阅读说明：本技术 一种基于奇异性李氏指数的刀具状态监测方法 (Cutter state monitoring method based on singularity Leersian index ) 是由 周长安 周德龙 张开兴 国凯 孙杰 孙智霖 于 2021-10-19 设计创作，主要内容包括：本申请公开了一种基于奇异性李氏指数的刀具状态检测方法,采集切削加工过程中的切削刀具全寿命周期内的声音、切削力及振动传感器信号；对采集的信号进行降噪预处理；对降噪预处理后的信号进行奇异性特征提取；对原始信号时域、频域统计信息、奇异性李氏指数和小波包系数进行与刀具磨破损密切相关深层次特征的自适应提取和融合；建立刀具破损状态的分类识别模型以及突变破损模型,从而分别实现对刀具磨损量和破损状态的在线监测。采用上述方法可相对稳定且可靠地实现对刀具状态的实时监测,不仅适用于大批量零件生产时的刀具状态监测,同时对小批量甚至单个零件的生产加工同样具有高度的适应性,适用范围更广。(The application discloses a cutter state detection method based on a singularity Leersian index, which is used for collecting sound, cutting force and vibration sensor signals in the whole life cycle of a cutting cutter in the cutting process; carrying out noise reduction pretreatment on the acquired signals; extracting singularity characteristics of the signal subjected to noise reduction pretreatment; carrying out self-adaptive extraction and fusion of deep features closely related to tool wear loss on time domain, frequency domain statistical information, singular Leersian index and wavelet packet coefficient of an original signal; and establishing a classification recognition model and a sudden change damage model of the tool damage state, thereby respectively realizing the online monitoring of the tool wear loss and the damage state. By adopting the method, the real-time monitoring of the cutter state can be realized relatively stably and reliably, the method is not only suitable for monitoring the cutter state during the production of large-batch parts, but also has high adaptability to the production and processing of small-batch or even single parts, and has wider application range.)

1. A cutter state detection method based on a singular Leersian index is characterized by comprising the following steps:

collecting sound, cutting force and vibration sensor signals of a cutting tool in the whole life cycle of the cutting machining process;

carrying out noise reduction pretreatment on the acquired signals;

extracting singularity characteristics of the signal subjected to noise reduction pretreatment;

carrying out self-adaptive extraction and fusion of deep features closely related to tool wear loss on time domain, frequency domain statistical information, singular Leersian index and wavelet packet coefficient of an original signal;

and establishing a classification recognition model and a sudden change damage model of the tool damage state, thereby respectively realizing the online monitoring of the tool wear loss and the damage state.

2. The singularity Leersian index-based tool state detection method according to claim 1, wherein the denoising preprocessing of the collected signals comprises:

determining that the mode maximum point is generated by signals or noise according to the change rule of the mode maximum point of the signals along the scale s in the (u, s) (space u, scale s) plane;

if the mode maximum value point is generated by noise, setting a screening threshold value on the maximum scale;

screening a mode maximum value point caused by noise through the screening threshold;

and setting the modulus maximum point with the value of the modulus maximum point wavelet coefficient smaller than the screening threshold value to zero.

3. The singularity Leersian index-based tool state detection method according to claim 2, wherein the determining that the mode maximum point is signal or noise generation according to the change rule of the mode maximum point of the signal along the dimension s in the (u, s) (space u, dimension s) plane comprises:

if the wavelet coefficient value of the module maximum value point is reduced along with the reduction of the scale s and finally converged, the maximum value line communicated with the module maximum value point corresponds to a signal point containing singular information and generates a signal;

alternatively, the first and second electrodes may be,

if the value of the corresponding wavelet coefficient of the modulo maximum point increases with decreasing scale, the modulo maximum point is noise-producing.

4. The singularity Leersian index-based tool state detection method according to claim 2 or 3, wherein the screening threshold is:

wherein Z is a constant and the discrete scale s is 2^jJ is the maximum value of the discrete scale coefficient, and M is the maximum value of all the modulus maxima on the maximum scale.

