A method for determining the morphology of milling operations based on MFCC feature analysis of acoustic signals

Abstract

This invention discloses a method for determining the morphological state of milling operations based on the MFCC (Multi-Frequency Compression Criteria) feature of sound signals. The method includes the following steps: detecting the roughness and waviness of milled surfaces with different process parameters; correlating the roughness and waviness data with the process parameters to obtain a dataset; collecting the sound signals generated during the corresponding machining process; using the roughness and waviness dataset characterized by the surface morphology feature, classifying the machining condition stability into stable and unstable machining states using mean-based clustering analysis; and correlating the machining condition stability states to classify the sound signals into stable and unstable milling signals. This invention is easy to install and has potential for industrial applications. It intelligently classifies the MFCC features of sound signals, replacing contact vibration sensors to determine the machining morphology. Then, by adjusting the milling process parameters, it simultaneously improves the surface quality and efficiency of milling operations.

Description

Technical Field

[0001] This invention relates to intelligent machining technology, and more particularly to a method for determining the morphological state of milling operations based on MFCC characterization of sound signals. Background Technology

[0002] Two types of vibrations are easily generated during machine tool processing: forced vibration and self-excited vibration.

[0003] Forced vibration can be controlled by isolating the forced vibration element, which is easier to solve than self-excited vibration.

[0004] Self-excited vibration, also known as chatter, 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 these, regenerative chatter induced by the dynamic cutting force generated by the effect of varying cutting thickness is the most harmful. This machining state is called unstable machining, which adversely affects the machining accuracy of the workpiece, tool and spindle life, and machining efficiency.

[0005] In traditional machining processes, conservative cutting parameters are typically selected offline to avoid unstable machining conditions. This severely limits machining efficiency and leads to production waste. Therefore, automatically detecting the machining condition and ensuring stable cutting processes are crucial for workpiece machining accuracy and efficiency. Thus, online identification of milling machining conditions is particularly important to guarantee machining accuracy and efficiency. Currently, most online machining condition identification methods use contact sensors such as force sensors and vibration sensors. On the one hand, these sensors require contact with the workpiece, significantly impacting actual machining; on the other hand, these sensors are expensive, making large-scale application in production environments difficult.

[0006] To address the difficulties in installing contact sensors and their impact on actual machining, existing technologies utilize datasets of surface roughness and waviness characteristics to digitally classify machining conditions. Mean value clustering analysis is then used to categorize machining condition stability into stable and unstable states. Sound sensors are used to collect sound signals corresponding to the process parameters. These signals are then correlated with the machining condition stability to classify them into stable and unstable milling signals. MFCC features are extracted from the sound signals, and a backpropagation neural network model is used to train the signals, enabling the discrimination of milled surface conditions. This method uses non-contact sensors, potentially replacing contact vibration sensors and possessing large-scale application potential. However, its milling condition identification is only applicable to planar machining and cannot meet the monitoring needs of complex curved surface machining. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a method for determining the morphological state of milling operations based on MFCC (Multi-Functional Calibration) characterization of acoustic signals. This invention solves the technical problem that existing technologies are unsuitable for monitoring complex curved surface machining by using acoustic signal features to replace surface morphological features.

[0008] This invention is achieved through the following technical solution:

[0009] A method for determining the topographic state of milling operations based on MFCC feature mapping of acoustic signals includes the following steps:

[0010] Step 1: By detecting the roughness and waviness of milled surfaces with different process parameters, the sound signals generated during the corresponding processing are collected, and the roughness and waviness data are correlated with the process parameters to obtain a dataset.

[0011] Step 2: Using the roughness and waviness dataset characterized by the surface morphology of the machined surface, the machining morphology state is digitally classified, and the mean clustering analysis method is used to classify the stability of the machining condition into stable machining state and unstable machining state.

[0012] Step 3: Associating the stability state of the machining conditions, the sound signals are classified into stable milling signals and unstable milling signals. The sound signals undergo pre-emphasis, framing, and windowing preprocessing. The windowed signals are then subjected to Fourier transform to obtain the energy spectral density. The energy spectral density matrix is ​​multiplied by the Mel filter matrix to obtain a new matrix H. A discrete cosine transform is performed on the new matrix H to obtain the basic MFCC parameters. The first-order and second-order difference coefficients are further calculated, and these three parameters are combined to obtain the complete MFCC feature coefficients. The MFCC feature coefficients of each sound signal are combined with the stability state of the machining conditions to obtain a feature matrix, resulting in a dataset of MFCC feature coefficients relating process parameters to the machining conditions.

[0013] Step 4: Select a portion of the dataset as the training set and the other datasets 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 surface state of the test set.

