Grinding state online monitoring method based on multilayer data fusion and mathematical combined driving model

Abstract

The invention discloses a grinding state on-line monitoring method based on multilayer data fusion and a mathematical combined driving model, and the method comprises the steps: carrying out the signal collection through a multi-source sensor, carrying out the fusion according to a corresponding weight, and carrying out the noise reduction through stacking self-coding. When signals are extracted, frequency domain multi-dimensional features are fed into the residual neural network and the bidirectional double-layer LSTM fusion model after being subjected to Pearson correlation screening, and single-working-condition grinding state recognition is achieved. 10 types of discrete grinding states are defined, and a physical driving type empirical formula for working condition and monitoring signals and working condition and state characterization is established based on working condition parameters such as grinding depth and main shaft rotating speed. And generating a training data set by using an empirical formula, and training a cross-working-condition monitoring model through a transfer learning mechanism to realize cross-working-condition identification. According to the method, online accurate recognition of the grinding state is achieved, the limitation of traditional single-working-condition monitoring is broken through, and the problem of repeated data collection training is solved through the physical law of an empirical formula.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent monitoring of tool state in precision grinding, and particularly relates to a grinding state online monitoring method based on a multi-layer data fusion and mathematical joint driving model. BACKGROUND

[0002] Precision grinding technology is widely used in high-end manufacturing fields such as aerospace, mechanical and electronic, and optical devices. In the machining process, there is a very high requirement for the surface roughness of the workpiece, generally in the micro-nano level, so it is necessary to detect the state of the grinding wheel and the workpiece in time. The existing detection method is generally offline in situ using a magnifying glass and a roughness meter for detection. According to statistics, the detection of the grinding state causes the machine tool downtime to account for 20%-40% of the total downtime, and there is a problem of not timely detection, which seriously affects the machining precision and efficiency, and it is urgent to perform online monitoring of the grinding state. The material removal amount of precision grinding is small, and the effective signal is submerged by strong background interference such as spindle frequency conversion noise and cooling liquid impact, so it is difficult to extract the effective features in the time and frequency domains, and the existing monitoring method has problems such as low accuracy and poor real-time performance. For the actual machining process, the grinding conditions change rapidly, and the mapping relationship of the monitoring signal is single, so the model is only applicable to single-condition monitoring. When cross-condition machining is performed, a large amount of data needs to be collected to train the model before accurate monitoring can be achieved, which wastes a lot of manpower and resources.

[0003] In summary, it has become a problem to be solved to realize precise and adaptive online monitoring of the grinding state in the precision grinding process. SUMMARY

[0004] In view of the difficulty in monitoring the grinding state of precision grinding, the low efficiency and poor accuracy of the existing monitoring method, and the difficulty in realizing cross-condition online real-time monitoring, the application provides a grinding state online monitoring method based on a multi-layer data fusion and mathematical joint driving model, which realizes precise and adaptive online monitoring of the grinding state by establishing a grinding state online monitoring model driven by grinding data mechanism.

[0005] To achieve the above purpose, the application provides the following scheme:

[0006] A grinding state online monitoring method based on a multi-layer data fusion and mathematical joint driving model, the method comprising:

[0007] Signals of the grinding wheel and the workpiece to be ground are collected by a multi-source sensor and fused according to corresponding weights, and noise reduction processing is performed by stacked auto-encoding;

[0008] Multi-dimensional features in the time and frequency domains of the fused and noise-reduced signals are extracted, and after being screened by Pearson correlation, a structured feature data set is obtained;

[0009] The structured feature dataset is fed into a residual neural network ResNet and a bidirectional double-layer LSTM fusion model to realize single-condition grinding state recognition.

[0010] The grinding states of the preset grinding wheel and the workpiece to be ground are discrete, and variable grinding parameter experiments are performed on the discrete processes of the two grinding states to obtain the changes of the monitoring signals and state representations under different parameters.

[0011] The grinding states of the preset grinding wheel and the workpiece to be ground are discrete, and variable grinding parameter experiments are performed on the discrete processes of the two grinding states to obtain the changes of the monitoring signals and state representations under different parameters.

[0012] Physical driving type empirical formulas are established based on the grinding shaft speed, feed speed and grinding depth working condition parameters, and the working conditions and monitoring signals, and the working conditions and state representations.

[0013] The training dataset is generated using the physical driving type empirical formulas, and the ResNet-SBiLSTM monitoring model is trained through the transfer learning mechanism to realize cross-condition recognition.

[0014] Preferably, the collected signals are fused according to the corresponding weights, and a double noise filtering algorithm is designed, which specifically includes:

[0015] The signal sensitivity ranking is obtained through grinding experiments: acoustic emission > vibration > current, and the linear fusion signal R' is obtained by fusing the signals according to the weight proportions of 60%, 30% and 10%, and the fusion signal is obtained by activating R'.

[0016] Each signal is analyzed in the time domain. First, the data is decomposed by WOA-VMD to obtain different intrinsic mode functions IMF. The components with different grinding state frequency spectrum differences are superimposed and reconstructed in the frequency spectrum, and the reconstructed data is obtained through inverse Fourier transform. Second, the data filtered by WOA-VMD is further deep feature extracted, and a stacked auto-encoding network model is built for deep feature extraction and representation learning, realizing signal enhancement and noise reduction.

[0017] Preferably, the extracted and fused and denoised signal time-frequency domain multi-dimensional features are screened by Pearson correlation to obtain a structured feature dataset.

[0018] The AE, vibration, current and fused signals are analyzed in the time domain, frequency domain and time-frequency domain, and multi-dimensional feature vectors are extracted.

[0019] The multi-dimensional feature vectors are deeply fused at the feature level to construct a feature space representing the grinding state.

[0020] The Pearson correlation coefficient PCC and the autocorrelation function ACF are used for feature screening, and finally a structured feature dataset is constructed through a sliding time window mechanism.

[0021] Preferably, the structured feature dataset is fed into a residual neural network ResNet and SBiLSTM fusion model to realize a single-condition grinding state recognition method, which comprises the following steps:

[0022] The STFT image format of the sliding window is converted into 299*299*3 as the model input; first, a 2D-CNN convolution layer is used to extract low-level spatial features, and after the convolution operation, the spatial size of the output feature map is:

[0023]

[0024] In the formula: K H and K W are the height and width of the convolution kernel, the input image size is HxW, H is the height of the image, W is the width of the image, the convolution kernel size is K H x K W , the step sizes in the height and width directions are S H and S W , the padding is P H and P W , H out and W out are the height and width of the output feature map, respectively.

