, and the optimal feature classification subset is selected to construct the dataset.

[0092] One possible implementation method is shown in the appendix. Figure 4As shown, BiLSTM is used to extract features from the dataset. The extracted features are then transformed to form a feature vector that XGBoost can recognize. XGBoost is then used for classification to obtain classification results for different wear states of the milling machine tool. The main steps include:

[0093] S4.1 For the constructed dataset, a bidirectional long short-term recurrent neural network (BiLSTM) is used to extract features from the dataset. The final output calculation formula of BiLSTM is shown below:

[0094] h i =f1(w1x i +w2h i-1 );

[0095] H i =f2(w3x i +w5H i+1 );

[0096] Y i =f3(w4h i +w6H i );

[0097] Where, x i (i = 1, 2, ..., t) corresponds to the time input data, h i (i = 1, 2, ..., t) represents the hidden state of the LSTM during the forward iteration, H i (i = 1, 2, ..., t) represents the hidden state of the LSTM during the backward iteration, Y i (i = 1, 2, ..., t) represents the corresponding output data, w i (i = 1, 2, ..., 6) represents the weights of each layer, where f1, f2, f3 correspond to the activation functions of different layers.

[0098] S4.2 reconstructs the feature vectors extracted by BiLSTM and then enters the Extreme Gradient Boosting Decision Tree (XGBoost) for tool wear state classification.

[0099] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the present invention without departing from its novel spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals, characterized in that, The main steps include the following: S1: Four noise sensors are used to collect noise signals from four directions, and the correlation degree of signals from different directions is obtained by using the gray B-type correlation method. Data-level fusion is then performed to obtain the fused sound signal. S1.1 The noise sensors are evenly arranged around the spindle of the gear milling machine spindle box, and the collected data is arranged in time series; S1.2 Using one point as a reference sequence and the other three points as comparison sequences, the correlation between the points in the reference sequence and the other points is obtained; S1.3 After obtaining the gray B-type correlation degree of each point relative to other points, the correlation degree energy of the measurement point can be calculated, and then the sound fusion signal can be obtained by using the relationship between correlation degree and weight. S1.4 Since the correlation degree is directly proportional to the weight, the sum of the weight values ​​is 1; S2: An accelerometer is used to collect vibration signals of the gear milling machine spindle. Then, time-frequency domain features are extracted from the fused sound signal and vibration signal. The extracted features are optimized using the ReliefF-mRMR joint algorithm, and the feature vectors with the greatest influence are selected to construct the dataset. S2.1 uses an accelerometer placed at the spindle position. After the tool goes through a full life cutting process, the vibration signal of each cut is collected. The vibration signals of the three wear states are marked as 1, 2, and 3 by the tool wear curve. S2.2 extracts the time-frequency domain features of the fused sound and vibration signals, including time-domain features: mean, root mean square, standard deviation, waveform factor, skewness, kurtosis, peak value, peak factor, and impulse factor. Frequency domain characteristics: power spectrum mean, frequency centroid, mean square frequency, and time-frequency domain characteristics, namely: wavelet packet energy; S2.3 First, the time-frequency features are filtered using ReliefF, and a sample R is randomly selected from the sample set S. i Search and R i k nearest neighbors H of the same class of samples j and the k nearest neighbors M of different class sample sets j； S2.4 After selecting the 15 feature subsets, the feature subsets are filtered again using the mRMR algorithm; S3: Use BiLSTM to extract features from the dataset, transform the extracted features to form a feature vector that XGBoost can recognize, and then use XGBoost for classification to obtain the classification results of different wear states of the milling machine tool.

2. The method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals according to claim 1, characterized in that: In step S1, four noise sensors are used to collect noise signals from four directions. The correlation between the signals from different directions is obtained using the gray B-type correlation method. Data-level fusion is then performed to obtain the fused sound signal. The main steps include: Using the grey B-type correlation method, the discrete signal is defined as y i (k),y j (k), k=1,2,3...,n, First calculate the displacement difference: ； Calculate the speed difference again: ； Next, calculate the acceleration difference: The formula for the grey type B correlation degree is: ； Where X i As a reference sequence, X j For comparison sequences; Assume the grey B-type correlation energy between the sound signal at each point and the signals at the other three points is E. ij Then the correlation signal energy of the i-th point is E i : ； ； The specific relationships are as follows: ； P1, P2, P3, and P4 are the weights of the four sound test points. After obtaining the weights of each point, weighted fusion is performed to obtain the sound fusion signal Nf at the data level. 。 3. The method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals as described in claim 1, characterized in that: In step S2, an accelerometer is used to collect vibration signals from the spindle of the gear milling machine. Then, time-frequency domain features are extracted from the fused sound and vibration signals. The extracted features are optimized using the ReliefF-mRMR joint algorithm. The dataset is constructed by selecting the feature vectors with the greatest influence. The main steps are as follows: Update the weight W of each feature according to the following formula: ； In the formula, m is the number of iterations; diff(A,X,Y) represents the distance between two objects on feature A; k is the number of nearest neighbor samples; P(c) represents the prior probability of target class c; Based on W(A), useless feature sets are eliminated, and 15 feature subsets are selected from the 26 feature sets; If x and y are two continuous random variables, P(x) and P(y) are the probabilities of x and y respectively, and the joint probability is P(x,y), then the formula for the mutual information (MI) between them is as follows: ； The formula for the mRMR principle is expressed as follows: ； ； In the formula, C represents the target category; S represents the feature set; is the number of features; x i x j Represent the i-th and j-th features; The mRMR principle is denoted as max By using the Mutual Information Qualifier (MIQ) standard The forward sequence search method is used to find the approximate optimal fault features. Assume that X is the original feature set, and there are already m-1 feature subsets S. m-1， In the remaining features The selection of the m-th feature must satisfy the following: ； Based on the results obtained from the mRMR principle, each feature is sorted from high to low, and the optimal feature classification subset is selected to construct the dataset.

4. The method for classifying the wear state of disc milling cutters for gear milling machines based on two-level fusion of acoustic and vibration signals according to claim 1, characterized in that: In step S3, BiLSTM is used to extract features from the dataset. The extracted features are then transformed to form a feature vector that XGBoost can recognize. XGBoost is then used for classification to obtain classification results for different wear states of the milling machine tool. The main steps include: S3.1 For the constructed dataset, a bidirectional long short-term recurrent neural network (BiLSTM) is used to extract features from the dataset. The final output calculation formula of BiLSTM is shown below: ； ； ； Where, x i i=1,2,…,t, corresponding to the time input data, h i Let i = 1, 2, ..., t, represent the hidden states of the LSTM during the forward iteration, and H... i， i = 1,2,…,t, represents the hidden state of the LSTM in the backward iteration, Y i， i=1,2,…,t, representing the corresponding output data, w i， i=1,2,…,6 represents the weight of each layer, where f1,f2,f3 correspond to the activation functions of different layers; S3.2 reconstructs the feature vectors extracted by BiLSTM and then enters the Extreme Gradient Boosting Decision Tree (XGBoost) for tool wear state classification.

Images