5. The singularity Leersian index-based tool state detection method according to claim 4, wherein the singularity feature extraction of the noise-reduced preprocessed signal comprises:

determining global singularity of a signal function through Fourier transform;

analyzing local details of the signal by utilizing wavelet transformation through a telescopic translation operation, and calculating a singular Leersian index of a signal function at a certain point or a certain interval;

calculating the convergence condition of the modulus maximum value of the wavelet coefficient after decreasing with the scale s to evaluate whether a singular point exists and the size of the singular Leersian index;

defining a maximum line within the (u, s) plane, points on the maximum line all being modulo maximum points;

and calculating the modulus maximum value on the wavelet transformation binary scale, and realizing the communication of the maximum value line and the evaluation of the Lee's index due to the non-discontinuity of the maximum value line.

6. The singular leersian index-based tool state detection method according to any one of claims 1-5, wherein the self-adaptive extraction and fusion of deep features closely related to tool wear damage to the time domain, frequency domain statistical information, singular leersian index and wavelet packet coefficient of the original signal comprises:

a multi-dimensional stacked sparse automatic encoder model MD-SSAEs based on feature fusion;

inputting the extracted singular Leersian index, time domain characteristics, frequency domain characteristics and wavelet packet coefficients serving as initial characteristics into MD-SSAEs;

through training of the new model, multi-dimensional depth features are obtained, and feature fusion is performed by using one-dimensional SSAE.

7. The singularity Lee's index based tool state detection method of claim 6, wherein the training of the new model results in multi-dimensional depth features, and feature fusion using one-dimensional SSAE comprises:

the MD-SSAEs comprise four one-dimensional SSAEs, and in the training process model of the MD-SSAEs, the structures of the four SSAEs are formed by one input layer X_iA plurality of hidden layers X^j(j-2, 3, …, n-1) and output layer y-XⁿComposition is carried out; the input layer contains unprocessed TD dataFD data obtained by applying fast Fourier transformSingular Lee indexAnd wavelet packet coefficients obtained by wavelet packet decomposition

Each hidden layer has four vectorsFour eigenvectors y are obtained by calculating all automatic encoders and establishing a new modeling framework₁,y₂,y₃,y₄；

Training the depth feature vector y through an MD-SSAEs model₁,y₂,y₃,y₄New feature vector X is fused into one_M＝[y₁,y₂,y₃,y₄]Vector X_MWill be used as input for one-dimensional SSAE for the extraction of deeper features.

8. The singular leek index-based tool state detection method as claimed in claim 6 or 7, wherein said feature fusion using one-dimensional SSAE comprises:

inputting the multi-dimensional depth features into the one-dimensional SSAE, and calculating corresponding numerical values through a weight matrix;

the multi-dimensional depth features are then used as input functions for non-linear regression to obtain output values.

9. The singularity Lee's index-based tool state detection method according to claim 1, wherein a classification recognition model of a tool breakage state and a sudden change breakage model are established, so as to realize online monitoring of tool wear loss and breakage state respectively, and the method comprises the following steps:

constructing a nonlinear regression analysis model of the data sensitive characteristics and the tool slowly-varying abrasion loss based on a convolutional neural network deep learning algorithm;

then establishing a state recognition model fusing heterogeneous characteristics and cutter sudden change damage based on a support vector machine;

and then, testing the model by using the depth characteristics obtained by identifying the sample set data, and comprehensively judging the health state of the cutter according to the output results of the two models to determine whether the cutter needs to be replaced.

10. The singularity Lee's index-based tool state detection method according to claim 9, wherein establishing a state recognition model fusing heterogeneous features and tool sudden-change damage based on a support vector machine comprises: the method comprises the steps of establishing an identification model fusing heterogeneous characteristics and a tool sudden-change damage state based on a SoftMax model and a support vector machine, identifying the tool wear state of the model on the whole by the design hierarchical structure of the model, and dividing the whole service life cycle of the tool into three types of wear states according to the tool wear degree and the damage state, wherein the wear states are initial wear, normal wear and rapid wear respectively according to the wear amount of 0-0.03 mm, 0.03-0.12 mm and 0.12-0.3 mm.

Technical Field

The application relates to the technical field of wear detection of numerical control machine tools, in particular to a cutter state detection method based on a singularity Leersian index.