[0025] Max pooling MaxPooling is introduced after the first layer convolution operation in each modal branch, and its calculation expression is as follows:

[0026]

[0027] u j i = σ(X j i );

[0028]

[0029] In the formula: i represents the index of the feature layer, j represents the index of the feature map, represents the jth feature map of the ith layer, x k i-1 represents the kth output feature map of the previous layer, N i-1 represents the total number of output feature maps of the i-1th layer, W k j and b j i are the convolution kernel weight matrix and the bias, respectively, σ is the ReLU nonlinear activation function, and uj i is the output after activation, is the output after pooling;

[0030] Data feature values X obtained by Pearson correlation coefficient analysis i =[X1,X2,X3,...,X N ] are input into

[0031] SBiLSTM, the network processes the input data time sequence features from two paths of "forward" and "reverse", the first layer forward and reverse path outputs are and The calculation formula of the first layer output is as follows:

[0032]

[0033] In the formula: is the output of the i-th LSTM gating unit in the forward direction, is the output of the i-th LSTM gating unit in the reverse direction;

[0034] The first layer output H i (1) is input into the second layer in the form of a sequence, and the output H i (2) of the second layer is as follows:

[0035]

[0036] In the formula: X i is the feature data after Pearson correlation analysis, f1(·), f2(·) are the SBiLSTM network hidden layer functions, and are described by a parameter set , is the weight and bias parameter of the first layer, is the weight and bias parameter of the second layer, H i (2) is the second layer output data.

[0037] Preferably, the grinding states of the preset grinding wheel and the workpiece to be ground are discrete, and variable grinding parameter experiments are performed on the discrete processes of the two grinding states, respectively, and the method for obtaining the changes of the monitoring signals and state representations under different parameters includes:

[0038] The grinding state is defined as a discrete space, wherein the grinding wheel wear grade is P, P [P1, P2, P3...P 10 ], and the workpiece surface roughness grade is Q, Q [Q1, Q2, Q3...Q 10]rpm, feed speed v∈[100, 300]mm / min, grinding depth a p ∈[5, 25]μm, record the changes of monitoring signals and detection characteristics under different parameters.

[0039] Preferably, the method for establishing the physical driving type empirical formula of the working condition and the monitoring signal, the working condition and the state characteristic based on the working condition parameters of the grinding shaft speed, the feed speed and the grinding depth includes:

[0040] The effective variation range of each variable is determined by single factor experiment, wherein each variable is the grinding shaft speed, the feed speed and the grinding depth;

[0041] Based on the response surface method of the central composite design, 1250 groups of experiments required by the full factor experiment are simplified into 200 groups of experiments; wherein the experimental levels of each variable are converted according to the coding formula, and the calculation formula is:

[0042]

[0043] The coding levels include five levels: the shaft low point level-α=-1.68, the cubic low point level-1, the center point 0, the cubic high point level+1, and the shaft low point level+α=+1.68; wherein the cubic point is combined with 8 coding coordinates of ±1, the shaft point is combined with (±α, 0, 0) (0, ±α, 0) (0, 0, ±α), and the center point is combined with 6 times of coding coordinates of (0, 0, 0);

[0044] The empirical formula of the physical driving model two between the working condition parameters and the label data set through the cross-working condition experiment is as follows:

[0045]

[0046] In the formula: P, Q are the grinding wheel wear and the workpiece surface roughness respectively, β0, is the model constant term, β1, β2, β3, is the first order coefficient, the square term coefficient of the model is β 11 , β 22 , β 33 , β 12 , β 13 , β 23 , is the interaction coefficient, ε2, ε2 (2) is the random error;

[0047] The empirical formula of the physical driving model two between the working condition parameters and the signal feature set through the cross-working condition experiment is as follows:

[0048]

[0049] wherein: F=[RMS, Spectral energy, Kurtosis, Variance] are the cross-condition signal feature vector data, ξ0is the model constant term, ξ1, ξ2, ξ3are the first-order coefficients, ξ 11 , ξ 22 , ξ 33 are the square term coefficients, ξ 12 , ξ 13 , ξ 23 are the interaction term coefficients, is the random error.

[0050] Preferably, the method for generating the training data set by using the physical driving empirical formula comprises:

[0051] Perform a cross-condition machining experiment by dynamically adjusting the process parameter vector G=[G1, G2, G3...G k ] and inputting it to the physical empirical model in step two:

[0052] G→(F', W', R');

[0053] wherein: (F', W', R') are the sensor signal feature set and the label data set generated by the model;

[0054] Adaptively obtain the signal feature vector under the current condition. Since the output of the empirical model is n discrete state points in a 4-dimensional feature space (n=10), and the deep learning model needs to input small sample continuous data, data augmentation is performed. First, the feature space is expanded, wherein RMS in F represents the time domain root mean square value, spectral energy represents the spectral energy, kurtosis represents the kurtosis, and variance represents the variance; for the generated F' which is a 10*4 feature vector, for the data generated for the 10 discrete points, uniformly sample in its neighborhood N μ (F) wherein μ=0.5, to generate F i =2000*4 random vectors;

[0055] F' μ =F'+ΔF, ΔF~U([-μ,+μ]);

[0056] wherein: ΔF is the generated random disturbance data, U is the random disturbance vector, and μ is the disturbance interval;

[0057] Continuously fit the label data set (W', R') and construct the continuous curves W' (g) and R' (g) of the conditions of the parameters by using the cubic spline interpolation method.

[0058] (W (g) ',R (g) )=a i +b i (g-g i )+c i (g-g i ) 2 +d i (g-g i ) 3 ;

[0059] wherein: W (g) ' and R (g) are continuous curves of the grinding wheel wear and the workpiece surface roughness varying with the working condition, g is the working condition parameter independent variable, g i is the i-th working condition parameter, i is the index value, a i , b i , c i , d i are constant term, first-order term, second-order term, third-order term coefficients respectively.

