CN110598299B - Surface integrity evaluation method based on sensor fusion and deep learning - Google Patents

Surface integrity evaluation method based on sensor fusion and deep learning Download PDF

Info

Publication number
CN110598299B
CN110598299B CN201910833046.4A CN201910833046A CN110598299B CN 110598299 B CN110598299 B CN 110598299B CN 201910833046 A CN201910833046 A CN 201910833046A CN 110598299 B CN110598299 B CN 110598299B
Authority
CN
China
Prior art keywords
integrity
sensor
signal
value
layer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910833046.4A
Other languages
Chinese (zh)
Other versions
CN110598299A (en
Inventor
焦黎
程明辉
冯吕晨
李晨旭
周天丰
解丽静
刘志兵
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Priority to CN201910833046.4A priority Critical patent/CN110598299B/en
Publication of CN110598299A publication Critical patent/CN110598299A/en
Application granted granted Critical
Publication of CN110598299B publication Critical patent/CN110598299B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computer systems based on biological models
    • G06N3/004Artificial life, i.e. computers simulating life
    • G06N3/006Artificial life, i.e. computers simulating life based on simulated virtual individual or collective life forms, e.g. single "avatar", social simulations, virtual worlds or particle swarm optimisation
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computer systems based on biological models
    • G06N3/02Computer systems based on biological models using neural network models
    • G06N3/08Learning methods
    • G06N3/084Back-propagation

Abstract

The invention relates to the field of machining, in particular to a surface integrity characterization parameter evaluation method based on multi-sensor fusion and a deep learning algorithm, which comprises a dynamometer sensor, a pressing plate, a vibration sensor, a clamp, an end mill, a temperature measurement sensor and a workpiece which are arranged on a milling machine workbench, wherein the dynamometer sensor is arranged on the milling machine workbench through the pressing plate; the method utilizes grey correlation analysis to convert a plurality of characterization parameters of the surface integrity into a single characterization parameter represented by grey correlation values, and realizes the evaluation of the surface integrity in the whole.