[0014] The above step one specifically includes the following steps: by detecting the roughness and waviness of the milled surfaces with different process parameters, a sound sensor is placed 50mm away from the workpiece to collect the sound signals generated during the corresponding processing, and the roughness and waviness data are correlated with the process parameters to obtain a dataset.

[0015] Step two above specifically includes the following steps: using a dataset of roughness and waviness characteristics of the machined surface morphology, the machining morphology state is digitally classified. The K-means clustering analysis method is used to classify the stability of the machining condition into stable machining state and unstable machining state;

[0016] Roughness and waviness data for each milling experiment x (i) =(Ra i Wa i ), i = 1, 2...8; Perform K-means clustering on this data set; Since we need to classify the processing state into two categories, stable and unstable, K = 2; The specific steps are as follows:

[0017] S1: Randomly select two cluster centroids, μ1 and μ2 ∈ R. 2 ;

[0018] S2: For each group of experiments x (i) Calculate the class c to which it belongs. (i) ;

[0019] c (i) =argmin j ||x (i) -μ i || 2 ;

[0020] S3: For each class c (i) Recalculate the centroid μ of this class;

[0021]

[0022] Repeat steps S2 and S3 until the algorithm converges; it can be divided into stable milling states and unstable milling states.

[0023] Step three above specifically includes the following steps:

[0024] The stability status of the associated machining conditions classifies the sound signals into stable milling signals and unstable milling signals;

[0025] The original sound signal is pre-emphasized to enhance the high-frequency information of the signal, resulting in a pre-emphasized signal;

[0026]

[0027] In the formula, α is the pre-emphasis coefficient, which is taken as 0.97; S(n) is the original signal; This is a pre-emphasis signal;

[0028] The pre-emphasized signal is divided into frames, and the frame length and frame shift are selected to divide a complete audio signal into several frames. These frames are then substituted into the Hamming window function to eliminate the discontinuities at both ends of each frame. This process is called windowing.

[0029] S(n) = s(n) * h(n);

[0030] Where s(n) is the framed signal; h(n) is the Hamming window function;

[0031] For the windowed signal, a fast Fourier transform is performed on each frame of the signal. The modulus of each frame's data points is taken and then squared to obtain the energy spectral density function.

[0032]

[0033] The energy spectrum is obtained by squaring X(k) and comparing it with the Mel filter H. p Multiply by (k) to calculate the logarithmic energy of the filter bank. The logarithmic energy C(p) of the p-th filter bank is:

[0034]

[0035] The calculated logarithmic energy can be used to obtain the Mel-frequency cepstral coefficients through discrete cosine transform.

[0036]

[0037] In the formula, M represents the dimension of the MFCC feature;

[0038] The obtained Mel-frequency cepstral coefficients are the basic parameters of MFCC. The first-order difference coefficients and second-order difference coefficients are further obtained, 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 the dataset of MFCC characteristic coefficients of different process parameters related to the processing condition state is obtained.

[0039] Step four above specifically includes the following steps:

[0040] Creating a BP neural network: Setting hyperparameters for the BP neural network, including the number of hidden layer nodes, maximum number of iterations, error threshold, and learning rate.

[0041] Network training and testing: Network training is a process of continuously adjusting weights and thresholds to reduce the network's output error.

[0042] Select training signals and perform state discrimination after training: For the feature matrix obtained in step three, randomly select several feature matrices as the training set and feed them into the BP neural network for training; use the remaining data as the test set and feed them into the trained BP neural network model to realize the discrimination of the state of the processed surface.

[0043] The training data of this invention consists of no fewer than 100 sets, and the dimension of the feature value matrix is ​​no less than 30 dimensions.

[0044] Compared with the prior art, the present invention has the following advantages and effects:

[0045] This invention detects the roughness and waviness of milled surfaces with different process parameters, collects the sound signals generated during the corresponding processing, and correlates the roughness and waviness data with the process parameters to obtain a dataset.

[0046] This invention uses a dataset of roughness and waviness characterized by the surface morphology of the machined surface to digitally classify the machining morphology state, and uses mean clustering analysis to classify the stability of the machining condition into stable machining state and unstable machining state.

[0047] This invention classifies sound signals into stable and unstable milling signals based on the stability state of the machining process. All sound signals are preprocessed, and then MFCC feature coefficients are extracted. The MFCC feature coefficients are then combined with the machining process stability state to obtain a dataset.

[0048] This invention divides the dataset into a training set and a prediction set. The training set is fed into a backpropagation (BP) neural network to train a model, and the trained neural network model is used to judge the surface morphology of the milled test set. This invention is easy to install and has potential for industrial applications. It can replace contact vibration sensors and improve both the surface quality and efficiency of milling by changing the processing parameters.