Background

With the continuous deepening of the concepts of industrial 4.0, intelligent manufacturing and the like, the requirements of various large enterprises on the intelligentization, automation and unmanned mechanical manufacturing are increased, while the machine tool cutter is used as an important execution part in the processing field, and the real-time state of the machine tool cutter is directly related to the fineness degree of a processed part, so that the online wear monitoring technology of the processing state is more and more concerned. However, the current tool wear detection mainly depends on manual measurement, the wear state and the need for replacement depend heavily on the experience of the detector, and therefore, the necessary technical support is lacked, which may result in too late tool change, thereby affecting the product quality, too early tool change, thereby resulting in waste of tools and economic reduction.

Although a machine tool cutter is a key execution end in a numerical control machining stage, the machine tool cutter is also a basic component which is most easily damaged and wasted, and particularly when high-value-added structural parts of various difficult-to-machine materials are machined, once the cutter is broken, broken or abraded to exceed the service life limit, the quality of parts is extremely easily reduced and even the parts are scrapped. According to data statistics, cutter damage is taken as a primary factor to bring high economic and time cost among all factors causing cutting process faults, in the total cost of part machining, the cutter and cutter changing cost can occupy 3% -12%, the downtime caused by cutter damage can occupy 7% -20% of the total downtime of a machine tool, and the influence on production efficiency can reach 25%. Therefore, the realization of the intelligent detection of the wear state of the cutting tool plays an important role in the aspects of improving the processing quality, saving the production cost, improving the production efficiency and the like.

In the traditional technology, the processing monitoring system which can be applied to commercialization generally has the obvious problems of high cost, complex installation, simpler signal processing, poor tool state identification precision, delay and the like. Therefore, the development demand for tool condition monitoring systems is becoming greater.

Disclosure of Invention

In order to solve the technical problems, the following technical scheme is provided:

in a first aspect, an embodiment of the present application provides a tool state detection method based on a singular leigh index, where the method includes: collecting sound, cutting force and vibration sensor signals of a cutting tool in the whole life cycle in the cutting process; carrying out noise reduction pretreatment on the acquired signals; extracting singularity characteristics of the signal subjected to noise reduction preprocessing; carrying out self-adaptive extraction and fusion of deep features closely related to tool wear loss on time domain, frequency domain statistical information, singular Leersian index and wavelet packet coefficient of an original signal; and establishing a classification recognition model and a sudden change damage model of the tool damage state, thereby respectively realizing the online monitoring of the tool wear loss and the damage state.

By adopting the implementation mode, noise is reduced, useful components in the signals are reserved to the maximum extent, then the singularity features of the signals are extracted, the deep-layer features of the signals are extracted and fused, finally, the classification identification model of the damage state of the cutter and the mutation damage model are used for identifying the wear state or wear amount of the cutter, the real-time monitoring of the cutter state is realized relatively stably and reliably, the method is suitable for monitoring the cutter state during the production of large-batch parts, and meanwhile, the method also has high adaptability to the production and processing of small-batch or even single parts, and the application range is wider.

With reference to the first aspect, in a first possible implementation manner of the first aspect, the performing noise reduction preprocessing on the acquired signal includes: determining that the mode maximum point is generated by signals or noise according to the change rule of the mode maximum point of the signals along the scale s in the (u, s) (space u, scale s) plane; if the mode maximum value point is generated by noise, setting a screening threshold value on the maximum scale; screening a mode maximum value point caused by noise through the screening threshold; and setting the modulus maximum point with the value of the modulus maximum point wavelet coefficient smaller than the screening threshold value to zero.

With reference to the first possible implementation manner of the first aspect, in a second possible implementation manner of the first aspect, the determining that the mode maximum point is signal or noise generation according to a variation rule of the mode maximum point of the signal along the dimension s in the (u, s) (space u, dimension s) plane includes: if the wavelet coefficient value of the modulus maximum point is reduced along with the reduction of the scale s and finally converged, the maximum line communicated with the modulus maximum point corresponds to a signal point containing singularity information and is generated for a signal; alternatively, if the value of the corresponding wavelet coefficient of the modulo maximum point increases with decreasing scale, the modulo maximum point is noise-producing.

With reference to the first or second possible implementation manner of the first aspect, in a third possible implementation manner of the first aspect, the screening threshold is:

wherein Z is a constant and the discrete scale s is 2^jJ is the maximum value of the discrete scale coefficient, and M is the maximum value of all the modulus maxima on the maximum scale.