Description

Surface integrity evaluation method based on sensor fusion and deep learning
Technical Field
The invention relates to the field of machining, in particular to a surface integrity characterization parameter evaluation method based on multi-sensor fusion and a deep learning algorithm.
Background
Surface integrity is a comprehensive indicator for characterizing, evaluating and controlling the changes that may occur in the surface layer of a machined part during the machining and manufacturing process and its effect on the performance of the final product. In the field of machining, the strength and fatigue life criteria of materials are the design basis of machining aviation parts and vehicle suspension systems, and the surface integrity characterization parameters of the parts are the most critical factors influencing the fatigue life of the parts. Increasing the fatigue life of aerospace and vehicle components by controlling the surface integrity of the components has become a focus of recent research. At present, most researches on surface integrity and the influence of the surface integrity on fatigue life still remain on basic cutting tests of cutting parameters, tool parameters and the like, and actually, cutting force, cutting heat and elastic-plastic deformation generated in the cutting process are the essential reasons influencing the characterization parameters of the surface integrity. In addition, the Evaluation and control of surface integrity characterization parameters mostly focuses on the study of surface roughness parameters and the prediction of residual stress variation trend, and international journal "material in technology je" discloses "Evaluation of the surface integration in the classification of a large alloy using an artificial neural network and a genetic algorithm" in 2018, volume 52, phase 3. Because the actual part is machined, the process features are often included, and the process features are not simply linear milling. If the evaluation of the surface integrity is to be carried out, it is far from sufficient to use only cutting parameters, and the characterization parameters of the surface integrity are numerous, more than one parameter of surface roughness. Therefore, the evaluation of the surface integrity characterization parameters at present has a plurality of limitations, and the evaluation of the machining surface is adversely affected.
Disclosure of Invention
The technical problem to be solved by the invention is to provide a method for predicting a plurality of characterization parameters such as surface integrity and the like based on a deep learning algorithm by considering a plurality of process characteristics and fully considering the influence of the coupling action of cutting force, cutting heat and vibration on the surface integrity characterization parameters on the basis of cutting parameters.
In order to solve the technical problems, the invention adopts the following technical scheme that the invention comprises the following steps:
firstly, establishing an X-Y-Z coordinate system and building a sensor cutting signal acquisition system;
the sensor cutting signal acquisition system comprises a dynamometer sensor, a pressing plate, a vibration sensor, a clamp, an end mill, a temperature measurement sensor and a workpiece, wherein the dynamometer sensor is arranged on a milling machine workbench through the pressing plate;
the T-shaped screw penetrates through a T-shaped groove of a milling machine workbench and a pressing plate and is fixed through the hexagon nut; the clamp is arranged on a dynamometer sensor, the vibration sensor is arranged on the clamp along the X direction, a workpiece is arranged in the clamp, and the temperature measurement sensor is arranged in the workpiece;
preprocessing the signals acquired based on the step I;
thirdly, based on the step two, performing feature extraction and selection on the cutting signal;
determining the overall evaluation index of the surface integrity;
establishing a surface integrity overall evaluation index prediction model.
The second step comprises:
201, carrying out noise reduction processing on time series signals collected by a dynamometer sensor, a vibration sensor and a temperature measurement sensor in the cutting process;
202 obtaining each mode function u through Hilbert transformk(t) analyzing the signal to obtain its single-sided spectrum, mixing the analyzed signals of each mode with an estimated center frequencyModulating the spectrum of each mode to a respective fundamental band:
203 calculates the square L of the gradient of the above demodulated signal2Norm, estimating the bandwidth of each modal signal, wherein the constrained variational expression is as follows:
wherein, { uk}:={u1,...,uKDenotes all mode functions, { wk}:={w1,...,wkRepresents the center frequency of the mode function;
204 introduces a secondary penalty factor alpha and a lagrangian multiplier lambda (t), changes the constrained variable problem into the unconstrained variable problem, and the expanded lagrangian expression is as follows:
205 by alternately updating by using a multiplicative operator alternating direction methodAnd λn+1Seeking to expand 'saddle point' of Lagrange expression, and converting the value problem of the center frequency into a frequency domain:
whereinIs the center of gravity of the current mode function power spectrum; to pairPerforming inverse Fourier transform to obtain a real partThen is { uk(t)};
206 optimizing parameters in the variational modal decomposition algorithm by using a particle swarm optimization algorithm;
207 comparing the decomposed cutting force signal and vibration signal with the original signal according to the amplitude and frequency characteristics of each order signal, eliminating the interference of random features, and reserving modal components simultaneously satisfying two conditions by using correlation and information entropy as the basis for signal screening, wherein the expressions of the correlation and the information entropy are respectively:
the third step includes:
301, extracting cutting force signals, vibration signals and temperature signals from a oscillogram, and extracting time domain characteristic quantities of a mean value, a standard deviation, a root mean square value, skewness and a kurtosis coefficient;
302, performing Fourier transform on the cutting signal to obtain a corresponding frequency spectrum, obtaining a power spectrum density function of the cutting signal according to a wiener-xinyou formula, calculating by adding a rectangular window on the power spectrum density function to obtain a power spectrum of the cutting signal, and finally extracting frequency band energy, variance and mean square frequency domain characteristic quantity;
303, performing fractal analysis on the one-dimensional time sequence signal, and extracting box dimension and correlation dimension fractal characteristic quantity of the cutting signal;
304, the extracted feature quantities are sorted by using a Caret package in the R language to determine the feature set.
The fourth step comprises:
401 taking the surface roughness, the axial residual stress and the radial residual stress as surface integrity characterization parameters, converting the three characterization parameters of the surface integrity into a single characterization parameter expressed by a gray level correlation value by utilizing gray level correlation analysis, and taking the single characterization parameter as a total evaluation index of the surface integrity;
402 form the surface integrity characterization parameters into a matrix as follows:
wherein m represents the number of indexes, and n represents the number of groups measured by the surface integrity characterization parameter;
403, performing non-dimensionalization on the index data in 402 by using an averaging method, wherein the expression is as follows:
404 determining a reference sequence, taking the optimal value of each characterization parameter of the surface integrity as the reference sequence, and calculating the absolute difference value of the index sequence of each evaluated object and the corresponding element of the reference sequence, namely | x |0(k)-xi(k) I (k 1., m, i 1., n, n is the number of objects to be evaluated), and determining the number of objects to be evaluatedAnd respectively calculating the association coefficient of each comparison sequence and the corresponding element of the reference sequence, wherein the expression of the association coefficient is as follows:
405, calculating the mean value of the correlation coefficient between each index and the element corresponding to the reference sequence for each evaluation object, wherein the expression is as follows:
the fifth step includes:
501, constructing a total evaluation index prediction model based on a deep belief network, training a limited Boltzmann machine (RBM) by using an unsupervised learning mode, receiving the output of a last layer of RBM (limited Boltzmann machine) by using a last BP (back propagation) network, and training the whole network by adopting a supervised mode;
502 regularization processing is performed before the selected features are delivered to the input layer, and the processing method is as follows:
wherein, Xi,jThe ith set of data values representing the jth characteristic quantity,means, σ, representing the j-th characteristic quantityiRepresents the variance of the jth feature quantity;
503, learning the feature vectors of the input layer by using a restricted boltzmann machine, and obtaining the value of the hidden layer node from the known visible layer node by using a particle swarm optimization algorithm according to the learning rule:
since RBM (restricted boltzmann machine) is a symmetric network, obtaining the hidden layer node yields the value of the visible layer node:
in the formula, viIs the value of the ith node of the visual layer, hjIs the value of the jth node of the hidden layer, b and c represent the bias values of the visual layer and the hidden layer, respectively, wjiThe weight value between the visible node i and the hidden node j is obtained;
504 the joint probability distribution of the feature vector of the visual layer and the feature vector of the hidden layer is:
where E represents an energy function, θ0={W0 ij,b0 i,c0 jThe parameter set of the model parameter is used, n represents the number of nodes of the visual layer, and m represents the number of nodes of the hidden layer;
505, training an RBM (restricted Boltzmann machine) model by using a CD (comparison hashing) criterion, and determining the node number m, the learning rate epsilon and the maximum iteration number T of the hidden layer of the RBM (restricted Boltzmann machine) model1And a network parameter set theta to RBM (restricted Boltzmann machine)0={W0 ij,b0 i,c0 jInitializing; gibbs sampling is carried out to enable the state vector V of the visible layer0The value of (A) is the value of the input layer, and the state H of each node of the hidden layer is calculated by using a formula0And the state V of the reconstructed visible layer and hidden layer1And H1The updating process of the parameters is as follows:
506, taking the output of the RBM (restricted Boltzmann machine) of the last layer as the input of a BP (back propagation) network, and calculating the predicted value of the surface integrity overall evaluation index through BP (back propagation), wherein the calculation process of the BP (back propagation) network is as follows:
wherein y ispreiIndicates the predicted value, wiRepresenting the weight of each neuron connection of the BP (back propagation) network, biRepresenting the bias of each neuron of the BP network, HiAn output value representing an RBM (restricted boltzmann machine) of the last layer;
507 selecting an optimal model, training the prediction model, selecting the model with the minimum error and the maximum decision coefficient as the optimal prediction model, and predicting the target function of the modelThe number being a mean square error function (MSE) and a decision coefficient (R)2) The following are:
wherein y isiAnd ypreiRespectively representing observed values and predicted values of the surface integrity overall evaluation index, and n represents the total number of samples.
When the workpiece is an arc workpiece, the temperature measuring sensor is arranged in the arc workpiece; when the workpiece is a milling groove workpiece, the temperature measuring sensor is inserted into the hole from the bottom of the milling groove workpiece.
The invention has the following positive effects: according to the invention, a set of multi-sensor fusion cutting signal measuring device is built by utilizing a force measuring instrument sensor, a vibration sensor and a temperature measuring sensor, so that the influence of cutting parameters and cutter parameters on surface integrity characterization parameters is considered, and the influence of cutting force, vibration and cutting temperature on the surface integrity characterization parameters in the cutting process is also considered; the invention simultaneously carries out noise reduction reconstruction processing on the collected cutting force and vibration time sequence signals, thereby enhancing the robustness of data; the method not only extracts the characteristics of the preprocessed signals in the aspects of time domain and frequency domain, but also extracts the characteristics of the preprocessed signals from the fractal angle, and selects the characteristics by utilizing a Caret packet in R language, thereby not only determining the characteristic quantity with higher correlation with surface integrity characterization parameters, but also sequencing the importance of the extracted characteristic quantity; the invention can simultaneously carry out linear milling, convex arc milling, concave arc milling and groove milling, and can obtain the influence rules between different process characteristics and surface integrity characterization parameters; the method utilizes grey correlation analysis to convert a plurality of characterization parameters of the surface integrity into a single characterization parameter represented by a grey correlation value, and realizes the evaluation of the surface integrity on the whole; the surface integrity overall evaluation index is predicted by using the deep belief network in the deep learning algorithm, and the prediction result based on the surface integrity overall evaluation index can be fed back in time, so that the adjustment of subsequent cutting parameters is facilitated.