[0049] This invention features easy installation and industrial application potential. It intelligently classifies the MFCC characteristics of sound signals, which can replace contact vibration sensors to identify the processing morphology. Then, by adjusting the milling process parameters, it can simultaneously improve the surface quality and efficiency of milling. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the processing system of the method of the present invention; where 1 represents the spindle box, 2 represents the milling cutter, 3 represents the workpiece, 4 represents the sound sensor, 5 represents the data acquisition card, and 6 represents the computer.

[0051] Figure 2 This is a schematic diagram of the processing.

[0052] Figure 3 This is a surface profile curve diagram under unstable processing conditions.

[0053] Figure 3 (a) Surface roughness curve.

[0054] Figure 3(b) Surface waviness curve.

[0055] Figure 4 This is a surface profile curve under stable processing conditions.

[0056] Figure 4 (a) Surface roughness curve.

[0057] Figure 4 (b) Surface waviness curve.

[0058] Figure 5 This is a clustering classification diagram for surface waviness and pre-roughness.

[0059] Figure 6 This is a time-domain diagram of the sound signal.

[0060] Figure 6 (a) Time domain diagram of stable audio signal.

[0061] Figure 6 (b) Time domain diagram of unstable sound signal.

[0062] Figure 7 This is a 3D plot of MFCC features.

[0063] Figure 8 The graph shows the prediction accuracy of the BP neural network. Detailed Implementation

[0064] The present invention will be further described in detail below with reference to embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0065] This invention discloses a method for determining the shape state of milling operations based on MFCC feature mapping of sound signals, which can be implemented through the following steps:

[0066] Step 1: Using a sound sensor mounted on the upper part of the CNC milling machine, at a distance of 100mm from the workpiece, collect the sound signal during the milling process. The sampling frequency is 4096Hz. Figure 1 As shown.

[0067] Design multiple sets of planar milling experiments, with a fixed feed rate of 600 mm / min, a radial depth of cut of 5 mm, a workpiece material of 45 steel, a workpiece length of 100 mm, and a stepped shape, with each step having a height of 0.3 mm and a length of 20 mm. The total machining time is 10 seconds. The spindle speed and depth of cut are varied. Figure 2 As shown.

[0068] There are a total of 40 processing parameter combinations. For each processing parameter combination, four audio signals with a duration of 0.4 seconds are selected, resulting in a total of 160 audio signals.

[0069]

[0070]

[0071] The three-dimensional contours of the microgrooves on the surface were characterized using laser confocal microscopy, and the surface roughness and waviness of each group of workpieces were detected.

[0072] Data visualization was achieved using the atomic force microscopy image processing software Gwyddion. The average roughness and waviness values ​​over a 1mm length were taken as the roughness and waviness data for the workpiece. For example... Figure 2 , Figure 3 As shown.

[0073] Step Two:

[0074] Using a dataset of roughness and waviness characterized by the surface morphology of the machined surfaces, the machining morphology status is digitally classified. The K-means clustering analysis method is used to classify the stability of the machining conditions into stable machining states and unstable machining states.

[0075] Roughness and waviness data for each milling experiment x (i) =(Ra i Wa i The data set consists of i = 1, 2, ..., 8. K-means clustering is performed on this data set. Since the processing states need to be divided into stable and unstable categories, K = 2. The specific steps are as follows:

[0076] Step S1: Randomly select two cluster centroids, μ1 and μ2 ∈ R. 2 ;

[0077] Step S2: For each group of experiments x (i) Calculate the class c to which it belongs. (i) ;

[0078] c (i) =argmin j ||x (i) -μ i || 2 ;

[0079] Step S3: For each class c (i) Recalculate the centroid μ of this class;

[0080]

[0081] Repeat steps S1 and S3 until the algorithm converges. Divide the workpiece surface into stable and unstable machined surfaces, such as... Figure 5 As shown.

[0082] Step 3:

[0083] The stability of the associated machining conditions categorizes sound signals into stable milling signals and unstable milling signals. These 160 sets of sound signals are further divided into stable machining state signals and unstable machining state signals. Stable and unstable sound signals are as follows: Figure 6 As shown.

[0084] The obtained sound signal has a length of 0.4s, a sampling frequency of 4096Hz, and a total of 1638 sampling points.

[0085] The original sound signal is pre-emphasized to enhance its high-frequency information, resulting in a pre-emphasized signal.

[0086]

[0087] In the formula, α is the pre-emphasis coefficient, which is taken as 0.97; S(n) is the original signal; This is a pre-emphasis signal.

[0088] The pre-emphasized signal is framed, with a frame length of 1103 and a frame shift of 441, dividing a complete audio signal into 4 frames, each frame consisting of 1103 points. These 4 frames are then substituted into a Hamming window function to eliminate discontinuities at the ends of each frame; this process is called windowing.

[0089] S(n) = s(n) * h(n);

[0090] Where s(n) is the framed signal; h(n) is the Hamming window function.

[0091] For the windowed signal, a fast Fourier transform is performed on each frame of the signal. The modulus of each frame's data points is taken and then squared to obtain the energy spectral density function.

[0092]

[0093] The energy spectrum is obtained by squaring X(k) and comparing it with the Mel filter H. p Multiply by (k) to calculate the logarithmic energy of the filter bank. The logarithmic energy C(p) of the p-th filter bank is:

[0094]

[0095] The calculated logarithmic energy can be used to obtain the Mel-frequency cepstral coefficients through discrete cosine transform.

[0096]

[0097] In the formula, M represents the dimension of the MFCC feature.

[0098] The obtained Mel-frequency cepstral coefficients are the basic parameters of MFCC. The first-order and second-order difference coefficients are then calculated, and these three parameters are combined to obtain the complete MFCC coefficients. The MFCC coefficients are correlated with the clustering results to obtain the feature matrix. The MFCC feature dimension has 39 dimensions and a length of 4. The average of the four data points in each dimension is taken as the MFCC feature coefficients for this processing signal. The results of the MFCC feature coefficients changing over time are as follows: Figure 7 As shown.

[0099] If the clustering result is unstable, set it to 1; if the clustering result is stable, set it to 2. Combine the MFCC characteristic coefficients of all parameter combinations with the processing condition stability results to obtain dataset R. This yields a dataset of MFCC characteristic coefficients relating process parameters to processing condition states.

[0100] Step Four:

[0101] BP neural network creation:

[0102] 1. Network creation: BP neural network hyperparameter settings: 5 hidden layer nodes, maximum iterations 1000, error threshold 10. -6 The learning rate is 0.01.

[0103] 2. Network training and testing:

[0104] Network training is a process of continuously adjusting weights and thresholds, so that the network's output error becomes smaller and smaller through training.

[0105] II. Select training signals and perform state discrimination after training.

[0106] The dataset R obtained in step 3 contains 160 data points. 110 data points are randomly selected as the training set and fed into a BP neural network based on a genetic algorithm for training.

[0107] The remaining 50 data points were used as a test set and fed into a BP neural network classification model to determine the condition of the processed surface. See details of the training and testing results. Figure 8 .

[0108] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for determining the morphological state of milling operations based on MFCC feature mapping of acoustic signals, characterized in that, Includes the following steps: Step 1: By detecting the roughness and waviness of milled surfaces with different process parameters, the sound signals generated during the corresponding processing are collected, and the roughness and waviness data are correlated with the process parameters to obtain a dataset. Step 2: Using the roughness and waviness dataset characterized by the surface morphology of the machined surface, the machining morphology state is digitally classified, and the mean clustering analysis method is used to classify the stability of the machining condition into stable machining state and unstable machining state. Step 3: Associating the stability of the machining conditions, the sound signal is classified into stable milling signals and unstable milling signals; the sound signal undergoes pre-emphasis, framing, and windowing preprocessing steps; the windowed signal is then subjected to Fourier transform to obtain the energy spectral density; the energy spectral density matrix is ​​multiplied by the Mel filter matrix to obtain a new matrix H; the new matrix H is subjected to discrete cosine transform to obtain the basic parameters of MFCC, and the first-order and second-order difference coefficients are further obtained; the basic parameters of MFCC, the first-order difference coefficients, and the second-order difference coefficients are combined to obtain the complete MFCC characteristic coefficients; The MFCC feature coefficients of each sound signal are combined with the stability state of the processing condition to obtain the feature matrix, and a dataset of MFCC feature coefficients of the process parameters related to the processing condition state is obtained. Step 4: Randomly select several feature matrices as the training set and the remaining data 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 surface state of the test set. Step four specifically includes the following steps: Creating a BP neural network: Setting hyperparameters for the BP neural network: number of hidden layer nodes; maximum number of iterations; error threshold; learning rate; Network training and testing: Network training is a process of continuously adjusting weights and thresholds to reduce the network's output error. Select training signals and perform state discrimination after training: For the feature matrix obtained in step three, randomly select several feature matrices as the training set and feed them into the BP neural network for training; use the remaining data as the test set and feed them into the trained BP neural network model to realize the discrimination of the state of the processed surface. The training set data shall contain no fewer than 100 sets, and the feature matrix shall have no fewer than 30 dimensions.

2. The milling machining topography discrimination method based on MFCC feature analysis of sound signals according to claim 1, characterized in that... Step one specifically includes the following steps: By detecting the roughness and waviness of milled surfaces with different process parameters, a sound sensor is placed 50mm away from the workpiece to collect the sound signals generated during the corresponding processing. The roughness and waviness data are then correlated with the process parameters to obtain a dataset.

Images