With reference to the third possible implementation manner of the first aspect, in a fourth possible implementation manner of the first aspect, the performing singular feature extraction on the noise-reduced preprocessed signal includes: determining global singularity of a signal function through Fourier transform; analyzing local details of the signal by utilizing wavelet transformation through a telescopic translation operation, and calculating a singular Lee index of a signal function at a certain point or a certain interval; calculating the convergence condition of the modulus maximum value of the wavelet coefficient after decreasing with the scale s to evaluate whether an odd singular point exists and the size of the singular Leersian index exists; defining a maximum line within the (u, s) plane, points on the maximum line all being modulo maximum points; and calculating the modulus maximum value on the wavelet transformation binary scale, and realizing the communication of the maximum value line and the evaluation of the Lee's index due to the non-discontinuity of the maximum value line.

With reference to the first aspect or any one of the first to the fourth possible implementation manners of the first aspect, in a fifth possible implementation manner of the first aspect, the performing adaptive extraction and fusion of deep features closely related to tool wear-out loss on time-domain, frequency-domain statistical information, singular li's index, and wavelet packet coefficients of an original signal includes: a multi-dimensional stacked sparse automatic encoder model MD-SSAEs based on feature fusion; inputting the extracted singular Leersian index, time domain characteristics, frequency domain characteristics and wavelet packet coefficients serving as initial characteristics into MD-SSAEs; through training of the new model, multi-dimensional depth features are obtained, and feature fusion is performed by using one-dimensional SSAE.

With reference to the fifth possible implementation manner of the first aspect, in a sixth possible implementation manner of the first aspect, the obtaining a multidimensional depth feature by training the new model, and performing feature fusion by using one-dimensional SSAE includes:

the MD-SSAEs comprise four one-dimensional SSAEs, and in the training process model of the MD-SSAEs, the structure of the four SSAEs is formed by one input layer X_iA plurality of hidden layers X^j(j-2, 3, …, n-1) and output layer y-XⁿComposition is carried out; the input layer contains unprocessed TD dataFD data obtained by applying fast Fourier transformSingular Lee indexAnd wavelet packet coefficients obtained by wavelet packet decomposition

Each hidden layer has four vectorsFour eigenvectors y are obtained by calculating all automatic encoders and establishing a new modeling framework₁,y₂,y₃,y₄；

Training through MD-SSAEs model to characterize depthEigenvector y₁,y₂,y₃,y₄New feature vector X is fused into one_M＝[y₁,y₂,y₃,y₄]Vector X_MWill be used as input for one-dimensional SSAE for the extraction of deeper features.

With reference to the fifth or sixth possible implementation manner of the first aspect, in a seventh possible implementation manner of the first aspect, the performing feature fusion by using one-dimensional SSAE includes: inputting the multi-dimensional depth features into the one-dimensional SSAE, and calculating corresponding numerical values through a weight matrix; the multi-dimensional depth features are then used as input functions for non-linear regression to obtain output values.

With reference to the first aspect, in an eighth possible implementation manner of the first aspect, the establishing a classification recognition model of a tool breakage state and a sudden change breakage model, so as to respectively implement online monitoring on a tool wear amount and a breakage state, includes: constructing a nonlinear regression analysis model of the data sensitive characteristics and the tool slowly-varying abrasion loss based on a convolutional neural network deep learning algorithm; then establishing a state recognition model fusing heterogeneous characteristics and cutter sudden change damage based on a support vector machine; and then, testing the model by using the depth characteristics obtained by identifying the sample set data, and comprehensively judging the health state of the cutter according to the output results of the two models to determine whether the cutter needs to be replaced.

With reference to the eighth possible implementation manner of the first aspect, in a ninth possible implementation manner of the first aspect, the establishing a state identification model fusing heterogeneous features and tool sudden-change damage based on a support vector machine includes: the method comprises the steps of establishing an identification model fusing heterogeneous characteristics and a tool sudden-change damage state based on a SoftMax model and a support vector machine, identifying the tool wear state of the model on the whole by the design hierarchical structure of the model, and dividing the whole service life cycle of the tool into three types of wear states according to the tool wear degree and the damage state, wherein the wear states are initial wear, normal wear and rapid wear respectively according to the wear amount of 0-0.03 mm, 0.03-0.12 mm and 0.12-0.3 mm.