Drawings
FIG. 1 is a general flow chart for implementing surface integrity evaluation
FIG. 2 is a diagram of a multi-sensor fusion cutting signal measuring device
FIG. 3 is a schematic view of temperature measurement during milling process
FIG. 4 is a flow chart of a cutting signal preprocessing process
FIG. 5 is a flow chart of deep belief network implementation
FIG. 6 is a schematic diagram of raw data of a collected cutting signal
FIG. 7 is a diagram illustrating the particle swarm optimization
FIG. 8 is a schematic diagram of a slicing signal decomposition and reconstruction process
FIG. 9 is a diagram showing the prediction results of the overall evaluation index of surface integrity
In the figure: 1 milling machine workbench, 2 dynamometer sensors, 3 pressing plates, 4 vibration sensors, 5 clamps, 6 four-edge end mills, 7 concave arc workpieces, 8 hexagon nuts, 9T-shaped screws, 10 milling groove workpieces and 11 temperature measuring sensors.
Detailed Description
The invention is described in detail below with reference to the drawings and the specific cutting examples. As shown in fig. 1, in order to realize the overall flow chart of the surface integrity evaluation, the main steps are as follows:
firstly, establishing an X-Y-Z coordinate system and building a multi-sensor cutting signal acquisition device;
as shown in fig. 2, the multi-sensor cutting signal acquisition system comprises a force measuring instrument sensor 2, a pressing plate 3, a clamp 5, a vibration sensor 4 and a temperature measuring sensor 11 which are arranged on a milling machine workbench 1, wherein the force measuring instrument sensor 2 is fixed on the milling machine workbench 1 through the pressing plate 3; the device is characterized by further comprising a T-shaped screw 9 and a hexagon nut 8, wherein the T-shaped screw 9 penetrates through a T-shaped groove of the milling machine workbench 1 and a waist-shaped hole of the pressing plate 3, and the pressing plate 3 is fixed through the hexagon nut 8, so that the dynamometer sensor 2 is pressed tightly; the fixture (5) is fixed on the dynamometer sensor (2) through an inner hexagonal countersunk head screw, the vibration sensor (4) is arranged on the fixture (5) along the X direction, a workpiece is arranged in the fixture (5), and the temperature measurement sensor (11) is arranged in the workpiece. As shown in fig. 3, the temperature measuring sensor 11 is fixed in a hole drilled in advance in the arc workpiece 7, and the temperature measuring mode during milling the plane is the same as that during processing the arc workpiece; when milling the groove type workpiece, the temperature measuring sensor is inserted into the hole from the bottom. And then, the cutting allowance reserved in advance is milled by using the four-edge end mill 6, and the collection and collection of cutting force, vibration and cutting temperature are completed in the milling process.
Cutting signal preprocessing process;
201 cutting force signals and vibration signals collected in the cutting process belong to time series signals, the time series signals are often interfered by noise, and the components of the time series signals are impure, so that the time series signals need to be subjected to noise reduction processing, and therefore, the decomposition and reconstruction of the cutting force signals and the vibration signals are realized by using a variational modal decomposition algorithm.
202 firstly, a variational problem is constructed, and each mode function u is obtained through Hibert (Hibert) transformationkAnd (t) analyzing the signal to obtain the single-side frequency spectrum. Mixing the analysis signals of each mode to obtain an estimated center frequencyModulating the spectrum of each mode to a respective fundamental band:
203, calculating the square L2 norm of the gradient of the demodulation signal, and estimating the bandwidth of each modal signal, wherein the constrained variation problem is as follows:
wherein, { uk}:={u1,...,uKDenotes all mode functions, { wk}:={w1,...,wkRepresents the center frequency of the mode function.
204, introducing a secondary penalty factor α and a lagrangian multiplier λ (t), and changing the constrained variable problem into an unconstrained variable problem, wherein the secondary penalty factor can ensure the reconstruction accuracy of the signal in the presence of gaussian noise, the lagrangian makes the constraint condition maintain strict, and the extended lagrangian expression is as follows:
205, then adopting a multiplicative operator alternating direction method, and alternately updatingAnd λn+1The 'saddle point' of the extended lagrangian expression is sought, by which is meant the critical point that is neither the maximum nor the minimum point. According to the same process, the problem of the center frequency is firstly converted into the frequency domain:
whereinIs the center of gravity of the current mode function power spectrum; to pairPerforming inverse Fourier transform to obtain real part of uk(t)}。
206 since the modal decomposition number K and the penalty factor alpha in the variational modal decomposition algorithm are determined in advance, parameters in the variational modal decomposition algorithm are optimized by the particle swarm optimization algorithm, the range of the modal decomposition number is set to [2,20] in advance, and the range of the penalty factor is set to [200,10000] in advance.
207, the decomposed cutting force signal and vibration signal contain multi-order information, and the amplitude and frequency characteristics of each order of signal are compared with the original signal to eliminate the interference of random features, so that the correlation and the information entropy can be selected as the basis for signal screening, and modal components meeting two conditions at the same time are reserved. The expression of the correlation and the information entropy is as follows:
the preprocessing process of the cutting signal is shown in fig. 4, after the optimal parameter combination is found by utilizing the particle swarm optimization algorithm, the decomposition of the cutting force and the vibration signal is realized by utilizing the variational modal decomposition algorithm based on the optimal parameter combination; and if the correlation coefficient of the decomposed component and the original signal is 0.6-0.9 and the information entropy meets the 3 sigma criterion, reserving the component for completing the reconstruction of the cutting force and the vibration signal, and discarding the rest components.
Extracting and selecting characteristics of the cutting signal;
the 301 time domain signal can accurately and intuitively reflect the change condition of each instant signal in the cutting process, the cutting force signal, the vibration signal and the temperature signal are extracted from the oscillogram, and time domain characteristic quantities such as a mean value (M), a standard deviation (RMS), a root mean square value (P), skewness (Skaew), a Kurt coefficient (Kurt) and the like are extracted.