[0060] Preferably, the method for realizing cross-condition recognition by training the cross-condition monitoring model through the transfer learning mechanism comprises:

[0061] The enhanced data set F μ ' generated by the physical experience model is input as the time-frequency feature value, (W g ', R g ') is the response output, and is input into the ResNet-SBiLSTM model, and the network grinding is improved through the following transfer learning strategy. Firstly, the source domain is pre-trained, the entire network is trained using the source domain big data, and a set of good initialization parameters is obtained, and the formula is as follows:

[0062]

[0063] wherein: θ1 represents the feature extraction layer parameter, θ2 represents the classification layer parameter, D source represents the source domain data set, and L loss represents the loss function.

[0064] After the source domain is trained, the target domain data generated by the experience formula is used to fine-tune the model, the feature extraction layer parameter θ1 is fixed, and only the classification layer parameter θ2 is fine-tuned. In the process of back propagation, the gradient is blocked in the feature extraction layer, and the parameter is not updated. The loss function includes cross-entropy loss and regularization term, and the calculation formula is as follows:

[0065]

[0066] wherein: L lossL represents the total loss function of the fine-tuning process CE L represents the cross-entropy loss function, and X represents the target domain sample target , Y represents the output of the model to the target domain sample target Y represents the target domain label; lambda is a regularization coefficient, and ||theta2|| 2 is a regularization constraint.

[0067] Compared with the prior art, the beneficial effects of the present application are:

[0068] 1. The present application obtains multi-source information data through multi-sensor data and signal layer data fusion processing; realizes deep filtering of noise signals by using WOA-VMD and stacked autoencoder network double noise filtering processing; builds a deep fusion network model and fuses in the decision layer, and the single working condition monitoring accuracy reaches more than 95%. In order to solve the problem of cross-condition adaptive monitoring, the mapping relationship among data-grinding wheel-workpiece is obtained through response surface experiment, and a physical model is established, a transfer learning mechanism is introduced to train the model, and cross-domain monitoring is realized through small sample data.

[0069] 2. The present application is highly integrated, simple to operate, and the entire hardware is integrated in the control box, easy to carry. The system software is built-in small window on the control box. When using, first arrange the sensor in different processing areas, open the system software to realize automatic online monitoring processing, and the monitoring accuracy is more than 95%.

[0070] 3. The present application can realize adaptive cross-condition processing monitoring, avoid self-learning setting before cross-condition use, the system is built-in mapping relationship theoretical model, small sample training data is automatically generated and input into the transfer learning model, and the model is automatically trained for monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0071] In order to more clearly illustrate the technical scheme of the present application, the following briefly introduces the drawings needed to be used in the embodiments, obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0072] Figure 1 It is a process diagram of the multi-source information fusion and transfer learning adaptive cross-condition monitoring system of the embodiment of the present application.

[0073] Figure 2 It is a signal layer fusion and double noise filtering algorithm diagram of the embodiment of the present application.

[0074] Figure 3 It is a visualization result diagram of t-SNE before and after data noise filtering of the embodiment of the present application.

[0075] Figure 4 A ResNet-SBiLSTM deep neural network hierarchical architecture diagram of an embodiment of the present application;

[0076] Figure 5 An online monitoring result diagram of a single working condition of an embodiment of the present application;

[0077] Figure 6 A response surface experimental design diagram of an embodiment of the present application;

[0078] Figure 7 An online monitoring result diagram of a cross-working condition of an embodiment of the present application. DETAILED DESCRIPTION

[0079] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0080] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0081] Embodiment one

[0082] The present application provides a grinding state online monitoring method based on multi-layer data fusion and mathematical joint driving model. The method collects data through multi-source sensors, combines multi-layer data fusion processing and deep neural network architecture, realizes data-driven grinding state single working condition accurate monitoring, and the accuracy is more than 95%.

[0083] When the grinding parameters change, the corresponding mapping relationship between the signal and the grinding state changes, and the data training model needs to be artificially collected again. This process has the defects of large redundant workload, time-consuming and laborious. Therefore, a physical model of the working condition parameters, the monitoring signal and the grinding state representation is established, the training data set of different working conditions is generated by using the model, and the transfer learning mechanism is introduced to realize network retraining with a small amount of generated data, so as to construct a cross-working condition self-adaptive monitoring method driven by data and mechanism. The specific implementation process is as follows:

[0084] As shown in Figures 1-7 The present application provides a grinding state online monitoring method based on multi-layer data fusion and mathematical joint driving model. The method comprises:

[0085] The signal acquisition of the grinding wheel grinding workpiece is performed through a multi-source sensor, and a signal three-layer fusion technology is innovatively proposed. The multi-source sensor signals are fused in the signal layer, the feature layer and the decision layer respectively, so that the internal connection of different data can be effectively captured, the information complementarity is enhanced, and the decision robustness of the monitoring system is improved.

[0086] A double noise filtering algorithm is developed. The signals [S, Z, D] directly collected by the sensor and the fusion signal R data are affected by a large amount of environmental noise such as grinding shaft rotation and grinding fluid spraying. Therefore, the signals are decomposed by using a variational mode decomposition, and a whale optimization algorithm is introduced to determine the number of decomposition layers and the penalty factor, so that the conditions of over-decomposition and under-decomposition are avoided, and the signal-to-noise ratio is improved by 18.3 dB. The data after preliminary noise reduction is output to a deep stack autoencoder network again to extract the internal connection of the data and perform data enhancement processing.

[0087] Multi-dimensional features in the time domain, the frequency domain and the time-frequency domain of the signals after double noise reduction are extracted, and a structured feature data set is obtained after screening by the Pearson correlation;

[0088] The structured feature image data set and the feature vector data set are respectively fed into a residual neural network ResNet and a SBiLSTM fusion model for parallel training, the features extracted by each module are fused, and single-condition grinding state recognition is realized;

[0089] Discrete grinding states of a preset grinding wheel and a workpiece to be ground are obtained, and variable grinding parameter experiments are performed on the discrete processes of the two grinding states, so as to obtain the changes of the monitoring signals and the state representations under different parameters;

[0090] Physical driven empirical formulas of the working conditions and the monitoring signals and the working conditions and the state representations are established based on the grinding shaft speed, the feed speed and the grinding depth;

[0091] The training data set is generated by using the physical driven empirical formulas, the ResNet-SBiLSTM monitoring model is trained by using a transfer learning mechanism, and cross-condition recognition is realized.

