CN118287728A - Milling morphology state discrimination method based on acoustic signal MFCC characterization - Google Patents

Abstract

The invention discloses a milling morphology state judging method based on acoustic signal MFCC characterization, which comprises the following steps: detecting roughness and waviness of milling surfaces with different process parameters, correlating the data of the roughness and the waviness with the process parameters to obtain a data set, and collecting sound signals generated in the corresponding processing process; classifying the stability of the processing working condition into a stable processing state and an unstable processing state by using a data set of roughness and waviness of the processing surface morphology characterization and using a mean value clustering analysis method; the associated machining condition stability state classifies the acoustic signal into a stable milling signal and an unstable milling signal. The invention has the characteristics of easy installation and industrial application potential, can intelligently classify the characteristics of the acoustic signal MFCC, can replace a contact vibration sensor, realizes the discrimination of the processing morphology, and then improves the quality and the efficiency of the milling surface by adjusting the milling process parameters.

Description

A method for distinguishing milling machining morphology based on MFCC characterization of sound signals

Technical Field

The present invention relates to intelligent processing technology, and in particular to a method for distinguishing milling processing morphology state based on MFCC characterization of sound signals.

Background technique

Two types of vibrations are easily generated during machine tool processing: forced vibration and self-excited vibration.

Forced vibrations can be controlled by isolating the forcing element and are easier to resolve than self-excited vibrations.

Self-excited vibration is also called chatter, which is caused by the interaction between the dynamic characteristics of the cutting process and the modal characteristics of the machine tool-tool-workpiece system. Among them, the regenerative chatter of the cutting system caused by the dynamic cutting force generated by the cutting thickness variation effect is the most harmful. This machining state is called unstable machining, which will have an adverse effect on the machining accuracy of the workpiece, the life of the tool and spindle, and the machining efficiency.

In traditional machining processes, in order to avoid the occurrence of unstable machining states, conservative cutting parameters are usually selected offline, which seriously limits machining efficiency and causes waste in production. Therefore, automatic detection of machining states and ensuring the stability of the cutting process are crucial to the machining accuracy and efficiency of the workpiece. Therefore, in order to ensure machining accuracy and efficiency, it is particularly important to realize online recognition of milling machining states. At present, most online machining state recognition methods use contact sensors such as force sensors and vibration sensors. On the one hand, they need to contact with the workpiece, which seriously affects the actual machining; on the other hand, such sensors are expensive and difficult to be used on a large scale in production sites.

In order to solve the problem of difficult installation of contact sensors and the impact on actual processing, the existing technology uses a dataset of roughness and waviness that characterizes the machining surface morphology to digitally classify the machining morphology state, and uses the mean clustering analysis method to classify the machining condition stability into stable machining state and unstable machining state. The sound sensor is used to collect the sound signal of the corresponding process parameters, and the sound signal is associated with the machining condition stability state to classify the sound signal into stable milling signal and unstable milling signal. The MFCC features of the sound signal are extracted, and the signal is trained using the BP neural network model to realize the discrimination of the milling surface state. This method uses a non-contact sensor, which can replace a contact vibration sensor and has the potential for large-scale application. However, the milling state recognition of this method is only applied to plane machining and cannot be applied to the monitoring needs of complex curved surface machining.

Summary of the invention

The purpose of the present invention is to overcome the shortcomings and deficiencies of the above-mentioned prior art and provide a method for distinguishing the milling machining morphology state based on the MFCC characterization of sound signals. The present invention solves the technical problem that the prior art cannot be applied to the monitoring needs of complex surface machining by using sound signal features instead of machining surface morphology features.

The present invention is achieved through the following technical solutions:

A method for distinguishing milling machining morphology state based on MFCC characterization of sound signals comprises the following steps:

Step 1: Detect the roughness and waviness of the milling surface with different process parameters, collect the sound signals generated in the corresponding processing process, and associate the roughness and waviness data with the process parameters to obtain a data set;

Step 2: Use the dataset of roughness and waviness of the machined surface to digitally classify the machining morphology state, and use the mean cluster analysis method to classify the machining condition stability into stable machining state and unstable machining state;

Step three, classify the sound signal into stable milling signal and unstable milling signal by associating with the stability state of the processing condition; perform pre-emphasis, frame division and windowing preprocessing steps on the sound signal, perform Fourier transform on the windowed signal, and obtain the energy spectrum density; multiply the energy spectrum density matrix with the Mel filter matrix to obtain a new matrix H; perform discrete cosine transform on the new matrix H to obtain MFCC basic parameters, further obtain the first-order difference coefficient and the second-order difference coefficient, and combine these three parameters to obtain the complete MFCC feature coefficient; combine the MFCC feature coefficient of each sound signal with the stability state of the processing condition to obtain a feature matrix, and obtain a data set of MFCC feature coefficients of the process parameters associated with the processing condition state;

Step 4: Select part of the data set as the training set and the other data set as the test set, input the training set into the BP neural network for training, and use the trained BP neural network to judge the machining surface state of the test set.

The above step one specifically includes the following steps: by testing the roughness and waviness of the milling surface with different process parameters, placing the sound sensor 50 mm away from the workpiece, collecting the sound signal generated in the corresponding processing process, and associating the roughness and waviness data with the process parameters to obtain a data set.

The above step 2 specifically includes the following steps: using the data set of roughness and waviness characterized by the machined surface morphology, digitally classifying the machining morphology state. Using the K_means clustering analysis method, classifying the machining condition stability into stable machining state and unstable machining state;

The roughness and waviness data of each milling experiment are x ⁽ⁱ⁾ = (Ra _i , Wa _i ), i = 1, 2...8; this set of data is clustered and classified by K_means; because the processing state needs to be divided into stable and unstable categories, K = 2; the specific steps are as follows:

S1: Randomly select two cluster centroids as μ ₁ ,μ ₂ ∈R ² ;

S2: For each set of experiments x ^(i), calculate the class c ⁽ⁱ⁾ to which it belongs;

c ^{( i )} = argmin _j || x ^{( i )} − μ _i || ² ;

S3: For each class c ⁽ⁱ⁾ , recalculate the centroid point μ of the class;

Repeat steps S2 and S3 until the algorithm converges; it can be divided into a stable milling state and an unstable milling state.

The above step three specifically includes the following steps:

The sound signals are classified into stable milling signals and unstable milling signals by correlating the stability status of the machining conditions;

Pre-emphasize the original sound signal to enhance the high-frequency information of the signal and obtain a pre-emphasized signal;

Where, α is the pre-emphasis coefficient, which is 0.97; S(n) is the original signal; is the pre-emphasized signal;

Divide the pre-emphasized signal into frames, select the frame length and frame shift, and divide a complete sound signal into several frames; substitute these several frame signals into the Hamming window function to eliminate the discontinuity at both ends of each frame. This process is called windowing.

S(n)=s(n)*h(n);

Among them, s(n) is the framed signal; h(n) is the Hamming window function;

For the windowed signal, each frame of the signal is subjected to a fast Fourier transform, and the data points of each frame are modulo and then squared to obtain the energy spectral density function;

The energy spectrum is obtained by squaring X(k), and then multiplied by the Mel filter _Hp (k) to calculate the logarithmic energy of the filter bank. The logarithmic energy C(p) of the p-th filter bank is:

The calculated logarithmic energy can be transformed into Mel cepstral coefficients through discrete cosine transformation:

Where M represents the dimension of MFCC features;

The obtained Mel-frequency cepstral coefficients are the basic parameters of MFCC. The first-order difference coefficients and second-order difference coefficients are further calculated, and these three parameters are combined to obtain the complete MFCC coefficients. The MFCC characteristic coefficients of each sound signal are combined with the stability state of the processing condition to obtain the characteristic matrix, and a data set of MFCC characteristic coefficients of different process parameters associated with the processing condition is obtained.

The above step 4 specifically includes the following steps:

BP neural network creation: BP neural network hyperparameter settings, number of hidden layer nodes; maximum number of iterations; error threshold; learning rate;

Network training and testing: Network training is a process of continuously correcting weights and thresholds, and through training, the output error of the network is made smaller and smaller;

Select training signals and perform state discrimination after training: Randomly select several feature matrices obtained in step 3 as training sets and pass them into the BP neural network for training; the remaining data is used as a test set and passed into the trained BP neural network model to realize the discrimination of the machining surface state.

The training data of the present invention is not less than 100 groups, and the dimension of the eigenvalue matrix is not less than 30 dimensions.

Compared with the prior art, the present invention has the following advantages and effects:

The present invention detects the roughness and waviness of the milling surface with different process parameters, collects the sound signals generated in the corresponding processing process, and associates the roughness and waviness data with the process parameters to obtain a data set.

The present invention uses a data set of roughness and waviness that characterizes the machining surface morphology to digitally classify the machining morphology state, and uses a mean cluster analysis method to classify the machining condition stability into a stable machining state and an unstable machining state.

The present invention classifies the sound signal into stable milling signal and unstable milling signal in association with the stability state of the machining condition. All the sound signals are preprocessed, and then the MFCC feature coefficients are extracted, and the MFCC feature coefficients are combined with the stability state of the machining condition to obtain a data set.

The present invention divides the data set into a training set and a prediction set, passes the training set into the BP neural network training model, and uses the trained neural network model to discriminate the milling surface morphology state of the test set. The present invention has the characteristics of easy installation and industrial application potential, can replace the contact vibration sensor, and improve the milling surface quality and efficiency by changing the processing parameters.

The present invention is easy to install and has the potential for industrial application. It can intelligently classify the MFCC features of the sound signal, replace the contact vibration sensor, realize the processing morphology discrimination, and then improve the milling surface quality and efficiency by adjusting the milling process parameters.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a schematic diagram of a processing system of the method of the present invention; wherein 1 represents a spindle box, 2 represents a milling cutter, 3 represents a workpiece, 4 represents a sound sensor, 5 represents a data acquisition card, and 6 represents a computer.

Figure 2 is a schematic diagram of the processing.

FIG3 is a graph showing the surface profile under unstable machining conditions.

Fig. 3(a) Surface roughness curve.

Fig. 3(b) Surface waviness curve.

FIG. 4 is a graph showing the surface profile under stable machining conditions.

Fig. 4(a) Surface roughness curve.

Fig. 4(b) Surface waviness curve.

Figure 5 is a surface waviness pre-roughness cluster classification diagram.

FIG6 is a time domain diagram of a sound signal.

Figure 6(a) Time domain diagram of stable sound signal.

Fig. 6(b) Time domain diagram of unstable sound signal.

Figure 7 is a three-dimensional graph of MFCC features.

Figure 8 is a graph showing the prediction accuracy of the BP neural network.

Detailed ways

The present invention will be described in further detail below in conjunction with examples and accompanying drawings, but the embodiments of the present invention are not limited thereto. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

The present invention discloses a method for distinguishing the milling machining morphology state based on the MFCC characterization of sound signals, which can be achieved through the following steps:

Step 1: The sound sensor installed on the upper part of the CNC milling machine is 100 mm away from the workpiece to collect the sound signal during the milling process. The sampling frequency is 4096 Hz. See Figure 1.

Design multiple sets of plane milling experiments, with a fixed feed rate of 600mm/mim, a radial cutting depth of 5mm, a workpiece material of 45 steel, a workpiece length of 100mm, a step-shaped workpiece, each step height of 0.3mm, length of 20mm, a total processing time of 10s, and change of spindle speed and cutting depth. As shown in Figure 2.

There are 40 sets of processing parameter combinations in total. For each processing parameter combination, 4 sound signals with a duration of 0.4 seconds are selected, and there are 160 sets of sound signals in total.

Laser confocal microscopy was used to characterize the three-dimensional profile of the microgrooves machined on the surface, and the roughness and waviness of the workpiece surface of each group were tested.

The atomic force microscope image processing software Gwyddion was used to realize data visualization, and the average roughness and waviness values of 1 mm in length were taken as the roughness and waviness data of the workpiece, as shown in Figures 2 and 3.

Step 2:

The roughness and waviness data sets that characterize the machined surface morphology are used to digitally classify the machining morphology states. The K_means clustering analysis method is used to classify the machining condition stability into stable machining state and unstable machining state.

The roughness and waviness data of each milling experiment are x ⁽ⁱ⁾ = (Ra _i , Wa _i ), i = 1, 2, ... 8. This set of data is clustered and classified by K_means. Since the processing state needs to be divided into stable and unstable categories, K = 2. The specific steps are as follows:

Step S1: Randomly select two cluster centroids as μ ₁ ,μ ₂ ∈R ² ;

Step S2: For each group of experiments x ^(i), calculate the class c ⁽ⁱ⁾ to which it belongs;

c ^{( i )} = argmin _j || x ^{( i )} − μ _i || ² ;

Step S3: For each class c ⁽ⁱ⁾ , recalculate the centroid point μ of the class;

Repeat steps S1 and S3 until the algorithm converges. The surface of the workpiece is divided into a stable machining surface and an unstable machining surface, as shown in FIG5 .

Step 3:

The sound signals are classified into stable milling signals and unstable milling signals by associating the stability state of the machining condition. The 160 groups of sound signals are divided into stable machining state signals and unstable machining state signals. The stable sound signals and the unstable sound signals are shown in FIG6 .

The obtained sound signal length is 0.4s, the sampling frequency is 4096Hz, and there are 1638 sampling points in total.

The original sound signal is pre-emphasized to strengthen the high-frequency information of the signal and obtain a pre-emphasized signal.

Where, α is the pre-emphasis coefficient, which is 0.97; S(n) is the original signal; is the pre-emphasized signal.

The pre-emphasized signal is divided into frames, and the frame length is selected as 1103 and the frame shift is 441. A complete sound signal is divided into 4 frames, and the length of the frame signal is 1103 points. These 4 frames of signal are substituted into the Hamming window function to eliminate the discontinuity at both ends of each frame. This process is called windowing.

S(n)=s(n)*h(n);

Among them, s(n) is the framed signal; h(n) is the Hamming window function

For the windowed signal, each frame of the signal is subjected to a fast Fourier transform, and the data points of each frame are modulo and then squared to obtain the energy spectral density function.

The energy spectrum is obtained by squaring X(k), multiplying it with the Mel filter _Hp (k), and calculating the logarithmic energy of the filter bank. The logarithmic energy C(p) of the p-th filter bank is

The calculated logarithmic energy can be transformed into Mel cepstral coefficients through discrete cosine transformation:

Where M represents the dimension of MFCC features.

The obtained Mel cepstral coefficient is the basic parameter of MFCC. The first-order difference coefficient and the second-order difference coefficient are further obtained. The three parameters are combined to obtain the complete MFCC coefficient. The MFCC coefficient is associated with the cluster classification result to obtain the feature matrix. The MFCC feature dimension has 39 dimensions and a length of 4. The average value of the 4 data in each dimension is taken as the MFCC feature coefficient of this processed signal. The result of the change of MFCC feature coefficient over time is shown in Figure 7.

If the clustering result is an unstable state, it is set to 1; if the clustering result is a stable state, it is set to 2. The MFCC characteristic coefficients of all parameter combinations and the stability results of the processing conditions are combined to obtain the data set R. The data set of the MFCC characteristic coefficients of the process parameters associated with the processing conditions is obtained.

Step 4:

BP neural network creation:

1. Network creation: BP neural network hyperparameter settings, the number of hidden layer nodes is 5, the maximum number of iterations is 1000, the error threshold is 10 ^-6 , and the learning rate is 0.01

2. Network training and testing:

Network training is a process of continuously correcting weights and thresholds, and through training, the output error of the network is made smaller and smaller.

2. Select training signals and perform state discrimination after training

The data set R obtained in step 3 has a total of 160 data. 110 data are randomly selected as the training set and passed into the BP neural network based on genetic algorithm for training.

The remaining 50 data are used as test sets and fed into the BP neural network classification model to distinguish the state of the machined surface. The training and test results are shown in Figure 8.

The above description is only a preferred embodiment of the present invention and does not limit the technical scope of the present invention. Therefore, any slight modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of the technical solution of the present invention.

Claims (6)

1. A method for distinguishing milling machining morphology based on MFCC characterization of sound signals, characterized in that it comprises the following steps: Step 1: Detect the roughness and waviness of the milling surface with different process parameters, collect the sound signals generated in the corresponding processing process, and associate the roughness and waviness data with the process parameters to obtain a data set; Step 2: Use the dataset of roughness and waviness of the machined surface to digitally classify the machining morphology state, and use the mean cluster analysis method to classify the machining condition stability into stable machining state and unstable machining state; Step three, classify the sound signal into stable milling signal and unstable milling signal by associating with the stability state of the processing condition; perform pre-emphasis, frame division and windowing preprocessing steps on the sound signal, perform Fourier transform on the windowed signal, and obtain the energy spectrum density; multiply the energy spectrum density matrix with the Mel filter matrix to obtain a new matrix H; perform discrete cosine transform on the new matrix H to obtain MFCC basic parameters, further obtain the first-order difference coefficient and the second-order difference coefficient, and combine these three parameters to obtain the complete MFCC feature coefficient; combine the MFCC feature coefficient of each sound signal with the stability state of the processing condition to obtain a feature matrix, and obtain a data set of MFCC feature coefficients of the process parameters associated with the processing condition state; Step 4: Select part of the data set as the training set and the other data set as the test set, input the training set into the BP neural network for training, and use the trained BP neural network to judge the machining surface state of the test set. 2. According to claim 1, the method for distinguishing the milling morphology state based on the MFCC characterization of the sound signal is characterized in that step one specifically includes the following steps: by detecting the roughness and waviness of the milling surface with different process parameters, placing the sound sensor at a distance of 50 mm from the workpiece, collecting the sound signal generated in the corresponding processing process, and associating the roughness and waviness data with the process parameters to obtain a data set. 3. According to the method for distinguishing the milling machining morphology state based on the MFCC characterization of the sound signal in claim 1, it is characterized in that step 2 specifically comprises the following steps: using the data set of roughness and waviness characterized by the machining surface morphology to digitally classify the machining morphology state; using the K_means clustering analysis method to classify the machining condition stability into a stable machining state and an unstable machining state; The roughness and waviness data of each milling experiment are x ⁽ⁱ⁾ = (Ra _i , Wa _i ), i = 1, 2...8; this set of data is clustered and classified by K_means; because the processing state needs to be divided into stable and unstable categories, K = 2; the specific steps are as follows: S1: Randomly select two cluster centroids as μ ₁ ,μ ₂ ∈R ² ; S2: For each set of experiments x ^(i), calculate the class c ⁽ⁱ⁾ to which it belongs; c ^{( i )} = argmin _j || x ^{( i )} − μ _i || ² ; S3: For each class c ⁽ⁱ⁾ , recalculate the centroid point μ of the class; Repeat steps S2 and S3 until the algorithm converges; it can be divided into a stable milling state and an unstable milling state. 4. According to the method for distinguishing the milling machining morphology state based on the MFCC characterization of the sound signal in claim 1, it is characterized in that the step three specifically comprises the following steps: The sound signals are classified into stable milling signals and unstable milling signals by correlating the stability status of the machining conditions; Pre-emphasize the original sound signal to enhance the high-frequency information of the signal and obtain a pre-emphasized signal; Where, α is the pre-emphasis coefficient, which is 0.97; S(n) is the original signal; is the pre-emphasized signal; Divide the pre-emphasized signal into frames, select the frame length and frame shift, and divide a complete sound signal into several frames; substitute these several frame signals into the Hamming window function to eliminate the discontinuity at both ends of each frame. This process is called windowing. S(n)=s(n)*h(n); Among them, s(n) is the framed signal; h(n) is the Hamming window function; For the windowed signal, each frame of the signal is subjected to a fast Fourier transform, and the data points of each frame are modulo and then squared to obtain the energy spectral density function; The energy spectrum is obtained by squaring X(k), and then multiplied by the Mel filter _Hp (k) to calculate the logarithmic energy of the filter bank. The logarithmic energy C(p) of the p-th filter bank is: The calculated logarithmic energy can be transformed into Mel cepstral coefficients through discrete cosine transformation: Where M represents the dimension of MFCC features; The obtained Mel-frequency cepstral coefficients are the basic parameters of MFCC. The first-order difference coefficients and second-order difference coefficients are further calculated, and these three parameters are combined to obtain the complete MFCC coefficients. The MFCC characteristic coefficients of each sound signal are combined with the stability state of the processing condition to obtain the characteristic matrix, and a data set of MFCC characteristic coefficients of different process parameters associated with the processing condition is obtained. 5. According to the method for distinguishing the milling machining morphology state based on the MFCC characterization of the sound signal in claim 1, it is characterized in that the step 4 specifically comprises the following steps: BP neural network creation: BP neural network hyperparameter settings, number of hidden layer nodes; maximum number of iterations; error threshold; learning rate; Network training and testing: Network training is a process of continuously correcting weights and thresholds, and through training, the output error of the network is made smaller and smaller; Select training signals and perform state discrimination after training: Randomly select several feature matrices obtained in step 3 as training sets and pass them into the BP neural network for training; the remaining data is used as a test set and passed into the trained BP neural network model to realize the discrimination of the machining surface state. 6. According to the method for distinguishing the milling machining morphology state based on MFCC characterization of sound signals in claim 5, it is characterized in that the training data is not less than 100 groups and the dimension of the eigenvalue matrix is not less than 30 dimensions.