The 302-cut signal is not only time dependent, but also frequency and period dependent. The method comprises the steps of obtaining a corresponding frequency spectrum through Fourier transform of a cutting signal, obtaining a power spectrum density function of the cutting signal according to a wiener-xinyou formula, calculating by adding a rectangular window on the basis to obtain a power spectrum of the cutting signal, and finally extracting frequency domain characteristic quantities such as Frequency Band Energy (FBE), variance (Var) and Mean Square Frequency (MSF).
The 303 fractal theory is commonly used for describing the irregularity and the self-similarity of an object, and useful characteristic information is extracted by using the 303 fractal theory, so that the motion state of the system can be qualitatively and quantitatively analyzed, and the signal analysis of a mechanical system is realized. In the fractal theory research, the fractal dimension is an important physical quantity and quantitatively describes the complexity of a research object, so that the fractal analysis on a one-dimensional time sequence signal is widely applied. For this purpose, a box Dimension (DB) and a correlation Dimension (DC) fractal feature quantity of the slice signal are extracted.
304 during the milling process, cutting forces Fx, Fy, Fz in three directions, vibrations Vx, Vy in two directions, and a cutting temperature Tx in one direction are collected. Feature set formed by feature extraction
X={Mi,RMSi,Pi,Skewi,Kurti,...,FBEi,Vari,MSFi,DBi,DCiAnd the vibration in the two directions and the cutting temperature in one direction are expressed, wherein i is 1-6.
305, the extracted feature quantities have strong redundancy and correlation, not only can increase the calculation amount of the algorithm, but also can interfere the prediction precision of the algorithm, therefore, the extracted feature quantities are subjected to importance ranking by using a Caret packet in an R language, the importance degree of each feature quantity is determined through analysis and calculation, and the feature set is determined.
Determining the overall evaluation index of the surface integrity;
and measuring surface roughness, axial residual stress and radial residual stress as surface integrity characterization parameters in the 401 test process. In order to realize the overall evaluation of the surface integrity, the three characterization parameters of the surface integrity are converted into a single characterization parameter expressed by the gray level correlation degree by utilizing gray level correlation analysis, and the single characterization parameter is used as the overall evaluation index of the surface integrity.
402 for the measured surface integrity characterizing parameters form the following matrix:
wherein m represents the number of indexes and n represents the number of groups measured by the surface integrity characterization parameter.
403, because the characterization parameters of the surface integrity have different dimensions, the index data is subjected to non-dimensionalization, and the common non-dimensionalization processing methods include an averaging method and an initial value method, where the averaging method is adopted. The expression is as follows:
404 determines a reference sequence, takes the optimal value of each characterization parameter of the surface integrity (the optimal value is equal to the minimum value in terms of surface roughness and residual stress) as the reference sequence, and calculates the absolute difference value of the corresponding element of each evaluated object index sequence (comparison sequence) and the reference sequence, namely | x |, one by one0(k)-xi(k) (n is the number of objects to be evaluated), and specifiesAnd respectively calculating the association coefficient of each comparison sequence and the corresponding element of the reference sequence. The expression of the associated coefficient is as follows:
405, calculating a correlation sequence, and calculating a mean value of the correlation coefficients of each index and the corresponding element of the reference sequence for each evaluation object (comparison sequence) respectively to reflect the correlation between each evaluation object and the reference sequence, wherein the expression is as follows:
and converting a plurality of characterization parameters into one characterization parameter through gray level correlation analysis. And finally, taking the obtained grey correlation degree as an overall evaluation index of the surface integrity.
Establishing a surface integrity overall evaluation index prediction model;
501, in order to realize the prediction of the overall evaluation index of the surface integrity, an overall evaluation index prediction model based on a deep confidence network is built. The deep belief network belongs to one of deep learning algorithms, and is composed of a series of limited Boltzmann machines (RBMs) and a layer of BP (back propagation) network. The training process of the deep confidence network can be divided into two steps: firstly, an unsupervised learning mode is used for training an RBM (restricted Boltzmann machine); secondly, the final BP (back propagation) network is used to accept the output of the last RBM (restricted Boltzmann machine), and the whole network is trained in a supervision way to carry out fine adjustment. The deep belief network used by the model consists of an input layer, two layers of constrained boltzmann machines, and a BP (back propagation) network (output layer).
502 because the selected features have different dimensions, the selected features need to be regularized before being sent to the input layer, and the processing method is as follows:
wherein, Xi,jThe ith set of data values representing the jth characteristic quantity,means, σ, representing the j-th characteristic quantityiThe variance of the jth feature quantity is represented.
503 followed by learning the feature vectors of the input layers using a constrained boltzmann machine. The limited Boltzmann machine consists of a visual layer and a hidden layer, the visual layer and the hidden layer are in full bidirectional connection, and the learning rule adopts a simulated annealing algorithm. Thus, the values of the hidden layer nodes are derived from the known visible layer nodes:
because RBM (restricted Boltzmann machine) is a symmetric network, the values of nodes of the visual layer can be obtained by obtaining nodes of the hidden layer
In the formula, viIs the value of the ith node of the visual layer, hjIs the value of the jth node of the hidden layer, b and c represent the bias values of the visual layer and the hidden layer, respectively, wjiIs the weight between the visible node i and the hidden node j.
504 the joint probability distribution of the feature vector of the visual layer and the feature vector of the hidden layer is:
where E represents an energy function, θ0={W0 ij,b0 i,c0 jIs the parameter set of the model parameters, n represents the number of nodes of the visual layer, and m represents the number of nodes of the hidden layer.