Drawings

Fig. 1 is a schematic flow chart of a tool state detection method based on a singular lees index according to an embodiment of the present disclosure;

fig. 2 is a schematic diagram illustrating a noise reduction effect of a vibration signal according to an embodiment of the present application;

FIG. 3 is a schematic diagram illustrating a noise reduction effect of a cutting force signal according to an embodiment of the present disclosure;

fig. 4 is a schematic diagram illustrating a noise reduction effect of an acoustic signal according to an embodiment of the present application;

FIG. 5 is a schematic diagram of a training process of the MD-SSAEs model provided by the embodiment of the present application;

fig. 6 is a schematic diagram of a learning process of a feature fusion structure according to an embodiment of the present application.

Detailed Description

The present invention will be described with reference to the accompanying drawings and embodiments.

Fig. 1 is a schematic flow diagram of a tool state detection method based on a singular ledebur index provided in an embodiment of the present application, and referring to fig. 1, the tool state detection method based on the singular ledebur index includes:

s101, collecting sound, cutting force and vibration sensor signals of a cutting tool in the whole life cycle in the cutting process.

Collecting indirect sensor signals of sound, cutting force, vibration and the like in the whole life cycle of the cutting tool in the cutting process, simultaneously recording the whole change process of the tool abrasion loss, and marking tool abrasion information for corresponding signals.

S102, noise reduction preprocessing is carried out on the acquired signals.

Milling is a typical intermittent cutting machining mode, in the machining process, a cutter rotates along with a spindle of a machine tool continuously, each cutter tooth of a milling cutter is cut in and cut out periodically, two or even more cutter teeth possibly participate in cutting at the same time, the complex dynamic cutting process causes that extremely unstable sensor signals are collected, meanwhile, uncontrollable factors such as machine tool systems and factory environments cause that a large amount of noise is doped in the signals, and therefore the noise in original signals needs to be removed efficiently and reliably before sensor signal feature extraction is carried out, and useful components in the signals need to be reserved to the maximum extent.

In processing and analyzing the sensor signal during milling, the effective component or the main energy of the signal is usually concentrated in the low Frequency part of the Tooth Pass Frequency (TPF) and integer multiples thereof. Noise in the sensor signal during cutting usually appears to have high frequency characteristics, and by utilizing the characteristics of the cutting noise, a low-pass filter, a band-pass filter or a wavelet filter is usually used, and the high frequency part in the signal is eliminated by setting a threshold value to complete the task of noise reduction. Fig. 2(a), 3(a) and 4(a) show waveforms of a raw vibration signal and a cutting force signal of a milling cutter in a feeding direction within one tool rotation period, and fig. 2(d), 3(d) and 4(d) show spectral curves corresponding to the raw signal, and it can be observed that the raw signal has more energy in a high frequency part, which indicates that more noise information exists in the raw data.

Firstly, performing extremum threshold denoising on an original signal by adopting common wavelet filtering, adopting a db3 wavelet, performing 5-layer wavelet decomposition, wherein waveforms after denoising are shown in fig. 2(b), 3(b) and 4(b), which can find that waveform changes of cutting force, vibration and sound signals are very obvious, and by observing corresponding spectrum analysis (fig. 2(e), 3(e) and 4(e)), noise energy of a high-frequency part can be well suppressed and effective information of TPF integral multiple in the signals can be kept, but the waveform of the signal after denoising is greatly different from that of the original signal, especially the waveform changes of the vibration signal and the sound signal can obviously find that the useful information in the original signal is blurred by the wavelet filtering algorithm, so that if the method is adopted to denoise the original signal, a large amount of singular point information is lost, thereby resulting in an inefficient analysis of the singular characteristics of the signal. Therefore, a noise reduction method capable of maximally preserving singular information in an original signal while reducing noise is needed.

According to the calculation of the Lee's index of different singular points of the signal, the Lee's index of the noise is usually negative, so that the maximum value point of the modulus can be judged along the scale s in the (u, s) plane (space u, scale s)The law of variation distinguishes whether the modulo maximum point is generated by noise or by a signal. If there is a modulo maximum point, its wavelet coefficient value decreases with decreasing scale s and eventually converges to u of the u-axis₀At the coordinate point, the maximum line communicated with the module maximum point corresponds to a signal point containing singularity information; on the contrary, if the value of the wavelet coefficient corresponding to the modulo maximum is significantly increased along with the reduction of the scale, the point is usually a point controlled by noise, so a threshold T (formula 1) is set on the maximum scale to screen the modulo maximum points caused by the noise, if the value of the wavelet coefficient of the modulo maximum points is less than T, the modulo maximum points are set to be zero, then the tower algorithm of Mallat is used to reconstruct the signal by using the wavelet coefficient to achieve the purpose of noise reduction, and the noise reduction algorithm based on the wavelet coefficient modulo maximum evaluation is referred to as the modulo maximum noise reduction method hereinafter.

Where Z is a constant, here taken to be 2, and the discrete scale s is 2^j(J-0, 1,2, … J), where J is the maximum value of the discrete scaling factor, and a relatively large scaling factor may cause some loss of locally significant singularity information, where J-5 is chosen, and M is the maximum value of all the modulo maxima on the maximum scale. Fig. 2(c), fig. 3(c) and fig. 4(c) show waveform curves of cutting force, vibration and sound signals obtained by wavelet coefficient reconstruction after a first-order gaussian function with a first-order vanishing moment is adopted for wavelet basis evaluation and noise mode maximum value points are screened, the waveform of the signals can be found to be closer and smoother than the original signals, and meanwhile, corresponding spectral analysis curve graphs 2(f), fig. 3(f) and fig. 4(f) are observed, so that the noise energy of the high-frequency part of the signals after noise reduction can be effectively suppressed, meanwhile, effective information on integral multiples of the TPF is retained, and the method can effectively and reliably remove noise and simultaneously retain useful components in the signals.

And S103, performing singular feature extraction on the signal subjected to noise reduction preprocessing.

For computing the global singularity of a function, the fourier transform is an efficient method if the function f (t) is bounded on the real space R and exists

The function f (t) is uniform in the Lee index α over R, where the Fourier transform of the function f (t) isWhile the uniform litz index singularity of the function f (t) on R depends heavily on the decay law of its fourier transform. However, the fourier transform cannot measure the local characteristic information of the signal, so that the signal cannot pass throughTo evaluate the singularity of f (t) at a certain point. In this case, the wavelet transform implements analysis of local details of the signal through a scaling translation operation, so that it can calculate the singular ledebur index of the function f (t) at a certain point or a certain interval.

The vanishing moment property of the wavelet basis functions is particularly important in order to be able to evaluate the local singularities of the signal. If wavelet basisExistence of n-order vanishing moment, then

When evaluating the Lee exponent α of a function f (t) using a wavelet basis with vanishing moments of order n (n > α) for a wavelet transform, the wavelet basis functions are orthogonal to Taylor polynomials of order n-1. Since n > α, the Taylor polynomial p_v(t) is of order n-1 at most, so that p_vThe wavelet transform of (t) is 0.

Due to the fact that

f(t)＝p_v(t)+ε_v(t) (5)

So that the wavelet transform Wf (u, s) of f (t) can also be written as

Wf(u,s)＝Wε_v(u,s) (6)

If any damping constant C exists_m(m. epsilon. N) so that

Then the wavelet basis is illustratedIs rapidly attenuated if the wavelet basisThere is an n-th order vanishing moment, when and only when there is a fast decay function θ (t), such that

Thereby to obtain

WhereinThe formula (8) shows that the wavelet base having n-order vanishing moment and rapidly attenuatingCan be expressed as the nth derivative of the fast decay function theta (t). Therefore, as shown in equation (9), the wavelet transform is equivalent to a multi-scale differential operator. If f (t) is n times in the vicinity of the point uCan be differentiated, then the formula (9) can be expressed as

Equation (10) shows that the singular ledeburite index of a signal can be estimated by the change rule of wavelet transform coefficients decreasing with the scale s.

By the above-mentioned relation between the calculation of the ledeburite index and the attenuation of the wavelet transform with decreasing scale, the ledeburite index can be calculated by analyzing the change rule of the wavelet coefficient in the (u, s) plane. In addition, whether a singular point exists or not and the size of the singular Leersian index can be evaluated by calculating the convergence condition of the modulus maximum value of the wavelet coefficient after decreasing along with the scale s.

Modulo maximum is the point if present (u)₀,s₀) So that the modulus | Wf (u) of the wavelet transform coefficient₀,s₀) I is a local maximum, i.e.

To avoid | Wf (u)₀,s₀) I is a special case of a constant, only the strict maximum value that is approximated from the left or right is considered. Meanwhile, a connected curve is defined in the (u, s) plane, all points on the curve are module maximum points, and the curve is called a maximum line.