[0092] In the embodiment, the data-driven grinding state single-condition monitoring method comprises the following steps:

[0093] Step one, multi-modal sensor data synchronous acquisition, the acoustic emission (AE) signal S, the vibration signal Z and the current signal D are synchronously acquired by using a high sampling rate data acquisition system, and the sampling rate fs is set to 500 kHz. The high sampling rate acquisition system (including a multi-source sensor, an amplifier, an acquisition card and the like) ensures that the vibration signal of the low-frequency vibration mode (<10 kHz) of the main shaft, the high-frequency elastic wave signal (>200 kHz) of the material fracture / dislocation, and the dynamic power characteristic current signal closely related to the output torque of the main shaft are effectively captured.

[0094] Step two, signal layer fusion and deep feature learning, the multi-sensor data is fused synchronously at the original signal layer to construct the fusion signal R. A dual filtering denoising method is designed, WOA-VMD filtering technology is used to preprocess the original data [S, Z, D, R] of different grinding states, and a stacked autoencoder network is designed to extract and learn the secondary deep features of the preprocessed data, realize signal enhancement and denoising, and update the original data set accordingly.

[0095] Step three, high-dimensional feature dimension reduction and data set construction, the data after dual denoising processing is analyzed in time domain-frequency domain-time-frequency domain, and multi-dimensional feature vectors are extracted. The heterogeneous features of multi-source sensors are fused at the feature layer to construct a feature space representing the grinding state. Pearson correlation coefficient (PCC) and autocorrelation function (ACF) are used for feature selection, and features with PCC>0.6 are selected to represent strong correlation, and high-redundancy features with ACF>0.9 are removed. Finally, a structured feature data set is constructed through a sliding time window mechanism.

[0096] Step four, multi-source decision fusion model construction, a multi-modal fusion ResNet-SBiLSTM neural network model is built. This model supports adaptive fusion of heterogeneous data at different levels, such as feature extraction layer and intermediate decision layer. The model is designed as a multi-target output structure to realize high-precision real-time monitoring of key states such as grinding wheel wear state and workpiece surface quality at any processing time during grinding.

[0097] Step five, model performance verification and evaluation, multi-condition experiments are carried out on a precision grinding machining platform. The real-time prediction accuracy, robustness and generalization ability of the constructed model for key indicators of grinding state, i.e. grinding wheel wear amount and workpiece surface roughness, in specific single-condition scenarios are analyzed, and its actual monitoring effectiveness in industrial environment is verified.

[0098] In step two, the data of each sensor is fused at the signal layer. Through grinding experiments, the sensitivity ranking is acoustic emission > vibration > current, so the linear fusion signal R' is obtained by fusing them with weight proportions of 60%, 30% and 10%. Since there are nonlinear coupling effects in the signal, the fusion signal R is obtained by activating R', and the calculation formula is as follows:

[0099] R' = 0.6 * S + 0.3 * Z + 0.1 * D;

[0100]

[0101] The time domain analysis of each signal shows that the data is seriously contaminated by noise and the fusion signal is also covered by a large amount of redundant information, so noise reduction processing is needed. The original data (data collected by multi-source sensors and reconstructed data, both of which are preliminarily filtered and twice filtered) is decomposed to obtain different intrinsic mode functions (IMFs), the changes of each IMF component of different grinding states are compared, the components with different spectrum changes of different grinding states are superimposed in the spectrum, and the reconstructed data [S, Z, D, R]' is obtained through inverse Fourier transformation, and the calculation formula is as follows:

[0102]

[0103] In the formula: δ(t) represents the Dirichlet function, u k represents the k IMF components obtained by decomposition, w k represents the center frequency of each component, is the time derivative operator, which represents the derivative with respect to time t. K is the number of decomposition modes, that is, the number of IMFs to be extracted. t is the time variable. j is the imaginary unit, f is the bandwidth of the original signal, u k (t) is the function u k at time t.

[0104] By introducing the Lagrange multiplier operator and the quadratic penalty factor λ(t), the constrained variational problem can be converted into an unconstrained variational problem. The augmented Lagrange function is as follows:

[0105]

[0106] In the formula: f(t) is the input signal, and λ(t) is the Lagrange multiplier.

[0107] In order to avoid over-decomposition and under-decomposition, the whale optimization algorithm is introduced to obtain the decomposition layer number K of VMD and the penalty factor α through iteration, and the main calculation formula is as follows:

[0108]

[0109] In the formula: X*(t) is the position after iteration, X(t) is the original position, D is the distance vector of the current individual and the optimal individual, and c and A are coefficient vectors.

[0110] The calculation formula of its main parameters is as follows:

[0111]

[0112] In the formula: r1 and r2 are random numbers between 0 and 1, t represents the current iteration number, T max represents the maximum number of iterations, and a is the attenuation coefficient.

[0113] The data after the WOA-VMD preliminary noise filtering is extracted again, and the stacked auto-encoder network model is built to reconstruct the deep signal features. The minimum reconstruction error is calculated as follows:

[0114]

[0115] In the formula: n is the number of training samples, R i and R i are the original data and the data after reconstruction respectively, λ is the weight attenuation term, n l is the number of network layers, s l and s l+1 are the number of neurons of the lth and l+1th layers of the neural network, w ji is the weight vector between the lth and l+1th layers, w and b are the weight and bias parameters of the neural network.

[0116] The input layer and the hidden layer of each auto-encoder are connected, and the minimum envelope entropy function of the reconstructed signal is selected as the loss function for backward fine-tuning parameters, to ensure that the reconstructed signal retains the main features of the original signal and has strong regularity. The weight and bias parameter update formula is as follows:

[0117]

[0118] In the formula: L is the loss function, L is the partial derivative, η is the learning rate, w1 and b1 are the weight and bias parameters of the first layer. h, h 1 , h 2 are the data after three-layer encoding, f1, f2, f3 are the corresponding encoding functions, x', x' 1 , x' 2 are the data after three-layer decoding, g1, g2, g3 are the corresponding decoding functions.

[0119] In step three, in order to further quantify the relationship between the online monitoring signal and the grinding state, the time domain analysis method is introduced. The statistical characteristics of the time domain signal are analyzed, and the mean, root mean square, standard deviation, peak-to-peak value, skewness, and kurtosis are calculated. The frequency spectrum of the time domain signal is obtained by fast Fourier transform, and the center frequency, mean square frequency, frequency variance, spectral width, and spectral energy are calculated. The time-frequency diagram is drawn by short-time Fourier transform of the time domain signal, and the variation law of the signal with the workpiece surface roughness is obtained. The sliding window is divided for 11 time-frequency characteristic values and STFT time-frequency diagram of multi-channel signal. The length of the signal is L1, and the characteristic calculation time window is L2. The calculation formula of the length of the characteristic vector l and the number of characteristic pictures m is as follows:

[0120]

[0121] The autocorrelation between the features and the cross-correlation with the classification results are screened by using the Pearson correlation coefficient, which can speed up the operation efficiency of the model. The core formula is as follows:

[0122]

[0123] In the formula: r is the Pearson correlation coefficient, and is the sample mean, x i and y i represent different feature value samples, and N is the sample point number.

[0124] In step four, the residual neural network and the long short-term memory neural network are combined to enhance the spatiotemporal feature extraction capability of the model. The STFT image format of the sliding window is converted to 299*299*3 as the input of the model. First, a 2D-CNN convolution layer is used to extract low-level spatial features. After convolution operation, the spatial size of the output feature map is:

[0125]

[0126] In the formula: K H , K W are the height and width of the convolution kernel, the input image size is HxW, H is the height of the image, W is the width of the image, the size of the convolution kernel is K H x K W , the step size in the height and width directions is S H and S W , the padding is P H and P W , and H out , W out are the height and width of the output feature map, respectively.

[0127] MaxPooling is introduced after the first layer convolution operation in each modal branch, which can effectively reduce the feature map resolution while retaining important response areas. The calculation expression is as follows:

[0128]

[0129] u j i =σ(X j i );

[0130]

[0131] In the formula: i represents the index of the feature layer, j represents the index of the feature map, represents the jth feature map of the ith layer, and xk i-1 the kth output feature map representing the previous layer, N i-1 the total number of feature maps output by the i-1th layer, W k j and b j i respectively the convolution kernel weight matrix and the bias, and is a ReLU nonlinear activation function, u j i is the output after activation, is the output after pooling.

[0132] Data feature values X obtained by Pearson correlation coefficient analysis i = [X1, X2, X3,..., X N ] are input into

[0133] SBiLSTM, the network processes the input data time sequence features from "forward" and "reverse" two paths, and the first layer forward and reverse path outputs are and The calculation formula of the first layer output H i (1) is as follows:

[0134]

[0135] In the formula: is the output of the i-th LSTM gate unit in the forward direction, is the output of the i-th LSTM gate unit in the reverse direction.

[0136] The first layer output H i (1) is input into the second layer in the form of a sequence, and the output H i (2) of the second layer is calculated as follows:

[0137]

[0138] In the formula: X i is the feature data after Pearson correlation analysis, f1(·), f2(·) are the SBiLSTM network hidden layer functions, and are described by the parameter set , is the weight and bias parameter of the first layer, is the weight and bias parameter of the second layer, H i (2) is the second layer output data.

[0139] In step five, the convergence speed and loss value of the model are used as evaluation indexes to determine the hyperparameters of the model. The optimization algorithm selects Adam algorithm, the batch size is 20, the learning rate is 0.05, and the Dropout value is 0.2.

[0140] In order to verify the accuracy of model identification, the absolute mean error (MAE), root mean square error (RMSE) and determination coefficient R 2 are compared, and the calculation formula is as follows:

[0141]

[0142] In the formula: l is the number of characteristic value samples, y i is the actual value, is the model prediction value.

[0143] By collecting real-time signals of model processing, the MAE of model identification grinding state is 4.6%, the RMSE is 5.4%, and the R 2 reaches 98%.

[0144] In this embodiment, the cross-condition adaptive monitoring method based on data and mechanism combined driving contains the following steps:

[0145] Step one, multi-parameter discrete state modeling, the grinding state is defined as a discrete space, where the grinding wheel wear grade is P, P ∈ [P1, P2, P3... P 10 ], the workpiece surface roughness grade is Q, Q ∈ [Q1, Q2, Q3... Q 10 ], the discrete process of two kinds of grinding states is respectively carried out for variable grinding parameter experiment, the spindle speed n ∈ [8000, 15000] rpm, the feed speed v ∈ [100, 300] mm / min, the grinding depth a p ∈ [5, 25] μm, the changes of monitoring signals and state characteristics under different parameters are recorded.

[0146] Step two, fusion of physical mechanism and empirical model, based on central composite experimental design to construct high-efficiency experimental matrix, for three grinding parameter factors, each factor has 5 levels, then 125 grinding experiments are needed, a total of 10 discrete states, 1250 experiments are needed, which requires large amount of work, high cost and long period. Through central composite experiment to reduce the number of experiments, including 2 k full factorial cubic points, 2k axial points, located on each factor axis at a distance of α = ± 1.682 from the center point, n0 center points, then the calculation formula of the total number of experiments n is as follows:

[0147] n = 2 k + 2k + n0

[0148] Wherein, the factor number k = 3, the center repeats n0 = 6, then the total number of experiments is reduced to 200 times, the second order response surface model is used to establish the physical mapping model one between the working condition parameters and the root mean square value, spectral energy, kurtosis, variance of each sensor signal, and the physical mapping model two between the working condition parameters grinding wheel wear state P and the workpiece surface roughness Q (the specific physical mapping model is mentioned in the specific explanation step two, that is, the three inputs of grinding depth, grinding shaft speed and feed speed, and the output is the model of signal feature set or workpiece state label set).

[0149] Step three, cross-condition data enhancement and generalization, cross-condition experimental processing is carried out, and the physical mapping model described in step two is used to establish 10 kinds of discrete state working condition parameters G = [n, v, a p ]n is the grinding shaft speed, v is the workpiece feed speed, a p is the mapping empirical formula between grinding depth and sensor signal feature set F = [RMS, Spectral energy, Kurtosis, Variance] and label set (P, Q). When cross-condition processing is carried out, the working condition parameters G are changed as independent variables, and the sensor signal feature set F and the label set (W, R) are adaptively obtained from the empirical formula. 10 groups of discrete state data generated according to the formula are interpolated and fitted in a small range, and each group of data is expanded into 2000 sample points.

[0150] Step four, transfer learning training of ResNet-SBiLSTM network model, freezing the feature extraction layer training parameters of the data processing network, using the small sample data and labels generated by the empirical formula, training the end full connection output layer, taking the minimum loss function as the target to continuously fine-tune the output layer parameters, and realizing the adaptive training of the network model.

[0151]

[0152] In the formula: L CE is the cross-entropy loss, and λ is the regularization coefficient of L2.

[0153] Step five, model verification and adaptive performance evaluation, continuous switching of different grinding conditions of grinding experiment, identification and monitoring through the model, in the experiment, the grinding force and the workpiece surface roughness are measured to measure the grinding wheel wear state, the actual detection value is compared with the model prediction value, and the error is 7.8%.

[0154] In step one, the time-varying curves of grinding wheel wear P(t) and workpiece surface roughness Q(t) are equally spaced and sampled, and N = 10 discrete time points t i(i = 1, 2, ..., 10) corresponding to state data points P and Q, under the premise of keeping the key state parameters of the grinding wheel relatively constant, with a controllable working condition parameter vector G = [G1, G2, G3...G... k As input variables, the corresponding output target vector (P, Q) is monitored and recorded, where P represents the wear state of the grinding wheel and Q represents the surface roughness state of the workpiece. That is, the states of the grinding wheel and workpiece are decomposed into 10 discrete points. In each discrete state, an empirical formula is established using the working condition as the variable and the signal characteristics and the wear state caused by the working condition as the output.

[0155] In step two, the operating parameters include three independent variables: spindle speed, feed rate, and depth of grinding. First, a single-factor experiment is conducted to determine the monitoring range of the grinding wheel and workpiece conditions, thus defining the variation range of the operating parameters: spindle speed n ∈ [8000, 15000] rpm, feed rate v ∈ [100, 300] mm / min, and depth of grinding a. p ∈[5,25]μm; Based on the response surface methodology of central composite design, the 125 groups (5) required for full factorial experiments are used. 3 The experiment was simplified to 20 groups; the experimental levels of each variable were converted according to the following coding formula:

[0156]

[0157] The coding level comprises five levels: axis low point level -α = -1.68, cube low point level -1, center point 0, cube high point level +1, and axis low point level +α = +1.68. Specifically, the cube points undergo eight combinations of coding coordinates with ±1, the axis points undergo experimental combinations of (±α, 0, 0), (0, ±α, 0), and (0, 0, ±α), and the center point repeats the combination of coding coordinates (0, 0, 0) six times. Therefore, the optimized number of experiments is 200.

[0158] The empirical formula for the second physical-driven model based on the cross-condition experimental parameters and the labeled dataset is as follows:

[0159]

[0160] In the formula: P and Q are the grinding wheel wear and workpiece surface roughness, respectively, β0, For the model constants, β1, β2, β3, These are the coefficients of the linear term, and the coefficients of the squared term in the model are β. 11 β 22 β 33 , β 12 β 13 β 23 , is the interaction term coefficient, ε2, ε2 (2) is the random error.

[0161] The empirical formula of the second physical driving model between the operating condition parameters and the signal feature set is as follows:

[0162]

[0163] In the formula: F = [RMS, Spectral energy, Kurtosis, Variance] are the cross-condition signal feature vector data, ξ0is the model constant term, ξ1, ξ2, and ξ3are the first-order term coefficients, ξ 11 , ξ 22 , and ξ 33 are the square term coefficients, ξ 12 , ξ 13 , and ξ 23 are the interaction term coefficients, is the random error.

[0164] In step three, a cross-condition processing experiment is performed, and the process parameter vector G = [G1, G2, G3...G k ] is dynamically adjusted and input to the physical empirical model in step two:

[0165] G→(F', W', R')

[0166] In the formula: (F', W', R') is the sensor signal feature set and the label data set generated by the model.

[0167] The signal feature vector under the current operating condition is adaptively obtained. Since the output of the empirical model is n discrete state points in a 4-dimensional feature space (n = 10), and the deep learning model needs small sample continuous data input, data enhancement is performed. First, the feature space is expanded, and for each state point F = [RMS, Spectral energy, Kurtosis, Variance], in which RMS in F represents the time domain root mean square value, spectral energy represents the frequency spectrum energy, kurtosis represents the kurtosis, and variance represents the variance. For the generated F' which is a 10*4 feature vector, for the data generated for the 10 discrete points, uniform sampling is performed in its neighborhood N μ (F), where μ = 0.5, to generate F i = 2000*4 random vectors.

[0168] F' μ = F' + ΔF, ΔF ~ U([-μ, +μ]

[0169] In the formula: ΔF is the generated random disturbance data, U is the random disturbance vector, and μ is the disturbance interval.

[0170] The continuous fitting is performed on the label data set (W', R'), and a continuous curve W' of the parameters of the working condition is constructed by using the method of cubic spline interpolation (g) and R' (g) .

[0171] (W' (g) ,R' (g) )=a i +b i (g-g i )+c i (g-g i ) 2 +d i (g-g i ) 3

[0172] In the formula: W' (g) and R' (g) are continuous curves of the wheel wear and the workpiece surface roughness with the change of the working condition, g is the working condition parameter independent variable, g i is the i-th working condition parameter, i is the index value, a i , b i , c i , d i are constant term, first-order term, second-order term, and third-order term coefficients respectively.

[0173] In step four, the enhanced data set F' μ generated by the physical empirical model is input as the time-frequency feature value, (W g ', R g ') is the response output, and is input into the ResNet-SBiLSTM model. The network grinding is improved by the following transfer learning strategy. First, the source domain is pre-trained, the entire network is trained using the source domain big data, and a set of good initialization parameters is obtained, and the formula is as follows:

[0174]

[0175] In the formula: θ1 represents the feature extraction layer parameter, θ2 represents the classification layer parameter, D source represents the source domain data set, and L loss represents the loss function.

[0176] After the source domain is trained, the target domain data generated by the empirical formula is used to fine-tune the model. The feature extraction layer parameter θ1 is fixed, and only the classification layer parameter θ2 is fine-tuned. In the process of back propagation, the gradient is blocked in the feature extraction layer, and the parameter is not updated. The loss function includes cross-entropy loss and regularization term, and the calculation formula is as follows:

[0177]

[0178] In the formula: L' loss represents the total loss function of the fine-tuning process, L CE represents a cross-entropy loss function, and the target domain sample is X target , represents the output of the model to the target domain sample, Y target represents the target domain label. Lambda is a regularization coefficient, and ||theta2|| is a regularization constraint. 2 is a regularization constraint.

[0179] In step five, a dynamic working condition switching verification experiment is carried out, and multi-modal real-time monitoring is performed. On the basis of changing 10 groups of grinding working conditions, the online monitoring system outputs a grinding wheel-workpiece state vector in real time wherein, is the grinding wheel wear and workpiece surface roughness monitored across working conditions, and the sampling frequency fs = 500 kHz. After the monitoring is completed, the machine tool is stopped, and a Fein roughness instrument and a telecentric lens are used to detect the state of the grinding wheel and the workpiece to obtain wherein, is the grinding wheel wear and workpiece surface roughness measured by the detection instrument. The normalized root mean square error of the model prediction and the actual detection is calculated. Experimental data show that the maximum error is 7.8%, which meets the processing requirements and improves the processing efficiency. Compared with offline detection, which requires stopping for about 30 minutes per time, the system realizes continuous monitoring in less than 5 minutes, improving the production efficiency by 50-60%.

[0180] In this embodiment, an online grinding state monitoring system is established, a system data acquisition box is designed, and a real-time acquisition and analysis software interface is developed, avoiding the previous mode of online acquisition and offline data processing, and realizing real-time monitoring and analysis of grinding processing. The system includes single working condition recognition and cross-working condition recognition functions, and internally has a physical mapping model between working conditions, signals and data. Inputting the current working condition data, the system can automatically generate training samples for cross-working condition recognition, thereby realizing accurate monitoring of the grinding state. Before processing, the operator can preset the state threshold of the grinding wheel and the workpiece. When the system identifies that the current state exceeds the threshold, it will send a signal to the machine tool numerical control system through the communication interface to trigger it to stop processing.

[0181] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A grinding state online monitoring method based on multi-layer data fusion and mathematical joint driving model, characterized in that, The method comprises: The signals of the grinding wheel and the workpiece to be ground are collected by a multi-source sensor, and are fused according to corresponding weights, and noise reduction processing is performed through stacked auto-encoding; Multi-dimensional time and frequency domain features of the fused and noise-reduced signals are extracted, and after Pearson correlation screening, a structured feature dataset is obtained; The structured feature dataset is fed into a residual neural network ResNet and a bidirectional double-layer LSTM fusion model to realize single-condition grinding state recognition; Discrete grinding states of the grinding wheel and the workpiece to be ground are preset, and variable grinding parameter experiments are performed on the discrete processes of the two grinding states respectively to obtain changes of monitoring signals and state representations under different parameters; Discrete grinding states of the grinding wheel and the workpiece to be ground are preset, and variable grinding parameter experiments are performed on the discrete processes of the two grinding states respectively to obtain changes of monitoring signals and state representations under different parameters; Physical driving type empirical formulas of working conditions and monitoring signals, and working conditions and state representations are established based on grinding shaft speed, feed speed and grinding depth working condition parameters; Training data sets are generated using the physical driving type empirical formulas, and the ResNet-SBiLSTM monitoring model is trained through a transfer learning mechanism to realize cross-condition recognition.

2. The method of claim 1, wherein, The collected signals are fused according to corresponding weights, and a double noise filtering algorithm is designed, which specifically includes: Through grinding experiments, the sensitivity of each signal is sorted as acoustic emission > vibration > current, and the linear fusion signal R' is obtained by fusing them according to weight proportions of 60%, 30% and 10%, and the fusion signal is obtained by activating R'; Time domain analysis is performed on each signal. First, the data is decomposed by WOA-VMD to obtain different intrinsic mode functions IMF. The components with different grinding state frequency spectrum differences are superimposed and reconstructed on the frequency spectrum, and the reconstructed data is obtained through inverse Fourier transform. Second, the data after WOA-VMD preliminary noise reduction is subjected to deep feature extraction again, and a stacked auto-encoding network model is built to perform deep feature extraction and representation learning, realizing signal enhancement and noise reduction.

3. The method of claim 1, wherein, The method for extracting multi-dimensional time and frequency domain features of the fused and noise-reduced signals, and obtaining a structured feature dataset after Pearson correlation screening comprises: Time, frequency and time-frequency domain joint analysis is performed on the AE, vibration, current and fused signals to extract multi-dimensional feature vectors; Deep fusion is performed on the multi-dimensional feature vectors at the feature level to construct a feature space representing the grinding state; PCC and ACF are used for feature screening, and finally a structured feature dataset is constructed through a sliding time window mechanism.

4. The method of claim 1, wherein, The method for feeding the structured feature dataset into a ResNet and SBiLSTM fusion model to realize single-condition grinding state recognition comprises: The STFT image format of the sliding window is converted into 299*299*3 as the model input. First, 2D-CNN convolution layers are used for low-level spatial feature extraction, and after convolution operation, the spatial size of the output feature map is: In the formula: K H , K W respectively are the height and width of the convolution kernel, the input image size is H x W, H is the height of the image, W is the width of the image, the size of the convolution kernel is K H x K W , the step sizes in the height and width directions are S H and S W respectively, the padding is P H and P W , H out , W out respectively are the height and width of the output feature map; MaxPooling is introduced after the first layer convolution operation in each modal branch, and its calculation expression is as follows: u j i = σ(X j i ); where i represents the index of the feature layer, j represents the index of the feature map, represents the jth feature map of the ith layer, x k i-1 represents the kth output feature map of the previous layer, N i-1 represents the total number of feature maps output by the i-1th layer, W k j and b j i respectively the convolution kernel weight matrix and the bias, σ is the ReLU nonlinear activation function, u j i is the output after activation, is the output after pooling; Data eigenvalues X obtained by Pearson correlation coefficient analysis i = [X1, X2, X3,..., X N ] are input into the SBiLSTM, and the network processes the input data time sequence features from two paths of "forward" and "reverse". The forward and reverse path outputs of the first layer are and The calculation formula of the first layer output H i (1) is as follows: wherein: is the output of the forward i-th LSTM gating unit, is the output of the backward i-th LSTM gating unit; The first layer outputs H i (1) The second layer inputs the sequence and outputs H i (2) The forward and backward features are concatenated, and the calculation formula is as follows: wherein: X i is the feature data after Pearson correlation analysis, f1(·), f2(·) are the hidden layer functions of the SBiLSTM network, and are parameter sets described, is the weight and bias parameter of the first layer, is the weight and bias parameter of the second layer, H i (2) is the output data of the second layer.

5. The method of claim 1, wherein, The preset grinding wheel and the discrete grinding state of the workpiece to be ground are respectively subjected to variable grinding parameter experiments on the discrete processes of the two grinding states, and a method for obtaining the changes of the monitoring signals and the state characteristics under different parameters includes: The grinding state is defined as a discrete space, where the grinding wheel wear grade is P, P ∈ [P1, P2, P3...P 10 ], and the workpiece surface roughness grade is Q, Q ∈ [Q1, Q2, Q3...Q 10 ], respectively. The variable grinding parameter experiments are carried out for the discrete processes of the two grinding states, the spindle speed n ∈ [8000, 15000] rpm, the feed speed v ∈ [100, 300] mm / min, the grinding depth a p ∈ [5, 25] μm, and the changes of the monitoring signals and detection characteristics under different parameters are recorded.

6. The method of claim 1, wherein, A method for establishing physical driving type empirical formulas of working conditions and monitoring signals and working conditions and state characteristics based on grinding spindle speed, feed speed and grinding depth working condition parameters includes: The effective variation range of each variable is determined through single factor experiments, wherein the variables are grinding spindle speed, feed speed and grinding depth; Based on the response surface method of central composite design, 1250 groups of experiments required by full factor experiments are simplified to 200 groups of experiments; wherein the experimental levels of each variable are converted according to the coding formula, and the calculation formula is: The coding levels include five levels: the axis low point level -α = -1.68, the cubic low point level -1, the center point 0, the cubic high point level +1, and the axis low point level +α = +1.68; wherein the cubic point is combined with 8 coding coordinates of ±1, the axis point is combined with (±α, 0, 0) (0, ±α, 0) (0, 0, ±α), and the center point is combined with 6 times of coding coordinates of (0, 0, 0); The empirical formula of the physical driving model two between the working condition parameters and the label data set through the cross-working condition experiment is as follows: where: P, Q are the grinding wheel wear and workpiece surface roughness, β0, is the model constant term, β1, β2, β3, is the linear term coefficient, the quadratic term coefficient of the model is β 11 , 22 , 33 , , 12 , 13 , 23 , is the interaction term coefficient, ε2, ε2 (2) is the random error; The empirical formula of the physical driving model two between the working condition parameters and the signal feature set through the cross-working condition experiment is as follows: In the formula: F = [RMS, Spectral energy, Kurtosis, Variance] are the characteristic vector data of the signal across operating conditions, ξ0 is the model constant term, and ξ1, ξ2, and ξ3 are the coefficients of the first-order term. 11 ξ 22 ξ 33 It is the coefficient of the squared term, ξ 12 ξ 13 ξ 23 It is the coefficient of the interaction term. It is random error.

7. The method of claim 6, wherein, A method for generating a training data set by using a physical driving type empirical formula includes: Cross-process machining experiments are carried out, and the process parameter vector G = [G1, G2, G3...Gn] is dynamically adjusted and input into the physical empirical model of step two: k ] G→(F', W', R'); In the formula: (F', W', R') is the sensor signal feature set and the label data set generated by the model; Adaptive acquisition of signal feature vectors under current working conditions, in view of the fact that the output of the empirical model is n discrete state points in a 4-dimensional feature space (n=10), and the deep learning model needs small sample continuous data input, data augmentation is carried out, first, the feature space is expanded, wherein R in F represents the root mean square value in the time domain, spectral energy represents the spectral energy, kurtosis represents the kurtosis, and variance represents the variance; for the generated F' which is a 10*4 feature vector, for the data generated for the 10 discrete points, uniformly sampling in its neighborhood N μ (F) wherein μ=0.5, F i =2000*4 random vectors are generated. F' μ = F + ΔF, ΔF ~ U([-μ, +μ]); In the formula: ΔF is the generated random disturbance data, U is the random disturbance vector, and μ is the disturbance interval; The tag data set (W, R) is continuously fitted, and a continuous curve W of the parameters of the working condition is constructed by using a cubic spline interpolation method (g) and R (g) : (W (g) , R (g) ) = a i +b i (g-g i )+c i (g-g i ) 2 +d i (g-g i ) 3 ; wherein: W (g) and R (g) are continuous curves of the wheel wear and workpiece surface roughness as a function of the working conditions, g is a working condition parameter independent variable, g i is the i-th working condition parameter, i is an index value, a i , b i , c i , d i are constant, linear, quadratic, cubic coefficients, respectively.

8. The method of claim 1, wherein, A method for training a cross-working condition monitoring model through a transfer learning mechanism to realize cross-working condition recognition includes: The physical empirical model generates an enhanced data set F μ For time-frequency feature values, (W g ', R g ') is the response output, input into the ResNet-SBiLSTM model, and the network grinding is improved by the following transfer learning strategy. First, pre-train the source domain, train the entire network using large data in the source domain, and obtain a set of good initialization parameters, whose formula is as follows: In the formula, θ1 represents a feature extraction layer parameter, θ2 represents a classification layer parameter, D source represents a source domain data set, L loss represents a loss function; After the source domain is trained, the target domain data generated by the empirical formula is used to fine-tune the model, the feature extraction layer parameter θ1 is fixed, and only the classification layer parameter θ2 is fine-tuned. In the process of back propagation, the gradient is blocked in the feature extraction layer and the parameter is not updated, and the loss function includes cross entropy loss and regularization term, and the calculation formula is as follows: In the formula: L loss represents the total loss function of the fine-tuning process, L CE represents the cross-entropy loss function, and the target domain sample is X target , represents the output of the model to the target domain sample, Y target represents the target domain label; λ is the regularization coefficient, and ||θ2|| 2 is the regularization constraint.

Images