505 is followed by training of the RBM (restricted boltzmann machine) model using the CD (contrast hashing) criterion. The calculation speed can be improved by using a CD (comparison hash) criterion, and the problem that the convergence speed cannot be ensured by using a Markov chain Monte Carlo method can be avoided. When the RBM (restricted Boltzmann machine) model is trained by using the CD (contrast hashing) criterion, the node number m, the learning rate epsilon and the maximum iteration number T1 of the hidden layer of the RBM (restricted Boltzmann machine) model are firstly determined, and the network parameter set theta of the RBM (restricted Boltzmann machine) is0={W0 ij,b0 i,c0 jInitializing; and Gibbs sampling is carried out, the value of the state vector V0 of the visible layer is used as the value of the input layer, and the state H0 of each node of the hidden layer and the states V1 and H1 of the reconstructed visible layer and hidden layer are calculated by using a formula. The updating process of the parameters is as follows:
the updating of the parameters is not performed after all samples are trained, but the parameters are updated once every 6 or 8 samples, so that the calculation amount can be greatly reduced, and the updating speed is accelerated.
506, taking the output of the RBM (restricted Boltzmann machine) of the last layer as the input of a BP (back propagation) network, and calculating the predicted value of the surface integrity overall evaluation index through BP (back propagation), wherein the calculation process of the BP (back propagation) network is as follows:
wherein y ispreiIndicates the predicted value, wiRepresenting the weight of each neuron connection of the BP (back propagation) network, biRepresenting the bias of each neuron of the BP (back propagation) network, HiAnd represents the output value of RBM (restricted boltzmann machine) of the last layer.
507, selecting an optimal model. And training the prediction model, and selecting the model with the minimum error and the maximum decision coefficient as the optimal prediction model. The objective function of the prediction model is the mean square error function (MSE) and the decision coefficient (R2), as follows:
wherein y isiAnd ypreiRespectively representing observed values and predicted values of the surface integrity overall evaluation index, and n represents the total number of samples.
The type of the dynamometer sensor is Kistler 9057B, the type of the vibration sensor is CT1050L ICP/IEPE, and the type of the temperature measuring sensor is a NANANANMAC quick response thermocouple, namely E12-3-K-U.
Example 1
The method is mainly based on the milling test developed above, samples with different characteristics are milled under different cutting parameters, cutting signals are collected in the milling process, and surface integrity characterization parameters of the samples are collected after the test is finished.
The material adopted in the test is 45CrNiMoVA, the cutting test is carried out by adopting an integral four-edge end mill Z-CARB ZAP 6x6x13x57, the test platform is DMU 80 mono BLOCK, the cutting force is acquired by respectively utilizing a dynamometer with the model of Kistler 9057B, the cutting vibration is acquired by a vibration sensor with the model of CT1050L ICP/IEPE, the cutting temperature is acquired by a temperature sensor with the model of E12-3-K-U, and the acquired original data are shown in figure 6.
After the cutting signal is acquired, because the cutting force signal and the vibration signal belong to a time series signal, the time series signal is often interfered by noise, and the components of the time series signal are impure, so that the noise reduction processing needs to be performed on the time series signal, and here, the cutting force signal in the 1 st direction is taken as an example to explain the noise reduction process of the cutting signal. Firstly, reading an original cutting signal, determining the optimal decomposition number and penalty factor combination in the variational modal decomposition algorithm by utilizing a particle swarm optimization algorithm based on the size of a peak state value, wherein the optimization result is shown in fig. 7. And then decomposing the input cutting force signal by using the optimal parameter combination [953,13] through a variation modal decomposition algorithm, and completing the reconstruction of the cutting force signal according to the correlation degree of the decomposed cutting force signal component and the original cutting force signal and the contained information entropy range, as shown in fig. 8.
And according to the cutting force decomposition and reconstruction process in the 1 st direction, the cutting force and vibration signals in other directions are decomposed and reconstructed. And then, calculating time domain, frequency domain characteristic quantity and fractal dimension of the decomposed and reconstructed cutting signal. Since the feature quantities obtained by calculation are numerous, not only redundancy of the feature quantities and reduction of prediction accuracy are caused, but also the amount of information is increased, and therefore, the selection of the feature quantities is realized based on a random forest algorithm by using a caret packet in R language. The top 8 important features after selection include: machining characteristics, Y-direction cutting force variance (Var2), total cutting force, Y-direction cutting force mean square error (Std3), maximum cutting temperature (Max6), X-direction cutting force skewness (Kurt1), Y-direction cutting force center of gravity frequency (VF2), and feed rate.
Due to the fact that the surface integrity characterization parameters are numerous, in order to achieve the overall evaluation of the surface integrity characterization parameters, the surface integrity characterization parameters are converted into overall evaluation indexes represented by gray-scale correlation values (GRG) by using a gray-scale correlation analysis algorithm, and partial conversion results are shown in table 1.
TABLE 1 results of surface integrity Gray level correlation analysis
After the feature extraction of the cutting signal and the conversion of a plurality of characterization parameters of the surface integrity, the overall evaluation index of the characterization parameters of the surface integrity is predicted by using a depth confidence network (DBN). The selected characteristic quantity is used as an input vector, a surface integrity overall evaluation index gray level correlation value (GRG) is used as an output quantity, a data set is randomly divided into a training set and a testing set in the prediction process, only the training set is used in the network training process, the testing set is used for testing the quality of the model, and the prediction result is shown in fig. 9. As can be seen from the prediction results, the surface integrity characterizes the feasibility of the parameter prediction method.
The embodiments described above are only preferred embodiments of the invention and are not exhaustive of the possible implementations of the invention. Any obvious modifications to the above would be obvious to those of ordinary skill in the art, but would not bring the invention so modified beyond the spirit and scope of the present invention.