According to the expression of equation (10), the wavelet transform can be considered as a multi-scale differential operator in which convolution operations are performedIt can be considered as "polishing" the original signal, if the wavelet basis used has only 1 st order vanishing momentThe modulus maximum point at this time corresponds to the step point of the original signal, which is also the signal f (t) passPeak point of first derivative after polishing.

However, when wavelet bases are usedA modulus maximum point (u) detected when wavelet transform is performed₀,s₀) It is not certain whether it is on a line of maxima converging to a smaller scale. If after the scale is decreased, | Wf (u)₀S) | at point u₀The absence of a modulo maximum point in the vicinity results in the inability to use the modulo maximum to evaluate the singularity of the signal.

Hummel's research rules indicate that when usedIs a wavelet basis function (As a gaussian function), for f (t) ∈ L²(R), the modulus maximum points of the wavelet transform Wf (u, s) are all positioned on a certain connected modulus maximum line, and the modulus maximum points are always kept continuous along with the decreasing of the scale. Due to the normalization of the gaussian function, all the detected modal maximum lines can be guaranteed to be extended to the minimum scale. Therefore, all wavelet basis functions used in the following of the present application are based on gaussian functions.

If f (t) is a point state or unity singular Lee's index α around v, the mode maximum point (u, s) in the cone of influence is satisfied only if a constant A > 0 is present

|Wf(u,s)|≤As^α+1/2 (12)

Logarithmic calculation on both sides of the formula, i.e.

log₂|Wf(u,s)|≤log₂ A+(α+1/2)log₂ s (13)

Thus, the singular Leersian index of point v can be calculated by calculating log₂I Wf (u, s) | is a dependent variable, log₂s is the maximum first derivative of the function curve composed of independent variables at the v pointNumber (slope of the curve).

Since the discrete wavelet transform based on the dyadic wavelet can completely and stably sparsely represent the original signal, only the wavelet transform dyadic scale { s ═ 2 is calculated in order to reduce the amount of calculation^jDue to the non-discontinuity of the maximum line, the communication of the maximum line and the evaluation of the Lee's index can be realized.

And S104, performing self-adaptive extraction and fusion of deep features closely related to tool wear damage on the time domain statistical information, the frequency domain statistical information, the singular Leersian index and the wavelet packet coefficient of the original signal.

In order to extract the characteristic value which is most relevant to cutter abrasion in the original signal and the singular Leersian index thereof, a novel multi-dimensional stack sparse automatic encoder model (MD-SSAEs) based on characteristic fusion is provided. Firstly, four SSAE models are designed to learn the characteristics of data, and the extracted Lee's index, wavelet packet coefficient, common time domain characteristics and frequency domain characteristics are respectively used as original characteristics to be input. To implement this structure, an improved loss function is employed to improve feature learning capabilities. Then, feature fusion and deep feature learning are performed using one-dimensional SSAE.

FIG. 5 represents a training process model for MD-SSAEs, with four SSAEs structured by one input layer X¹A plurality of hidden layers X^j(j-2, 3, …, n-1) and output layer y-XⁿAnd (4) forming. The input layer contains unprocessed TD dataFD data obtained by Fast Fourier Transform (FFT)Singular Lee indexAnd wavelet packet coefficients obtained by wavelet packet decompositionEach hidden layer has four vectorsFour eigenvectors y are obtained by calculating all automatic encoders and establishing a new modeling framework₁,y₂,y₃,y₄。

The loss function of the MD-SSAEs model proposed in this application is modified as:

where K is 1,2,3,4 and j is 1,2, …, n_kRespectively, the kth SSAE and the jth autoencoder, and x is the input to the autoencoder. c (w) and r (w) are the number of columns and rows of the weight matrix w.

Weight matrixThe optimal solution can be calculated by equation (15), and by training the new model, y is obtained₁,y₂,y₃,y₄The multi-dimensional depth feature is then subjected to feature fusion and deep feature learning by utilizing one-dimensional SSAE, and the depth feature vector y is trained by an MD-SSAEs model₁,y₂,y₃,y₄New feature vector X is fused into one_M＝[y₁,y₂,y₃,y₄]Vector X_MWill be used as input for one-dimensional SSAE for deeper feature extraction. The tool has good continuity in the abrasion process, and is connected to an output layer of the one-dimensional SSAE by utilizing the non-stationarity and the complex non-linear characteristic of the non-linear regression function, so that the capability of predicting the abrasion of the progressive tool is improved.

FIG. 6 illustrates a learning process for feature fusionStructure, first, characteristicsInput into one-dimensional SSAE and hide table quantityMay be determined by a weight matrixAnd calculating corresponding numerical values. Table amount of deeper levelMay be determined by a weight matrixAnd H₁And calculating a result. Subsequently, feature X_FObtaining an output value Y by using an input function F (X) as a non-linear regression_pre. Absolute error E between final tool wear and actual tool wear₁A calculation will be made.

E₁←|Y_pre-Y_act|＝|^F(X_F)-Y_act| (17)

Where A, B, and C are constants, matrices, and vectors of nonlinear regression functions, respectively.

To reduce E₁Is the parameter level of the update equation theta ═ { A, B, C } is

Wherein [ theta ]]_(q)And [ E₁]_(q)Respectively representing a parameter set and a q-th iteration, wherein q represents the current iteration number. Eta_(q)Is the rate of change, R_lAnd R_hThe empirical values of the decreasing and increasing coefficients, respectively, are set in the present application to R_lE [0.2,1) and R_h∈(1,5]。

In performing error minimization E₁Thereafter, the parameters A, B, C will be trimmed to a, b, c. And yield value Y_preIs converted to be based on X_FAnd X of a, b, c_F', feature X_FWill pass through Y_pre' and a, b, c calculate the corresponding result. Finally, for the purpose of error back propagation, between X_FAnd X_FError E of `₂Input to the one-dimensional SSAE.

When the error E is detected₂After input to one-dimensional SSAE, W₂Will be based on E₂Is fine-tuned to w₂Then W₁Will be based on w₂Upgrade to w₁. Hidden table quantity x₂Will be based on X_MAnd w₁Conversion to x₂', then by x₂' and W₁' the correlation calculates the feature vector X_M', last between X_M' and X_MError E between₃Will be input into the MD-SSAEs model. Following the rule of error back propagation, a weight matrixWill be based on the error E₃Fine tuning is performed. By now the modeling framework has been designed and the construction of the proposed model, which can be used for tool wear prediction, can be done by iterative training.

And S105, establishing a classification recognition model and a sudden change damage model of the tool damage state, so as to respectively realize the online monitoring of the tool wear loss and the damage state.

A nonlinear regression analysis model of data sensitive characteristics and tool slowly-varying abrasion loss is built on the basis of a convolutional neural network and a cyclic neural network deep learning algorithm, a signal characteristic input matrix is built, extracted deep characteristics are input, then a mapping relation between a detection signal and a tool state is built, and real-time output of the tool abrasion loss is achieved through a trained deep learning model.

The method comprises the steps of establishing an identification model fusing heterogeneous characteristics and a tool sudden-change damage state based on a SoftMax model, a support vector machine and the like, identifying the tool wear state on the whole, dividing the whole service life cycle of the tool into three types of wear states according to the wear extent and the wear state of the tool, wherein the wear states are respectively initial wear, normal wear and rapid wear, and the wear amount is 0-0.03 mm, 0.03-0.12 mm and 0.12-0.3 mm. The two models are integrated, the tool abrasion loss and the damage state under the drive of real-time data are monitored on line, and whether the tool is replaced or not is determined through the output result.

In order to solve the problem of online monitoring of the wear state of a cutter in the milling process, the invention provides a cutter state detection method based on a singular Lee index, which is characterized in that a signal noise reduction algorithm based on modulus maximum evaluation is established for noise reduction, useful components in signals are retained to the greatest extent, the Lee index in the signals is obtained by a Lee index calculation method based on a wavelet singularity analysis theory, deep-layer characteristics of the signals are extracted and fused by a multi-dimensional stack sparse automatic encoder model, finally, the wear state or wear amount of the cutter is identified by a depth learning calculation method, the real-time monitoring of the cutter state is relatively stably and reliably realized, the method is suitable for monitoring the cutter state during large-batch production of parts, and has high adaptability to small-batch production and even single part production and processing, the application range is wider.

It is noted that, in this document, relational terms such as "first" and "second," and the like, may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.