Claims (5)

1. A surface integrity evaluation method based on sensor fusion and deep learning is characterized by comprising the following steps:
firstly, establishing an X-Y-Z coordinate system and building a sensor cutting signal acquisition system;
the sensor cutting signal acquisition system comprises a dynamometer sensor (2), a pressing plate (3), a vibration sensor (4), a clamp (5), an end mill (6), a temperature measurement sensor (11) and a workpiece, wherein the dynamometer sensor (2) is arranged on a milling machine workbench (1) through the pressing plate (3), and the milling machine workbench (1) is provided with a T-shaped groove;
the milling machine is characterized by further comprising a hexagon nut (8) and a T-shaped screw (9), wherein the T-shaped screw (9) penetrates through a T-shaped groove of the milling machine workbench (1) and the pressing plate (3) and is fixed through the hexagon nut (8); the clamp (5) is arranged on the dynamometer sensor (2), the vibration sensor (4) is arranged on the clamp (5) along the X direction, a workpiece is arranged in the clamp (5), and the temperature measurement sensor (11) is arranged in the workpiece;
preprocessing signals acquired based on the step I:
201, carrying out noise reduction processing on time series signals collected by a dynamometer sensor (2), a vibration sensor (4) and a temperature measurement sensor (11) in the cutting process;
202 obtaining each mode function u through Hilbert transformk(t) analyzing the signal to obtain its single-sided spectrum, mixing the analyzed signals of each mode with an estimated center frequencyModulating the spectrum of each mode to a respective fundamental band:
203 calculates the square L of the gradient of the above demodulated signal2Norm, estimating the bandwidth of each modal signal, wherein the constrained variational expression is as follows:
wherein, { uk}:={u1,...,uKDenotes all mode functions, { wk}:={w1,...,wkRepresents the center frequency of the mode function;
204 introduces a secondary penalty factor alpha and a lagrangian multiplier lambda (t), changes the constrained variable problem into the unconstrained variable problem, and the expanded lagrangian expression is as follows:
205 by alternately updating by using a multiplicative operator alternating direction methodAnd λn+1Seeking to expand 'saddle point' of Lagrange expression, and converting the value problem of the center frequency into a frequency domain:
whereinIs the center of gravity of the current mode function power spectrum; to pairPerforming inverse Fourier transform to obtain real part of uk(t)};
206 optimizing parameters in the variational modal decomposition algorithm by using a particle swarm optimization algorithm;
207 comparing the decomposed cutting force signal and vibration signal with the original signal according to the amplitude and frequency characteristics of each order signal, eliminating the interference of random features, and reserving modal components simultaneously satisfying two conditions by using correlation and information entropy as the basis for signal screening, wherein the expressions of the correlation and the information entropy are respectively:
thirdly, based on the step two, performing feature extraction and selection on the cutting signal;
determining the overall evaluation index of the surface integrity;
establishing a surface integrity overall evaluation index prediction model.
2. The method for evaluating the surface integrity based on the sensor fusion and the deep learning according to claim 1, wherein the step (c) comprises the following steps:
301, extracting cutting force signals, vibration signals and temperature signals from a oscillogram, and extracting time domain characteristic quantities of a mean value, a standard deviation, a root mean square value, skewness and a kurtosis coefficient;
302, performing Fourier transform on the cutting signal to obtain a corresponding frequency spectrum, obtaining a power spectrum density function of the cutting signal according to a wiener-xinyou formula, calculating by adding a rectangular window on the power spectrum density function to obtain a power spectrum of the cutting signal, and finally extracting frequency band energy, variance and mean square frequency domain characteristic quantity;
303, performing fractal analysis on the one-dimensional time sequence signal, and extracting box dimension and correlation dimension fractal characteristic quantity of the cutting signal;
304, the extracted feature quantities are sorted by using a Caret package in the R language to determine the feature set.
3. The method for evaluating the surface integrity based on the sensor fusion and the deep learning according to claim 1, wherein the step (iv) comprises the following steps:
401 taking the surface roughness, the axial residual stress and the radial residual stress as surface integrity characterization parameters, converting the three characterization parameters of the surface integrity into a single characterization parameter expressed by a gray level correlation value by utilizing gray level correlation analysis, and taking the single characterization parameter as a total evaluation index of the surface integrity;
402 form the surface integrity characterization parameters into a matrix as follows:
wherein m represents the number of indexes, and n represents the number of groups measured by the surface integrity characterization parameter;
403, performing non-dimensionalization on the index data in 402 by using an averaging method, wherein the expression is as follows:
404 determining a reference sequence, taking the optimal value of each characterization parameter of the surface integrity as the reference sequence, and calculating the absolute difference value of the index sequence of each evaluated object and the corresponding element of the reference sequence, namely | x |0(k)-xi(k) I (k 1., m, i 1., n, n is the number of objects to be evaluated), and determining the number of objects to be evaluatedAnd respectively calculating the association coefficient of each comparison sequence and the corresponding element of the reference sequence, wherein the expression of the association coefficient is as follows:
405, calculating the mean value of the correlation coefficient between each index and the element corresponding to the reference sequence for each evaluation object, wherein the expression is as follows:
4. the method for evaluating the surface integrity based on the sensor fusion and the deep learning as claimed in claim 1, wherein the fifth step comprises:
501, constructing a total evaluation index prediction model based on a deep belief network, training a limited Boltzmann machine (RBM) by using an unsupervised learning mode, receiving the output of a last layer of RBM (limited Boltzmann machine) by using a last BP (back propagation) network, and training the whole network by adopting a supervised mode;
502 regularization processing is performed before the selected features are delivered to the input layer, and the processing method is as follows:
wherein, XijThe ith set of data values representing the jth characteristic quantity,means, σ, representing the j-th characteristic quantityiRepresents the variance of the jth feature quantity;
503, learning the feature vectors of the input layer by using a restricted boltzmann machine, and obtaining the value of the hidden layer node from the known visible layer node by using a particle swarm optimization algorithm according to the learning rule:
since RBM (restricted boltzmann machine) is a symmetric network, obtaining the hidden layer node yields the value of the visible layer node:
in the formula, viIs the first visible layerValues of i nodes, hjIs the value of the jth node of the hidden layer, b and c represent the bias values of the visual layer and the hidden layer, respectively, wjiThe weight value between the visible node i and the hidden node j is obtained;
504 the joint probability distribution of the feature vector of the visual layer and the feature vector of the hidden layer is:
where E represents an energy function, θ0={W0 ij,b0 i,c0 jThe parameter set of the model parameter is used, n represents the number of nodes of the visual layer, and m represents the number of nodes of the hidden layer;
505, training an RBM (restricted Boltzmann machine) model by using a CD (comparison hashing) criterion, and determining the node number m, the learning rate epsilon and the maximum iteration number T of the hidden layer of the RBM (restricted Boltzmann machine) model1And a network parameter set theta to RBM (restricted Boltzmann machine)0={W0 ij,b0 i,c0 jInitializing; gibbs sampling is carried out to enable the state vector V of the visible layer0The value of (A) is the value of the input layer, and the state H of each node of the hidden layer is calculated by using a formula0And the state V of the reconstructed visible layer and hidden layer1And H1The updating process of the parameters is as follows:
506, taking the output of the RBM (restricted Boltzmann machine) of the last layer as the input of a BP (back propagation) network, and calculating the predicted value of the surface integrity overall evaluation index through BP (back propagation), wherein the calculation process of the BP (back propagation) network is as follows:
wherein y ispreiIndicates the predicted value, wiRepresenting the weight of each neuron connection of the BP (back propagation) network, biRepresenting the bias of each neuron of the BP network, HiAn output value representing an RBM (restricted boltzmann machine) of the last layer;
507 selecting an optimal model, training the prediction model, selecting a model with the minimum error and the maximum decision coefficient as the optimal prediction model, wherein the objective function of the prediction model is a mean square error function (MSE) and a decision coefficient R2The following are:
wherein y isiAnd ypreiRespectively representing observed values and predicted values of the surface integrity overall evaluation index, and n represents the total number of samples.
5. The method for evaluating the integrity of the surface based on the sensor fusion and the deep learning according to claim 1, wherein: when the workpiece is an arc workpiece (7), the temperature measuring sensor (11) is arranged in the arc workpiece (7); when the workpiece is a milling groove workpiece (10), the temperature measuring sensor (11) is inserted into the hole from the bottom of the milling groove workpiece (10).
CN201910833046.4A 2019-09-04 2019-09-04 Surface integrity evaluation method based on sensor fusion and deep learning Active CN110598299B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910833046.4A CN110598299B (en) 2019-09-04 2019-09-04 Surface integrity evaluation method based on sensor fusion and deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910833046.4A CN110598299B (en) 2019-09-04 2019-09-04 Surface integrity evaluation method based on sensor fusion and deep learning

Publications (2)

Publication Number Publication Date
CN110598299A CN110598299A (en) 2019-12-20
CN110598299B true CN110598299B (en) 2021-04-30

Family

ID=68857530

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910833046.4A Active CN110598299B (en) 2019-09-04 2019-09-04 Surface integrity evaluation method based on sensor fusion and deep learning

Country Status (1)

Country Link
CN (1) CN110598299B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021174525A1 (en) * 2020-03-06 2021-09-10 大连理工大学 Parts surface roughness and cutting tool wear prediction method based on multi-task learning
CN111558849B (en) * 2020-05-11 2021-10-22 内蒙古工业大学 Disc milling cutter machining parameter optimization method and device, electronic equipment and storage medium
CN112975574B (en) * 2021-04-22 2021-07-30 四川大学 Surface quality on-line detection system for aluminum alloy thin-wall part in milling process

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108536938A (en) * 2018-03-29 2018-09-14 上海交通大学 A kind of machine tool life prediction system and prediction technique

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108536938A (en) * 2018-03-29 2018-09-14 上海交通大学 A kind of machine tool life prediction system and prediction technique

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
58SiMn高强度钢车削表面完整性的试验研究;罗智文等;《表面技术》;20170131;第46卷(第1期);全文 *
Artificial intelligence for automatic prediction of required surface roughness by monitoring wear on facemill teeth;D. Yu. Pimenov等;《Intelligent Manufacturing》;20171215;全文 *
Evaluation of the surface integrity in the milling of a magnesium alloy using an artificial neural network and a genetic algorithm;Madhesan Pradeepkumar等;《Materials and technology》;20180630;全文 *
The Evaluation of Surface Integrity During Machining of Inconel 718 with Various Laser Assistance Strategies;Szymon Wojciechowski等;《www.researchgate.net/publication/321053456》;20170131;全文 *

Also Published As

Publication number Publication date
CN110598299A (en) 2019-12-20

Similar Documents

Publication Publication Date Title
CN110598299B (en) Surface integrity evaluation method based on sensor fusion and deep learning
Kilundu et al. Tool wear monitoring by machine learning techniques and singular spectrum analysis
Beyca et al. Heterogeneous sensor data fusion approach for real-time monitoring in ultraprecision machining (UPM) process using non-parametric Bayesian clustering and evidence theory
Zhang et al. Prediction of surface roughness in end face milling based on Gaussian process regression and cause analysis considering tool vibration
CN109766930B (en) Method for predicting residual life of mine mechanical equipment based on DCNN model
CN110303380B (en) Method for predicting residual life of cutter of numerical control machine tool
Bull et al. Probabilistic active learning: An online framework for structural health monitoring
Yang et al. Differential evolution-based feature selection and parameter optimisation for extreme learning machine in tool wear estimation
Silva et al. Damage detection in a benchmark structure using AR-ARX models and statistical pattern recognition
Wang et al. Random forest-bayesian optimization for product quality prediction with large-scale dimensions in process industrial cyber–physical systems
Fadare et al. Development and application of a machine vision system for measurement of surface roughness
Hui et al. Milling tool wear state recognition by vibration signal using a stacked generalization ensemble model
CN112684012A (en) Equipment key force-bearing structural part fault diagnosis method based on multi-parameter information fusion
CN111881574A (en) Wind turbine generator key component reliability modeling method based on distribution function optimization
Lopes et al. Neural network models for the subjective and objective assessment of a propeller aircraft interior sound quality
Sallehuddin et al. Forecasting small data set using hybrid cooperative feature selection
Ferguson et al. A data processing pipeline for prediction of milling machine tool condition from raw sensor data
Vaxevanidis et al. 5. FEM Analysis and ANN Modeling for Optimizing Machinability Indicators during Dry Longitudinal Turning of Ti–6Al–4V ELI Alloy
Kundu et al. PCA-ANN Based Approach for Remaining Useful Life Prediction for Roller Ball Bearings
CN112247674B (en) Cutter wear prediction method
Gao Statistical Learning for Lightweight Materials Manufacturing
Rai et al. Development of Predictive Model for Surface Roughness Using Artificial Neural Networks
Wang et al. Tool Wear Volume and Residual Life Prediction Based on Extreme Learning Machine
Xu et al. Analyzing the Solution of Chemotaxis Equations with Logistic Source Term
Griffin et al. Application of machine learning for acoustic emissions waveform to classify galling wear on sheet metal stamping tools

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant